A $150 million model, $5 million six months later: Benchmark’s AI investment math

Nathan Labenz frames a short conversation on two tracks: how to invest venture capital into the AI wave, and how to apply AI to hardware. His guests are Eric Vishria, general partner at Benchmark, the firm behind eBay, Twitter, and Uber, and Sergiy Nesterenko, founder and CEO of Quilter. Regular listeners have already heard Sergiy on a cross-post from Will Summerland’s Autopilot feed; this hour is mostly new ground, including follow-ups on superhuman circuit-board design. Vishria’s opening provocation, which Labenz puts on the table before anyone sits down, is that foundation-model companies are pouring huge sums into some of the most remarkable — and most rapidly depreciating — assets in venture history. Benchmark, as of this recording, has not invested in any of them.

There is no shortage of AI startups. Each Benchmark partner is probably meeting four or five a week. Many of the ideas sound interesting, and the revenue traction, Vishria says, is tremendous — unlike anything the firm has seen for speed. Early-stage investing, though, looks out five or ten years, toward the handful of exceptional companies the industry creates a few times a decade. The job is to tell a real business from a sugar high of revenue, developers, hype, or investor dollars. That is hard at the beginning. It is also what they are paid to do.

The framework Vishria uses often has nothing to do with the company itself: what is changing in the world that is large enough for this firm to become huge? “AI” is a wave, not an answer. You need something more substantive and specific to build something durable. Quilter, for him, had that extra layer, which is why a firm that is still skipping foundation models made a hard-tech exception.

Electronics are soaking into ordinary objects. A light switch that was simple not long ago now has a circuit board in it. The number of products and companies that need boards is expanding fast. What Vishria did not understand until he met Sergiy is how manual the next step is. After the schematic — which parts, connected how — comes place-and-route: thousands of components and tens of thousands of traces laid onto a board. A complicated board takes a human two to four months. Faster iteration makes better products; that friction, in a world drowning in electronics, is a problem.

Six months before this conversation, copilots were the fashion: a copilot for lawyers, doctors, developers. Sergiy’s line was the opposite of everything Vishria had been hearing. This is not a copilot problem. You do not want a human-assisted person doing printed-circuit-board layout. If an auto-layout tool does 90 percent, the last 10 percent will kill you. The board is a dense mesh of thin lines. Ask a human to draw one more without wrecking what is already there, and it is like stuffing a new route into a rush-hour subway map. Quilter sits in the middle of interfaces the industry already has: take the schematic in, spit a manufacturable board out. Start simple — not an iPhone A14 on day one — and climb.

The second surprise was the model. Everyone was talking about large language models. Quilter uses none of them. They design a game, give the user knobs on the objective — cheap and large, or dense, expensive, and power-efficient — and return a range of solutions. Vishria knows almost nothing about PCBs; he had to go validate. What excited him at the macro level was Chris Dixon’s old post *The Idea Maze*. You want a founder who has been rolling around in the maze: down a path, dead end, back up, another dead end. Every question comes back “I thought about that, here’s why it doesn’t work, here’s why this might.” Sergiy had designed boards at SpaceX, tried a string of AI approaches, and converged on reinforcement learning. That, to Vishria, is someone who has lived in the maze long enough to have a shot at walking out.

Labenz hears a textbook disruption pattern. On the earlier episode, SpaceX’s internal board team could not even serve other teams inside SpaceX. If a specialist group cannot feed its own company, the underserved market is broad. He parks that for a moment and tries the idea maze himself: what is the data landscape? Deep learning usually wants a lot of it. Is there an open-source corpus of published boards, or did the lack of data force another AI route?

It is the question Sergiy gets most. The Copilot picture is GitHub plus Stack Overflow, fed to an LLM, predicting average human behavior. Quilter does not do that, for two reasons. First, the data barely exists. GitHub does not hold many open-source boards; the best designs sit in Apple, Google, and SpaceX repositories. Second, and more compelling to him: people are not good at designing boards, full stop. A complicated layout that takes two, three, four months is a pile of margin-on-margin decisions. The board comes out bigger, with more layers, more expensive than it needed to be. Supervised learning on that data, even if you could get all of it, would top out around a high-level human designer.

