Meter: The Internet Utility

A Deep Dive on Vertical Integration, Networking, and How to Win the Internet

Welcome to the 419 newly Not Boring people who have joined us since 2024! If you haven’t subscribed, join 239,249 smart, curious folks by subscribing here:

---

Today’s Not Boring is brought to you by 2025 Presenting Sponsor… Ramp

I don’t typically write Deep Dives on any company more than once, other than Ramp. I’ve written about Ramp four times. There’s no company I know or trust more, so I’m particularly psyched that Ramp is joining us as a 2025 Presenting Sponsor.

From the beginning, Ramp has been on a mission to save finance teams time. In my December 2020 Deep Dive, I wrote: “By letting Ramp handle all of the work they don’t want to be doing anyway, the finance team is freed up to focus on work that adds value to the business and helps it grow.”

While the mission remains the same, Ramp keeps leveling up the tools it uses to achieve it.

Now, Ramp isn’t just a smarter corporate card—it’s a fully integrated finance automation platform. It streamlines approvals, automates expense management, delivers real-time insights, and seamlessly syncs with ERP systems. Your finance team gets the control and visibility they need, without the busywork.

With Ramp, your finance team has the best engineers in fintech working on your behalf. And every day, that little card in your wallet keeps getting smarter and taking more of the grunt work off your plate.

That’s why Ramp is trusted by 30,000+ companies big and small, including Shopify, AirBNB, Stripe, Anduril, CBRE, and of course, Not Boring. Join us.

Time is money. It’s time your team upgraded to Ramp: the Official Business Card of Not Boring.

---

Hi friends 👋,

Happy Thursday! I’ve been wanting to share today’s Deep Dive with you for a long time.

Normally, when people ask me “What are you writing about?” and I say the name of the company, at least half the people say something like, “Oh yeah, I’ve heard of them, they’re the ones who…” or “That’s so and so’s company, right?”

When I’ve mentioned Meter, nine times out of ten, I’ve been met with: “Huh? Who’s that?” These are people in tech, other investors, people whose job it is to know today about the companies that will be worth the most money tomorrow. Blank stares.

That is what I live for. Meter has the widest gap between promise and awareness of any company I’ve covered, and I’m pumped to be the one to get to tell you about them.

Because the one out of ten who does know about Meter knows. The Meter Fan Club, small but ravenous, has been a secret society. Today, it’s no longer secret.

Note: the essay is too long to fit in an email; click to read the whole thing online.

Then print it out. Get it bound. Highlight it. Save it. Come back to it. Enjoy.

Or if you prefer your Deep Dives in your ears, the audio edition will be up soon — check back here in an hour or so, or go to the ElevenReader App.

Let’s get to it.

---

Meter: The Internet Utility

At the turn of the 20th century, back when electricity was new and before it was a utility, the New York City skyline looked like this.

Illustration from Harper's Weekly July 27, 1889

Back then, when a tenant of a building wanted electricity, they had to go through a long and complicated process.

They’d call the electric company, who would send an engineer to assess the building’s location and suitability for electric installation. Assuming suitability, electrical engineers would design custom electrical plans for each building, which included the placement of wiring, fixtures, switches, and other electrical components. Then they would install wiring and appliances, like light bulbs and chandeliers. Once the building was wired, engineers would run wires from the nearest utility pole to the building, hence the blocked city sky.

Then, the electric company would install a meter to measure electricity consumption for billing purposes, and send representatives to train tenants on the safe operation of electrical systems.

The utilities you don’t think about at all today, like water and power, did not always have the relatively seamless interface that we’re all familiar with: pay bill, access the fruits of an incredibly complex network at the turn of the faucet or the flip of a switch. Nothing ever starts out that way.

When J.P. Morgan asked Thomas Edison to make his Italiante New York City mansion the first private residence in the city to be illuminated solely by electricity, Edison’s crew of workers:

Clopped up in a horse-drawn wagon to Morgan’s nearly renovated mansion at 219 Madison on the northeast corner of 36th Street. They laboriously excavated a large earthen cellar beneath the wooden stable, their shovels rhythmically slinging dirt and rocks into a growing pile. Within the musty space of the dirt cellar, they installed a squat steam engine and boiler to power two electric generators, all of which displaced Morgan’s carriage horses to a nearby stable. The men also dug a ditch connecting the new cellar to the house, lined it with bricks, laid in the electrical wires, and bricked it over. Inside the mansion, decorator Christian Herter supervised the snaking of insulated electrical wires up through the elaborately wood-paneled and plastered walls where ordinarily the gas lines would have gone. These wires were then threaded through every space in the mansion, and new electrical fixtures were installed. In some rooms, electrical wires hung straight down every few feet from small holes in the tall ceilings, sprouting at their tips several small lightbulbs.¹

After all of that, the generators made so much noise as to disturb the neighborhood peace, and the steam engine spit out fumes and smoke that tarnished neighbors’ silverware. Then, a broken spud set fire to the Morgans’ library.

That was 1882. Two decades later, in 1901, the dream of omnipresent, effortless electricity was still that, as captured by Emile Zola in Travail:

The day must come when electricity will be for everyone, as the waters of the rivers and the wind of heaven. It should not merely be supplied, but lavished, that men may use it at their will, as the air they breathe.

The buildout of the electric grid that we know today was a multi-billion, multi-decade process that Brian Potter needed four parts of a series on The Grid (I, II, III, IV) to cover. In addition to the grid itself, each individual home, factory, and office had to be wired to receive and utilize the electrons that the grid delivers. That takes investment, buy-in, and time. Lots of time.

The project continues, and will continue. Infrastructure projects are endlessly expansive.

But the current state of things is that when you move into a building – a home, an office, a warehouse, a store, any building really – it’s already wired up for electricity. The electricity comes with the building. You just move in, sign up, and turn on.

That is not how the internet works for companies, yet.

While each building has internet from one or more ISPs running into it, a working network doesn’t just come with the building.

The network is customized from scratch each time: new ISP contract, new network design, new hardware, new security, new software. IT teams work with a fragmented set of vendors. One company provides the internet (the ISP), another designs the network, another sells the hardware (either directly or as a value-added reseller), another installs, and yet another provides the software to monitor, manage, and configure the network as a company’s needs inevitably change (managed service provider, or MSP).

And the system works well enough! You’re probably reading this from office WiFi that’s working just fine.

There is some really good networking hardware out there. Cisco’s Meraki is the leader in the space. Meraki generates somewhere between $5-10 billion, and is growing fast. Ruckus makes great access points. Palo Alto Networks makes great security hardware and software; it’s worth $125 billion. Arista makes great switches; it’s worth $136 billion.

That hardware can be CapEx intensive, sure, but as an IT leader at Dollar General told a Tegus interviewer, retailers often prefer CapEx over OpEx because they can amortize the costs over time.

Managed Service Providers do a robust book of business, too. The global MSP market for networking alone is something like $70-80 billion and expected to grow at an 8% CAGR, much faster than the 3% growth CIOs expect for internal IT teams. It’s more fragmented, but huge.

Certainly, setting up and managing a network can still be a pain for IT teams. It’s more expensive than it probably should be. A MSP can’t build the kind of software to manage networks that a company could if the company also built the hardware.

All of which is to say, the internet is not yet a utility, but it does work.

But the internet can and should work much better. It must, as the internet inevitably grows and more and more of what we do gets done in packets.

That is the bet that Sunil and Anil Varanasi made a decade ago, continue to execute against today, and plan to execute against for a very long time.

For the better part of the last decade, they’ve been building Meter.

Meter builds internet infrastructure. That image of the New York City skyline comes from Meter’s onboarding docs, because Meter’s mission is to build the next modern utility, one in which turning on internet, networking, and WiFi is as easy as turning on water or electricity.

“Every building has a power meter and water meter,” the company writes. “We’re called Meter because we believe internet, networking, and Wi-Fi should be just as ubiquitous.”

Meeting Meter

I first spoke with Anil on December 15, 2023.

I later learned that, in anticipation of our call, he had “literally tried to read everything [I’d] ever written.” I learned that this is something he does.

Within two minutes of meeting, he asked how I thought my writing had changed over Not Boring’s four years, then he asked if he could give me his opinion. I said of course.

It had gotten more confident, he observed, but had lost its vitality and freshness. More recently, after we’d gotten to know each other, he recalled the spirit of his message: “The first thing I said was, ‘Yo, what happened to you?’”

Somehow, the feedback came off as true and refreshing, because I’d been feeling it myself without putting it into words. He was right. Damn.

Then, at the end of that same call, Anil said that he would love for me to write about Meter.

Put yourself in my shoes.

Here’s this person who I’d just met, who told me that my writing wasn’t as good as it used to be, then turned around and asked me to do the deepest dive on his company, the one that he and his brother Sunil had been building for a decade in near-silence and that they will probably spend most of the rest of their lives building, that anyone’s ever written, by far.

Now here we are, thirteen months later. I’ve spent more time researching and writing about Meter than I’ve spent on any other Deep Dive.

The subject deserves the treatment. In a short 2022 blog post on Meter, Brie Wolfson wrote:

To understand the space that Meter is operating in and the future they are building towards is to contemplate just how much surface area we are talking about; things held in hands to things fixed in walls, humans to machines, work to play, the ocean floor to in-orbit satellites, city centers to remote neighborhoods, public policy to personal preferences.

By way of agreement, this is the longest Deep Dive I’ve ever written, too. It’s 100 pages. That’s a small book.

After all this time and all these words, I’ve come away convinced that Meter has a shot at becoming one of the most valuable companies in the world.

Today, I’m going to explain why I believe that.

It is both simple and complex, like when someone asks you how you were able to magically see and speak with someone halfway across the world.

You might say, “Oh, I just used Zoom.”

Or you might describe what a packet is, how it is born, the journey it takes, and all of the infrastructure that supports it along the way.

In this piece, I’ll attempt to do both, because if you, like me, believe that there is an opportunity and a necessity for startups to rebuild the infrastructure on which the world runs, then understanding exactly how one of the best to do it does it, in real-time, is valuable information.

So we will talk about what a packet is, how it zips across the globe, and why it matters.

We will of course discuss Meter’s products. It builds hardware (switches, security appliances/firewalls, wireless access points, cellular access points), software (OS, firmware, Connect, and Command), and operations (from design to installation to ongoing support).

We will talk about why, despite the half a trillion to ten trillion in networking company market cap, depending on how you define a networking company, there’s a need for a vertically integrated new entrant. And why, despite all the market cap available, there aren’t many new entrants - maybe one or two legitimate newcomers a decade. This isn’t software.

We will talk about the challenges and benefits of vertical integration, deeply, because vertical integration is the key to Meter’s present and future.

We will get into the nitty gritty details on everything from architectural decisions to go-to-market to software that’s actually soft. Because a successful vertical integrator is dependent on getting thousands of details right and orchestrating them all.

All of this has remained core to the Meter story since the first time I talked to Anil:

Some notes from my first call with Anil Varanasi, December 15, 2023

Some things have changed since we started talking, too, or more accurately, have grown.

First, while the importance of networking was kind of clear to me then – no one likes spotty internet, amiright – it has become abundantly clear over the past year that networking is up there with energy as one of the main bottlenecks to the AI data center buildout. In the process, better, faster networking has gone from a “nice-to-have” to strategically critical. Everyone understands the need for power, few understand the need for networking.

Second, my appreciation and understanding of Vertical Integrators has grown, too. Certainly thanks in part to spending so much time learning from Sunil and Anil, my entire investment thesis is that startups have the opportunity to overtake sclerotic incumbents with vertically integrated offerings that are better, cheaper, faster and higher-margin. Meter is a living masterclass in vertical integration.

Third, I have a working thesis that AI supports winner-take-most vertical integrators in previously fragmented hardware markets. Meter is as good of a supporting example as I’ve found.

And finally, Meter’s master plan is entering its endgame, even if it has decades left to go. Specifically, by vertically integrating hardware, software, data pipelines, APIs, operations, and applications, Meter has what it needs to do two things:

1. Enter the Data Center market, which it plans to do this year using the same architecture it uses for local networks to reap the benefits of scale and compounding.

2. Develop autonomous networks by building a foundation model for networking. No other company has all of the ingredients and talent needed to do this. Meter does.

The nature of vertical integration, done right, is that it takes longer to get going, and then it zips. Done really right, by the time the incumbents know you’re coming, you have them in checkmate. To the extent that data is more valuable today, this is more true today than ever.

We will talk about all of this. This is the kind of stuff that I could, and will, nerd out on for thousands of words.

Foundational to all of that, though, or as a necessary precondition to it, is the je ne sais quoi that’s made me want to spend over a year writing this one piece.

The “Oh, I just used Zoom” of Meter is Sunil and Anil.

The Brothers Varanasi

When I wrote in Vertical Integrators: Part IV that the hard-to-define characteristic of truly exceptional founders is that “they make world-class people want to do their best work for them,” I had Anil, and that conversation, in mind. I am not alone.

Nitan Shalon, who leads Meter’s Command product, told me that Anil was one of the main reasons he joined a networking company over the many other opportunities available to someone with his skillset (PhD candidate in EECS from MIT, where he was a researcher in its Computer Science and Artificial Intelligence (CSAIL) Lab):

“Anil does this weird thing. He’s both a killer and a joking friend at the same time,” Nitan explained. “And he’s extremely ambitious. Scarily ambitious. Getting to work with him directly every day was a big pull.”

I’m one little example. Nitan is one little example. At Meter, there are 100 little examples of exceptional people who joined the company to work with Anil and his complementarily and equally (Anil wants me to say “more”) impressive brother, Sunil.

