Loading...
Etched - Building AI Hardware to Make Inference Faster and Cheaper - [Invest Like the Best, EP.480]
This episode explores how Gavin Uberti and Rob Lockett built Etched, an AI hardware company focused on making inference faster and less expensive. They discuss why conventional assumptions about semiconductor companies and young founders proved limiting, and how challenging industry constraints led them to pursue a different architecture. The conversation covers prefill and decode workloads, low-voltage inference, memory bandwidth, custom interconnects, and the importance of designing complete systems rather than individual chips. They also explain why they believe inference infrastructure will become one of the largest markets in the world.
Ramp is the only platform built to make your finance team leaner, faster, and better, saving businesses 5% on average so you can focus on growth.
Ramp just opened for business in the United Kingdom. So, if you're running a business in the UK, you can now use Ramp's AI-powered finance platform to manage cards, expenses, bills, approvals, and accounting all in one place.
I run my business on Ramp, and so should you.
Learn more at ramp.com/invest.
Rogo is the AI platform purpose-built for financial institutions, serving tens of thousands of bankers and investors at hundreds of leading firms worldwide.
Rogo's AI agents autonomously execute large chunks of your firm's workflows. They can screen deals, draft CIMs, run buyer outreach, and conduct due diligence on data rooms, all with full security and compliance handled.
Every action feeds Rogo intelligence, the firm's context layer, turning years of deal experience into a secure, governed institutional memory that makes each new transaction stronger than the last.
To learn more, visit rogo.com/invest.
OpenAI, Cursor, Vercel, and Vercel all have something in common: they all use WorkOS.
To achieve enterprise adoption at scale, you have to deliver on core capabilities like SSO, SCIM, RBAC, and audit logs. Instead of spending months building these mission-critical capabilities yourself, you can use WorkOS APIs to gain all of them on day zero.
Visit WorkOS.com to get started.
Patrick O'Shaughnessy: Hello and welcome, everyone. I'm Patrick O'Shaughnessy, and this is *Invest Like the Best*.
This show is an open-ended exploration of markets, ideas, stories, and strategies that will help you better invest both your time and your money. If you enjoy these conversations and want to go deeper, check out Colossus, our quarterly publication with in-depth profiles of the people shaping business and investing. You can find Colossus, along with all of our podcasts, at Colossus.com.
Patrick O'Shaughnessy is the CEO of Positive Sum. All opinions expressed by Patrick and podcast guests are solely their own opinions and do not reflect the opinion of Positive Sum. This podcast is for informational purposes only and should not be relied upon as a basis for investment decisions.
Clients of Positive Sum may maintain positions in the securities discussed in this podcast. To learn more, visit PSUM.vc.
My guests today are Gavin Uberti and Rob Lockett, the founders of Etched.
A few years ago, when they set out to build a better AI chip than the largest companies in the world, almost everyone I called told me it could not be done. They've since done it, taping out a working chip on their first attempt and becoming the first hardware company founded after ChatGPT to do so. They now have more than a billion dollars of customer demand for their first product, and Etched is raising $800 million to build it.
Etched builds chips and systems designed to run AI models faster and at lower cost. They started the company in 2023, and the product is a complete rack for inference: the chip, along with the boards, the power delivery, the interconnects, and all the manufacturing to produce it.
We talk about the technical bets behind their architecture, how they hired industry legends and paired them with elite 22-year-olds, and why they believe inference will become one of the largest markets in the world.
I think you will find the story of what they've built hard to forget.
Please enjoy my conversation with Gavin and Rob.
Patrick O'Shaughnessy: Gentlemen, it's been three years or so, Gavin, since you and I last talked, which is nuts. At the time, I was just wildly intrigued by your story and what you were going to build. I didn't know a lot about chips at the time. I was considering investing in the company, and so I was calling everyone I could see about that could give me an opinion or something.
At the time, basically the consensus was these kinds of companies are not built by young people. In the semiconductor world, the best companies are founded by 40- or 50-year-old people that have had whole careers where they've experienced and learned all the problems and shipped multiple chips. Two 21-year-olds are not going to do this. It's just not going to work.
It was indicative of a theme, which was, “Nobody believes in us.” That's obviously changed a lot now. You can just walk the halls and talk to the people that have chosen to come work here. But in the early days, it felt like this was something that you had to face down.
What was that like, facing that down, when a set of incumbents and industry veterans and investors and everyone else sort of didn't believe in you? What did that anneal in you to build the company the way that you are? What was the impact of that?
Gavin Uberti: I think there's a sort of level of naivety required to think that you could build a chip better than every other AI chip ever built and build a company to do it way faster than ever has been done. We had the naivety.
There were many times where we would say, “Why isn't this possible?” and really push on it.
It turns out that everybody's answers are extremely siloed to a set of constraints that aren't true anymore.
The reality is the entire semiconductor and data center industry is built on buffer. What I mean by that is every part of the stack—from the EDA tools, to the power modules, to the circuit boards, to the chip design and standard cells—everything is built to be general-purpose for everything, not just in the data center, but IoT on the edge and so forth.
When you have a specific use case you're really trying to design for, you can change the constraints a lot.
I'll give you just a very simple example, which we're not the only ones who do. One of the things you care a lot about is the clock speed of your chip. It's proportional to the throughput of your system.
When you are doing finite analysis for different timing, what clock speed are you actually going to be able to run at when you tape out your chip? There's this concept called corners, which is what temperatures you're going to be able to run at this clock speed.
The default configurations for a lot of these EDA tools assume that you're going to be running your chips in freezing temperatures. Now, I don't know about you, but I've never seen an AI data center with ice in it.
So we can feel pretty confident that our chips don't need to run at full speed at zero degrees Celsius. In fact, they're never really going to be running below 80 degrees Celsius anyway.
Just by knowing that that's a constraint that doesn't matter, we can make a ton of changes throughout the entire system. That's a very simple one, but there are many more that get you 20% here, 50% there, 2x here, and these compound into a system that can be radically better for inference.
Rob Lockett: There are two kinds of people.
There are some folks who go purely on heuristics. They think a claim that young founders can go beat the incumbents in a world of performance cannot happen. There is nothing you could say to me that would make me change my mind.
But there's also people out there who are, of course, skeptical, but willing to go ahead and say, “I'll spend the time. I'll do the work. It is actually possible.”
For example, one of our earliest supporters was Mark Ross. Mark was a very prestigious semiconductor expert. He used to be CTO at Cypress Semiconductor, which sold for $9 billion.
When we met him, we were just a couple of guys in a dorm room. We came to him and said, “Hey, do you want to go build hardware for inference? We can be much faster than Nvidia.”
Mark said, “No, you can't. It will not work. But if you want to convince me, write a white paper, build a functional simulation, and show me.”
After a lot of very long nights, we went back to Mark and said, “Here's a simulation. What do you think?”
He was like, “Huh, this works.”
But he said, “To build a company like this, you'll need a large amount of capital. You only have $3 million. Even if you get up and started, that way you hadn't raised $5 million or even a lot more after that.”
Then he was again surprised, but got more involved. Then he became an advisor, then a half-time advisor, eventually a full-time CTO, as he saw more and more of the development and progress.
I think, in general, this gap acts as a filter to really heavily filter folks who want to go ahead and be right regardless from those who want to be truth-seeking and say, “Sure, I'm skeptical, but I will go ahead and work with the numbers myself. If I can figure out why this is possible, well, let's go build it.”
Patrick O'Shaughnessy: The specifics that you've made bets on, the way that you built this system, are immensely interesting to me. Because so many people are trying to do this now—build new chips that will do a better job of serving inference at massive scale—the world is interested in the research approaches and the different architecture approaches that people are taking to building a new AI chip.
I'd love to just start by describing what this thing is and what it does. But maybe more interestingly and more importantly, the process that you went through to decide what bets to take, what technologies to invent, and comparing and contrasting those with what you've seen the rest of the marketplace try to do.
Gavin Uberti: I think you have to start with the product. We're not just building a chip. We're building a full inference solution, and that means a rack.
That means the chip. That means the power delivery into the chip. That means the board. That means the interconnects for the chips to talk to each other. That means the production for mass-volume racks.
Really, the production is the product.
We think about how we get an advantage. There are two key parts of running inference: there's prefill and there's decode.
We have two key technology bets across both of these things.
Prefill is reading in a huge volume of text, and decode is then using that data to generate output tokens.
When you run prefill, your job is not to predict tokens. You already know the text. Your job is to get the model's memory—we call this the KV cache—into the right state.
Then you can go ahead and run decode with that same KV cache.
What we will often do is what we call PD disaggregation: prefill-decode disaggregation. You will have one cluster of servers running these prefills. You'll then transfer those model memories, that KV cache, over to the decode cluster, and then use that cluster to generate the next tokens.
It's sort of like loading the gun and then firing it. I think about it in super-simple terms.
Patrick O'Shaughnessy: It's getting the model's memory into the right state and then using those things to go do tasks.
Patrick O'Shaughnessy: Can you really think about this market a bit lazily? Is this a prefill chip or a decode chip? If you're a decode chip, is it an HBM chip or an SRAM chip? Are you using 3D DRAM? Are you using optics? Are you using copper?
Gavin Uberti: When we started this, we just wanted to understand why extremely smart people were working on these different directions.
We seriously looked at architectures like having a bunch of DDR memory and a shared memory pool. We looked at advanced packaging to basically break out of the shoreline. We looked at things like whether there was a way to put memory dies on top of compute dies.
