Andrew Feldman, Cerebras Co-Founder and CEO: The AI Chip Wars & The Plan to Break Nvidia's Dominance

Summary

• Feldman’s core attack on NVIDIA: the GPU’s off-chip HBM memory is a fundamental architectural limitation for generative inference — serving one word from a 70B-parameter model means moving ~140GB of weights from memory to compute, again for every next word. Wafer scale let Cerebras use fast SRAM at capacity, and “what used to be their advantage is now weakness… it can be beaten and I think they know it.” His market-structure call: NVIDIA goes from “approximately all” of the market today to 50-60% share in five years — between Uber’s 90/5 and cloud’s even split.
• The CUDA moat is “not real at all” in inference — “you can move from OpenAI on an Nvidia GPU to Cerebras… with 10 keystrokes.” The real, rarely discussed moat is market-share leadership itself: Intel made “nearly a decade of catastrophic decisions” and still holds ~75-80% of x86. “I can make a bunch of bad decisions for a decade and only lose 20% share — the moat was just unbelievable.”
• Scaling-law gains continue — Feldman flatly rejects the “far along on compute, algorithms and data” consensus: “I think they’re wrong. I think we are early in all of them.” A GPU doing inference runs at 5-7% utilization — “95 or 93% wasted” — OpenAI’s o1 shows inference scaling laws “fully functional,” and we won’t be as dependent on Transformers in 3-5 years (“100%”). In five years training data is “almost all synthetic.”
• The inference market equation: users × frequency × compute-per-use — all three growing simultaneously, a rare condition. AI flipped from “novelty” to useful in Q4 2024, the market in five years is “way over 100 times bigger,” and faster/cheaper only expands it: “no examples in compute in 50 years in which by making things cheaper, faster, the market got smaller.”
• Chip providers will be worth more than model providers on a 5-year view. Today’s model valuations are option pricing — “uncertainty is a friend of the value of the option” — but Buffett’s weighing machine eventually kicks in. On models generally: competing on release cadence four months ahead of rivals carries “not a lot of value”; staying top-decile for years does.
• Cerebras is cash-flow positive where peers hemorrhage cash — “gross margins were a measure of your technical differentiation… if you’re running a negative gross margin business, you’re selling commodity.” The flip side: G42 is 87% of revenue (a deal estimated north of $1bn), which Feldman defends as a learned muscle: “the way you catch three large customers is to catch one first,” with several relationships targeted in 24 months.
• On China: Cerebras refused to sell — the deal “wouldn’t be used for good” (facial recognition of minorities, military) and failed his mother test — yet he thinks the US underestimates China “100%,” export controls may not be “a tractable problem,” and the attempt to cut off EDA tools just spawned US-VC-backed EDA startups in Shenzhen.

Deep dive

1. The bet in 2015: AI’s hard problem is moving data, not computing

• Feldman’s framing of chip design: a chip “does calculations and it moves data” — and AI inverted the difficulty. The math is trivial (“a matrix multiplication and an FMAC can be developed by any second-year electrical engineering student”); the hard part is that results and intermediate results must be moved constantly — to memory, from memory, among GPUs. Cerebras bet that solving data movement would yield a faster, lower-power AI computer.
• His one admitted miss, notable from a fifth-time founder: “the first time I underestimated the size of the market by a lot.” What the team got right was that AI would pressure memory bandwidth and communication structure — the dimensions GPUs weren’t built for.
• Training and fine-tuning are computationally “approximately the same”; generative inference is the outlier. To generate one word from a 70B-parameter model at 16-bit weights, you move ~140 gigabytes from memory to compute — then again for the next word. “That’s called memory bandwidth, and if you have an architecture like the GPU, that is your fundamental limitation.”

2. Wafer scale and the yield problem nobody solved in 70 years

• The memory trade: HBM is “phenomenal… but it’s slow” — high capacity, built for graphics where you rarely go back to memory. SRAM is “unbelievably fast but has low capacity” — a normal-sized SRAM chip serving a 400B-parameter model needs ~4,000 chips; a DeepSeek 671 needs “six or 8,000… what an administrative nightmare.” Wafer scale gets SRAM’s speed with enough capacity on one, two, or ten wafers — and less power, since off-chip I/O is among the most power-hungry operations on a chip.
• Why nobody had done it: yield. His analogy — a wafer is cookie dough, flaws are M&Ms your mother throws blindfolded; “the bigger the cookie, the higher probability you hit an M&M,” and traditionally you binned or trashed flawed chips. Cerebras’s answer, borrowed from memory-making: build the processor from hundreds of thousands of identical tiles with redundant rows or columns, shut down a flawed tile and route around it. “Nobody had ever been able to do that in the 70-year history of our industry” — likely Gene Amdahl’s company, Trilogy, “crashed and burned trying.”