Reinforcement learning has a higher ceiling. DeepMind’s Go work is the example he always comes back to: train on human games, reach grandmaster; then throw the human data out, learn from win and loss, and get to a level nobody can match. Labenz asks whether Quilter started from literally zero human data, and how they survived the sparse-reward problem — the same trap the Eureka project hit when success for a robot hand was so rare you could barely score it. Winning a whole board is so rare that random exploration never finds it. People already broke layout into steps and heuristics, so Quilter does the same: place components, check manufacturability and collisions, compute a heuristic for routing success, and pay dense little rewards that correlate with the final sparse one. The long-term dream is a single yes or no from physics. For now, on boards where they are not yet as good as humans, they still use the heuristics humans would use.

After the break, Labenz draws the language-model analogy. RLHF leans on a human reward that is inconsistent, sometimes wrong, and easy to game with sycophancy. Are Quilter’s checkers rock-solid physics simulators, or a pile of heuristics that can quietly lead the search astray? Vishria zooms out even further: he thinks generative models are interpolators, good inside the hull of human-created data, while humans remain the extrapolators. That is why he finds AI-scarcity talk misguided. Labenz is less sure the bounds will hold once self-play hits language models, but he wants the circuit-board answer first.

Sergiy’s luxury is that humans are not in the loop. Humans are bad at looking at a board and judging whether it will work; about 80 percent of boards that get built are faulty in some way. The physics — Maxwell’s equations, thermodynamics — has had convergent numerical methods for more than a hundred years. Numerically solving those equations, carefully, is “completely trustworthy,” far more than a human eye. The only problem is speed: twenty minutes for one physics solution, millions of them during training, and you cannot wait. So they use conservative approximations. You may not get the arrangement physics would allow at the limit. You still get one that works, that competes with or beats humans, because the approximations were built to be safe. “We don’t have to negotiate with humans. We just have to make sure that it’s manufacturable and that it’s going to work, and physics tells us the answer, and physics is unambiguous about that.”

Search depth is a user choice they do not yet expose. A sparse, cheap board that only needs to implement the schematic should come back in minutes to an hour. An iPhone motherboard at a hundred million units, where a dollar saved per board is a hundred million dollars, could search for a month — as long as a human would spend on one design, never mind the billions of variants a machine can try. Today the product constraint is overnight: upload at the end of the day, result within twelve hours, review in the morning. Compute is a different species from LLM training. No 10,000-GPU runs across a trillion tokens. Most of it goes into playing the game, still mostly CPU, some GPU. They train during production: when a user uploads a board, the system is not only doing inference.

Physics cost will rise as they chip through problem classes. For now, low-speed boards, currents around four amps, straightforward math. Next is high-speed digital: cheap bounding methods, then one or two full-wave validation runs that Sergiy guesses will cost a few GPU-days per board. The brute-force electromagnetic method is finite-difference time-domain: grid the world into something like 100 million cells, apply Maxwell for 100,000 steps. Cleverer options include finite elements, the method of moments on surfaces, and, at low frequency, a quasi-static approximation that factors time out and only computes capacitance and mutual inductance. Labenz wonders about a diffusion model, denoising a board the way Stable Diffusion denoises an image or protein models denoise a fold. Sergiy kills it on data quality. Three times out of four, a board sent to a manufacturer looks right and is manufacturable — and the physics does not work. Factory files are a mask set: where to etch copper, which parts to glue down. They do not tell you what signals run through the board, and without signals you cannot score physics. Self-play into a diffusion model is valid only after you already have first-principles-verified good designs.

What keeps Quilter from superhuman layout sooner? Talent, Sergiy says. You need people who are excellent at neural nets and at cutting-edge RL — you cannot grab the latest paper and hope — and also at C++, CUDA, and physics, all in one company. His crystal ball is foggy past a couple of months, same as Labenz’s. A guess: a few years for small boards, maybe five for something like a motherboard. Labenz asks whether there is a sheer cliff they do not know how to climb, or graduated stairs. Sergiy is confident in the latter. This is “100 percent a solvable problem.” Self-driving is do-or-die: you are confident you will not crash, or you are not. Quilter has checkpoints — board size and complexity, which physics, how much better than humans — and can treat each as 20 percent better, compounding. Hard, lots of people, lots of effort. Nothing about it looks unsolvable.