In doing so, we realized that there's no free lunch. Everything has a trade-off, right?
With 3D DRAM, you have a thermal issue. You have a supply chain issue. You have to figure out hybrid bonding. You have to figure out the flop scenario for the decode chip.
So we went through everything, both on the prefill and the decode side.
In doing so, we realized there were a few design spaces that nobody had seriously tried to explore because they had never been done in AI chips.
We asked ourselves, “What are the actual metrics that are going to matter the most?”
Gavin Uberti: On the prefill side, the thing that matters is FLOPS and FLOPS density.
People talk about FLOPS often as a headline number, but in reality, you should care about the FLOPS you're getting when you're running real workloads.
There's this concept called MFU, or model FLOPS utilization, which is, for every peak FLOP advertised, how many cents on the dollar are you actually getting?
On GPUs, you often get somewhere between 20% and 50%, depending on the workload.
Actually, you probably cannot run at 100% because you have a thermal issue.
As you increase the FLOP utilization, you have more transistors going on and off, you draw more power, and the chip will self-regulate and lower its clock speed to make sure it doesn't overheat.
So, as we looked at inference, we said, “If we want way more FLOPS because we want to run at way higher throughput, we fundamentally need to solve the thermal problem before we even think about adding FLOPS to the chip.”
If I just add more FLOPS to a GPU today, or another AI chip, I'm not actually going to get more performance because it's just going to thermal throttle.
Fundamentally, the essence of that is this concept of Dennard scaling, which is that voltage is quadratically proportional to power.
If I 2x my voltage, my power goes up by 4x. If I cut my voltage in half, my power goes down by a quarter.
So we asked ourselves, “How could we run voltages lower than GPUs?”
We talked to a lot of people about this. We flew out to Silicon Valley after dropping out, and we asked dozens of people in semiconductors at all these different chip companies how they did it.
The answer we got was, “You can't. You can't run at voltages lower than GPUs.”
This was very dissatisfying because there were many different industries of chips that run at voltages lower than GPUs. Bitcoin miners run at under a quarter of the voltage of GPUs.
So this is obviously physically possible. The question is, are there issues with GPU architectures that make them unable to run at these voltages?
When we looked at the problem for a long time, we were able to create a new mechanism of running at much lower voltages—a new type of power delivery that we call low-voltage inference.
We think all AI chips in the future are going to be low-voltage chips. They're going to have to cram way more FLOPS into the same silicon area and, without thermal throttling, run at way lower voltages.
That's pretty simple.
Gavin Uberti: For decode, it is all a memory game.
More memory bandwidth means you can load the model faster, load the KV cache faster, and serve more tokens per second per user.
We think people ask the wrong question here. People often ask, “How much memory bandwidth is on your chip?”
You should be asking how much memory bandwidth is on your full-scale-up cluster.
What we are able to do is add way more bandwidth at much lower latency from chip to chip. That allows us to serve at much higher speed because you can use the SRAM and the HBM from the full scale-up cluster as a single pool.
That's a technical bet we call cluster-scale memory.
On GPUs today, the cluster memory can be very badly utilized because the time to hop from one GPU to another is extremely long.
For example, on Blackwell systems, it can be about 4,000 nanoseconds point-to-point. That means that if you go ahead and go to an 8xTP setup, you will get way less than an 8x improvement in tokens per second per user.
What we did is build our own totally custom interconnect stack.
Everything about the second layer is custom-built. Now we can get far better latencies and advance this way, too. We can cut this by more than a factor of 5x.
That allows us to use the memory of other chips much more effectively.
As you scale the world size, your time per token goes down proportionally.
Patrick O'Shaughnessy: It's not that surprising, given all these architectures were built before ChatGPT. If we're trying to build a chip for modern workloads, it's going to look very different.
Gavin Uberti: The way we organize our FLOPS, the way we do our voltage domains, the way we do our power planes—it's super different.
The way we do the packaging is going to look super different. The way we do the board design is going to look different. Then, inside the decode, the way we connect everything is going to look very different.
We're now bringing forward our first generation of this low-voltage inference technology, which is running at under half the voltage of any other AI chip.
Patrick O'Shaughnessy: If you zoom all the way out, why is this so important? Why is the delivery of much higher throughput, much lower cost per token, and better tokens per watt—all of these metrics that the industry is talking about more and more—so important?
Everyone knows the supply side of the equation is a big problem right now. Why is this, in the bigger picture, looking at a decade, the bottleneck in the technology world?
Gavin Uberti: I think it comes down to productivity.
We're at this extremely interesting moment in the history of civilization, where there is real artificial intelligence—not science-fiction stuff, but models that can solve problems that most humans can't.
It's going to create new scientific discoveries. It's going to create instant access to medical care and instant access to education.
And now it’s just about how many people can use it at the same time, how many products can serve this at the same time, and also the speed of doing different tasks.
So when you think about wall-clock time, if we can take an agent that can run at a certain quality and can take a year to solve a certain task using inference-time compute, if you have way faster decode speed, you can compress that into a month.
So the amount of scientific innovation and the amount of actual proliferation of technology will happen much faster.
And then the second part is concurrency, where today it’s just not possible for a billion people to use these models concurrently. Ultimately, some people are going to get downgraded, some people’s models are going to be slower, and some people just won’t be able to access the hardware.
A few years from now, there are going to be giant models serving billions of users.
We’re very much in the early innings of AI today, where the paid plans have only a few million users in the world using paid plans of AI models. So we’re at one one-thousandth of the global population actually using this stuff.
So if you want to serve a giant scale, a lot of things change. One of them is the number of chips that communicate together, where people usually think about this in the context of training.
You have these giant training clusters. You have clusters with over 100,000 GPUs that are all networked together. On the inference side today, people usually think about it as an eight-chip cluster, or maybe an NVL72—the scale of domain. But very quickly, this is going to become thousands of chips and tens of thousands of chips.
And the way to get the most performance there, where the time between sending data from one chip to another—that primitive matters way more than is getting cut right now.
So when we think about optimizing memory bandwidth for the system, you have to think about how fast these chips can communicate together, because they can only communicate really quickly with themselves and very slowly with other chips.
They’re not going to actually be able to serve giant models at 10,000 or 20,000 tokens per second. So we need multiple orders of magnitude of infrastructure built out throughout the entire stack, from the watt to the transistor to the token to actually bring this stuff to the world.
And I think that, if you look at most other goods—like the iPhone 4 example—they’ve gotten these economies of scale as a result. More money does not really buy a better iPhone. If you’re a billionaire or if you’re just the average American, you buy the same phone.
And tokens aren’t like that yet. We’re still in the very early days, where the general-purpose systems, relatively small ones, are kind of hand-crafting these tokens, like the image-generating tokens, back and later the Renaissance.
And I wanted to live in the world where you have the same economies of scale for token-making that you do for making, say, iPhones or cars or anything else.
I think that is one of the huge unlocks that allows a huge group of people to use the best-quality models. Economies of scale have all made capitalism very—I don’t know—fair. I think that allows you to go ahead and have the same product in many, many different hands.
And you’re able to go then to serve way more users than a single-scale-up cluster allows. You’d get closer to that point for concurrent serving, too.
Yeah, and also just certain products aren’t usable if they’re slow. So if you want to serve coding models and you want people to actually use them, there’s a certain number of tokens per second you need to hit.
So the question is: while maintaining that per-token speed, how many users can I serve at the same time?
And you can basically decide, “I’m going to shut off a bunch of the world from using this stuff,” or, “Everyone’s going to get a worse experience.”
So fundamentally, you need to find ways to push out the curve, and that’s why there’s such a pressure for new hardware.
Vanta automates compliance, so your team can spend less time on security reviews and more time getting customers.
Vanta cuts audit prep by 82% and gives you instant, up-to-date proof of trust without all the manual work, so you can close deals faster and with less friction.
Customers report a 526% return on investment, and more than 16,000 companies use Vanta, including Ramp, RV, and Snowflake. Get started with $1,000 off at vanta.com/invest.
Ridgeline is the first end-to-end system of record with embedded AI for investment management firms, running portfolio accounting, reconciliation, reporting, trading, and compliance on one unified platform.
Firms are moving off legacy technology and onto Ridgeline because of how far ahead Ridgeline’s AI features are compared to anything else in investment management software.
I’ve been hearing from a lot of investment managers about AI, and they fall roughly into two camps, with some unsure where to even start, and others convinced they can build their own order management system over a weekend.
The reality is that running an investment firm will always require governance controls and a single source of truth for your data, and no amount of AI enthusiasm changes that requirement.
Ridgeline is built exactly for that foundation, which is why I believe that the firms that come out ahead in the AI era will be the ones running on Ridgeline’s Unified Platform.
If you’re serious about your firm’s AI strategy, Ridgeline should be part of that conversation, and you can request a demo at Ridgeline.ai.
I like to take some time to step back and hear both of your stories for how you came to this idea and this company, and then kind of walk through what it’s been like to build it.
Because I think in so doing, we’ll understand the system that you’ve built for the company itself that will then be able to power subsequent generations of products like this one for this crazy inference feature that we’re staring down.
Rob, maybe starting with you, just take it however far back you want. But what I’m curious about, and your personal story was actually the very first thing I ever heard from either one of you many years ago now, which really kind of blew me away, I’m most interested in your motivation ultimately for being here doing this thing.
It starts back actually in high school for me. I’ve been very unlucky and lucky at different points in life. This is one of the tougher times.