The Varanasis are like the Forrest Gump of Silicon Valley if Forrest were intentional. They’re somehow involved in the important stuff, quietly and behind the scenes.

Dan Romero, the Farcaster co-founder, met Anil when, back in 2016 or 2017, he doesn’t remember exactly (although he told me to ask Anil, who would remember exactly (he did, it was December 5th, 2015, actually)), he tweeted about some “esoteric crypto-oriented DNS” and got a “random cold inbound from an anon account” saying, “I’d like to meet with you.”

Dwarkesh Patel recently wrote about Anil asking how much it would take him to podcast and blog for six months, back when he was a relative unknown. He said $10,000, Anil gave it to him, he became a really great podcaster, and Anil told him (somewhat) jokingly that he regretted doing it, because it meant that Dwarkesh was wasting his talents podcasting.

Link to Tweet

Kevin Kwok quote tweeted Dwarkesh and said, “There are many examples of @acv doing this, under the radar. And it's one of the most lovely and impactful things one can do. In general, many things Anil and his brother do that should be more widely emulated.”

One of the reasons that Sunil and Anil Varanasi are not more widely emulated is that they have flown relatively under the radar.

Sunil (left) and Anil (right)

Sure, there was the Ashlee Vance piece in Bloomberg in 2022, the Alex Konrad piece in Forbes when the company raised $35M, and the Fortune piece more recently, which is the kind of targeted PR you get if you know what you’re doing. And finally, after years of asking, Aarthi and Sriram finally convinced their friend Anil to join them on the Aarthi and Sriram Show:

But for the most part, the brothers keep their heads down.

Anil has 1,185 X followers (and 0 tweets) and Sunil, as far as I can tell, has no profile. They stay out of the public spotlight. And yet, everyone in the know seems to know them.

Meter’s investors include Sequoia Capital, Sam Altman, Scott Belsky, John Bicket, Sanjit Biswas, John and Patrick Collison, Egon Durban, Daniel Gross, Diane Greene, Lachy Groom, Sam Hinkie, Reid Hoffman, Sanjay Jha, Tobi Lutke, Jay Parikh, Geoff Ralston, Dan Schulman, David Solomon, Allen and Co, WndrCo., which is the kind of cap table you put together if you know what you’re doing.

From this all-star team, too, Sunil and Anil are able to draw out the best work. At the end of my first conversation about Meter with Sam Hinkie, he told me, “I have more interest in helping you with this piece than you would know. Among my favorite pastimes is to brag about my friends and explain how extraordinary they are before even they know it.”

Brie Wolfson, excellent in her own right – Stripe, Figma, Positive Sum CMO, and the author of the defining essay on Taste, among other accomplishments – titled her Meter piece, “Why I’m rooting for Meter.”

I even found a three-year-old Will Manidis cameo when trekking through the G2 comments. Among the other 22/22 5-star reviews, he wrote: “Meter changed my life.”

G2

Maybe it shouldn’t have surprised me that Will was a Meter fan, too. Unintentionally, he’s recently been tweeting both the thesis of this essay:

Link to Tweet

And the reason why, despite and because of the fact that Meter is still a young company that you’ve never heard of, I think it’s worthy of such an in-depth study:

Link to Tweet

If I’m right, then setting out to learn about the Varanasis before it’s obvious that they’re the greats will allow us to access the truth only available during the journey.

After the fact, when they’re billionaires, it will be totally unsurprising that people want to do great work for them. People are inclined to want to do great work for really rich and successful people. They see it as their ticket. Drawing that out of people before most even know about your company, on the other hand… that’s a skill worth studying.

Because this ability to pull the best from the best people, to inspire them to contribute and root, transcends any particular product. There’s a sense you get from the Varanasis that whatever they chose to do, they’d be able to pull it off. Chips, cancer cures, fusion, whatever.

But given how I’ve described them so far – high-performance, reliable, operating quietly in the background, getting the best out of anyone connected to them – maybe what they decided to do was inevitable. Networking is a natural fit.

Meter is not, of course, the first networking company. Cisco, the 40-year-old networking giant, is worth more than $235 billion. It was, at one point, the largest company in the world. The top five companies in the networking space are valued at more than half a trillion dollars combined. Meter is valued at less than half a billion.

So Meter isn’t the biggest, and it certainly isn’t the first, but its ambition is to be the last networking company.

“If there’s a packet moving in the world,” Anil told me, “we hope it will be moving through Meter hardware and software.”

Today, roughly 20 quintillion (20 x 10^18) packets move in the world each year. That unfathomable number is only going to grow, a fundamental bet the Varanasis staked their careers on a decade ago that is now a foregone conclusion.

Meter wants to move all of them.

That is an odd goal, both enormous and specific. Believing that it’s even possible requires a deep technical understanding of how that world works today, and an informed view of how it will work in a couple of decades. It speaks to the brothers’ Brobdingnagian ambition, and their willingness to patiently lay the groundwork on all of the small things that might compound in specifically-unpredictable but aggregately-inevitable ways to realize it.

The story of any startup worth writing about can’t be untangled from the story of its founders, but the Meter-Varanasi knot is tied particularly tightly. When I wrote about Stripe in 2020, I wrote, “It’s hard to take a Stripe bear case too seriously, because there is no way that the Collisons haven’t thought much more deeply about the challenges it faces than anyone else in the world.” There is a similar circular logic at play here.

To understand Meter, we need to understand Sunil and Anil. To understand Sunil and Anil, we need to understand Meter.

It will take the deepest of Deep Dives, a masterclass in Vertical Integrator strategy, the alpha in a well-constructed network, and what it takes to turn infrastructure into a utility. Of course, we’ll cover Meter’s products, how they fit together and compound on each other.

The journey will be winding, splitting and branching off in many directions before, hopefully, coming back together at the end to paint a coherent picture of one of the companies (and two of the founders) I now believe will be among the most consequential of this generation.

In that way, this Deep Dive might look like the journey of a packet itself. Zoom zoom.

Consider the Packet

What have you accomplished in the last 200 milliseconds?

I don’t mean to suggest that you’re lazy – you’re in the middle of reading thousands of words on networking and business strategy, you’re one of the special ones – but, and again, unfair comparison, consider the packet.

Sending Bits, The Way Things Work Now, David McCauley

Chances are, in the past 24 hours, you’ve been on a Zoom. As you sit there, bored or self-conscious, your camera captures light through its lens and focuses it on a sensor, which converts the light into electrical signals, which are processed by an image signal processor that corrects color, reduces noise, and encodes signals into a format like YUV or RGB. This happens up to 30 times per second.

When you speak up – “Sorry.. yes … can you .. ok, good well I was just going to say…” – the sound waves you emit cause your microphone’s diaphragm to vibrate. These vibrations, too, are converted into electrical signals, and an analog-to-digital converter (ADC) samples these signals, typically at 44.1 kHz or 48 kHz for high-quality audio. The samples are ever-so-briefly stored in a buffer.

Zoom’s software takes this raw video and audio data and compresses it – using a codec like H.264 or VP9 for video, or like Opus or AAC for audio – to reduce the size of the data while maintaining quality. This compressed data is split into smaller chunks, each of which is wrapped in Real-time Transport Protocol (RTP) headers that include sequence numbers and timestamps so that all of the chunks can be properly reassembled on the other end.

At this point, somewhere between 6 and 26 milliseconds into the process, the data first becomes a packet, of sorts: an “RTP packet.” These packets are encapsulated in User Datagram Protocol (UDP) datagrams, which add another layer of information about where the data is heading and who can access it.

Our new packet is one of many, over at least 100 created every second on your Zoom call, separated at birth to find their own way across the world only to be reunited on the other end. For this architecture, we have the Soviets to thank, at least in part.

Communications networks used to be “circuit switched”: a connection was established between two points, and data flowed continuously from one to the other. For a while there, when you wanted to call someone, a switchboard operator would manually connect your line to the recipient’s line to create a dedicated circuit for the call.

This was inefficient for obvious reasons, but a bigger concern was that it was vulnerable. A nuclear war could wipe out circuits connecting military leaders, leaving them unable to communicate at the absolute worst time. So the government tasked RAND Corporation’s Paul Baran with designing a network that could survive.

Baran proposed a distributed network with “packet switching,” in which messages would be broken down into small chunks (packets), each of which could take multiple different paths through nodes in the network to their destination.

A few years later, in 1966, the Advanced Research Projects Agency (ARPA) set out to create a network of computers as robust and fault-tolerant as Baran’s telephone network design. They built ARPANET, which went live in 1969 with four nodes, and grew node-by-node, mainly university-by-university, over the next decade.

ARPANET

In 1974, UCLA’s Vint Cerf and ARPA’s Bob Kahn published A Protocol for Packet Network Interconnection, in which they introduced the Transmission Control Program (TCP) protocol to more reliably and efficiently break data into packets, send it across various routes, and reassemble it at its destination. In 1978, the protocol was split into two pieces – TCP, for reliable data transfer, and IP (Internet Protocol), for routing between different networks – to form the standard TCP/IP protocol on which the modern internet is still largely based.

So that is how the idea for packets was born and grew. But our packet is not quite a real packet yet. It needs its IP header.

Zoom passes the UDP datagram to your computer’s operating system via a socket API, and your OS wraps it in another header, the IP header. The IP header gives routers the information they need to send the packet to the right destination, including (but not limited to) the source and destination IP addresses.

The packet is born.

A packet, Julia Evans

Like a neonatal nurse, the network interface driver takes the newly born packet and swaddles it in an Ethernet frame. Thus wrapped, the packet is sent to the network interface card (NIC), which converts the digital frame into analog signals – for copper wires, electric signals; for fiber-optic cables, light pulses; for WiFi, radio waves – and transmits them through the wires or air - for transmission over the local network (like your office’s WiFi).

The idea of a local network is newer than ARPANET – computers spoke to each other over long distances before they learned to speak to their neighbors. With mainframes, there was no need for local area networking; people brought their tasks to the machine. Some had dumb terminals attached to them – kind of like computer screens and keyboards with no processing power of their own – but all of the terminals in one location simply interacted with the central machine.

That changed in the early 1970s with the minicomputer. Instead of mainframes that cost millions of dollars, out of reach of most small organizations and departments within larger ones, minicomputers cost anywhere from $10,000 to $50,000. All of a sudden, one company, even one office building or floor in an office building, could have multiple computers, and peripherals, like printers, that worked with them. Which meant: machines needed to learn to talk to each other.

Enter Bob Metcalfe (the Metcalfe of Metcalfe’s Law).

It’s easy to forget this now, now that we take their inventions for granted, but people like Bob Metcalfe were just people – “Metcalfe hailed, as though at the top of his lungs, from Brooklyn and Long Island. He had graduated from Harvard and MIT—bitching every step of the way, to hear him tell it—with degrees in electrical engineering, business, and applied mathematics,” wrote Michael Hiltzik in Dealers of Lightning – with problems to solve.

In 1973, at Xerox PARC, Metcalfe invented Ethernet, not to change the world, but simply to get PARC’s Alto computers to connect with each other. The network had to be cheap – “less than 5% of the cost of the computers it was connecting,” simple, reliable, expandable, and fast. Nothing Metcalfe came up with met all of those requirements, until he remembered a paper he’d read months earlier by a professor at the University of Hawaii on a radio network called ALOHAnet, which was designed to let computers talk to each other across the Hawaiian archipelago, but which, Metcalfe realized after correcting some of the professor’s mistaken assumptions and calculations, and replacing radio transmission with inert physical lines, a passive medium, like the “‘luminiferous aether’” once thought to pervade the universe as the medium for the propagation of light,” would also work for letting computers speak to each other within the same office.

And Bob said, let there be Ethernet.

With Ethernet, computers could share the same coaxial cable to communicate. Instead of serial connections and point-to-point links, CSMA/CD (Carrier Sense Multiple Access with Collision Detection) meant that devices could "listen" to the wire before sending data to avoid collisions and retransmit if a collision was detected. Metcalfe’s first version could carry 2.94 Mbps, which seems hair-pullingly slow today but was groundbreaking at the time.

Like many technologies invented at Xerox PARC, Ethernet was commercialized somewhere else. Metcalfe left to start 3COM – Computer, Communication, Compatibility – in 1979, where he brought on Bob Krause as President in 1981 and elevated him to CEO in 1982.

Krause gave an incredible interview on the company’s history on the Seeking Truth in Networking podcast. He said that in the beginning, the cost of Ethernet – the digital card (NIC), transceiver, and cable – was $3,500. He knew the company needed to get the cost down to $1,000, then to $100. They drew an exponential declining curve from $3,500 to $100 over 7 years, and realized that they’d just need to follow Moore’s Law’s 50% cost declines per year.

To do that meant putting Ethernet on a chip – a VLSI, or Very Large Scale Integration – with the NIC and the transceiver on a single card, which they did. Originally, computers with Ethernet Chips were daisy-chained together with a single coaxial cable used as a backbone and each computer connected to the backbone through T-connectors with BNC connections.

That might be the way we did things today if it weren’t for another famous borrower of Xerox PARC technology.

See, Apple was 3COM’s second or third customer. When Krause and the team showed Steve Jobs the daisy-chained setup, he… didn’t love it.

“Who’s the braindead idiot asshole who came up with this?” he asked. “This is shit. This is dreck. If you want to make the thing easy to use just plug the thing into the goddamn telephone jack in the wall.”

Success has many fathers, as does our packet. Steve Jobs is one, partially responsible for the path it travels on its local network, not along a daisy-chained path from computer to computer, but over twisted pair cabling that plugs into a switch as easily as a telephone jack into the wall.

So our packet, just a few milliseconds old, begins its journey across the network. The NIC transmits its signals as radio waves to the access point – here, we first encounter Meter’s hardware.

