Oct 6, 2025 · 1h 20m · news
Cerebras CEO, Andrew Feldman on Why Raise $1BN and Delay the IPO & Why NVIDIA’s Worried About Growth · 20VC with Harry Stebbings
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this deep-dive interview, Cerebras Systems CEO Andrew Feldman joins host Harry Stebbings to discuss their landmark $1 billion funding round, the physical limitations of current GPU memory architectures, and the realities of competing directly with NVIDIA. Feldman outlines why hardware companies require seasoned leadership, how AI will realistically reshape workforce productivity, and the intense, obsessive discipline needed to build a legendary tech giant.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 14.3% of the talking time here. How this is scored →
Harry as informed peer 3.0 Guest teaching 4.5 Guest disagreement 3.0 Harry pushing back 2.3
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 24:13 Direct rejection of training vs inference market split
Andrew Feldman flatly shuts down Harry Stebbings' premise that Nvidia is superior at training while Cerebras takes inference, stating 'No, we're faster on both'.
Hardest push from Harry ▶ 1:03:03 Challenging guest over UAE revenue concentrationHarry Stebbings directly challenges Andrew Feldman on Cerebras' S-1 filing, asking if he is obligated to speak favorably of the UAE due to 80% revenue concentration and questioning if it was a planning mistake.
Biggest teaching moment ▶ 48:31 Historical analysis of software firms building hardwareAndrew Feldman breaks down decades of tech industry history to explain why software companies systematically fail when trying to design custom chips.
Harry holds his own ▶ 42:05 Host provides economic rationale for AI researcher payHarry Stebbings counters Feldman's hesitation regarding executive researcher pay by laying out an enterprise value rationale that justifies billion-dollar pay packages.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| Trailer: Key Takeaways & Teasers | 1 | 1 | 1 | 0 | Introductory clips and light banter regarding Cerebras' $1 billion funding round. Harry Stebbings plays a welcoming host role with minimal pushback or technical debate. | |
| Choosing a Pre-IPO Round Over a Direct Listing | 2 | 3 | 2 | 2 | Stebbings asks why Cerebras raised a late-stage pre-IPO round rather than going public immediately. Feldman reframes public tech announcements as option contracts on unquantifiable future demand. | |
| Strategic Planning and Risk Management in Hyper-Growth | 2 | 3 | 2 | 2 | Feldman discusses corporate planning under rapid exponential growth. Stebbings asks broad questions about sustainability, while Feldman notes that betting against market leaders yields no alpha. | |
| Giant Stumbling and Big Tech Growth Anxiety | 3 | 4 | 4 | 2 | Stebbings quotes Grok CEO Jonathan Ross predicting Nvidia reaching a $10 trillion market cap. Feldman dismisses public stock picking and aggressively critiques Nvidia's strategy of predatory pre-announcements. | |
| Deciphering Complex Deals & Lighthearted Banter | 3 | 6 | 4 | 3 | Stebbings references an 18-month chip depreciation model. Feldman explicitly rejects this as empirically wrong, educating Stebbings on how chip retirement is driven by speed-to-power unit economics. | |
| Performance Improvements and System Bottlenecks | 4 | 5 | 2 | 3 | Stebbings cites a hardware founder who claims SRAM cannot scale for large models. Feldman validates the premise's core truth while schooling Stebbings on how wafer-scale manufacturing bypasses traditional SRAM limits. | |
| Training vs. Inference and Software Barriers | 3 | 6 | 5 | 3 | Stebbings asks if Nvidia is inherently better at training while Cerebras excels at inference. Feldman forcefully rejects the premise, asserting Cerebras is faster at both and explaining the software adoption barriers. | |
| Geometric Growth in the Inference Market | 3 | 6 | 1 | 1 | Stebbings references Sam Altman's quote on ChatGPT usage patterns. Feldman delivers a detailed historical lecture on economist Paul David's computer and dynamo paper regarding productivity adoption curves. | |