At the end of my sophomore year of high school, I got injured at a martial arts tournament. The next day, I couldn’t walk for some reason. I thought it was something wrong with my SI joint or something.
I went through physical therapy and different types of scans. They couldn’t figure it out. Eventually, they found this big bump on my back. An MRI told me it was a tumor.
I was told I had stage-four bone cancer and under a 30% chance of survival. It was like a two-year, crazy chemotherapy, surgery, and learning-to-walk-again experience.
You go through something like that, it changes the Overton window of human experience and makes you appreciate what actually matters. And you also ask yourself, what are you going to do if you have the chance to live?
If you actually want to get through something like that, you need to be hoping for something. And I always knew I wanted to do something very impactful if I had the chance to get through it.
It took me a couple of years to figure out what that was going to be. At the same time, as I got to college and met a bunch of other people building cool tech, I got extremely excited by AI models, especially once GPT-3 came out.
I was like, wow, this is the first model that can kind of speak English. These things aren’t really smart.
What happened was, when GPT-4 came out, there was GPT-4V, which was the first model with image uploading. So I went through my camera roll and found a picture of my back with this bump on it before I was diagnosed.
And I said, “Hey, ChatGPT, pretend you’re an expert doctor. A patient comes in and says they have this bump on their back. What could it be?”
And it immediately said, “This could be a tumor. You should get an MRI and immediately go to the doctor.”
And I just kind of sat there still. It was like, it took me six months—and yesterday this feature wasn’t there, today it’s here. I go to show my parents, and I get this notification being like, “You’re all out of image credits today. You need to see if you can get a Pro plan.”
And I was like, holy crap. This is going to change everything. And we clearly don’t have the infrastructure to serve it.
There are very few things you can work on that can actually bring this technology at scale to the world faster. There are plenty of people who are super smart working on models. The fabs seemed maybe unreachable to work on, but it seemed like the hardware was all designed before ChatGPT.
Every GPU, every AI chip that was serving these models was fundamentally built before this and retrofitted to serve these modern models. There’s going to be an entire new wave of architectures that come out.
And what a more exciting thing to work on than bringing this to everybody?
From a very different angle, at the same time, I was running a startup incubator called Founders Fund, which incubated a bunch of different companies. Some of the earliest ones being Cursor and Anysphere, which merged, and Mercor and Etched went through it, and a handful of others.
At the time, as these models were getting smarter—it was 2022—I was realizing all of these companies are spending all the money they raised on compute.
And I had this realization, as I was working on some of my own stuff, that all the products I want to build are going to cost tens of millions of dollars a year in inference.
This is not going to be tenable. The cost structure of every software company in the COGS is not going to be zero anymore for an incremental user. It’s going to be quite high, and it’s going to be a function of inference.
And the economics of every business are going to also be in front of people using more and more coding agents.
So fundamentally, it seems like inference is going to be really important. And it feels like we are on a decade march for inference to become the biggest market in the world.
When you think about that, 10 years from now, if there are going to be these giant projects, everything in that data center fundamentally hasn’t been designed today.
We should go pick something and work on it. That’s kind of how it got started.
I’m really excited for you to go back about as far, probably early in high school, maybe even earlier, and tell your favorite hash marks on the timeline that ultimately led to your ambition to drop out of Harvard and start this company.
But first of all, that was at a company called Exa, where I did kernel development. I was 17, and a 17-year-old can’t sign legally binding contracts.
So I went ahead and did a traditional CYA. Then we had to sit me down and said, “Gavin, don’t wish I had this information.”
And Exa was one of the only companies that saw it and said, “Hey, maybe this is a good trade.” And afterward, building kernels, a number of other companies sent—it’s not about Bob—out for $400 million.
They did the same thing at Octo. They got bought by NVIDIA for $100 million.
But when you do this kernel work, what you realize is that the math is relatively easy. But to get high-speed decode, I think the math is data movement.
Almost all the work that you do is optimizing how you move data around a single chip or across multiple chips.
That’s why we went ahead and built this cluster-scale memory tech. We bring that interconnect time way, way, way lower. You can do way more movement and, as a result, get a much faster time to generate each subsequent token.
And build these crazy things we’re all talking about for doing a year’s worth of work in a month, or more than that in the future.
Can you talk about the competitive drive that’s evident in some of the high-school competitions that you participated in and won?
We did a couple.
For example, I was very active in FIRST Robotics. I was lucky to have a very talented partner, Safra. For a long time, we had a professional school team where we were about 20 guys, all working together as an authentic school at FIRST Tech Challenge.
The goal is to build a robot that scores the most points, among other things too.
In FIRST, they put a lot of emphasis around collaborating with other teams, trying to do really good documentation, and trying to get others inspired to go do the same thing.
For an inside-out approach, we were going to go ahead and do it this way: we’re going to win. And we did nothing else besides build the machine to score the most points.
As a two-person team, a two-person player, rather than a 20-person team, we were much, much smaller than almost every other team in the competition.
We figured that if we were going to specialize, if we were going to go out and do this, and when the game was really well, we would need to advance based on the quality of our documentation or our outreach.
We were just going to go win. And so we did. We embraced being a two-person team.
Both of us decided we were going to redesign it every three months. And we did. We actually had the world record for the highest score during this competition at one point.
We were rated by OPR third in the world for software development. And it was a damn good machine.
What from that episode can I translate as an analogy for how you built the company?
We think how you want to go ahead and do a full product like this for the couple key ideas.
One of them is velocity, velocity, velocity, velocity. You win by shipping. You’re not going to win by having the best outreach or the best communications. You win by building the best product.
And similarly, we think you can do it with a lot fewer folks. If you’re willing to just focus on product, product, product, and parallelize relentlessly, you don’t need 20,000 people like the big companies have.
You can build the best product in the world with far fewer people.
There’s just that saying: the best part is no part. I think for us, it’s also the best vendor is no vendor.
As much as possible, we want to vertically integrate the entire product, both because we get more performance and because we can move way faster.
So everything from the chips to the boards to the cold plates and the interconnects to even the production, we want to do all of it as in-house as possible.
I think we’re the only startup right now that’s building its own rack as well as its own chips. And we did it all at the same time.
A couple of years ago, the last time we were public, at that point we had just started building our rack. And we brought over an engineer who built all of NVIDIA’s HGX and DGX systems, which is 80% of the revenue.
And we said, “We’re going to build the rack at the same time.”
We actually went through multiple iterations of the rack before the chips even came back. Before the chips came back, we made thermal chips that had the exact same hotspots as we expected our chips to have, so we could build the cold plates, over-pressurize them, and blow them up.
We haven’t had a single leak since our chips came back with the cold plates, because we already validated them.
We have a factory in Taiwan, only a few dozen people out there. We built a clone of a bunch of the test stations in our office.
We have a two-megawatt data center on this floor, and we did 24/7 development cycles. People were doing day shifts and night shifts to actually get the hardware up and running as quickly as possible.
But it’s that extreme vertical integration and extreme parallelization of the schedule that lets you get products to market way faster.
If you think about the building of the early team and what it required as two young guys building this company, there are lots of very talented young entrepreneurs out there.
Maybe, for the first time of this scope or magnitude in a long time, all of whom probably can benefit from the lessons that you’ve learned getting very sophisticated talent and people to come join you even after careers at the other great companies.
If you were teaching this as a class—here’s how to get elite talent when you’re young and inexperienced and naive—what would be the advice?
We’re pretty bimodal in our talent philosophy.
It starts with what we call the legends. When we’re trying to solve an incredibly hard technical problem and generally do something that hasn’t been done before, we need to find the very best person in the world.
And often, the number-one guy in the world versus the number-10 guy versus the number-100 guy makes a huge difference in whether it’s actually possible to solve the problem.
We created this system we call project-based recruiting, where we map out all of the hardest technical problems across all industries that anyone has ever had to solve.
We look at temporality. So who are the people who did the zero to one? Who is in charge, quote-unquote, who actually did the work?
We talk to as many people as possible, and then we just track it. And you’d be surprised by the amount of people who say yes after the first conversation is pretty low.
But the amount of people who say yes after the 20th conversation is surprisingly high. You really gotta keep that up.
When you hear no from somebody who really is the best in the world, then that really means, hey, she doesn't come back when you had a few more miles to get up. I think it's one of the most convincing things to see: hey, we make bold claims, and when you hit those again and again and again, that is really belief-inspiring.
When we decided we wanted to build a rack and not just a chip, we were looking at this and saying, how many products have actually shipped that scale? For a rack-scale system that actually has the power density that we're trying to solve.
And we just kind of said, if we were going to wave a magic wand, what would the best possible person in the world look like?
And we're like, well, if we could find somebody who started that in Nvidia and built the entire rack team through all their different generations, learned all this different stuff, but still scrappy, still understands startup culture, but at a huge scale—like, that would be the best possible person.
So we mapped all of the different teams that related to all of the different rack-scale products at Nvidia. And we found three people that we thought could fit the bill. We talked to all of them, and two of them had just retired. One of them was planning to do one more generation for Nvidia and then retire.
That same guy is Brian.
And over time, we convinced him to join.
Brian started the HGX and DGX team at Nvidia, which was a majority of Nvidia's revenue—tens of billions of dollars a quarter.
And the other two guys ended up investing, by the way.
When you have somebody like that, they just know what good looks like, because they've seen it. And there's so many times where we'd talk to Brian, and he'd point to us and be like, “That's a billion-dollar lesson I learned. A billion-dollar lesson I learned.”