The access point performs a crucial transformation: it receives these wireless radio signals, converts them back into digital data, and then sends that data to the switch – also provided by Meter – via Ethernet cable (with Power Over Ethernet, PoE, which powers and connects in the same cable).

Not to Scale, Very Bad Representation, You Get the Point

The switch reads the MAC Address within the ethernet frame to figure out where to send the packet next, whether to another device on the network or to a gateway router that will send it out into the internet. The switch forwards the packet only to the specific port needed to reach its destination, avoiding congestion and increasing efficiency in the network.

Let’s assume that our packet is ready, that it’s time to venture out to the internet, the network of networks. Typically, the switch sends the packet to the router, which, as its name suggests, routes the packet to the destination it reads from its IP header.

In Meter’s case, the security appliance, which is a firewall and router in a single device, serves as the physical router. Its routing stack is a combination of its software-defined router and hardware, which work together to manage traffic, handle IP addresses, and determine the best path by which to send our packet across the internet.

From the security appliance, the packet leaves the Local Area Network (LAN) out into the Wide Area Network (WAN). From the device on the network, it travels through a series of point-to-point links, like fiber-optic cables that use light pulses to transmit data at near light-speeds, to your Internet Service Provider’s (ISP) backbone.

At this point, our packet’s journey looks almost like the journey of an electron from generation facility to service panel. In this analogy, the backbone is like the transmission line – it’s the big bundles of fiber-optic cables, buried underground or under the sea, over which data travels very long distances. It reaches the ISP’s border router, and enters the realm of core network infrastructure.

The core network is a place where the internet truly becomes the "network of networks." Here, routers belonging to Tier 1 ISPs—the largest internet providers in the world—exchange data using Border Gateway Protocol (BGP), the postal system of the internet that decides which paths data packets should take based on a combination of policy, economics, and efficiency. The border router uses BGP to determine the packet’s next hop, typically along one of the high-speed fiber-optic transmission lines that interconnect data centers and ISPs across the country or across continents.

Here, our packet, still barely even a twinkle in its parents’ eyes on human timescale but nearly halfway through its packet-scale life, comes to its first fork in the road.

• If it’s traveling to a destination in its own geographic region or on the same ISP (or an ISP with a peering agreement with its own), it might take a direct route, hopping from the ISP’s network to an Internet Exchange Point (IXP), a physical location where networks connect and packets switch lanes.

• If it’s traveling to another country or on another network, it’s likely that it will have to pass through at least one data center.

Ours is traveling across the globe, say to India, the reverse of the journey that the Varanasis made as pre-teens. So it will need to travel through a data center.

Data centers do many things. They are famous now for their role as the training grounds of digital gods, but other, more traditional data centers act as aggregation points, helping to manage and direct the immense amounts of data traveling across countries and continents. It is in one of these more traditional data centers that our packet finds itself, before being shot out into an undersea cable.

Something Like This

Our intrepid packet, now a seasoned traveler at the ripe old age of 50 milliseconds, plunges into the depths of the ocean. It's carried along undersea fiber optic cables, a network of glass threads crisscrossing the ocean floor like the neural pathways of a global brain. These cables, each no thicker than a garden hose, carry 97% of intercontinental data traffic. Our packet is in good company.

As it hurtles through the darkness at nearly the speed of light, our packet might pass by the occasional curious fish or, if it's particularly unlucky, a shark. Sharks sometimes bite undersea cables, though no one's quite sure why.

Upon reaching India’s shores, our packet is shot to a local data center to go through the digital equivalent of customs. The data center’s routers check the packet’s IP address to figure out where to send it next, through India’s domestic internet infrastructure, on to the ISP of its destination. These days, chances are that’s Jio, which offers both fiber and wireless connection.

Less than a decade old and boasting a 52% Indian ISP market share, Jio stands as a testament to the idea that although many of us take internet access for granted, better internet infrastructure remains a work in progress. Expected to IPO this year at a market cap of around ₹10 lakh crores ($120 billion), it is a testament to the massive value that building better internet infrastructure can create.

But our packet cares nothing for IPOs or rupees. It’s almost home. It makes the final sprint over Jio fiber to the local network of the person on the other end of the Zoom.

At last, after a journey of about 200 milliseconds (an eternity in packet time, but less than the blink of a human eye), our packet arrives at its destination and makes the original local network journey in reverse: switch to access point to radio signal then over WiFi to the computer.

Today, in India, our packet likely travels across Cisco hardware. One day, it may be Meter’s.

Finally, our packet begins to disappear.

The receiving computer's NIC greets it and strips away its Ethernet frame. Its operating system peels off the IP header, then passes the UDP datagram up to the application layer, where Zoom's software awaits.

Here, our packet is reunited with its siblings – other packets that took different routes but arrived in the same place within milliseconds. Zoom reassembles them in the correct order, using the sequence numbers in their RTP headers like a jigsaw puzzle guide.

Finally, the data our packet carried is decoded, decompressed, and converted back into audio and video signals. The recipient's speakers vibrate, their screen flickers, and voila! Your awkward interjection is heard and seen, as if by magic, on the other side of the world.

Our packet's job is done. In less time than it took you to read "Sorry.. yes …", it traveled halfway around the world, bringing a tiny piece of your presence along with it. And as you continue your Zoom call, countless more packets embark on the same journey, in a never ending cycle of birth, light-speed travel, and decomposition.

This is the way the world works now, and the way the Varanasi brothers bet the world will work more and more in the future, in the same way that Metcalfe and Krause bet that an exponential curve would continue to run until they could connect all computers in local networks that in turn connected to wider networks that have connected the world.

So among all of the many things that the Varanasis could be doing, they chose to move packets, for three reasons:

1. The internet is phenomenal and there’s merit to making it great.

2. Everything is packets.

3. We all will use the internet, and packets, more than we currently do.

“This call right now is packets,” Anil told me. “When you order an Uber, send an email, it’s packets. The money in your bank is packets. Everything is just packets moving left to right. That’s it. That’s the world.”

Move the packets, move the world.

“We had very particular opinions about how to do that,” Anil said, “and about the business model.”

Why Packets?

Nearly 100% of the world’s people rely on packets, and now you know more about packets than 99% of them.

One question you might be asking, now that you know so much about packets, is why the hell are Sunil and Anil working on moving them?

The system works. You’re reading this right now – on a phone or a computer or maybe even on a TV screen or Vision Pro – because somehow, over the past half-century or so, thousands of technical whizzes have pieced together a system that can take my words and deliver them to your device at any time, anywhere in the world.

The network of networks that moves packets is brain-breakingly complex. We traced just one path, but our packet could have been beamed up to a Starlink satellite and back down to your phone, or over a series of cell towers and to your phone, or through our ISP’s wires and straight to your computer, if you happen to be on the same ISP as me. Increasingly, internet traffic isn’t humans talking to humans, but machines talking to machines.

“Already, there are more devices on the internet than people by orders of magnitude,” Sunil points out. “A lot of communication has been human-to-human or human-to-machine. The future will be machine to machine.” Sunil and Anil see it as “high-frequency trading for everything.” Networking will only get more complex and more challenging.

So why work on it? One simple answer Anil gave me when I asked is simply that “it seemed incredibly fun to work on this.”

He mentioned a blog post, taken down but archived, that PayPal co-founder Max Levchin wrote after working on Slide and selling it to Google for $228 million. Slide made apps for Facebook. It wasn’t work that Levchin ended up being proud of. So he wrote this essay about how to decide what to work on.

It came down to building something incredibly hard that should be valuable to the world if you solve it. And because it’s so hard, and because if it’s actually valuable it will take a long time to create that value, it should be something that’s personally fun for you. Hard, valuable, fun. Levchin ended up naming his fund HVF: Hard Valuable Fun.

Clearly, networking is hard. That packet journey should give you a clue about the magnitude of the challenge, as should the fact that, fifty years in, it’s nowhere close to solved.

It is valuable, too. Because while I said the system works, and it does, it doesn’t work as easily or smoothly as it could or should. Internet is not yet a utility, in the same way that water or power is. As everything becomes packets, ensuring their smooth flow will become increasingly valuable. We will go into this in more detail, because this is Meter’s mission: to build the next modern utility.

I think the real reason that the Varanasis chose networking, though, is because, it is fun.

To a very small group of people that includes them, there is nothing more fun than working on the type of neverending challenge the internet presents.

In some ways, this is personal. There may not be anyone who uses the internet as well as Anil. His personal goal, which mirrors the company’s, is that “If something good is happening on the internet, I should know about it.”

Which sounds ridiculous – the internet is so vast – but… actually seems to be happening. He found Dan Romero on the internet, and Dwarkesh Patel, and countless others in fields as far-flung as biology and filmmaking and hardware and economics.

How it’s happening is illuminating. After I talked to Dan, I texted Anil to ask if we could talk about his funnel, to answer the question of how, when he’s running a company and just had a kid, he even has time to spelunk the internet.

The short answer is that he’s made the good parts of the internet come to him.

The longer answer is that there are two parts, and I’m spending so much essay on this particular side quest, when there’s so much more to write, because it’s something really important I’ve noticed in people who seem to be particularly adept at navigating the modern world but that I haven’t yet put into words.

It is, if you’ll pardon the pun, a form of leveraged networking.

So the two parts:

1. Put what you want into the world, over and over, for a long time.

2. Be a person other people want to do things well for.

Anil has been telling people, “If something good is happening on the internet, I should know about it” for a long time, and over time, it compounds. “Now hundreds of people know that this is how I think.” That part – the consistent repetition of a goal - is straightforward, if not easy.

The second part – being a person other people want to do things well for – is harder. It’s related to that characteristic of truly exceptional founders I mentioned earlier. It means that whenever someone who knows Anil sees something good on the internet, their first thought is to send it to Anil. That alone could be overwhelming, no better than taking the internet’s firehose straight to your face. What makes it work is that he’s cultivated a group of people with their own strong filters, and somehow makes them want to do the extra bit of work to apply those filters, to only send him the good stuff.

I’m not sure I can capture quite how he does that. It’s not a tit-for-tat thing. It’s some combination of high expectations, reciprocal curiosity, and a sense that introducing something or somebody really excellent to Anil is like a game, made all the more fun because the rules are not explicit.

It is a form of alchemy, or at least sifting at scale. In either case, you end up with gold.

I guess there is one thing replicable: a lot of outbound. A couple of months ago, he came across someone doing really interesting work on YouTube with just hundreds of views and sent him a note. “This happens a lot on GitHub, too,” he said. “Someone will write really good open source libraries and 20 people will download.” Send them a note. The author of your favorite book? Chances are, fewer people than you’d expect have even read the book, and only a small fraction of them would ever send the author a note, let alone one with a thoughtful question. Anil reached out to Mark Robichaux, the author of one of the best business books ever written, and one particularly relevant to Meter, Cable Cowboy, and it turned out that very few people had.

The whole thing compounds. These people feel appreciated, and they too start sending inbounds, which sparks more outbound, and on and on, in an ongoing game of creating alpha on the internet.

Two of Sunil and Anil’s professors at George Mason University, Tyler Cowen and Alex Tabarrok, wrote a paper in 2015 titled, The End of Asymmetric Information? The internet, they argued, is replacing the age of asymmetric information with a world of ubiquitous information. If asymmetric information is alpha, where does alpha come from in an age of ubiquitous information?

The answer, Anil thinks, is to create asymmetric information for yourself, through this constantly churning network of inbound and outbound.

“Alpha for what?” I asked.

Second, “just to learn as much as possible because it’s fun.”

But first, always first, is Meter.

“Dan,” who he reached out to cold to discuss decentralized DNS nearly a decade ago, “has been invaluable to Meter with intros, people, ideas, and customers.” Dan is how Anil met Julia DeWahl, Dan’s wife and my Age of Miracles co-host. “I almost recruited her, she is incredible operationally,” but Julia went to Starlink before doing her own thing, so it didn’t happen, but “through Julia, I’ve met Opendoor people who are now at Meter and have been instrumental.”

This is one of many examples. Imagine hundreds of these, all unpredictably but surely compounding.

“On a long enough time horizon, we don’t know what the outcome will be,” he explains, “but it seems plausibly brilliant if we can keep doing it for decades.”

Which brings us back to why Meter is so fun for Sunil and Anil. It’s the type of problem that they can work on, together, for decades and never run out.

Whenever I interview someone for a Deep Dive, my last question is always, “If you read the final piece and it’s missing X, what’s the X that would make you think I missed the whole point?”

Sam Hinkie thought for a second before he said, “These are not people looking to turn their insights into dollars in their pockets soon. Their interest in being acquired or getting to IPO to put money in the bank or go do another thing would really stun people if they realized how low that is.”

When the answer is surprising, like Sam’s was, it’s not actually my last question, because I can’t help but ask follow-ups. In this case, I asked what he thought motivates them.

This time, he didn’t pause:

They’re driven by working together. If they were going to work alone, it would be a lot less interesting. They want to solve very hard problems, whether technical, business design, or empire design. They care a lot about progress and the speed of innovation. They care a lot about power law people; they want to attract those people, partner with those people, learn with them, employ them. And they want to unlock them where they can.

If that’s the byproduct of doing other things they care about, solving the multi-decade puzzle of building an enduring, interesting business they can run together, they’d be stoked about that. They do have some progress-heavy and acceleration-heavy interests that they could wave a hand and say what they’re doing influences that, but having their impact as founders and leaders be in a thing they built together that they run and harvest for a very long time is the thing.

Surely they won’t be driving things Buffett-style at 94, but they might come to the office and be additive if they could. They really might be interested in something that longish-term.

That may be the main reason they chose networking, of all of the things to choose. It is bottomless.