3. Speed is not a spec — it’s what creates product categories

• Feldman refuses a single ranking of fast/cheap/accurate: for a cancer diagnosis, “93% accuracy is just plain not as good as 94%” and you’ll pay and wait; for batch work like Llama 405B generating tuning data for 70B, cheapest may matter. But in interactive use “milliseconds matter” — Google showed years ago “you can destroy your user’s attention with milliseconds of delay.” “There is no search if you’ve got to wait eight minutes.”
• The analogy he leans on: when the internet was slow, Netflix mailed DVDs; broadband arrived and “suddenly Amazon’s a studio. It changed everything. Speed in inference does the same thing” — new applications open up that couldn’t exist at GPU latencies. Cerebras claims the fastest inference “across a whole set of models” since its August 26 launch, per Artificial Analysis.

4. The inference equation: all three multipliers growing at once

• His sizing formula: inference market = people using AI × frequency of use × compute per use. “We are in this rare time” where all three are growing simultaneously — hence off-the-charts growth. Five years out the market is “way over 100 times bigger.”
• The turning point wasn’t technical: “ChatGPT was not really a technical innovation, it was a user interface invention.” Until mid-2024 AI was “a novelty… whoa, this is cool.” Starting Q4 2024 it became useful — “if your marketing team isn’t on an LLM, each person several times a day, they’re not doing their jobs” — reaching “my dad, my brothers, the doctors,” and “when you get them, the market is ripping.”
• On power: he concedes the industry “consumes an enormous amount of power… and some water,” so “the burden is on us to deliver exceptional value” — cures, societal problems. The US problem isn’t supply but geography and process: “we have power in Niagara… what we don’t have is power where you want to build data centers” and no national mechanism to override local regulation and installed interests.
• Pushback on Jonathan at Groq’s “tourist data centers” claim — Feldman mostly disagrees: the early movers were Bitcoin miners like TeraWolf and Crusoe, “certainly not tourists… extremely sophisticated data center builders” now leading gigawatt-scale projects. “Sure there’s some tourists,” but many facilities will be fine.

5. Scaling-law gains continue — “I think they’re wrong”

• Against the refrain that we’re far along on compute, algorithms and data: “I think they’re wrong. I don’t think we’re very far along… I think we are early in all of them.” Exhibit A: a GPU doing inference is “5 or 7% utilized — that means it’s 95 or 93% wasted.” Costs fall via cheaper computers, lower-PUE data centers, and better algorithms compounding together.
• On the scaling-laws debate itself: there’s genuine argument about whether we’ve “run out of mojo” on data, but “OpenAI’s work on o1 shows me that the scaling laws certainly for inference are fully functional — the more compute you put on inference, the better answer you get.”
• The algorithmic headroom, as told: many models are still all-to-all connected — “connections that don’t produce anything that we still end up doing math over.” His analogy: to learn something you can read 50 books, or the three that matter, or summaries of those three — “the problem is we don’t know which they are at the beginning.” MoE, dropout, sparsity are early steps. And on architecture: “we won’t be as dependent on Transformers in three years or five years as we are now — 100%. They’re not the end-all be-all” — he won’t guess the successor (“I don’t know whether they’re going to be state-based models”), but the attention head’s quadratic effect is a known weakness people are “desperate to overcome.”
• Synthetic data in five years: “almost all synthetic,” with utility he thinks is equal to human data. His pilot analogy, worth keeping whole: real driving data is “people driving straight on a freeway — not difficult”; what you want is “an unprotected left turn in the snow… thousands of different ways, millions of different ways.” Like simulators for pilots and rare cases for surgeons, synthetic data fills in exactly what’s expensive or painful to gather.

6. DeepSeek was “focused engineering” — and the distillation complaint fails a consistency test

• What impressed him: “they weren’t confused about being model intellectuals… they were interested in being better. From an invention standpoint that’s a little boring, but from an engineering standpoint that was sweet effort.” DeepSeek proved “you don’t need 5,000 people and billions of dollars of gear — you can do it with 200 smart people and more gear than DeepSeek said they had, but less gear than others had.” The inauguration-timed announcement he files under politics.
• On distillation: “is summarization wrong?… I don’t think distillation is wrong, and if distillation is wrong then certainly using people’s copyrighted data is wrong. You’ve got to be a little bit consistent.” And on impact: “there are very few examples of an open-source anything having the sort of immediate impact that model had… this had a loud boom.”