And, yeah, that just saves the cycles.
And you pair someone like Brian with somebody like Sanford.
Host: Do you have any name for them?
Guest: Brian's a legend.
Host: What's Sanford?
Guest: Yeah, we say, “Chips on shoulders, put chips in data centers.”
Sanford and Gavin, in high school, were world robotics champions. Sanford was finishing his senior year of college, and we called him up a couple years ago and said, “Hey, can you come check out what we're doing? We need some help on the platform side.”
He comes for a week, and we say, “Can you build a cold plate this week?”
If you asked any thermal engineer anything like that, they would think you're just totally naive. I mean, these things take months to do.
And to be clear, they do. But you can make real progress in a week if you put your mind to it and you think it's possible.
And he built a contraption in a week that actually de-risked a pretty key power question we had.
You put those two together, and they've done incredible things.
It's really about the other part, because you need the extremely driven people who just keep asking why and don't know where the bodies are buried to take tons of aggressive risks. And then you need the people who've seen scale and still have the startup-scrappy mentality to help them along the way.
It's really the legends plus some naive, “A-plus principles”-type talent. It's not just that you have both in the company; it's that they're working together.
Host: That's right.
If I think about that funnel, is there anything else more interesting to say about how much better you've gotten at recruiting and why those metrics keep getting better?
Guest: One of the shocking things is—and I'm sure it means a different theory in practice—it kind of self-selects.
The kind of person who is opportunistic and should go join wherever the hot company is, or think, “Go ahead and do the diligence; you will not come work here.”
And just one of the things I've worried about as we announce more and more of the product and its specs, we may have loosened it. So we're not very careful.
You kind of have to be sick in the head to join our company.
But you think about it on paper. It's like, you—a person who is probably a very accomplished engineer—they can make a good amount of money, it's liquid, it's predictable somewhere else.
You're going to convince your family to move to San Jose and live in this apartment on this housing program for the semiconductor company run by two 24-year-olds, now that's pre-product, that is going against the biggest companies in the world and the most supply-constrained environment ever created, with a design that they're saying is not going to be 10% better, but it's going to be 10x better.
Something must be wrong with you to do that.
People are just wired differently here. They really want to not prove people wrong who don't believe, but prove people right who do believe. They just take it personally.
And that's really fun to find those people.
And frankly, the nature of the company makes it very easy to whittle out people who aren't like that.
Host: One of the very first things you and I talked about, Rob, was—I started asking about Servo, which is the name of the first product here. And you said we can talk about that in great detail. But the thing you should know is that what we're really focused on is building a machine that can, at scale, produce these things and generations of them as efficiently and at the highest possible quality levels.
We want to build the company—or the machine that is the company—that is the thing that will produce this thing and subsequent things.
So I'd like to talk about a few principles or cornerstones of the company. We've alluded to some of them. You said velocity. You said vertical integration is becoming a more popular topic. Parallelization is something maybe that we should talk about.
But I'm especially interested in your guys' willingness to take huge risks to go faster. Maybe tell your favorite story about why this is the philosophy, what it's allowed you to do that maybe other companies haven't done.
Guest: One of my favorites is there was a time where we were getting close to taking out the chip. We realized, wait a minute, one of our vendors is way, way behind schedule, and we have two very bad options.
One option: just keep the current vendor and push our deadlines out by on the order of a year.
Another option: if we switch vendors, start over and also push timelines out by a year.
Neither of these was a good option. So we had to look for option number three.
And what that was, was we said, “We're out there all in Bangalore.”
But in the end, it's actually going and doing the work. We went out into the ship—I'm sorry, into the chip—and sent our top engineers across the world to Bangalore for six months.
I was there in Bangalore for four and a half months personally. Every morning, we'd go ahead and walk across crazy, busy Bangalore streets into the office. We'd be the first ones in.
We'd go out and put a wide variety of tools—both things like adding a huge amount of the code that was going in, building a bunch of tools to make this go even faster, making sure we're making the right design decisions on the spot right there—to allow our backend force to make decisions immediately.
And then, at one end, we'd walk back through the now-empty Bangalore streets and do it all again the next day.
We still had a bunch of the team in the U.S. We ran these 12-hour-on-each-side handoffs, where we had 24-hour development. At 8 a.m. and 8 p.m. every day, we'd all get on Zoom, we'd share all the data, and we'd say, “When I wake up, this must be done. We must get this chip out.”
And it was extremely intense.
At the same time, we saw other chips at the same stages as us with that same vendor that ended up taking years. They still aren't out today. They still haven't even taped out today.
And it's that level of extreme urgency that's required to bring products to market.
Host: What is the key to doing this well? This has become a trope because it will be Elon Musk's special skill, and others that seek to emulate him would try to do this too. He's figured out what the binding constraint is and just floods the zone personally on that thing, which is kind of like going to Bangalore or something.
It seems like this is a central tenet of the business, and of any business that's going to do this kind of vertical integration. What's the key to doing that well? What have you learned about that specific act?
Guest: For me, I think there are two key tricks to this.
The first one is that you can't build a chip alone. It's got to be a team problem, and your most important job is to go get great people to go with you.
And if you go ahead and are inspired and excited to do crazy things like this, that is a huge ask to say, “Hey guys, uproot your lives for six months,” or, in one case, 12 months.
We sent one guy out well ahead. It sucks, but we're lucky to have team members who are in it for the right reasons.
The second big thing, too, is being able to make decisions very fast.
One of the worst things is when there's a factory, or there is a vendor, who is waiting for you to go ahead and make some call and is just stalled. And this happens all the time, even for very small things.
You have to send folks to get a big amount of responsibility and say, “Make a reasonable call.”
Okay, if you're wrong every now and then, but I would have much, much rather be right most of the time and give an answer immediately than wait every time for the perfect response.
Speed wins.
Host: What about spending money to go faster?
There's this learn-by-doing, which has become so interesting. As the world has gone away from software and towards more hardware again, in the world of technology, we outsourced so much of the learning by doing—by shipping stuff overseas and effectively just being the idea guys here in the U.S.
That obviously is reversing, and you've adopted this way of learning by doing. You want to be in that iteration-learning loop. And part of that is a willingness to spend and take risks with dollars.
Can you talk about that a little bit?
Guest: I think it was a great quote: “The biggest risk is not taking risk.” Very similar here, which is, every day there's over a billion dollars of revenue in this category, and a lot of it's inference. So every day we don't ship, we're just leaving tons of opportunity on the table.
So your willingness to spend money should be extremely high if you can get a very clear ROI out of it.
We have this concept that we call pre-fetching, which is: when you're waiting for one thing to get done, when you know you're going to do other things once you have it, are there ways that you can parallelize the entire schedule?
For example, we know our chip is going to come back at a certain date. We wanted everything possible that could be done without the chip to be done before the chip lands.
And this costs a lot of money.
This means that we want to build our entire software stack beforehand. We shipped racks to customer data centers without our chips in them, with all the networking, all the CPUs, all the storage, all set up, so we could bring all those data centers online before the chips came back.
It meant that we took over 700 FPGAs and put the entire, full medical chip on an FPGA cluster and ran a dozen different models with our full inference stack on them before the chips came back.
It means that we built a thermal chip to mock the thermal profile of our chip and built cold plates based on that before the chip came back.
It means we had the entire production line ready. It means we did many revisions of the circuit board. It means the entire product was ready to go before the chips came back.
And this is what it gives you.
There's another very famous AI chip company that took 10 months to go from getting their silicon back to having them running inference in a rack. And this was publicly announced to their investors and was a really big deal.
We were able to do it in 40 days.
And it's because, by the time the chip came back, everything was boring. The software was already written. The rack was already there. The production line was already set up.
We were just go, go, go, go, go. Get everything together.
You don't always have everything. You make some tweaks on the fly, and then off you go.
Guest: In that particular case, the shift made a big difference, too. We went out and literally had a day shift and a night shift.
There were team members who would come in at 10 a.m. and leave around midnight. Others would come in at midnight and leave at 10 a.m.
You run it around the clock to get to those 40 days.
I mean, over half the company lives next to the office, so it makes it easier to do that type of thing.
Host: You pay them to do that, right?
Guest: Pay them extra, and you still do that.
Host: The invisible hand does wonders.
Guest: I mean, hey, it works for me, too. We're both there.
Host: Let's take one big step back and talk a bit about the broader ecosystem here.
The amount of shortages on the supply side, the exposure of risks in the global system, and the supply chain around this stuff has become, every day, Wall Street Journal front-page news.
The stocks that people are watching and investing in and excited about—the memory stocks—these were boring commodities, like nothing-burgers, five years ago. And now they're at the center of global attention.
If you just assess, because you've been building in it, the global connected supply chain that's required to make stuff like this possible, just riff on it.
What scares you? What's working well? What needs to change? What do you hope you change by virtue of how you build this thing? What's your assessment of this story right now?
Guest: I like that one of the most undervalued pieces of a supply chain story is that almost none of these things are—you buy them and they don't talk to them to begin. You have to go collaborate.
That is the most important part of the process of creating a system with TSMC or memory vendors. You need that partnership.
I think that, for TSMC in particular, people don't understand why it is so valuable. People look at the tech, and the tech is the best in the world. But for me, the real value is all in the service.
TSMC customer service is way, way better—and I have to hand it to them—than any other company in any other industry.
It's the kind of thing where, if you say, “Hey, we're going to improve your yield by making this change,” you can make them a recommendation, and they will go and run an experiment on their own dime, in our case, to see if they could actually get the higher yield.