Certainly, they have expertise and experience in the category. In college, Ashlee Vance wrote, “They started a company to install networking equipment for businesses, and had 50 employees and 100 customers by the time they graduated. That experience convinced them that what really had to be done was a redesign of the entire idea of networking equipment and the software that runs it.”

So when they sold the business in 2014, to Sam’s point, they didn’t bank the money or go wild like most 20-something millionaires would have done. They moved to Shenzhen, China, where they shared a single room in a hostel while searching for manufacturing partners for their routers and switches.

“We’d spend all day at the factory and sleep on the factory floor, because we wanted to learn everything about the manufacturing process,” Sunil told Vance. “I didn’t have a bed, because it was covered in circuit boards, so that we could test our software.”

Sunil and Anil spent five years, just the two of them, in China and back home, doing R&D. Learning, tinkering, thinking. They didn’t hire Meter’s first employee until five years in, until they knew what they wanted to build, and how.

A decade into the journey, there’s as much to learn and even more to do than there was in 2014.

So when I asked Anil what makes this fun, the word he used was “expansiveness.”

“It is so cool to build hardware, design operating systems and APIs, design software, figure out how to deploy the network in the real world – supply chain, ops, regulation, legislation – then take the data that comes out and build models,” he said. “It’s everything from ISPs to wired to wireless to data centers. The expansiveness intellectually just makes it so cool to work on this day to day.”

Building infrastructure also means opportunities to expand above and below, for a very long time.

Infrastructure, once in place, tends to provide the infrastructure, stability, and cashflow to build things on top, even things that seem unrelated.

“There’s a non-zero chance that if you move the clock 50 or 100 years, people might not remember that Google was a search company,” Anil predicted. “But it’s possible they’ll remember that [Alphabet subsidiary] Isomorphic Labs discovered every drug in the world.”

It also provides the opportunity to build deeper and deeper in the stack.

“Take our own example,” he says. “Say we’re successful in cellular, in data centers, in WiFi and wired. We could go down the stack and build our own ASICs (application-specific integrated circuits, or custom chips). I just don’t see how we don’t do it if we’re successful. Just think, it’s 2032 or 2030 or something, and I get to go learn about building ASICs. That’s amazing. That is so cool.”

That is the future, though. Infrastructure takes decades to build, decades made up of thousands of days filled with supply chain, APIs, software, manufacturing, and the rest.

If Sunil and Anil want to earn the right to keep playing this game for decades, they need to be successful in cellular, in data centers, in WiFi and wired first.

So we should talk about what Meter is building today.

The Vertically Integrated Internet Networking Company

Meter is building an internet utility by pursuing the Vertical Integrator strategy:

1. Integrate multiple cutting-edge-but-proven technologies.

2. Develop significant in-house capabilities across their stack.

3. Modularize commoditized components while controlling overall system integration.

4. Compete directly with incumbents.

5. Offer integrated products that are better, faster, or cheaper (often all three).

In the place of a previously fragmented smorgasbord of options, Meter makes its own hardware, designs the network, installs the hardware, builds software to control, monitor, and manage the network, provides ongoing security and support, and even refreshes hardware when new versions come out. It works with traditional partners, like VARs and MSPs, but the Meter team quarterbacks the whole process so that, to the customer, it’s seamless.

Customers simply provide an address and a floorplan, and Meter takes care of everything else.

1. Meter designs, builds, and deploys all of the hardware inside the space: switches, access points, and security appliances.

Meter Hardware

When Sunil and Anil started the business a decade ago, networking hardware had become commoditized. Nobody was working on it anymore. But they knew that they would need to own the hardware to build the best software, so they built hardware. In the process, they honed a differentiated view:

“One of our diametrically opposed views with the industry,” they write in an onboarding memo, “is that hardware isn’t and shouldn’t be commoditized. Better hardware can lead to much better software, creating a virtuous cycle.”

And because Meter only sells the full stack or nothing at all, each piece of hardware it makes focuses on only its specific part of the journey - routing, switching, wireless, power – and can be optimized for that function and with the software integration in mind.

2. Not only does Meter develop in-house capabilities across the stack, it is turning those capabilities into software.

Today, it manages all network design, installation, configuration, management, and support. It also develops the software that connects and protects devices in the space, like the operating systems, APIs, and firmware. It owns its full data pipeline end-to-end, and has collected data from software, hardware, and operations in a structured way since the beginning.

Recently, it launched Command, which lets users talk to their network to understand, monitor, and control it.

Meter Command

We will go deep on Command, but for now, what’s important to understand is that it was only possible to build in a reasonable timeframe because Meter makes its own hardware and APIs and trains its own models. This is true for much of Meter’s software-defined network, and where the company thinks it will earn an edge over incumbents.

Over time, more of Meter’s in-house capabilities will get turned into software, including its upcoming networking foundation models. Doing so allows Meter to scale faster, more consistently, and more efficiently, in a way that competitors can’t.

3. It modularizes both the components inside of the devices – like the chips – and, importantly, the ISPs. With Connect, customers can enter their address and get quotes and recommendations from the Meter team. It’s like Kayak for internet. Meter wanted to make it possible to buy the internet on the internet.

I just signed a new office lease and used Connect to figure out which ISP to use. I had a personalized recommendation from Gayla on the Meter team within 2 minutes of submitting. Connect gives Meter and its customers leverage over ISPs at no cost to the user.

Email from October 15, 2024

4. Meter competes directly with incumbents, from Cisco’s Meraki to Juniper to Palo Alto Networks, and Arista.

It isn’t providing components or aiming to serve a small niche. It’s going head to head with the aim to win. Why Sunil and Anil think they can win is a masterclass in strategy we will touch on shortly.

5. Meter can offer a product that is better, faster, and cheaper.

With Meter, the whole network, hardware included, comes at one monthly price. It’s as simple as a utility, removes the costs that would otherwise go to a series of intermediaries, and improves as Meter introduces new products.

To make that more tangible: one customer was recently looking at paying an incumbent north of $100 million over five years; Meter will save that customer around 50% of the cost.

This is the point of vertical integration: to offer a product that is better, faster, and cheaper than what the incumbents offer, or can offer without burning everything down.

One theme that Meter makes clear is that while vertical integration takes a lot of time and effort to get right, once in place it lets you do things you couldn’t, do everything faster, and provide and capture more value.

In the past few months, Meter has released Command and Cellular, which gives customers perfect cell coverage inside of buildings for “5x cheaper install, management, and support than leading DAS (distributed antenna system) solutions.”

Each new product connects to the others. Customers will be able to find the ISP required for both their wireless and in-building cellular networks, including Starlink, through Meter Connect. Cellular signal will flow natively through the same Meter hardware powering the wireless network. And they’ll be able to manage everything through Command. And all of it can be set up in the same install, and included in the monthly fee.

Crucially, because Meter is vertically integrated, customers can spend 3-4x as much with Meter as they would with legacy vendors. There are a number of reasons for this:

• Meter sells customers more parts of the stack than a traditional vendor would, so they keep more of the value.

• It can increase margins by cutting out layers up and down the stack.

• Customers get more value from the network – they save time, for example, which has a value that currently sits unmonetized.

• And Meter can continue to add products that fit into the stack – like Cellular.

The equation is simple, Sunil told me: over a 5-7 year customer contract, the customer might get 10x more value than they would with legacy networking, and Meter gets 3-4x more value from a customer than a legacy networking company would.

It’s a win-win, “paid for” with the margin and drag vertical integration cuts out of the system.

Meter has “paid for” the right to earn that value with persistence. The fact that you’re just reading about Meter now, ten years in, is an intentional decision.

“From the beginning, we wanted to do the whole thing,” Anil explained, “So compared to legacy vendors, we should at least be on par before coming out. We didn’t want to bring any attention until we weren’t just a point solution.”

Now, they feel that they’re at par, and that now that they are, the advantages to owning the full system will only compound.

“After this year – I don’t think we’re there yet, but after this year,” Anil believes, “We will be the best networking company in the world. Then our advantages will accumulate on top of that.”

Accumulating Advantages on Long Time Horizons

There’s this quote from Stripe’s Patrick Collison that I love:

There’s something quite deep about the notion of using time horizons as a competitive advantage, in that you’re simply willing to wait longer than other people and you have an organization that is thusly oriented.

I’ve used that quote in maybe half a dozen Not Boring pieces, but something I didn’t pick up on until talking to Sunil and Anil is that the quote makes the advantage sound a lot more passive than it actually is.

The advantage doesn’t come from simply waiting longer than others. It comes from making decisions that work well enough to earn the right to keep playing in the present with an eye towards their future benefits, and then fighting a brutal, non-stop battle against entropy until that future comes to pass.

As John D. Rockefeller wrote of his partner Henry Flagler in his autobiography, covered on a recent Founders podcast episode, “He followed his convictions of building as though the trade was going to last, and as a result, he laid strong foundations for later years.” Sunil and Anil operate on a similar principle in a modern context.

There’s something like a matrix for thinking about what and how to build with an orientation towards the present or the future on one axis and whether you’re building a product or a system on the other.

To build a future-oriented system, like Meter is, you have to make trade-offs in the present that are also bets on the future.

In Vertical Integrators, I wrote that “A deep tech company designs a product for optimal performance, taking on the risk of unproven science or technology to do so, while a Vertical Integrator designs a system for optimal performance, taking a risk on the combination of proven technologies.”

After studying Meter, I think that’s half-right. It’s not just about designing a system for optimal performance in the present, but one that, if things play out the way you predict, will perform better than anything else in the future.

In other words, you might choose to do something that’s sub-optimal, or at least much harder, today in order to checkmate competitors tomorrow.

Moving from the theoretical to the applied, let’s look at Meter through this lens.

Meter started building hardware – even though it was potentially commoditized and certainly more expensive upfront – in order to build both software – the operating system and APIs – and a data pipeline with an eye towards being able to build a variety of apps on the same stack.

Over time, Sunil and Anil bet, there would be more and more packets, and with the right data and integrations, software would be able to handle more and more of the work of moving them.

On one of our Zoom calls, Anil drew me a graph:

How Meter Works (with one “application” coming soon)

The data, hardware, and software Meter has built over the past decade span all of Meter’s “applications”: ISP, wired, wireless, cellular, and soon, one more. That is critical.

Sticking to that plan has required short-term trade-offs.

Sunil explained that it’s harder to build software for all of the products than it is to build software for each product individually in the moment, but that if everything is part of the same OS and the same data pipeline, the advantages compound over time.

The hardest thing is keeping hardware, software, operations, and data moving at the same pace in the same direction. It is very easy for the balance to get lopsided.

“That happens a lot,” he explained, “The hardware will be incredibly ahead, then software is catching up, then you’re like oh shoot, from what we did on all these other sides, we want hardware to be different, but fuck, we already put that in production.”

The game becomes: how long can you have hardware, software, and data span all five applications? “Having all three span the entire platform gives us true power,” Sunil said.

Anil compared it to Tesla and SpaceX, both of which have had to keep a bunch of different things in sync and both of which are now dramatically benefiting from accumulating advantages two decades in the making.

But, having spent six years building and managing a network of distributed physical spaces at Breather, I realized that they were dealing with an extra level of complexity.

Tesla owners don’t get a new car for free every time a new model comes out, and even if they did, at least the car is a standalone thing that doesn’t have to work within an existing network and doesn’t have to be “installed.” It can just drive. SpaceX makes rockets in one place, and launches them in one place.

Meter, on the other hand, installs hardware in locations all across the country. They replace the hardware with newer models when they release newer models. The more they grow, the more complex it becomes to keep the entire network in sync. How do you deal with the inherent complexity of a network like that?

“This is the one thing I was hoping you wouldn’t ask,” Anil joked, “because it’s one of the secrets to vertical integration that I’m not sure I want others to learn.”

One of the Keys to Vertical Integration

I mean, that’s catnip. That’s why I wake up in the morning. The secret is a combination of business model, technology, and operations working in harmony, and of a bet gone right.

“Not only do we sell to customers,” Anil revealed, “but we buy out their existing hardware, install Meter, and then upgrade all of their hardware for free.”

That seems expensive and generous. Anil calls it selfish.

Here’s what can happen at hardware companies, even vertically integrated ones, according to Sunil and Anil:

There will be 10, 15 different versions of the hardware that date back 5, 6, 7 years. You’ve got different people maintaining different versions of firmware, different versions of operating systems, different versions of data pipelines. Then you’ve got support teams that you have to coach, every time a new support person comes, to the older stuff, and you have to maintain documentation for the new products you might be building on top of how they interact with it. And then you’re having to figure out how all of this is going to price for customers that are on different pieces.

And doing things this way actually makes sense if you’re selling hardware! If you’re Cisco, and you sell Meraki networking equipment, your job is to sell hardware to customers at a nice margin, and convince them to buy your new hardware to replace their obsolete stuff as frequently as you can get away with. Cisco has ~64% product gross margins and a refresh cycle of ~5 years. It’s a good business if you can get away with it!

Perplexity

The challenge is, you can’t force customers to buy your new stuff, so many of them will be running old hardware that you need to continue to support. Over time, the whole thing sprawls out of control, the organization slows, and it becomes harder (if not impossible) to build excellent software products to support the whole suite.

Meter, on the other hand, doesn’t sell hardware. It sells a Network. Its product is a system that continuously provides excellent internet, networking, and WiFi for a monthly fee. It doesn’t have to charge itself a margin on hardware. It locks in stable and predictable monthly cash flows against which to build and plan.

That leads to different choices, ones you’d only make if your model wasn’t selling hardware.

“When we release new hardware,” Anil said, “we have a greedy algorithm at Meter that we want to go replace all the old stuff.”