After another sponsor break, Labenz asks about the mature future. Can a tinkerer who wants a talking stuffed animal, and who knows nothing, still get a board? Specifying what you want is an expert problem today. Sergiy draws a hard line: Quilter is layout, one of two problems. Schematic is writing code — a block diagram of inputs, logic, outputs, abstract on purpose. Layout is compiling: making atoms do what the schematic said. They are targeting layout because a compiler should exist for electronics and does not. Compilers in software led to higher-level languages, to Python, to models that write the Python. Electronics will climb that stack too, until a person who never learned C++ can live in Python, or in block diagrams, and still get a board.

Labenz has just done an episode on a young man who set out to design his own GPU from scratch in two weeks and finished in four. That was possible, in part, because the chip world’s equivalent of a layout compiler already exists. Circuit boards do not yet have that courtesy. The compiler analogy is also how Sergiy answers the ignorant-founder question: you do not skip schematic forever, but once layout is a compiler, the rest of the stack can move up.

Zooming out to the Benchmark portfolio, Labenz asks the blunt question: how do I make money in AI? Vishria is bullish because disruption is primordial soup. A lot of crap gets created; a few good things do too; the work is finding them. He loosely sorts the market into three categories. First, foundation-model companies, with techniques, proprietary data, a claim to something special. Second, infrastructure: Cerebras, an AI chip-and-systems company Benchmark backed in 2016, where he sits on the board; Groq; Fireworks, an inference provider, another board seat. Third, vertical applications — lawyers, doctors, accountants, and Quilter — applying a model to a job.

Each category has a large opportunity and a large problem. Foundation models are incredibly expensive to develop and “extremely quickly depreciating.” Vishria repeats the line from the top of the show: they may be the fastest-depreciating asset in the history of humankind. You spend $150 million; six months later someone can build the same thing for $5 million. That has not historically been a good way to make money in venture. The capital tends to get incinerated. He could be totally wrong; someone may build a durable edge. General-purpose models, so far, have been expensive and have decayed very quickly. That is the books behind Benchmark’s empty foundation-model column.

Infrastructure companies have done pretty well. Real businesses, revenue quickly, they enable the rest of the ecosystem. The open question is duration. A lot of AI development, he says, looks like software development circa 2002 in its tools and abstractions. LLM observability is the example Labenz offers and Vishria takes: it is a thing today; he does not know if it is a thing long term. Verticals have the insane revenue traction, and the sharpest criticism. If you can ship in three months, six months, or a weekend hackathon, ten competitors can too. Where is the moat? In a bunch of cases, nobody knows. You have to have a theory. Quilter, Sierra, and a few others are bets that something will compound into a barrier even though traction came fast.

Another pattern Vishria cannot stop noticing: so much AI work is done by researchers. An entrepreneur told him she has a rule for code they write. It works as a proof of concept. You take it into production to scale, and it is not written for that. Her rule, as he remembers it, was `rm -rf` — delete it, rewrite. That research-to-product divide, he says, has not existed in software in his adult lifetime. Maybe it existed around databases in the 1970s and 1980s. The things people now do to serve PyTorch and CUDA are things ordinary software developers have not had to touch for at least twenty years. Labenz hears the same note from Demis Hassabis on why DeepMind is merging into Google proper: a lot of the research is done, the bridge to engineering is the huge remaining challenge, and that bridge is its own thing.

Out of time, Labenz asks for a call for startups, anything Vishria wishes someone would bring. Vishria refuses. If a venture capitalist has the idea, with very few exceptions — maybe Mike Speiser can pull it off — that is a bad situation. Having ideas is not the job, and not the thing they are generally good at. He wants to meet people who have been working on something for a long time and are obsessed with it. Benchmark’s job is to be involved in the best companies as early as possible. Labenz’s coda: with things moving this fast, there is no substitute for obsession.