| Energy Demands and Power Grid Mismatch | 2 | 6 | 4 | 2 | Stebbings asks about immense power grid constraints for AI. Feldman directly corrects the common narrative, clarifying that the US has abundant energy, but it suffers from a spatial mismatch relative to fiber and population. | |
| Risks and Concentration in the S&P 500 | 3 | 5 | 3 | 2 | Stebbings probes S&P 500 concentration risk in Big Tech. Feldman educates the host on how modern market indexing misleads investors into taking on hidden sector exposure. | |
| Talent Shortages,TSMC and Hardware Bottlenecks | 5 | 3 | 2 | 4 | Stebbings challenges Feldman's questioning of AI researcher salaries by offering a clear economic rationale based on enterprise value added per scientist. Feldman agrees and compares it to entertainment pay. | |
| The Realities of Building Data Centers | 3 | 4 | 2 | 2 | Stebbings asks how real estate investors lose money building data centers. Feldman breaks down the construction, permitting, and cost-per-megawatt operational realities. | |
| Vertical Integration vs. Software Companies Building Chips | 4 | 7 | 4 | 3 | Stebbings quotes a guest advocating for AI labs to build custom silicon. Feldman provides a comprehensive historical lesson detailing why software firms repeatedly fail when trying to manufacture chips. | |
| Chip Market Structure and public company Margins | 3 | 6 | 5 | 2 | Stebbings asks if the semiconductor market will turn into a 90% monopoly. Feldman rejects the question's premise, interrogating the definition of market segments and citing historical specialization. | |
| Sovereignty, European Labs, and China's AI Strategy | 3 | 4 | 3 | 2 | Stebbings asks about sovereign AI labs and China-US competition. Feldman discusses walking away from a 2019 Chinese deal due to national security and technical ethical concerns. | |
| Talent Acquisition, Visa Policy, and US Grid Constraints | 3 | 4 | 3 | 3 | Stebbings brings up arguments against H1-B visa programs. Feldman defends high-skilled immigration using personal family history and critiques local municipal permitting bottlenecks. | |
| Quick Fire: Middle East Peace, UAE Revenue, and the Ultimate Wafer-Scale Bet | 5 | 4 | 4 | 6 | Stebbings presses Feldman hard on S-1 figures showing 75-80% revenue concentration in the UAE, asking if Feldman is forced to praise the region and questioning if it was a resource planning error. | |
| Pitfalls in Tech Investing & Why Young CEOs Fail in Silicon | 2 | 5 | 3 | 1 | Stebbings asks about venture capital misallocations. Feldman explains why 25-year-old CEOs consistently struggle in silicon manufacturing compared to consumer software domains. | |
| The Durability of the Data Provisioning Market | 4 | 6 | 5 | 2 | Stebbings names data vendors and references a prediction of short-term labor shortages. Feldman explicitly calls the prediction 'absolutely wrong' and details how AI will reform education and entry-level work. | |
| The Thrill of Battling Goliaths like NVIDIA and Cisco | 1 | 1 | 1 | 0 | Feldman describes the motivation of competing against incumbents like Nvidia and Cisco daily. Stebbings enthusiastically agrees and praises Feldman's perspective. |
Statements from this episode (54)
No Company Ever Went Bankrupt by Overpaying Extraordinary Talent
“No company ever went bankrupt by paying extraordinary people too much.”
Cerebras Closed AI Chip Sector's Largest Raise at Highest Valuation
“Well, look, it was the largest raise ever done in, in our category. It was done at the highest valuation and with the premier investors.”
Feldman: Cerebras added five US data centers in 2025
“We added five this year in the US to add more data centers, and we have more big ideas.”
Feldman: Quantization from 8-bit to 4-bit won't solve AI scaling
“I think incremental improvements make believe gains achieved by dropping from, you know, eight bit to four bit. Those aren't gonna get us to the promised land in AI.”
Feldman: Cerebras still intends to go public despite pre-IPO round
“We still have every intention of going public.”
Feldman: Cerebras customers request between 5M and 40M queries per second
“We have customers coming to us and saying we would like between five and 40 million queries per second.”