Because then we can go delete all the old code bases. We can go delete all the old APIs, all the different versions, and always singularly point to one thing.

Half the trick is that because Meter has software-like margins on recurring revenue, it can replace hardware as needed without substantially impacting its economics. Put differently, the long-term product and economic impacts of having everything operating on the same plane outweighs the short-term impact of replacing hardware.

The second half, though, is that Meter shouldn’t actually have to refresh its hardware very often.

Our incentives get aligned with the customer, because we just want to build the most lasting hardware, and they tend to be the beneficiaries of that. And then that kind of gets us this circular advantage. The kind of loop is that “Oh, man!” because of this great hardware, Meter has better software, Meter has a better product. And that kicks off this fortuitous loop for us because the incentives are aligned, because we’re not selling hardware to customers.

“We tend to overinvest in hardware way sooner than anybody else,” Anil told me. “Way sooner.”

Playing the long game, along with the right business model and architecture, gives you the opportunity to make bets that others might not. But it’s still a bet.

The Architecture Bet

I’m afraid that, despite all the complexity I’ve described, I’ve made this sound too easy, like Meter’s success to date and potential success in the future is based on a formula as straightforward as: put smart people to work on a hard problem with a vertically integrated solution and the right strategy for a long time.

In some ways, that’s true. Like I said earlier, there’s a sense you get from Sunil and Anil that they’d succeed in whatever they chose to do. But it misses the whole story, which, no matter how good the founders are, involves a willingness to risk it all, the combination of fortitude and educated luck it takes to make a bet and be right.

The history of success in technology is the history of betting on the right curve at the right time in the right way.

There’s that famous Jeff Bezos clip, in which he talks about betting on the internet’s growth in 1994 by selling books:

In this piece, we’ve already talked about another such bet: Metcalfe and Krause betting on the continuation of Moore’s Law in the early 1980s by putting Ethernet on a chip.

Meter itself was a bet: that there will be more packets zipping around the globe in the future than there were in 2014, when Sunil and Anil started a networking company.

Both Moore’s Law and the growth of the internet are general bets that many have made and won. But there are other, more specific bets that require an architectural understanding so deep that you can predict which specific technology a group of people will adopt years before they do.

When Sunil and Anil made a bet on its chipset architecture six or seven years ago, they made the latter kind of bet. It’s a very specific, technical bet, deep in the weeds, in the guts of the hardware itself. But it says so much about Meter’s strategy and capability that it’s worth going this deep.

Here’s why architecture choice matters:

The architecture you choose dictates the firmware, operating system, and software you create. Meter’s technical strategy has been to have the same hardware, software, and data span all of its applications for as long as possible to build compounding advantages. Since it sells a working Network and not just hardware, it wants to change out its hardware as infrequently as possible, but even if it has to change the specific devices, say to add new ports or a new generation from the same family of chips, the whole thing is predicated on building on the same architecture for as long as possible.

Taking a long view – a future-oriented, system view – might mean choosing an architecture that’s suboptimal today on the informed expectation that the architecture will become ubiquitous and cheap in a few years.

Think of the trade-offs. A young company, with far fewer resources than its incumbent competitors, chose to build its hardware on an architecture that was more expensive than the one its competitors were using. Immediate cost disadvantage. If you were trying to compete with Meraki by selling hardware apples to apples tomorrow, you couldn’t go with the expensive, unproven architecture. You would lose.

But if you take a longer view, in the context of the product you want to build and the way you make money, betting right can turn into a nearly untouchable advantage as the chips come down the cost curve and developer adoption grows.

The architecture they chose had a big advantage in the short-term, too, one that would pay off over time: it allowed them to write software faster, develop features faster, get to market faster, and continue to move faster. Speed has a compounding effect of its own.

So that was the early bet in its simplest form: higher unit costs in exchange for speed.

So how do you make the bet?

“To talk to them together is to get 90% of the words from Anil,” Hinkie told me, “and even the other 10% is a little bit of a dance that they do.” You may have noticed that in this piece to this point. There’s a lot more Anil than Sunil.

“But Sunil is deeper in a few areas, like architecture,” he continued. “He has this surprising backbone, this real, real ruggedness about what’s required.”

When we started talking architecture, Sunil took over, showing off a mind as comfortable with technical specifications as it is with markets and human behavior. There are benefits to reading all of the good stuff on the internet.

Because the bet doesn’t just come down to which chip is technically superior.

Superior in what ways, for which uses, first of all?

But as importantly, it’s a bet on what other people will adopt, this reflexive thing where you’re betting that other people will bet that this is the winning chip.

“What if you’re right that it’s the right architecture, but nobody else buys it and it doesn’t come down the cost curve?” I asked.

That, they said, is the nightmare scenario. The LiDAR scenario: all-in on the wrong technology.

Here’s how Sunil thinks about avoiding it, in his own words, because it’s such a fascinating glimpse into how to predict the future:

We take a few things into account, right? Which verticals the chips are deployed across, are these the highest end chips, what’s the power budget for them, what’s the cost per unit today, and what’s the software ecosystem built around them. Many things dictate the choice of the platform of the chip that we choose.”

We have to be directionally correct. It doesn't have to be accurate. We need to know where the architecture is going. It's knowing that, “Okay, so because the power budget is actually lower, and the performance is higher, we can bet that this is the chip that the market will adopt.”

And given that the market will adopt it, like which verticals will the market adopt it? Is it going to be mobile? Is it going to be data center? Is it going to be desktop? Where will that happen?

And given that, okay, if data centers are adopting it, then what would the total adoption rate look like? What is the total budget for that industry?

If it's, let's say, $1, and then you have a budget of, you know, desktops being $2. Then it's, “Okay, we can bet on the desktop ones, because we know that it will ramp up faster, and then we'll get pricing benefits for economies of scale.” So we take a few of these heuristics about how to choose a chip or a platform.

Anil jumped in to highlight that if you’re making a bet on a semiconductor, you’re doing it because you want the price to drop precipitously in five years as adoption continues.

“And why?” Sunil jumps back in, “Why you would want the price to drop is not just because it gets cheaper for buying the chips.”

It’s actually the amount of talent required to build on top of that, it’s actually smaller. It gets smaller and smaller, which is what we want, right? Having like 10 people building a product is better than having 100 people building a product because the cost of managing teams is very high.

So when the price drops, more adoption happens, you have more talent density, and then you can build faster.

I was fascinated. This was a masterclass, in real-time. So I asked if they had some skill or process that most people don’t have that helps them weigh everything differently than other people.

“Well, you see the face he’s making right now?” Anil asked. “You need somebody who can make that face. Basically, somebody who can shoot their shot. I mean this very seriously.”

Not the actual face, which was more intense

He went on, describing his brother’s uncanny ability to come to a hard decision:

We have an incredible amount of arguments, and I was totally wrong about the platform. And I know when he says something, and he says it a certain way, to take it incredibly seriously. Six, seven years ago, same thing on the Intel platform, even yesterday! I was arguing with him about a new version of Ethernet. I called him, he was at his son’s friend’s birthday party, and he went to the side because I was arguing with him about Ethernet ports.

But you need someone who can actually keep all of the context in their head and actually make a decision.

Can that be taught? Is that something innate?

Humbly, Sunil suggested that it just takes a lot of conversations, a lot of experimentation and just building, a lot of mistakes. He described the process as building an intuition.

Anil was more willing to point out something rare in his brother: “I think a core personality trait is that you shouldn’t actually care what other people think, in a respectful way, not in a very disrespectful way, you just shouldn’t care. Whatever genotype or phenotype or whatever somebody has around needing external validation just shouldn’t exist. Like zero external validation.”

“I think that can be taught,” he said, “But I just don’t know.”

Still, I pushed, it seemed like the power and performance of the chip is something that could be reduced to a formula. You can just make a decision based on those specs without caring what anyone else thinks.

Not quite. You have to simultaneously not care what people think of your decision in the present while predicting how other people will think in the future. This is the magic.

“There are a lot of chips you have to choose from, that are kind of similar” Anil said. “Then what he’s saying is the real thing to select from is the architecture of the chip:

What is the architecture of the chip itself?

What are the company’s [chipmaker’s] KPIs and their own roadmaps and company politics?”

Sunil jumped back in:

You need to look at the architecture and examine, do I understand this in one sitting? Yes or no? When you look at something, you want to make sure you can grasp it faster.

The reason is not because you’re smart or not, that doesn’t matter. It’s: “Will a group of people be able to do the same thing and communicate it faster?” And only then does adoption happen.

There are many technologies other than Ethernet that were really amazing, but Ethernet happened because it was so easy to understand, and it was cheap, and it got even cheaper.

So you need to make sure that when you're understanding a new technology or architecture, you want to make sure that it's easier to understand. It can be complicated, but if it's easier for you to communicate to other teams, other people on the team as well, then that option becomes faster, and that's when it becomes the winner.

All of this – technical expertise and an understanding of what other people will understand – went into Meter’s decision, six or seven years ago, to bet on the platform and ability to build software, even though the chips were more expensive and the platform was less popular in networking.

And the bet paid off! Sunil nailed it.

Betting right, early has meant “for the really expensive hardware, like our security appliances and switches, they can last 6, 7, 8 years and still be advanced compared to everybody else in the market.”

That duration matters more to Meter than it does to a company selling hardware, because Meter is responsible for replacing hardware, and because it means 6, 7, 8 years of compounding development of software, data pipelines, and applications on the same architecture, so that even if other companies end up adopting the same chips, your system is years ahead of what they can build.

It all fits together. Hardware, software, operations, business model. A whole greater than the sum of individually precise parts. The thing you come to realize studying Meter is that everything is handled with this level of thoughtfulness.

Meter’s operations, for example, are inspired by trips Sunil and Anil made to Japan a decade ago, when they were living in Shenzhen, and became so obsessed with figuring out how the country’s 5.5 million vending machines stayed refreshed enough to deliver hot meals that they traced the whole operation through to its source. That inspires how Meter’s hardware stays fresh, even as its footprint grows.

Bokksu

But that is a story for another time, because the point of the operations or the chips aren’t the operations or the chips, but the whole product they allow Meter to build.

There’s no better manifestation of the benefit of operating everything from hardware to software to operations on one plane than Command.

Internet on Command

Some trade-offs are simply trade-offs. There are benefits and drawbacks.

As I did my own internet spelunking on Meter, one thing that came up a few times is that Meter is great for more straightforward uses – like startups – and not customizable enough for large enterprise use cases. Serious IT teams and network engineers, the feedback went, still want to do things their own way: pick the hardware, design the network, and configure it with more precise control.

Certainly, while Meter’s customers include large, sophisticated organizations like Bridgewater, its customer list is dominated by technology company logos instead of large enterprises.

Some Meter Customers

To deliver a simple product that a smaller, faster-growing company wants, the trade-offs you make might include hiding away the complexity that real power users need (or at least want).

Meter’s goal, however, is to move every packet in the world, from the smallest companies to the largest. That means serving the customers with the most complex needs, too.

As Anil said on the Built for Trust podcast, “Simplicity should be in the delivery of the products, not the products themselves. Simplicity should not stop you from having the complexity that you and your customers demand.”

This is where Command comes in.

While the trade-off between simplicity and complexity appears in many ways, the clearest may be in the software people use to get information about their networks and take action on them.

For most of the history of networking, IT professionals interacted with their network via the Command Line Interface (CLI), uncovering its secrets and coaxing it to do better with arcane spells known only to the initiated few.

How Does Internet Work

The CLI is great because it has so many degrees of freedom and is so configurable, but it comes with a steep learning curve, it’s pretty ugly, and because it’s so configurable, you end up doing the same thing over and over again. So networking has shifted most of the work to dashboards, which are prettier, have a softer learning curve, and come preconfigured, but which trade off configurability and customization.

Cisco AppDynamics Network Dashboard

One of the points of technology is to eliminate trade-offs. That is the point of Command.

Nitan Shalon, who leads Command, said that the point of Command is to get the best of both worlds: the power of the CLI with the beauty and simplicity of a dashboard.

With Command, the IT team can create dashboards to monitor network, usage, health, and security, get real-time information at the device level, and take action on the network, all via natural language.

Meter Command

Command is the manifestation of the accumulating advantages of building everything on the same data, hardware, and software. It’s trained on Meter’s data, hardware, and software.

By simply typing, IT teams and even employees can ask questions of the network down to the specific device or client level: why is @Packy’s computer not connecting to the internet? They can build custom dashboards just as easily. It feels like a little preview of where software is heading.

Link to Tweet

Beyond understanding, though, customers can take action with Command. Tell the network to adjust, in English, and it will actually adjust the network.

Nitan gave me the example of one IT veteran who needed to configure a network for an event with a lot of people crowded in one space. He needed to figure out utilization on the access point in the space, how it changes over time, and potentially configure the traffic to make it more manageable. It’s not hard - he knew what to do – it’s just a laborious pain in the ass. If he gets it right, no one notices. If he gets it wrong, everyone’s mad at him.

Normally, that would take something like 45 minutes. With Command, it takes 5 seconds.

Command isn’t out to replace network engineers, but to let them do more. Like many professions we talk about here – from manufacturing to construction – the number of trained network engineers isn’t keeping up with demand for their skills. People who are interested in computers these days have moved up the abstraction stack, just like Nikola Tesla was an electrician but a modern Nikola Tesla would build something that takes electricity as an input for granted.

Finally, Command lets Meter provide even better customer support with a small team, which means that support costs won’t have to scale with growth.

Evan Jackson, the Director of Technology at Black Pine School in Berkeley, said that prior to Command, he would pull up the WiFi specs if there were an issue and send them to the Meter support team to fix the problem. “Now, they just look at Command and say, ‘Hey, this is where the problem is.”