And when we found that we were right, in the history of our work, they moved it over to the rest of the line.
That kind of thing doesn't happen in most industries.
If I go to a steelworks plant and say, “Hey, I want you to change the composition of the steel,” they'll say, “Screw you.”
Not TSMC.
It is why they are number one and why they're going to win.
One of the things that matters a ton is power availability and time to power. And the problem is, the more power you want, the more shortage there is.
It's actually very similar to chip clusters, which is like, why is Colossus charging $12 an hour for Blackwells? It's because they're the only place you can buy 20,000 of them at once, right?
Why is a 500-megawatt data center so hard to find? It's the exact same reason.
One of the things we need to think about is: how do we get way more juice out of each megawatt?
People are looking throughout the entire stack, whether it's improving the PUE, but also entirely new hardware, to get the most tokens per megawatt to solve this problem.
But fundamentally, building new buildings is hard. It's much easier to go from 100 megawatts to a gigawatt than from a gigawatt to 10, and 10 to 100.
And we are pushing the limits of what's possible on these timelines. So there are a lot of people trying to scale in their data centers as much as trying to scale them out.
Host: One of the interesting things about a system like this is what it replaces.
If I think about a rack like this versus, I don't know, the set of Blackwells or something—Rubins or whatever's coming next—how should I conceptualize that?
It's not just physical space. Reversal—talk about this. There recently was a call that literally said this is a problem: there's literally no space to put the systems.
How should I conceptualize what this represents or replaces in terms of other units of compute?
Guest: Here's how customers think about deploying models generally.
When I'm building a data center or I'm building a cluster, it's not in the abstract of, “Oh, I like these chips, and this is the power footprint,” and so forth.
It's like, I have a real production workload I'm trying to serve, and for my product to be useful, there's a certain speed I need to serve at.
For certain products, it's really fast, and for certain products, it's really slow.
Whatever the speed is, this is my speed. The question is, in a given amount of power, how many users can I serve while guaranteeing that speed? So another way to put it is: what is my interactivity? What is my throughput?
We are just finishing kind of the early innings of the AI infrastructure boom, where people really just cared about speed. GPUs were not able to reach a lot of the speeds of other types of chips, like all these SRAM chips—thousands of tokens per second—and that enabled tons of new use cases that got people very excited.
There’s an entirely new wave of AI chips, us being one of them. They’re all going to be able to hit these speeds. The question then is, if you’re hitting these speeds, what is the number of users you can serve at the same time?
And by proxy, if I have a 100-megawatt data center, how many software agents can I run at the same time?
So when people are doing that evaluation, our hardware is generally going to be able to get you an order of magnitude more concurrency at a given level of interactivity. That directly translates into tokens for what? Tokens for dollar? All the things people care about when they’re actually serving these giant mixture-of-experts models.
There are these now-famous interactivity curves. Not many people publish them, but you can see a Blackwell curve. You can see an AMD curve, which is a little bit worse than Blackwell’s, and it’s still an $800 billion company.
So if you think about what then the impacts are of shifting that curve, not just a little bit further out, but much further out—what are the things that most excite you about what this will enable?
I want to go out and solve some of the hardest problems, and I want to go solve these in less and less time.
There were things growing up that I was not sure I’d be able to live to see. For example, the unitary cubic conjecture—one of the things we talked about in college—and I was not sure I’d see that proven in my life. This was done by an AI model, and it was not over a long period of time.
If you’re able to run the same model ten times faster, you can shrink the time to go have these breakthroughs. And there’s a huge number of other problems in math like this as well that I worry will take thousands of years to go through.
Do you either have a much smarter model, or a model of the same intelligence running much faster? You can shrink that, and I can see it. It’s so cool seeing these breakthroughs get made. I am so excited to see much more of this happen.
I think too often people think about tasks and applications and stuff in these very short time horizons. Doing a chat and it’s 50% faster is nice, but it’s not game-changing.
As these agents go to longer and longer time horizons and the models get more and more capable, you’re going to see gigantic bodies of work that would take months of compute.
We think about this in wall-clock time. If you talk to a pre-training researcher at a lab, they’ll tell you that wall-clock time often is one of the most important things that matters.
What wall-clock time means is the time from starting your run to finishing it, to actually get data back. If you can shrink this time from a six-month run to a two-month experiment, you’re going to be able to do many more iterations, and people will make changes on the model or the architecture to actually improve the wall-clock time.
It’s very similar here in terms of how we think about the use cases. The exciting part about super-low-latency decode is that wall-clock time on long-horizon tasks becomes much shorter.
So a year-long compute build would now take a month, and a month-long compute build will now take three days. A three-day compute build will now take seven hours, and so forth and so forth.
That’s the thing that I think is really hard to internalize, because the models are just getting capable enough to do this stuff.
I thought it was really cool months ago when Cursor published that they had a bunch of coding agents build an entire browser from scratch in a week. Totally nuts. And that will soon happen in under an hour.
There are going to be many of those types of things that are going to happen, with these massive parallel agents all working on a given task.
What are the ultimate limitations of these systems? Is it just a physics question—how many times faster and cheaper can we get, theoretically?
There’s a lot of room at the bottom of this. If you think about chip-to-chip latency on an NVIDIA product, it’s looking at 4,000 nanoseconds to go from one chip to another. We’ll be able to do much better than that.
What’s the mathematical limit? It’s the speed of light. You can do it in just a handful—two or three nanoseconds.
And there are 4,000 today.
There is a lot of room at the bottom. The same thing is true of power efficiency. Sure, we’re able to go and bring the power down by so much, but you can go lower. You can go much, much lower.
It’s very challenging, but when I think about 20 or 30 years in the future, I think it’s inevitable.
And also, for economies of scale, for cluster scale-up—for a long time, HGX was the biggest scale-up. Then they had the NVL72, bringing it to 72. But you can be way, way bigger.
You look at a fab, for example. You have a fully billion-dollar, single, monolithic building with only a handful of lines running through it. You could have the same kind of thing for some futuristic, megacluster—$40 billion, $100 billion—as a giant megafactory serving one or a handful of models for a massive number of users.
To get that same economies-of-scale thing: the same model, a massive number of people.
Kernel engineering was your first job. That has emerged as a thing that nobody had ever heard of in their lives, to now something they hear about all the time—the importance of it, to eke more performance out of the bare raw metal.
When will that just be something that AI does entirely as well? Are humans still the best kernel engineers? Are they doing it with the assistance of AI systems? How far down will humans still be in the loop of designing these things? When will that go away?
Today, it’s all very hybrid. The best kernels are still written by human-AI collaborations.
Also, any AI model is built with these fundamental primitives, like matrix multiplications, convolutions, and chip-to-chip operations—collectives. Making these overlap and making these really fast matters enormously.
In the kernel engineer’s job, it takes a huge amount of skill: how do I allocate memory? How do I verify that if there’s some issue, like every transmit, it doesn’t stall the whole pipeline?
These things are very challenging, but they can go and make your model, say, 3% better. And you can do so many.
When we thought about our software stack, we wanted to go to the state of where the puck is going to be. Three years ago, there were kind of two ways you could build software.
One was effectively graph compilers. These things are not very performant, but they work out of the box. They don’t require a human to come in and tweak all the kernels.
But we went the opposite direction. We are kernels-first programming, and that means that for a long time it did not work out of the box. But if you work with kernel experts, you can get incredibly, incredibly high performance.
And the thing about this is that now, as the coding models get better and better, they’re doing more and more of the kernel-generation task. And when the models keep getting smarter, they’ll eventually do all of it.
All of it. It will become superhuman. So we’re going to build before the world is going.
Even today, we think about our profiling tools and debugging stack. We think about it from how the model will use these tools, more than we think about how humans use these tools.
We sometimes run experiments internally, and we had Codex actually get GPT-OSS running from scratch, just based off of our docs, completely by itself. It did it overnight.
We think about gain selection a lot, and what we mean by that is making sure we’re investing our energy in the right bets. Because regardless of what you choose to work on, it will take tremendous effort.
One of the things that we started with was the decision explicitly not to build an arbitrary graph compiler, not to support arbitrary PyTorch, not to support arbitrary CUDA, and not to support arbitrary ONNX graphs.
Instead, we envisioned a world where there were going to be under 100 models that actually mattered. They were all going to look very similar from the underlying mathematical perspective, and we were going to build primitives using physics that were going to accelerate these as much as humanly possible.
We were going to allow the most sophisticated customers to have direct access to the hardware and do whatever they want.
That has saved us a tremendous amount of time by not having to build a compiler, and that has allowed us to actually get much more performance.
And finally, when we started, a lot of people dismissed this idea. The only people that took it seriously were in high-frequency trading.
They hate compilers too. They’ll write their own kernels.
We’ve had dozens of people from high-frequency trading join the team because they saw this philosophy too.
What are the limits to vertical integration? How do you know where to draw the line?
The circumstances of the broader market are really interesting to me, where the vast majority of AI chips get bought by a very small set of customers. Many of those customers are themselves trying to design their own AI chips—OpenAI and Google, for example.
It seems like a very funny circumstance for the most valuable thing in the world: it all kind of flows through a couple of chipmakers and a couple of chip buyers. They all seem to be thinking about doing each other’s job.
And then you’ve got the circumstance where these things go in a data center, and you’ve got neoclouds and inference providers and this other part of the stack. You’ve got model builders and providers.