Multi-Billion-Dollar AI Deal Announcements Are Just Options on Future Compute
“I think you should think about sort of these announcements as sort of options on the future. Right. That's really the way to think about it is in an unknown environment, how can I take an option on the future? I don't know if I'll use it all, but I'll pay some…”
Cerebras making multi-year, multi-hundred-million-dollar supply chain and data center commitments
“We're making five and seven year investments in, in a data center capacity. We are making hundreds of millions now on term of billions of dollars of bets in supply chain.”
Feldman: Investors generate no alpha by betting on dominant incumbents
“Truth is most things don't work, and most of the time Goliath wins. Right? But there's no alpha in that. There's no money made for you or me betting on the biggest of the big dogs to continue not to lose.”
Feldman: Global economy will reorganize around AI within five years
“However is it possible that our economy looks very different? In five years. Is it possible that, ah, the things we value, ah, are very different? That we have reorganized around AI. We've seen a major bump in, in labor productivity. We have benefited dramatic…”
Nvidia Conceals Massive Chip Field Failure Rates to Keep Customers Waiting
“You don't talk about the field failure rate of your products, which are massive. Rather you sort of keep talking about the future in an effort to convince people to wait, to make a good decision rather than go with technology that's, that that's better and pre…”
Nvidia Invested in OpenAI Primarily to Lock Up Chip Demand
“Beyond the fact that, that Nvidia has chosen to try and lock up a portion of OpenAI's demand. By investing in them. That, that, that's about as much as you can say.”
Feldman: Claims of a two-year AI chip depreciation cycle are wrong
“I think people are clearly still getting value from H-one hundreds, and that's more than two years, right? So you know If you say it's a two-year depreciation, you're empirically wrong, right? I mean, they are and I think people are still getting value from A-…”
Feldman: Chip depreciation is driven by generational performance jumps, not calendar time
“The question of depreciation is how much faster are future generations than the current generation? That's the actual question on depreciation. Because what you're, with depreciation, you're saying, at some point, it's no longer worth using a part that's fully…”
Real AI Chip Performance Gains Are Only 2x to 2.5x Per Generation
“A little bit of engineering digging probably leads you to the conclusion that you're getting two, two and a half X per meaningful generation move. Not more, right?”
Feldman: Chip FLOP increases are wasted without proportional memory bandwidth
“If you make one part fast and the other part, you don't move, doesn't move as far forward. It becomes the new bottleneck. And your, it doesn't matter how, how many flops your chip has. If you can't get data onto and off of the chip, Those are wasted.”
Feldman: Memory bandwidth, not compute speed, limits GPU AI inference
“And that includes memory, which has, for inference, is the fundamental limiter for the GPU architecture. And so it doesn't matter how much faster the chip goes. It matters how much faster the memory bandwidth is.”
Feldman: Prior to Cerebras, no chip exceeded 840 square millimeters
“Nobody in the history of the computer industry had been able to build the chip bigger than about 840 square millimeters.”
Elon Musk and Tesla Failed at Building Wafer-Scale Chips for Dojo
“And after we did it Elon tried Dojo and they failed.”
Feldman: Cerebras chips are faster than Nvidia at both training and inference
“No, we're faster on both, but the software challenges In training are real.”
In AI Inference, Nobody Cares About Nvidia's CUDA or PyTorch
“In inference, the truth is, nobody cares about CUDA. Nobody even cares about PyTorch. All right. What they want is an API.”
Feldman: Vastly more people do AI inference than AI training
“To move people off GPUs in inference, and the number of people doing inference is vastly higher than the number of people doing training.”
Feldman: AI inference growth compounds across users, frequency, and compute per query
“The greater growth of inference is the number of people who use it, Times the frequency of use. Times the amount of compute needed per use. Right? It is three different variables multiplied by each other. The problem is they're all growing fast.”
Feldman: The tech industry is 100% underestimating the AI inference market size
“I think the reason I'm a hundred percent sure that, that we are underestimating the market is because of that premise.”