Command is only possible, Nitan said, because “everything is on a single plane.”

“When you talk to a phone or an AP or a switch or a controller, very different things go on under the hood,” he explained. “It’s very different engineering to be able to hit ‘@’ anything, it masks a lot of different behaviors. If you want to be able to talk to everything, you need to be able to talk to everything.”

I asked Nitan if it would even be possible to build Command without owning all of the hardware and software. “Anything is possible – we put a man on the Moon,” he said, “The question is how easy it is to get to a certain level of performance, reliability, and cohesion across the stack. It’s a time curve.”

For another company, it might take a big team five or ten years. For Meter, it took 2-3 engineers - Nitan with frontend and backend help working full-time to build Command.

Nitan is exceptional, like so many of Meter’s 100 or so employees. Before joining Meter, he was pursuing his PhD in MIT’s CSAIL Lab working on applying AI to life sciences. A surprising amount of people at Meter – including Sunil, who’s a geneticist – were trained in biology, which isn’t that surprising when you think of Meter’s network as a living, growing, and (soon) self-repairing thing but is remarkable nonetheless.

His point in mentioning how few people it took to build Command wasn’t to talk about his or his co-workers’ talent, though: “It just speaks to the velocity with which, once the ecosystem of Meter is established, things can be grown and built, sprout, continue to iterate, evolve, and hit exponential curves.”

It’s an example of the seeds of the future-oriented systems approach that Anil, Sunil, and the Meter team have been building towards for a decade beginning to bear fruit. And it’s a hint that, while Meter is by no means the biggest networking company today, its place atop the stack in a decade may be an inevitability.

Why Won’t Cisco Just Do This?

Nothing is inevitable, really. Meter is a minnow in the networking space, and it is competing with very large, established, well-capitalized players who are full of very smart people.

A core piece of my investing thesis is that many incumbents in very large markets, as smart as they are, are locked into an old paradigm. As Jerry Neumann writes, “Existing businesses have a techno-economic paradigm of their own, the one associated with the previous technological revolution, and they have a lot invested in the old way of doing business.”

In this case, the old paradigm would be: sell hardware.

The new one might be: sell software-defined networks.

These companies are not dumb, though, and underestimated.

Sometimes old dogs learn (or acquire) new tricks.

So the moment I knew for sure that Meter was going to win was one afternoon, a few months ago, in a conference room in Meter’s San Francisco office, when I asked Anil some version of the question: “Why won’t Cisco just do this?” and he went up to the whiteboard.

I thought I was recording the conversation, but I never hit “Record,” which haunts me to this day. It was a masterclass in applied strategy that mirrors Sunil’s Intel bet, a combination of technical and organizational (down to the people involved) understanding of an entire ecosystem. It went something like this (reconstructed with his help):

Cisco, founded 41 years ago in George Orwell’s favorite year, remains the biggest player in networking. Its networking business does roughly $30 billion in revenue per year. As mentioned, it’s the most valuable networking company, with a $235 billion market cap. But it is also positioned in such a way that makes it hard to respond to Meter.

The main challenge is this: it acquired Meraki 15 years ago. Meraki now represents a meaningful and growing share of Cisco’s revenue in local networks. Now, because of internal politics, the company is looking to consolidate and deprecate the Meraki brand and move everything under Cisco. And not just the brand, but all of the hardware and all of the software. That is a “crazy painful thing,” and it’s leading customers to look for alternatives.

The upshot is that Cisco is spending the majority of its time on the integration as opposed to developing new products or systems.

Even if Cisco were to focus on new products or systems, it would have to fundamentally change its business model – from selling hardware to selling a network, hardware included – as a stable public company with a network of partners dependent on the current model.

My favorite of Hamilton Helmer’s 7 Powers is counter-positioning: “​​A newcomer adopts a new, superior business model which the incumbent does not mimic due to anticipated damage to their existing business.” Meter is counter-positioned to Cisco. If Cisco wants to compete, it will have to hurt its existing business, at a time when its focus is divided anyway.

Finally, Cisco is just not a software company, and it’s certainly not an AI company. As networks become more software-defined and AI-defined, Meter’s software chops and ability to recruit top-tier technical talent will provide a growing advantage.

In short, there’s $235 billion up for grabs over time.

But Cisco is not the only player in the space. He went through each one in turn.

Aruba/HP Enterprise, which owns 20% of the enterprise networking equipment market and is growing fast, acquired Juniper Networks for $14 billion last January. Juniper has about 5% market share, but it was early on marketing its AI capabilities before most other players. It bought Mist Systems in 2019 and announced that the combined companies would “blaze the path to AI for IT™ in this era of digital transformation!” Juniper had also acquired Apstera and an AI Ops platform to automate the management of data center resources as part of a push into AI data centers that now seems prescient. As TheNextPlatform wrote in January:

It is abundantly clear that networking is becoming more important in the datacenter and central to the efficient use of compute and storage, and it is also clear that the idea that networking will only be 10 percent or less of the budget of a distributed computing cluster is also in the past. Networking represents around 20 percent of the cost of a modern AI cluster, and we would not be surprised to see it go higher as low latency and bandwidth become more important. This is a good reason for HPE to contemplate an acquisition of Juniper.

While the acquisition makes sense, Anil thinks that given the shininess of AI, HPE finds itself in a situation in which the 5% market share acquiree is running things at the 20% acquirer, and that there will be huge challenges in how they integrate over the next 4-5 years.

That same logic that TheNextPlatform laid out for the acquisition has also pulled other players who once competed in Meter’s market to the data center, and rightly so.

Arista, which Sunil and Anil call the strongest technically and very well-run, has grown to a $126 billion market cap thanks to its focus on data centers, and particularly AI use cases. Fortinet, which has grown to become a $74 billion company, and Palo Alto Networks, a $124 billion company, both offer popular security products and are “shining bright in datacenter,” alongside Arista, according to Sunil and Anil. Each is up over 400% over the past five years.

Perplexity as of January 29, 2025

Those three, he said, are “entirely focused on what’s happening between models, between clusters, and things surrounding that.” They are much less focused on the enterprise.

Even so, he thinks there is a gaping hole that they’re not addressing that Meter will hit when it goes to the data center.

The last big incumbent competitor is Ruckus, which Anil readily admits makes really excellent hardware. Ruckus’ parent company, Arris, however, was acquired by CommScope for $7.4 billion in April 2019. CommScope is one of the oldest technology companies in the country – it laid a lot of the copper and fiber across which packets travel – but it’s in a bad spot. In 2019, when rates were low, it took out a $9.5 billion secured term loan, then rates shot up and its market cap tanked.

Perplexity as of January 29, 2025

Ruckus is great, but CommScope has nothing to invest in it.

Across the board, the set of incumbent competitors each have either split time and attention in a moment when focus is crucial, or a full focus on AI data centers. Very few are focused directly on networking, and fewer on enterprise networking.

All of this, of course, is Anil’s perspective on the situation. It is well-informed but clearly biased. That said, he and Sunil are betting the company on the accuracy of their assessment.

But to hear him tell it, at this point, ten years in, given the choices that the company has made and what they’re able to build now because of those choices, the opportunity is theirs to lose. Checkmate.

What we didn’t talk about that day, maybe because Meter doesn’t view them as true competitors, were the other startups. They’re worth touching on briefly, because they illuminate the trade-offs associated with different approaches.

The most direct competitor on paper is Nile, founded by ex-Cisco execs, which says they offer a Network-as-a-Service. Nile does combine more products and services into a single offering than legacy providers, but they aren’t as vertically integrated as Meter. They don’t make all their own hardware, and hand off ongoing network management to partners. WiFi blogger (yes, there are WiFi bloggers) Lee Badman highlights the challenge: each layer of abstraction – from customer to partner to Nile – adds potential delays and friction.

On the other end of the spectrum, Lightyear, which offers a product like Meter Connect, is compelling on paper because the company does software with a sprinkle of finance, both high margin, fast-growing things. It’s nearly impossible for me to see how they would use their position to move into Meter’s territory as easily as Meter moved into theirs. One of Meter’s customer acquisition channels is Lightyear’s business.

Both Nile and Lightyear, and other startups can build good businesses. In the short-term, by taking on less, they might be able to build businesses that grow out of the gate more quickly and efficiently.

In the medium- to long-term, however, my core belief is that most vertically integrated wins.

I think vertical integration is the only credible way to attack strong but increasingly modularized incumbents and potentially win. As I wrote in Better Tools, Bigger Companies:

For the past few decades, as software has eaten the world, entrepreneurs and incumbents alike have attempted to fix these industries with software. They’ve made them more efficient, perhaps, but they’ve only nibbled around the edges. Giving a power plant a CRM only does so much.

The fact is, physical challenges require physical solutions.

In other words, for networking physical spaces, most vertically integrated wins.

So the plan is simple:

First: build a vertically integrated product offering incumbents can’t match and startups can’t catch.

Then: grow it. Surprisingly, that might be the hardest part.

The Distribution Challenge

One of the painful lessons from studying technology businesses is that the best product doesn’t always win.

From my perspective, as bullish as I am on Meter, the biggest threat to Meter’s becoming “the last networking company,” or to achieving the $100 billion+ valuation I think is possible, is distribution.

Let me throw some names at you: Oracle. Bill.com. SAP. Microsoft. Salesforce.

Founded nearly half-century ago, Oracle is a $450 billion company that somehow remains at the center of the tech universe – as a participant in The Stargate Project and Trump’s preferred US buyer for TikTok with a founder who is the 4th richest person in the world – despite looking like this:

The one time I ever wrote a bearish piece was on my least favorite piece of software, Bill.com, in January 2021’s Bill-a-Bear. The stock nearly tripled in my face. Turns out, while I hated it, accountants actually liked the software, because they were used to it. (Thankfully, it’s fallen off since).

BILL stock, Google Finance

In that Bill.com piece, I wrote that its “Switching costs are so high that Hamilton Helmer used the company as the example for switching costs in his book on moats, 7 Powers. As Flo Crivello summarizes it, SAP’s switching costs mean, “Switching to another solution can be a months-long effort costing several millions of dollars, and much more in missed profits if done wrong,” and shared this screenshot.

Microsoft, while its software is in a league of its own compared to these other three (see: Excel Never Dies), its distribution is even more impressive. I was a long-suffering Slack bull, and I thought it had a much better product than Microsoft Teams. Then Teams did this:

Slack got bailed out by Salesforce in a $28 billion December 2020 acquisition. The fact that Salesforce was the acquirer (and currently sports a $334 billion market cap) and Slack the acquired is proof positive that product elegance is not a pre-requisite for success. In my recent Deep Dive on Rox, I wrote of Marc Benioff’s company:

And Salesforce, for all of its flaws, has staying power. Just because I don’t like the software doesn’t mean that the company is doomed to fail. Painful software plus good distribution and high switching costs is a hard combination to kill.

I could have just written those five names – Oracle. Bill.com. SAP. Microsoft. Salesforce. – and you probably would have gotten the message, but I really want it to sink in. The most common mistake I’ve made in analyzing startups is underestimating the importance of incumbent distribution and the strength of their switching costs.

I bring it up now, because I could write this same sentence about Cisco:

“Painful software plus good distribution and high switching costs is a hard combination to kill.”

Cisco has built one of the most sophisticated distribution engines in enterprise technology. They have over 50,000 channel partners pushing their products globally. Those partners are split into tiers, which creates healthy competition. They have extensive partner certification programs, which partners invest heavily in, creating high switching costs. An astonishing 90% of Cisco’s revenue runs through its partners.

As I was chatting with Claude about this piece and about Cisco’s distribution in particular, it sung Cisco’s praises – “Their partners become deeply embedded in customer operations, creating strong barriers to entry for competitors” – before adding something crucial: “The downside is that Cisco sometimes struggles with rapid transitions (like the shift to cloud) because their highly optimized, large-scale distribution model makes transitions more challenging compared to more nimble competitors or newer entrants.”

So two things can be true at the same time:

1. Cisco will be much harder to beat than it appears.

2. It is possible to beat Cisco with the right timing.

Timing is one of Meter’s core strengths.

Like its product, the company’s growth strategy takes the long, compounding view while leaving space to seize on rapid transitions.

Growing a Utility

First time founders think about product. Second time founders think about distribution. Exceptional founders think about both, and how they work together.

Meter grows in three main ways that take advantage of its product and business model:

1. It acquires new customers in existing spaces when they move in.

2. It moves to new spaces with existing customers when they do.

3. It expands the products that they can offer customers.

Building a vertically integrated utility is a very hard thing to do, as we’ve discussed, but one of the benefits is that every customer that Meter acquires is really both a customer and real estate that Meter acquires.

Any given customer rents the network while they rent the space, but when they move, Meter’s hardware and wires stay in place. Because customers don’t buy the networking hardware upfront, they don’t take it with them. The next tenant can simply call Meter and ask them for the internet to be turned on.

There might be tweaks, of course, as there still is with power. The new tenant might have more employees, or connected devices, or robots. Meter might install more hardware, or use Command to reconfigure the network using the existing hardware. But to the new tenant, it’s practically as simple as flipping a switch.

When the tenant moves in and becomes a customer, they’ll inherit a Meter support team – and Command – that is already familiar with the network. Evan at Black Pine told me that if there’s one thing I need to mention in the piece, it’s how excellent Meter’s support team is. They’ll inherit the accumulated knowledge of the space, and Meter’s accumulated knowledge of how modern networks should work in any space. Much of that knowledge will be captured in software that just adjusts automatically going forward.

So that’s one way Meter grows over time: it’s so easy for a new tenant to turn on Meter in their new space that they hardly consider anything else.