I can imagine a world where, because you have the best hardware, you design models and you build data centers. What’s unique outside of your current vertical? How do you think about where to draw the lines for the business?
We have a saying that production is the product.
Ultimately, what matters here is we know inference is going to be the biggest market in the world. Whoever produces the most tokens is going to be the most valuable company in the world.
So all the decisions we make are: how do we get the most token capacity online as possible?
Part of that is building a really good product that has way more throughput, that can run way better latency, and so forth, so we can, per chip we make, get way more tokens online.
Another part of it is not doing parts of the stack unless we absolutely have to get to the giant scale.
There are parts that we decided to do because it was absolutely required to get the scale, like building the rack instead of just building the chips, and doing a CMO model instead of doing a JDM model.
But there are parts of it that are kind of noise to us right now. We’re not going and building our own data center today. That doesn’t actually help us get more capacity online.
In general, our customers are actually making power and moving their clusters around to get our chips online because they’re such high-throughput.
If there was a world where other things were constrained, we would totally go and integrate with them. But the reality is we’re just purely focused on getting as many tokens online as possible.
I think this comes down to economies of scale again. I sort of part the stack, and there are huge economies of scale in some areas and others there aren’t.
For example, on designing models, there are huge economies of scale there. For chip fabrication, same story.
But if you think about building some small metal part of that rack, there’s not that same effect.
We think the natural boundaries are on the chip side, on the bottom, and on the model side, at the top. And we’ll fill the whole gap between.
A few weeks ago, there was a guy who was running a next-generation AI chip for one of the frontier companies. He was trying to recruit one of our architects, and I actually kind of did a reverse card.
Instead of recruiting the person, trying to recruit our guy, within a week we hired him.
I was going on a walk as we were finalizing the offer. I said, “You’re leading the super-important project. Why are you deciding to join?”
His answer was super interesting. He said, “Fundamentally, it’s not existential for my company for this product to win.”
For Google with TPUs, the revenue comes from search. Google won’t fail if TPUs fail.
Meta won’t fail if MTIA fails. Microsoft won’t fail if Maia fails. OpenAI won’t fail if their chip fails.
Ultimately, this is our product. It is completely unsurprising that the best chip in the world is built by a company that only builds that chip.
It’s NVIDIA, right? For us, it is completely existential for us to get as much token capacity online as possible.
It recruits a set of talent and attracts support from suppliers and from customers that view it with the level of intensity that we do.
Look at their FLOP specs. You compare any of these chips built by the labs or by the hyperscalers. The FLOP density—for say, FP4 times FP8—is lower than the Blackwell B300.
And that makes sense. They don’t have to take the risk. They just have to go and build a similar-enough product and not pay the NVIDIA tax.
As I think about you guys building the solution, the process of doing so is solving a sequence of really hard challenges. What has been the single episode that was the hardest to overcome?
We were designing the chip. We built this massive FPGA cluster to go and verify the full chip work.
FPGAs are digital entities. You can get digital logic, but not analog logic.
It turns out that when the chip came back, we began to see issues in our tensor-flow adaptation, with incorrect results. We realized, wait a minute, there’s a problem with the back-pressure logic across a clock-domain crossing. It’s failing.
This is a good way to produce wrong results, and it’s very hard to solve.
We realized there was one and only one way to solve it: we had to line up two clock signals on our chip to within 50 picoseconds. That is literally 50 trillionths of a second.
We had to get these signals aligned to this super-small granularity and do it on every chip two billion times a second.
A lot of people said this was impossible. We had people quit. People literally were like, “This problem is unsolvable.”
Well, when you have a problem like that, step one is: okay, let’s assume the problem is solvable. How would it be solved?
First, we realized what we had to be able to do was find a way to move our clock phase by a picosecond—ten picoseconds. And we had an idea: what if we had these two clocks adjust just a little bit apart from each other?
We figured out that if you go and figure out the phase, and then use a drifting mechanism to wait for just the right amount of time to get that always lined up, we can do this extremely reliably and lock the phases exactly where they have to be.
We can guarantee this never happens.
People were, I think, blown away that this worked and that it worked as well as it actually did.
But we made it work.
“How long did that take?”
That was actually about two weeks.
“It was three weeks.”
It was a very scary two weeks. But it was the kind of thing where, when that kind of thing happens, that is the most important time to go ahead and be investing effort.
That is the hardest time to go do it—when you feel like things are hopeless. But the sooner you solve that problem, you can get back to building and scaling up production to mass volumes.
I think a lot of our story is, as Gavin says, assuming it is possible.
Assuming it is possible to have a chip with way more FLOPs. Assuming it is possible to have a system with way lower latency between chips. Assuming it is possible to create a shared memory pool that can run at way higher bandwidth.
How would one do it?
A lot of the time when we do experiments, we’ll do dozens of experiments and all of them will fail. But we only need one to work.
There was one time—I mean, Gavin, I think, was leading the charge in our chip bring-up—with, I think, 30 different board experiments. Three of them worked, and all three of them were worth their weight in gold.
This is one of the things that people are going to be like, “Gavin, almost none of this works.”
You’ve got to get lucky once.
One idea for one of these stories that I’m asking about—difficult moments in the company’s history—is around the ability to raise capital to fund the thing.
I think when you started it, you needed capital, but you did not know you needed the quantum of capital that you’ve ultimately raised and are spending to build the solution.
And you hadn’t raised money before. These are all new things, right?
There were moments where it was really, really difficult because I was there. I saw it.
Moments were extremely difficult to raise the money that you did, without which the company would not exist. It would have died. And, like many great stories, there are many near-death moments.
But money specifically in this new world—this isn't software. You don't just need a little bit of money. Maybe you could tell the story about the true hardest part about raising money early on, before you had something that you could show people and be so proud of, and so you could show them other socks off.
It was just you guys talking about an idea, but talking about the early difficulties raising money, because it was pretty hardcore.
We've had some intense moments.
It reminds me of probably early 2024, before we raised our Series A. We were at this point where we had done enough of the architecture, done enough of the design, that we knew that the chip architecture was sound. We had to go build it. There was a lot more to do.
We were ready to go into what's called the physical design stage. We needed to sign an agreement with the physical design vendor, which would cost you at least $40 million.
And then we had this realization as the models were getting bigger and bigger. You were seeing these giant ML models come out, and we were going to need to build the entire cluster—not just the chip, but boards. We were going to need to build interconnection, build cold plates. We were going to need to figure out all of the networking and everything.
This was going to cost a lot more than the $15 million we had in the bank.
And, man, that was scary.
You're sitting in that moment thinking, “Holy crap. We can't afford this.”
I'd be looking at it like, “Huh. How hard is it to go back to Harvard?”
At the end of 2023, we put together this memo. We spent 100 hours on this because we had no idea how people were going to believe us when we asked for the amount of money we were about to ask for.
It was 30 pages. It was extremely technical and in-depth: all the different things we needed to build, all the milestones we needed to hit, how the market was going to evolve, all the new use cases, the cost per token, and all this modeling.
We went and talked to investors—every major investor in the Valley.
They were just like, “Okay, two kids that just finished Harvard, haven't taped out a chip, no test chip, inference—who knows if this is going to be a big market? Everything's going to be training. The model's still hallucinating. This could all be a bubble.”
At the time, the biggest semiconductor fundraiser for a Series A was around $40 million.
We were looking at this, and we were just tallying the bill. I'm thinking, “We're going to spend $100 million in the next 12 months. We really want to do this. If we want to actually get to scale and actually get the performance we're talking about, this is going to be extremely capital-intensive. How the hell are we going to pull this off?”
I think one of our key ways we got started was process. We thought to ourselves, “Where's the cheapest possible way we do this?”
We decided that if we ate nothing but ramen, then we would go ahead and spend basically just the money for the masks that we'd have for our first tape-out. If we did that, we'd go on doing it on $30 million.
I'm insanely low-number.
We actually went out and got a debt provider. They wanted to lend us the money we needed to cross this threshold.
I think that was the first thing to catalyze a series of other things.
Maybe we can do one more thing.
One more thing.
We were at this moment where we thought, “If we really want to build this company, we're not going to half-ass it. We're not going to go do a test chip and spend years on it, and let the entire AI market boom while we could be building the product.”
We're going to do it. We're going to go all the way. We're going to need to find a way to get $100 million.
I remember Gavin and I were sitting down in the office in San Francisco late at night, just looking at each other. We were like, “Could we cut $500,000 here? Could we cut $100,000 here? How long could we convince everyone not to take a salary?”
And we're like, “Holy fuck. The math is not going to close. We really need to solve this.”
There was a period of a few weeks where you kind of just go into survival mode. You call every person that could possibly know and invest, and you're like:
“We need $100 million to do this. If we do this, we think this could be one of the most important companies of all time. Do you know somebody that wants to take an aggressive bet, that wants to believe in us?
“Here's all the information. We're an open book. Here's the team. It's great people. We've been working super hard. We've done these things in record time. But we have these 100 things to go. Do you want to do this?”
And the snowball starts. You get a million here and $2 million here, and you're like, “Okay, we're not going to run out of money this month.”
You get a $5 million check, a $10 million check. “Okay, maybe I can buy those FPGAs.”
The snowball happened, and we were very lucky that we ended up putting it all together.
We had a board meeting. I showed you the spreadsheet. We looked at it, and it was $133 million. These were all soft commits.
We looked at each other and said, “We're going to take it.”
That was the Series A, and it's been much easier since then. We've raised almost a half-dozen rounds since, many of them from those investors just doubling and tripling down.