7. Where value accrues: chips over models, and the moat nobody names

• CUDA lock-in “in inference, it’s not real at all… none — you can move from OpenAI on an Nvidia GPU to Cerebras to Fireworks to Together with 10 keystrokes.” Most AI is written in PyTorch; compilers are “hard but tractable.” NVIDIA’s real moats are being the default — and Feldman thinks challengers under-study this: Intel made “nearly a decade of catastrophic decisions” pre-Lip-Bu and still holds ~80% of x86 while AMD clawed to 25-30%. “That’s a moat… as a challenger we have to think about it exactly, because it’s exactly that we need a bridge for.”
• The five-year market structure: between Uber (90/5/5) and cloud’s shared oligopoly — “Nvidia is going to have somewhere between 50 and 60% of market… right now they have approximately all of it.” He’s emphatic NVIDIA won’t “roll over and play dead in inference” — “one of the great decades of any company in history,” from ~$10bn in 2014. But NVIDIA’s long wait times are “a huge opening”: “when the bully falls, everybody wants to give him a kick.”
• Chip providers larger than model providers in enterprise value at five years: yes. His mechanism for why models look expensive now: “when you price an option, variance and uncertainty increases the option’s value… part of these extraordinarily high prices is this wild variance.” Long-run, Buffett applies: markets are a voting machine short-term, “a weighing mechanism” long-term — “at some point the weighing kicks in, and usually it’s in the public markets.”
• On model-company defensibility: “you’re competing against other people’s release cadences — you’re four months ahead, they’re six months. If that’s really where you are, there’s not a lot of value. But if you can stay top-decile over years while the people above you are changing constantly, I think there’s a lot of value.” Hardware endures — Cisco, Juniper, Apple, NVIDIA — because “what they do is hard. That’s why it’s worth challenging.”

8. The business: cash-flow positive, one giant customer, and why go public

• On being cash-flow positive while rivals bleed: “traditionally your gross margins were a measure of your technical differentiation… if you’re running a negative gross margin business, you’re selling commodity — your value creation isn’t being recognized.”
• G42 at 87% of revenue (estimated north of $1bn when announced) is “both” strength and weakness: “the way you catch three large customers is to catch one first… being a strategic partner is a learned skill.” The proof points: tens of exaflops deployed — “vastly more than anybody that isn’t AMD or Nvidia” — software hardened on some of the largest AI clusters, manufacturing scaled “2x and 5x and 2x.” Target: “several” relationships in the next 24 months.
• On Harry’s challenge that the IPO filing seemed preemptive and hands competitors asymmetric information: “we have nothing to hide… we’ve got asymmetric technology.” The affirmative case: first in category, and “some of our largest targets would have a stated preference for doing business with public companies.”

9. China: refuse the sale, respect the rival, doubt the controls

• On export controls, a hardware-native’s distinction: a 500-600lb server “arrives on a pallet — you can have somebody from the embassy visit it, take photos once a month. It’s not going anywhere.” Software and open source are “a whole other level.” He notes DeepSeek “probably did use chips in Singapore.” His deeper doubt: “I don’t know if it’s a tractable problem to delay another nation’s progress on a technical trajectory” — the EDA restrictions just meant “US venture capitalists backed tons of Chinese companies in Shenzhen to build EDA tools.” This administration is “probably net a fair bit better” for AI than the last, which “lined itself up against big tech.”
• Yet Cerebras itself refused a China deal — his test: “just ask yourself, would my mother be proud?” What he saw or couldn’t see — facial recognition “to identify minorities for persecution,” military equipment — failed it. “It’s more important than money.” He holds both positions without smoothing the tension: controls may be futile, but this sale was his choice.
• Do we underestimate China? “100%… one of the most obvious and frequent errors in judgment.” The evidence as he lists it: extraordinary infrastructure investment, exceptional engineering-talent generation, state-backed VCs, national champions, “a belt-and-suspenders strategy to make much of the third world dependent on them.” Shenzhen’s economic zones were “clearly a visionary move” — and America has done the same when it wanted to (Trump-1 vaccine rule relaxation). His uncomfortable questions: why can’t the US build trains, why are “our bridges and our freeways in disarray”?

10. Quick fire: being wrong, tiny chips, and what experience buys

• His best documented error: fighting co-founder JP’s 2016 water-cooling plan — “I fought so hard and I was so wrong.” Google announced water-cooled TPUs a year or two later; “now Nvidia’s only selling water-cooled parts. I was dead wrong and JP was right.” The generalization: “if you’re not prepared to be wrong a fair bit, you ought not to be making a lot of decisions.” A CEO must be “mostly right most of the time” — unlike VCs, where “on average you’re wrong all the time and what they care about is the occasional time you’re really right.”
• Underappreciated corner of the chip market: sub-milliwatt inference chips living next to sensors that “only send back useful data” — enormous volume, “fundamental for robotics” — though not his market (“I like to build bigger things and sell them to the data center”).
• On Dario’s predictions: he rejects living to 150 and 90% of code machine-written this year, but expects AI penetration approximating cell phones within a year or two. Contrarian belief: Middle East peace is closer than believed, on the “rise of a moderate, business-focused Arab state” across UAE, Qatar, even KSA. And on fifth-time founding: where a business has manufacturing, supply chain, and hundreds or thousands of engineers on a schedule, experience compounds — “nobody would say with a straight face: what I’m looking for is an engineering leader with no experience.”