Another way that Meter grows is that old tenants become new tenants somewhere else. As a company outgrows one space and leaves its Meter network behind, it grows into a new one and brings Meter with it. Again, they didn’t buy networking hardware, so they’re not bringing it with them to their new space. They’re just asking Meter to run it back. This is predicated on Meter providing a great service at a good price – as long as it works, companies will carry Meter with them, like bees pollinating real estate.

This is true for individuals, too. The IT professionals I spoke to, and those who gave Tegus interviews, said that if they went to a new space or a new company, they would look to bring Meter in with them.

Meter’s approach might not be the fastest way to get to scale, but like everything else that the company does, it compounds over time.

If the average lease is five years, then in two decades, it will have gone through four generations of this compounding process, and it will scale exponentially, assuming it continues to deliver.

That, though, is very slow by itself, but faster if it continues to expand the base off of which it compounds. Of course, Meter will continue to acquire new customers, and will have more resources to do so as it grows. That will allow it to compound on a larger and larger base.

For a product that makes IT teams’ lives easier and more productive, word of mouth is one important growth channel: multiple people interviewed by Tegus said they heard about Meter from another IT professional, although Evan told me that they use Meter because he met Anil at a basketball game.

Another is Meter’s products themselves, and specifically Connect. If Connect becomes the place that companies turn to figure out which ISP to use – often the first step in the network setup process – Meter has the opportunity to sell them on doing the rest of the network with Meter, too.

It’s pretty elegant: a customer acquisition tool that makes Meter money upfront without costing customers anything.

Each new customer is both a new customer a new space, and each accelerates the compounding.

Each customer is also a potential customer for more Meter products as it grows the number of applications it builds on top of its hardware, software, and data.

Cellular, for example, is a separately-billed product that costs much less than installing a separate DAS, works with Connect, and can increase Meter’s share of wallet with existing customers while improving their connectivity.

The beauty of building infrastructure is that there are so many ways to expand the offering, as Anil highlighted to explain why he was so excited about building Meter, and as Nitan echoed when we spoke.

“There are so many things that are adjacent,” he said. “Networking is the core pillar, but there are so many adjacent things that are ginormous businesses only potentiated if you own the core pillar.” He gave me one example, on which I’m sworn to secrecy, and added that “the number of those examples is hundreds long.”

So, for its core enterprise networking product, this is how Meter grows:

1. It acquires a customer, through Connect, word of mouth, or good ol’ fashioned sales and marketing, and sets up their space.

2. When the customer moves, they can bring Meter with them.

3. The new tenant in the customer’s old space can become a Meter customer.

4. As Meter adds new products, it can sell them to customers new and old.

Given time and continued effort, that alone could make Meter a very valuable company. Cisco is worth a quarter-trillion dollars, a majority of which comes from its core enterprise networking business. There is a path for Meter to win that business with a better product without the upfront cost, capture more value per customer over time, grow more quickly, and become the internet utility. In that scenario, an eventual market cap north of $100 billion seems achievable.

There is an opportunity for Meter to build a business much larger than that, though, with the next step in Meter’s journey: the Data Center.

How Meter Works

Earlier, I wrote:

The data, hardware, and software Meter has built over the past decade span all of Meter’s “applications”: ISP, wired, wireless, cellular, and soon, one more. That is critical.

The one more is Data Center, and it means another way for Meter to grow off of the same platform: not just to new enterprise spaces or new products within those spaces, but into a new set of products for a new category of spaces.

Networking the Data Center

Data centers represent both a major test of Meter’s thesis that unified architecture and vertical integration can create compounding advantages in networking, no matter the space, and an enormous market at the heart of the AI wave that will be rebuilt multiple times in the coming years.

Last week alone, and granted, it was a particularly big week, all of this happened:

• OpenAI, SoftBank, Oracle, and GMX announced The Stargate Project and a $500 billion commitment to build data center infrastructure for OpenAI.

• Microsoft CEO Satya Nadella confirmed that “I’m good for my $80 billion. I’m going to spend $80 billion building out Azure.”

• Meta CEO Mark Zuckerberg posted that his company is building a 2GW+ datacenter so large it would cover a significant portion of Manhattan, and plans to spend $60-65 billion in capex this year alone.

Zuck on Threads

Assuming that China’s DeepSeek didn’t kill the demand for data centers (I don’t think it did²), this is going to be an unprecedented buildout with unique challenges.

“If you talk to the folks building data centers and ask about the big bottlenecks,” Anil told me, “the first thing they talk about is power. But people don’t ask what’s after that, and it’s networking: designing, deploying, and maintaining networks.”

“Power is a given,” he said, “but networking is a different beast.”

It’s the kind of beast that Meter has learned to tame over the past ten years. The other good thing about taking the longest view in the room is that every once in a while, you end up in exactly the right place at exactly the right time.

Like right now. Meter plans to enter data centers this year.

To start, it doesn’t plan to go after AI data centers. There are a lot of normal data centers that competitors have neglected in order to pursue the AI opportunity. Meter wants to serve them.

Still, data center networking is competitive, getting more competitive, and it’s not what Meter has historically done. Half of the competitive analysis earlier said something like: “They’re too focused on data center to compete with Meter in the enterprise.” And now Meter is taking the fight to them. How and why?

The secret is that Meter’s hardware and software will span the entire networking stack from the enterprise to the campus to the edge to the data center.

Not its current chipset, though. They’re making a new bet, one that they expect to last for the next decade.

This time, the decision started with the decision to go to the data center, and then worked backwards: “What does DC performance look like? Can we use that same chipset for campus? Etc…”

The biggest difference between the two is that in local networks, you have tens or hundreds of something. In data centers, you have hundreds or thousands. That means that power adds up quickly, which means that power consumption ends up becoming a much bigger factor in the decision.

I can’t give you a table on the different options this time, because the decision is still secret, but the takeaway is this. Meter has a lot more to lose now when it bets the company on a new chipset, and they’re going with a chipset that hasn’t been around nearly as long as their current architecture. They’re betting a much more valuable company on a much less proven chipset!

We have established that Sunil and Anil don’t gamble blindly, so what’s going on? A few things.

First, using a newer chipset isn’t as risky as it would have been in the past, thanks in large part to Zuck and Co. As he’s currently doing with Meta’s open source Llama models, Zuck open sourced its custom-designed data center technologies way back in 2011 in order to commoditize its complements. The project was called Open Compute, and it is still thriving today.

Thanks to Open Compute, the Varanasis told me, the abstractions in data center networking have standardized.

“All of the switching software and platforms that teams have built we can now use to test the new chip itself,” Sunil said. “That would have taken years in the past. Today, you can take the entire stack, put it on this new chip, and ‘instantly’ test how the stack is performing.”

In short, Open Compute allows Meter to move fast and break things in software before wasting years testing hardware, which means it can entertain “weirder” options.

Second, scaling laws apply here as well. The same architecture works for a local network on a 1 Gbps switch all the way up to an 800 Gbps switch in a data center. If Meter gets the stack right, it should also be able to scale down the DC version and bring that type of performance to local networks.

That the same architecture could handle both DC and local by simply scaling ports and bandwidth – “same chipset, just tuned down for power usage and port differences in software” – is the “dream of every networking company.”

The traditional approach requires different platforms for different speeds, like having separate car platforms for sedans vs trucks. Meter's approach is more like Tesla using the same platform for Model 3 through Semi. It matters because seamless scaling from local networks to data center has been a "holy grail" in networking - attempted through acquisitions (Cisco/Meraki, HPE/Juniper), but never achieved organically due to architectural differences.

Third, capturing the local network to data center holy grail gives Meter another set of compounding advantages. It’s a simple law of scale economies that you want to buy more of the same SKU to bring down its cost. By using the same small number of chipsets across offices, campuses, warehouses, and data centers, Meter can create more demand for each SKU and drive down costs more rapidly than a competitor that only handles one use case, or even handles both but with different chips.

Plus, all of the same less obvious advantages we discussed earlier apply here. Meter won’t have to maintain a lot of different versions of old hardware. Like enterprise, Meter will offer an as-a-service model in the data center and replace hardware as new hardware is built - selfishly, not altruistically. The same chipsets also mean that Meter can run the same software, APIs, and data pipelines in data centers that they run in local networks. Expanding the number and type of locations all built on the same platform just accelerates the compounding advantages.

For example, “We can bring Command right into the DC when we launch,” Anil told me:

The algorithms and the operating system, on top of that is the API, on top of that is apps like Command. And hopefully, in the next six months, large models.

Large models? From a networking company?

Yes. This is the last layer of the Meter stack, and the one Sunil and Anil believe will make Meter the last networking company.

The Meter Stack

Autonomous Networks

Meter is training a foundation model for networking to automate network design, configuration, and management. They will start with the enterprise.

Meter already does networking design, configuration, and management with a combination of humans, software, and small models (Command) today. It has been collecting metadata on each piece, and how they work together, the whole time. Now, it is training large models that will do more and more of the work autonomously.

Think of the design as the high-level blueprint for the network. It includes decisions like network topology, hardware components, IP address schemes, protocols, redundancy, and planning for future growth. Send in a floorplan and requirements, get network design.

Configuration, then, is the process of implementing the design by setting up and customizing hardware and software components. It includes things like assigning IP addresses and subnet masks to specific devices, configuring routing protocols, implementing security rules (firewalls, VPNs, access control lists), setting up quality of service (QoS) policies for traffic prioritization, and applying the right settings to switches, routers, and other networking hardware devices.

As it stands, each of these chunks of the process takes weeks or months, depending on the specifics of the space and the network requirements. A 100,000 GPU data center is different than an ISP colocation facility, which is the type of DC Meter plans to start with.

Management is the ongoing process of monitoring, controlling, and optimizing a network to ensure it operates efficiently, reliably, and securely. This is what Meter does today with its support team and Command.

All of these things Meter does manually today, it wants to automate over time, in order to create networks that behave almost like the biological organisms Sunil and Nitan are used to studying.

The idea of networks that can self-configure, self-optimize, self-heal, and self-protect has been a vision of the networking industry for decades (very similar to how everyone has wanted autonomous vehicles for a long time). There’s been some progress: software-defined networking decouples network management from hardware, intent-based networking allows operators to define business outcomes and let the network self-configure around them, and cloud-native architectures can scale and adapt dynamically.

But the big dream – the idea of feeding a computer a floorplan and getting back a network design in minutes, then installing the hardware and having it configure automatically and continue to manage itself over time – has stayed out of reach until now.

Meter has the path to make this a reality. The Why Now and Why Meter should be fairly obvious.

Why Now: large models have gotten much better generally. Good data and a lot of compute are pretty much all you need.

Why Meter: Meter’s vertical integration and data pipeline mean that they have the right data to train a networking foundation model right as it is becoming possible to do so.

So Meter is building models that could design networks, whether for offices or data centers, configure them, and manage themselves, like a networking architect and engineering agent.

This is one of those rare but massive opportunities: a solvable problem that can deliver substantial value.

Networking is one of the largest job categories in technology, behind software engineering, and demand for their talents is growing just as the supply of qualified network engineers shrinks. Like Cursor or Devin, Meter’s hope is that with its models, the best network engineers can do 100x more than they could have before.

To start, Meter’s large model will make suggestions through Command and human operators will approve or modify recommendations. The models will “just use Command to write software and say, ‘Here’s my suggestion on what to do.’” The reason Meter built Canvas into Command in the first place is so the model could eventually write its own software. Long game.

Canvas

Command, Anil says, is Meter’s Level 2 Autonomy (Additional Driver Assistance) and they’re working to bring greater automation in the future.

As they progress, keeping humans in the loop, they’ll build trust with network engineers who will rightly be skeptical and worry that automation could lead to security flaws and extended loss of business. It will be up to Meter to build the capabilities and trust step-by-step.

Whether they will actually be able to train the model that does what they hope it will is still an open question until it’s trained and operating to design, configure, and manage networks. While the problem is solvable, it’s hard, and it’s not solvable by just anyone.

Sunil and Anil believe that vertical integration is the only way to do this.

By owning the full technology and operational stack – its own hardware, software, data pipelines, APIs, design and installation data, support tickets, logs, and configurations – Meter owns unique and valuable data that it can use to train models.

This applies to pretty much everything Meter does. Meter’s installation partners, for example, go and survey spaces and feed data into one platform: floorplans, photos, videos, configurations, before and after snapshots.

By combining this data with operational data – what went wrong and what worked well after that initial design and configuration – Meter can train the model on how to design and configure networks better than it could with the install data alone. In the other direction, by knowing exactly how everything was installed and configured and tying those data points to support tickets, Meter can train its models to provide better ongoing management.

Anil compares Meter’s advantage to Tesla’s in self-driving. Tesla controls everything from hardware to software to data pipelines. Tesla today, he said, “is a car, a bunch of cameras, and then models taking action on driving, and learning as it does.”

It is an approach that takes a long time to pay off, but may be the only approach that gets to full autonomy in a way that makes any financial sense at all.

As Ben Thompson wrote in Elon Dreams and Bitter Lessons:

The Tesla bet, though, is that Waymo’s approach ultimately doesn’t scale and isn’t generalizable to true Level 5, while starting with the dream — true autonomy — leads Tesla down a better path of relying on nothing but AI, fueled by data and fine-tuning that you can only do if you already have millions of cars on the road. That is the connection to SpaceX and what happened this weekend: if you start with the dream, then understand the cost structure necessary to achieve that dream, you force yourself down the only path possible, forgoing easier solutions that don’t scale for fantastical ones that do.

Anil also believes that similar to driving, the networking solution space is finite. While the journey of the packet takes many steps, each individual choice is relatively contained. From here, go there or there. From there, go there or there.