That's allowed us to get to market so quickly. This rack would not be possible had we not been so aggressive.
That's the thing. Suppliers, too. I think there's a little bit of a combination here.
TSMC was willing to work with us back before we raised any of that $100 million, back when it was still really, really scary.
Actually, when they had let us get some of their emulators on extremely favorable terms—we'd pay over many years—basically, a big loan.
It takes a lot of belief from your partners to do this. But I think when you come out with this very strong team, all the folks who back you are not in it just out of pure financial sense. They believe.
Why don't you assume they see the way you think?
This is a great story, even for you and Jordan.
They were at a conference, a semi event, and I was one of the only young CEOs. I think that's kind of a novelty. They asked me to come in there and speak.
I get to the SEMI event, and I am the only speaker there under 40 and the only person there under 30. Out of a time of 22, I go up and I speak.
There was a speaker's dinner afterwards. By pure luck, happy serendipity, a very senior TSMC VP was at this very nice dinner. There was the former CEO of Arm there. It was very boozy. Everyone was in suits.
I'm there with this VP, and it turns out we both did math in college. We get into a little piece of effort. We begin talking in great detail about how modern AI models work at the actual tensor-by-tensor level.
The guy gets it. We begin talking about, “Hey, how do you run this very effectively? Why is global connectivity critical technology to make this work?”
The following day, I get an email from TSMC saying, “Gavin, I want to work with you. Trying to make it happen.”
They've been a great partner ever since.
It's amazing to think about some of the tropes. Obviously, I should break the fourth wall here: I'm a big-action investor. I've been involved for a long time. I think the world of you guys, so I'm incredibly biased in this conversation.
I'm trying to ask questions that are broader and interesting and could be objections to what you're doing, and we'll keep doing that.
But it's so interesting to me that when you read about investing, everyone cites this idea of contrarian and right as the quadrant that makes all the money, and it sounds really nice.
But contrarian means everyone else thinks you're stupid.
When you go and you get the “no” from literally everybody, it is a fascinating quadrant to exist in before you become consensus. What was it like for you? I'm super curious.
Well, it's interesting. At the time, it was the largest by a lot—the first check that I had written. I used to say I'm not a math expert or a semiconductor expert or an AI expert, really, at the time.
It was much more of a belief in the concept of this market potentially being huge, you having made very, very clear bets on how the future was going to look, having positioned the company in order to attack those things in a hardcore way.
And then just the two of you and what I felt about you was the majority of the reason why we made the bet when we did, in 2023 or whatever it was.
At the time, it was the biggest. And I think the same thing you said about naivete applies to investing as it does to maybe building a semiconductor startup, which is: I didn't know what I didn't know.
When I called experts, they were basically like, “This is stupid.” They laid out in very logical terms why this wasn't going to work and why it was such a low-probability bet.
I think one of the things I've learned from it is that you kind of have to damn the base rate. If you invested on base rates, you should do something other than what we do.
It's always the index fund.
Yeah, there's always something: index fund, exactly.
So it's actually never been scary for me. I think probably most of that is because there's a lot I don't know. If I knew more about what you guys have done and the difficulty, I probably wouldn't have done it.
I don't know what that says about maturing as an investor. Maybe I don't want to know a lot more and have some of that healthy naivete. I don't know.
A lot of the traditional semiconductor funds missed the entire AI chip space—all the AI chip companies—and all the coding experts missed all the coding companies.
I think it's very hard to realize the constraints have changed. When you've looked at tape-outs for 20 years and you've seen so many of them not work on the first try, the second try, the third try—when you couldn't even run a workload—you totally forget that EDA tools are way better and that FPGAs have existed in a way that they didn't exist before.
All the types of validation you can do today just weren't possible before.
A lot of our believers were on two sides of it. They were either believers in the market and the team, or they were building chips today and were extremely technical.
The high-frequency trading firms would literally audit everything, from the microarchitecture and the RTL to the board designs, the schedule, and the software stack.
We would sit down with 10 of their people who built their own chips, and they were asking us such detailed questions that we were wondering, “Are they going to build the chip?”
It was really on either of the sides. If you were anywhere in the middle, you just wouldn't understand it.
In the investing world, they often talk about a very interesting perception—something that you see or believe that others don't. That perception creates the opportunity.
I've invested five or so times in this, and every time when you do it, the stakes are getting bigger and bigger.
It does get a little scarier and scarier. Because you guys have been so quiet in the marketplace, I think it's very easy to dismiss you.
As the stakes get bigger and bigger, those dismissals are harder to hear.
I do think betting on something that you see, when what you hear from the outside world is very different—that perception gap equals opportunity.
Exactly.
The last thing I would say is that the accumulated evidence of your guys' and your team's ability to solve seemingly impossible problems is one of the most interesting things a company can have.
Companies do this or they don't. That's the thing. It's a big advantage of people who've been here for a long time.
You get some new joiners who are scared of shitlessness. You see a thing like this, and there are old-timers who have been here for all of two years.
This is the kind of trench I've seen.
Another one.
Yeah, there's definitely a find-a-way mentality. If you're here, you're here because you assume it's possible. So we can't be saying it's impossible.
Everything is solvable, and we're just going to work at it until we figure it out.
A favorite story I have is about this guy who's kind of a legend in silicon validation who joined our team.
We were doing the early stages of what's called wafer sort. When your chips are coming out of fabrication, you get a lot of these wafers.
You have this thing called a probe card that attaches to the wafer before you dice it. With these probe pads, you send electrical signals to basically test which chips are good and bad.
When I slice the wafer into a bunch of chips, I can package it and only package the good ones.
We go through our first wafer. It's like two or three in the morning because we're doing it with TSMC over the phone. We have this screen with the wafer. It's all gray, and each chip is gray.
Then, as you start running the patterns, the squares are supposed to turn green or red.
And they all turn red.
I'm like, “Fuck. This is really bad.”
Everybody's like, “Guys, take a breath.”
You eat the back of your eyes like the puzzle begins.
You have to have the attitude of: yes, you will go stare into the abyss. You will go see scary things. And we'll solve them.
When did you see the first green square?
I don't know what it took.
But in the moment, it's extremely scary. And there's a certain type of person who is just addicted to that feeling—feeling the fear and solving it.
And we have a lot of people here.
If you think about applying all of this earned know-how from the last several years and now thinking ahead to Gen 2, Gen 3, and beyond, what will you be doing most differently as a result of everything that you've learned?
Conceptually, the way that you will attack designing and producing this next one, based on what you learned during the first time.
It took us a while to get to the principles that we think really matter for scaling inference.
We tried a bunch of things early on, from compilers that would turn different models into FPGAs to burning weights in silicon to splitting HBM into KV cache and weights, and all of these different things.
There were a lot of cycles of learning until we got to the point where we realized that, fundamentally, if you want to run a majority of the tokens in the world, you need to do three things.
You need to build a chip with the most FLOPs given the power budget.
You need to build a chip that has the lowest latency between other chips, so the biggest scale-up becomes possible.
And you need to produce as much of it as possible.
I think probably in the first half of our journey so far, we learned the first two, and that informed the design a lot. That informs a lot about the bets we're making in the future, with the low-voltage in front of the cluster-scale memory.
But the production part, I think in the past year, has become extremely obvious.
How much people want to deploy this stuff if you can have it available today.
The best ability is availability.
If I have 1,000 chips today, someone's going to use them.
We need to build a chip that's not just way better than what's been built before. It needs to be available at many-gigawatt scale.
We need to be building a product that is producible at gigawatts per month in the limit.
As we think about that, a lot of the design decisions we're making with our next generation, which you've seen already, are just about simplicity: removing tons of parts, trying to assemble and disassemble a thing again and again, learning how to make it as quick and the cycle time as possible in production, making sure it's going to be reliable, making sure it's going to be serviceable, and making sure it's going to be producible at gigantic scales.
What about other problems in the ecosystem that are outside of your control, such as capacity at the leading nanometer at TSMC or availability of HBM4 memory, or some of these other things where everyone is fighting for scarce units of capacity?
How do you face up against those realities when you're trying to produce as much as humanly possible?
If people deploying the most compute in the world, do you think about supply a bit?
There's only so many wafers being produced on a given nanometer node, in a given fab. And there's only so much memory being produced.
And that's why, actually, for our first-gen product, we built it on a different supply chain than the ribbons.
But we're on 4 nanometer. We're on a different HBM, and so forth. So it actually is not a zero-sum. It's a positive-sum, where more is more.
Often, when we're talking to people deploying at scale, it's not a decision between a gigawatt of a GPU and a gigawatt of us. It's two gigawatts.
And I think, as much as possible, thinking about supply chain early in the design decisions—because if you have a high-performance product and you can't produce it, then you're just a podcast.
That's the other big thing about breadth of integration, too. For certain things, like chips or memory, you have to go ahead and partner. For most of the other stuff, those are also very highly in demand. The more that you build yourself, the more stuff you can do on top of what the world can currently build.
It's not all you're taking on the build. It's not all you're taking on the build. It's just like you're adding way, way more. I think that's how you win.
One of the things that we really haven't talked about at all is the models themselves, which is kind of crazy—the things behind all of this demand.
Anything interesting that you would say about the way that you see models progressing, based on what we've seen so far? I guess I'm more interested in how you as thinkers about hardware think hardware might impact where the models themselves go in the future.
One of the most important ideas that we believe in is that machines don't think like people think.
You look at airplanes, for example. An airplane is slightly like birds flying. But when you think about how mechanical devices have to work, it's often very different.
And much the same way, for people, storing data in long-term memory is very cheap, and doing math is relatively expensive.
And it is the exact opposite for chips. Generally, moving data is very expensive, and doing math is very cheap.
As time goes on, then, you'll end up finding that math gets cheaper at a rate that is faster than memory gets cheaper. There's just this fundamental limit on any kind of DRAM device.
You should think about: How can I make my model use a huge, huge amount of compute?
What if I had, for example, many copies running at the same time? What if I activated a huge number of experts? What if I had gigantic experts that could run on multiple servers at the same time?
That is how you'll build models that are the next generation of intelligence.
In context, too, there's been a lot of work on very efficient inference. What if I don't load the full context in the memory?
And most of the time, I think they make a lot of sense. The thing about building superintelligence is: Why can't it go look at a billion tokens of context? Why can't it spend a huge amount of compute to go ahead and read all of those super fast?
I would love to go talk to a machine that was able to go attend to every book ever written in its short-term memory. I'd like to have to go into a point where you can.
A theme in models right now is this focus on something called dynamism, which is the ability to control the level of computation and memory spent at a per-token or per-user level when doing attention, as well as the ability to dynamically, in your chip, on the fly, send data to other chips for different AI models doing certain types of operations.
And the reason is fundamentally, as we are scaling context length, as we're scaling model size, as we're scaling the amount of computation per user, we're looking for ways to be more efficient.
So the first thing is, like Gavin says, mixture-of-experts architecture, where maybe we don't need every parameter being used for every token.
But maybe there are things where, even at a token level, we can say, well, this token needs this context, and this other token—they can share that memory. So we don't have to have the overhead of using as much memory.
Maybe this token is really important, so we should spend more compute. We should have longer context on that token.
So hardware that really accelerates these types of very dynamic computations is extremely important.
And you can imagine current hardware that was designed before the size of these architectures has lots of overheads in doing them. So you basically end up in these really bad worlds where you have inefficient hardware at doing this dynamism. Therefore, you can't run it very well.
Or you have these very blocky architectures that are kind of applying blunt force to many different tokens that all need more or less computation.
I have two questions about the future. We talked a lot about what you built so far and how you built it.
The first is about the new ways that people might start using these systems—the raw technology, longer runtimes, things of this nature. When inference gets much cheaper, faster, and more accessible, there's more total supply and it's better, what are the things that you think people will use that capacity to do that are the most interesting and exciting to both of you?
Yeah, that was the viral tweet by Noam Brown. We said the agents and models are having a longer, longer time for inference. They need to ask the take six months, and there's often not enough time to go into the I/OA for such a long period of time, because by that point you'll have a new model out. You'll want to go to the I/OA instead.
With tech like we built—our cluster-scale memory—you can go ahead and run that six-month job much faster.
But there's a second piece of this, too. I'm talking to them about it now as well. It's not just the time; it's also the number of people or agents who are working on this.
If you're trying to go and evaluate, can a human build a rocket? You will find that the answer is no. No one person can build a rocket. Instead, you have to put a team together.
And I believe the same thing will be true of agents, too. If you want to go ask, can an AI go out and build some crazy futuristic piece of software? You'll probably need a very large team.
Maybe that's ten. Maybe that's a million. It has to go and have this enormous amount of cluster-scale memory to go ahead and have that very short time per token, and a huge amount of FLOPs to be able to go run that whole fleet.
I'm going to be a little futuristic. I firmly believe we are on a global march of inference becoming the majority of global GDP.
You may think more than ten years, but it's going to happen. And right now, we measure productivity as a society as GDP per capita. But really, it's going to look much more like agents per megawatt, or it may be agents per gigawatt by then.
And, while being futuristic, I think this is the second-to-last year that a majority of the workforce is going to be human. I think in 2027 you're going to see there's going to be more agents doing knowledge work than humans, and it's going to be extremely interesting to see what happens.
You could imagine a world where, for countries, a majority of their energy ends up going into data centers doing inference, and the energy efficiency of those data centers basically governs how many agents are in their workforce.
So you're going to see, like Gavin is saying, one agent or a team of five to ten agents working on group projects for a couple of days. They can do pretty cool stuff because they're smart, but it's not going to be civilization-scale.
What happens when you have countries that can have literally a billion concurrent agents, like a billion people in the workforce, working 24/7 concurrently on the same stuff?
It's just kind of unfathomable what's going to happen. It's going to be the biggest proliferation of technology humanity's ever seen.
When you have these huge, huge amounts of demand, you get this idea of economies of scale again.
We're thinking about people. I have a brain. I'm not using the whole thing all at the same time. That's only part of us going to be active, and this is the way a healthy brain works.
And the way models have worked is much the same way. On an end-to-end model, only a small fraction of the parameters is being used for any given token at any given moment.
But if you have a large number of users on a piece of hardware, you can kind of take that brain, cut it up into many different experts and many different servers, and run a huge amount of volume through it.
So you'll have a bunch of different pieces of the traffic. You'll have many of them using each part of the brain at a given point in time, and you'll have to make the cost per thought, cost per token, way, way lower.
So you're going to end up with these giant-scale distributed brains as the form factor. This is a big data center with a bunch of chips, a huge amount of FLOPs, and a huge amount of scale-up and interconnect.
You think we'll see a trillion-dollar individual data center?
Absolutely. It is a matter of time.
It's like asking, will you see a billion-dollar fab, or a ten-billion-dollar fab, or a hundred-billion-dollar fab? It's inevitable that the economies of scale don't stop at $40 billion, the current number for fabs.
The cost per wafer keeps going down as you keep spending more money.
And the same thing will be true of plants that go out and make steel, or plants that go out and make tokens.
A very smart alien lands on Earth and wants to know from each of you how you would frame up this opportunity that you guys are tackling. What do you say to them?
I would say that thinking is really valuable. Every company in the world runs on thinking.
And we are entering this really unique moment in time where you have a machine that can go think almost as good, and soon as good, and soon better than the best humans can.
The little thing of these machines is going to be a huge opportunity. But more important than that, the way in which you go ahead and run this kind of thinking is going to be very, very different as demand goes higher and higher and higher and higher.
It's a unique moment right now to build a new set of solutions, a new roadmap for how you run future quadrillion-parameter models for a billion people all at the same time on a gigantic scale-up cluster.
We are in a new era of intelligence where the cost of producing intelligence is dramatically so much cheaper than the value of the intelligence, that we are in a many-year, probably many-decade, supply shortage of these tokens.
Basically, any chip or any system that can produce tokens is likely to be extremely valuable. And you should find some part of the supply chain of the token.
It can be everything from model training down to what we're doing in silicon, in other ways, to spend time on and push the frontier.
The companies that are the largest are, frankly, going to be the companies that produce most of the global supply of tokens and own the majority of the supply chain of that token.
And importantly, it's people who build systems that, as they get more and more put together, get cheaper.
The way you want this to scale is not that if I want to serve ten times more tokens, I buy ten times more servers. There must be some solution where, if I only have ten times more tokens, then I get some economies-of-scale benefit with my set of cluster-scale memory tech.
That allows me to then not charge as much as ten times more for those tokens.
What are the ridiculously exciting futures that you guys are building to enable?
When I did this with Gatom last time, I asked him my traditional closing questions. So this time I'll ask you: What is the kindest thing that anyone's ever done for you?
During my cancer treatment, there was a big decision I had to make.
The doctors came to me and said, “It's time for you to decide. Do you want to get surgery, or do you want to get radiation? Here's the trade-off. If you get surgery, you're more likely to live, but you have to assume you'll never be able to walk again. If you get radiation, you'll be able to walk again, but it's not the same probability that you'll live. You may die. What do you want to do?”
And I was 16, and my parents said, “You have to make this decision for yourself.”
I thought a lot about it for a long time and decided I'm going to do the surgery.
I got the surgery. One of the things you do when you get a tumor resection is they do something called a necrosis analysis, where they look at all the different cells and say, is a cell dead or alive? Because if you have a bunch of cancer cells that are alive, you have a problem.
And they looked at it and they said, “You usually want 98–99% necrosis for us to say you're in the clear. You're below that. You should go get radiation.”
And there were only a few machines in the world that actually could do the type of radiation I needed. I wanted them to be in Boston. I was in a wheelchair, and I needed to move to Boston for multiple months.
And both of my parents decided to move out and drop everything they were doing and live with me.
And I'm eternally grateful.
Who's beautiful?
Nice, guys.
You know how small advantages compound over time? That's true in investing and just as true in how you run your company.
Your spending system is your capital allocation strategy. Ramp makes it smarter by default: better data, better decisions, better economics over time. See how at ramp.com/invest.
As your business grows, Vanta scales with you, automating compliance and giving you a single source of truth for security and risk. Learn more at vanta.com/invest.
The best AI and software companies, from OpenAI to Cursor to Perplexity, use WorkOS to become enterprise-ready overnight, not in months. Visit workos.com to skip the unglamorous infrastructure work and focus on your product.
Ridgeline is redefining asset management technology as a true partner, not just the software vendor. They've helped firms 5x and scale, enabling faster growth, smarter operations, and a competitive edge. Visit ridgelineapps.com to see what they can unlock for your firm.
Every investment firm is unique, and generic AI doesn't understand your process. Rogo does. It's an AI platform built specifically for Wall Street, connected to your data, understanding your process, and producing real outputs. Check them out at rogo.ai/invest.