Because of how networking works, Meter has a shot at “solving” networking with more data. And because Meter is vertically integrated, it can spin its own training loop by putting the model into production and learning from the results of the actions it takes within those finite bounds.

Taken together, Meter can train models on consistent real-time, time series data, use those models to design and configure networks from a limited and consistent set of options, deploy models at the push of a button, and manage networks autonomously.

It should make Meter’s better, faster, cheaper offering even better, faster, and cheaper. That is the point of vertical integration, and it’s accelerated by models.

As mentioned earlier, Meter has largely been adopted by smaller, faster-growing tech companies who value Meter’s simplicity and speed over the limitless configurability of incumbent solutions. With its models – already with Command, but even more so with more powerful models – Meter should be able to offer simplicity in the front with infinite complexity and configurability in the back.

And it should keep getting smarter, learning from new security threats and network issues, and immediately pushing updates through fully-controlled hardware.

This is the big bet, the same as Tesla’s: while writing handcrafted, hard-coded software for network security or configuration might have yielded better results in those categories in the short-term, more generalized networking AI that can learn will be better at everything in the long-run.

This is the Bitter Lesson, and it’s one networking point solutions will taste in the coming years.

As far as enterprise networking is concerned, it’s hard to see how another competitor could match Meter’s capabilities here. It feels like checkmate.

But what about the data center?

Meter Models Take the DC

The point of all of the hard, long-term decisions that Meter has made and continues to make is that it should be able to do anything that it does across its entire offering.

If Meter can design and configure an office network faster, it should also be able to design and configure a data center faster. While it is yet to be proven, that would be a particularly valuable capability to have.

In October, NVIDIA CEO Jensen Huang spoke to BG2 co-host and Altimeter Capital GP Brad Gerstner and Altimeter partner Clark Tang about, among other things, the buildout of xAI’s data center in Memphis:

From the moment of concept to a data center that’s ready for NVIDIA to have our gear there to the moment that we powered it on, had it all hooked up, and it did its first training…19 days.

When I spoke to them recently, Sunil and Anil were just as blown away as Jensen.

“Normally, if you want to bring up a cluster like this,” Anil explained, “the network design takes 6 months, installs take time, configuration takes time. It all adds up. It takes years. Musk did it in 19 days.”

By automating network design and configuration, Meter might shave valuable months off the DC launch process. As it stands, each of these chunks of the process takes weeks or months, depending on the specifics of the space and the network requirements. Meter hopes their model will trim this to minutes instead of months.

The big question Meter faces now is this: how applicable is all of the enterprise data it’s built up to the data center? Meaning, do all the advantages we discussed in the last section transfer to the DC?

This remains an open question, and it may be a trillion dollar one.

First, how do Meter’s models work in traditional data centers?

Then, maybe, next year or beyond, how do they work in AI data centers?

For AI data centers, the ones filled with racks and racks of GPUs, there are three kinds of networking:

1. Networking between GPUs

2. Networking between clusters of GPUs

3. Networking between the DC and the rest of the internet

Most of xAI’s brilliance in Memphis had to do with networking between clusters. NVIDIA’s Infiniband and NVLink, and the Ultra Ethernet Consortium cover networking between GPUs and between clusters. To the extent Meter touches AI data centers, it will focus on the third category, a slightly modified version of its current bread and butter.

To start, though, the question will be how well Meter’s models work in traditional data centers, like the ones that our packet traveled through, and that so many packets travel through every millisecond.

Frankly, if your goal is to touch every packet, these data centers are the honey pot.

And Anil is optimistic that its models will port well, both because Meter runs on a consistent platform and the internet runs on consistent protocols.

Ethernet runs on IEEE 802.3 in both environments. TCP/IP, DNS, HTTP/HTTPS, VLAN, BGP, OSPF, EIGRP, MPLS, IPSEC, NTP, and so many more protocols operate the same in both environments.

This has to be true. Our packet is able to complete its journey around the world, no matter what’s inside of it, where it’s coming from, or where it’s going, because of these open protocols.

There are differences, to be sure. There’s more wireless in offices than there is in data centers. And Meter will have to use its full stack to capture data on network design, configuration, day-to-day operations in the data center. It will start more manually at first, and then let models do more and more.

But this is the key insight, and one that made no sense to me when Anil first told me that Meter wanted to touch any packet, anywhere: a packet is a packet.

If you can move a packet in the office, you can move it in the data center.

And the more places you can apply the same hardware and software to moving packets, the better you get at moving packets than anyone else can. You buy more hardware, so your costs drop. You collect more metadata, so your models improve.

This is how you build the last networking company by touching all the packets.

The Last Networking Company

A decade into Meter’s journey feels like milliseconds into the journey of our packet.

On the one hand, it’s still incredibly early. Meter is doing double-digit millions in revenue, and although it expects to 5x this year, it is still three orders of magnitude shy of Cisco. Meter’s capital raised is in the double-digit millions, too. Its team size will soon hit triple digits.

It seems far too early to make the kind of claims I’ve made today, and maybe it is!

Meter has a tremendous amount of execution ahead of it. It has to launch into a market, Data Center, that’s new to Meter but familiar to a lot of smart, well-capitalized competitors. It has to actually finish training its large model, and then implement it. And if it gets all of this just right, it still has to overcome the distribution advantage that incumbents have built up over decades, and that is often much more powerful than I give it credit for.

On the other hand… although our packet is still safely ensconced in its local network when your computer’s operating system wraps it in the IP header that assigns its destination, though it may have thousands of miles left to travel, across prairies and airwaves and under the sea, that IP address comes with a sense of inevitability: whatever route it takes, it will get where it’s supposed to go.

That feels like where Meter is, doesn’t it?

The exact route is still uncertain. Will it dominate the enterprise or will it shift focus to the DC? Will it do both, because it turns out they really are pretty similar, or will it find its attention divided, even if its architecture remains unified? Will its models really work? Will it really be the one to build the self-driving network after all these years? Will network engineers trust it, even if it does? With time, the answer to all of these may be “yes,” but time is a luxury most startups don’t get.

Despite the inevitable uncertainty, though, to me, Meter’s success does feel inevitable in a way that it doesn’t for most companies, even ten year old ones.

That is in part because in the beginning of a new Techno-Economic Paradigm, the most vertically integrated often wins. And Meter is the most vertically integrated. Meter, in fact, reveals why vertical integration is so key in these times of turbulence: it is the best-positioned of all to build this new TEP’s biggest prize, a foundation model that makes everything better, faster, cheaper.

Certainly, the fact that Meter has taken the longest view in the room, has spent a decade sleeping on factory floors and making short-term painful trade-offs for the right to win over the long term, the only term that really matters, contributes to the feeling. Like a well-planned chess match, moves that made little sense in the moment start to make sense in the endgame. And then, suddenly, checkmate. Or maybe Go is a better analogy. A series of Move 37s bearing fruit.

In the end, though, it comes back to where it all started: with Sunil and Anil.

They are the ones who decided to vertically integrate. They are the ones who recruited the all-star team that could pull it off. They are the ones who have taken the longest view while executing violently and taking big bets in each of the days that make up a decade. They are the ones who have scoured the internet and the globe for little scraps of potential advantage that might, one day, pay off, and even compound.

I told you earlier that there are two ways to tell the story.

You might say, “Oh, I just used Zoom.”

Or you might describe what a packet is, how it is born, the journey it takes, and all of the infrastructure that supports it along the way.

The former story is easier to write. I could have saved 20,000+ words by just telling you, “Sunil and Anil are exceptional. Trust me.”

But to understand what makes them exceptional, you need to see the company they’ve built and the decisions they’ve made to this point. You can judge those moves for yourself. You don’t need to take my word. And maybe, if I’ve done my job right, you’ll get the same sense that I’ve gotten getting to know them over the past 13 months, that the decisions they’ll make from here will be the right ones too, even if they take a little time to pay off.

“Truly excellent founders are a compression algorithm,” I wrote, with Sunil and Anil among those at the top of my mind. “If you judge that a founder is truly excellent – has the potential to be an Edison, Rockefeller, Ford, Jobs, Huang, or Musk, which sounds outlandish until you remember that theirs are the magnitude of company we’re after – that tells you almost everything else.”

I didn’t come into that first call with Anil thinking that networking was an interesting category, but I quickly decided that Anil was a really interesting founder, so I figured it must be. Circular logic that turned out to be true.

To believe that Meter might be the last networking company, that it could really move every packet, takes a little bit of that circular logic, backed up by some proof.

As it stands, whether a packet is traveling through an office (Wired + WiFi), an ISP (Connect), or through the air to your phone (Cellular), Meter has a product to help it travel more smoothly. Soon, it will be able to tap into the packet motherlode, the data center (DC), where moving packets is like shooting fish in a barrel.

Of course, an office is not all offices and the data center is not all data centers. With the pieces in place, Meter will need to grow. Scale economics depend on growth. Moving every packet depends on growth. And becoming the last networking company depends on growing so much that you consume everything in your path.

The trick is, as with so much of the Meter story, recognizing that all of that growth doesn’t have to happen overnight. It can happen slowly but surely.

If you play this current trajectory out over a really long time horizon, say the next decade or two, you get a world in which Meter isn’t just bigger than today’s biggest networking companies, but bigger than any networking company has been.

What excites me so much about Vertical Integrators is that the ones that win will be more valuable than the incumbents they replace because new technology lets them do more, better, faster, cheaper, and at higher margins. Meter is the most mature expression of that thesis of any company I’ve covered. Vertical Integration and AI really should lead to winner-take-most; Meter reveals the mechanism.

How valuable is Meter if it pulls that off?

Meter owns all of the hardware, software, and operations in that process, so it both creates and captures more value than any other networking company theoretically could. It is expanding the uses of its hardware and software – from just local to local and DC – to bring down unit costs and increase performance.

It grows radially, with each new customer and their space becoming a node for further expansion. And as it adds more products, it can sell each of those customers more. It does this through a business model that is both more digestible for customers and higher-margin for Meter.

Now, it’s bringing all of the advantages of vertical integration to bear to build foundation models that make all of this work even faster and better at even higher margins.

So each of Meter’s products is getting better and cheaper as the company grows, and Meter offers more of them, so it captures more value while charging less, and now it’s making all of those products smarter, too.

Most importantly, because of the bets it’s made, the platform it’s built, and the time horizon it operates on, Meter can build faster for longer, can accelerate its compounding. Over time, slope beats intercept.

All of which is to say: even though Meter is small, barely 0.1% of Cisco’s market cap, it feels like checkmate. That statement might sound crazy for a few more years, until it doesn’t.

And what’s the prize if they win?

When Thomas Edison lit JP Morgan’s library on fire in 1882, it was already clear that electricity was valuable. The world’s most powerful banker put up with inconvenience, noise, smoke, and danger in order to have it. At the time, though, electricity was mainly good for one thing: lightbulbs.

Then, over the next few decades, electricity became a utility. Now, it powers everything from toasters to televisions to Teslas. Air conditioners run on electricity. Modern medical devices run on electricity. Semiconductor manufacturing runs on electricity, and so too do the data centers where the most advanced chips end up.

Those data centers require lots and lots of power, power in quantities Edison couldn’t have imagined. A single 1 GW data center under planning by a hyperscaler today will consume 10,000x more power than Edison’s Pearl Street Station. When Tesla and Westinghouse opened the Niagara Falls hydroelectric plant in 1895, there wasn’t enough demand for the 75 MW it produced; the City of Buffalo, its biggest customer, used just 1%. Today, we can’t get enough.

Which is to say, if Meter owns the internet utility, it will be quite valuable in the obvious way. Becoming the largest networking company is a quarter-trillion-dollar prize. Becoming the last networking company is more valuable still. But it will also create and capture value inconceivable from today’s vantage point.

The Varanasis have spent more than a decade contemplating and building for a world in which packet consumption grows like electricity consumption has grown. Everything in our world that is dumb, will become smart. We’ll hold ongoing conversations with the devices around us. Words to packets, packets to words.

Today, a handful of cars drive themselves some of the time. In two decades, all cars will drive themselves all of the time. As robots take on an increasing share of labor, work will run on electrons and packets.

As the price of intelligence falls, we will infuse our world with abundant intelligence, streamed in packets. Quadrillions will give way to quintillions will give way to sextillions.

Meter, if it has its way, will move all of them, wherever they move. The world will be blanketed in networks that spring up and adapt in real-time. Meter plans to be the one blanketing, as the possibilities for packets branch and compound in unpredictable ways to create unimaginable marvels.

Branching and compounding Meter for decades is exactly what Sunil and Anil want to do. One day, whether this decade or in three decades, I think Meter could be a trillion-dollar company, a number dwarfed only by the number of packets they move.

---

Thanks to Sunil, Anil, Sarah, Nitan, Swarit, Dan, Sam, Tom, Bunpa, Evan, Aman, Sehaj, and everyone else who taught me how the internet works and how Meter plans to make it work better.

---

That’s all for today. We’ll be back in your inbox tomorrow with the Weekly Dose.

Thanks for reading,

Packy

1. Empires of Light, Jill Jonnes

2. No sooner than those announcements were made, DeepSeek threw a wrench in the market with its R1 reasoning model, trained on much less compute for much less money than America’s leading models, and open sourced. This has temporarily deflated expectations for chip and DC demand, although many have pointed out that Jevons Paradox would suggest that cheaper intelligence will create the demand for even more intelligence, requiring even more chips and DCs. It will be interesting to watch it play out with regards to AI DCs, but one thing is almost certainly true: cheaper intelligence will mean more packets zipping around the globe, which will mean more traditional ISP DC capacity will be needed.