Feldman: Massive AI productivity gains require reorganizing workflows around AI
“If you use OpenAI in various of their competitors, the way you use Google, you'll see a very modest jump. In productivity. If you use them in a fundamentally different way, that was Sam's point, you'll see a huge jump. And if we use, if we reorganize ourselves…”
Feldman: Younger users treat AI as an operating system for life
“Younger users are using it in a different way than older users. Older users are replacing something they already had. Younger users are using it in a way that never existed before as an operating system for life.”
US AI Power Constraint Is Geographical Mismatch, Not Total Supply
“People often say we don't have enough power in the US, and this is strictly wrong. We have plenty of power. It's in the wrong places, right? It's not where we have people or where we have fiber optic cable.”
Biden Administration AI Policy Was Misguided and Afraid Compared to Trump's
“I think it's confusing. And on net, it's probably more to help. The Biden administration was misguided and afraid. And I think the Trump administration, I think to his credit, they surrounded himself with some smart people and in the AI space. And I think they…”
Feldman: Nuclear power is not unavoidable for powering AI infrastructure
“No, it's not unavoidable. It's a very reasonable decision for countries that don't have lots of alternatives. I mean, you know, Canada has more falling water than anywhere else on Earth, right? The opportunity for Canada to develop the cheapest power on Earth …”
Feldman: S&P 500 Investors Face Heavy Tech Concentration Risk, Not True Diversification
“If people continue to think the S&P is sort of an index of the global economy or, and it's not, It's 30% or 50%, seven companies. Then they're exposed to sector risk that they weren't signing up for. They thought they were diversified. And in fact, they're hea…”
Nvidia Is the 21st Century's Greatest Company; Valuation May Be Too Low
“I think they've proven themselves to be the greatest company of the first quarter of the 21st century, right? 21st century. They've proven themselves to be the extraordinary company in the first quarter of the century, and I don't know if four trillion's right…”
Feldman: TSMC fab expansion speed bottlenecks global AI chip supply
“TSMC can't build fabs fast enough. I think the truth is, is that these are both for TSMC and Samsung. These fabs are the most amazing manufacturing plants on the planet. And, you know, these are 30,000,000,050 billion dollar factories. And the, their ability t…”
Feldman: Elon Musk builds facilities in 6-8 months versus 18+ months for others
“You know, somebody like for Elon, who's the fastest in the world and maybe the best at building plants and large construction projects, it takes six months, eight months. And for the rest of the world, it takes a year and a half. Maybe longer.”
Feldman: Data center financing attracts capital markets by mimicking real estate bonds
“It has a structure that they really understand, right? It looks like a bond to them. It looks like a piece of real estate. You get a tenant. They pay rent every month. You can loan against that. You get an investment grade tenant. That's basically a bond. It h…”
Feldman: CoreWeave's financial engineering popularized debt-backed data center financing
“And I think CoreWeave and some of their sort of financial engineering and innovations there helped the world to see that.”
Fast-Moving Software Development Culture Fails in Chip Manufacturing
“Modern software does not fit well In a chip making framework. I mean, weekly sprints don't work well on two year long projects. You know, move fast, break things often is not the way you think in the chip world. The way you think in the chip world is measure t…”
Feldman: Tech giants achieved chip design success primarily through acquisitions
“Where there's been success, it has frequently been acquired. Apple got into the chip business through buying PA Semi. Amazon got into the chip business through acquiring Annapurna.”
Feldman: Cerebras achieved higher valuation because rival chip startups had negative margins
“I think the reason we were able to raise it a higher valuation and from better investors and more money is because we had them. Right? And others who were out looking for mark, looking for money had negative margins.”
Nvidia Charges Gross Margins Between 78% and 85% on High-End Chips
“Well, because they want to get rid of the 78% gross margin that NVIDIA is charging. That's why it makes sense. Just like that. It might be, on the high-end chips, 85%.”
Feldman: Cerebras passed on a major China deal in 2019 before export controls
“We had a huge opportunity in, in 2019 to do a deal in China, and I decided to pass because I didn't think it was the right thing to do. Long before the Department of Commerce limited exports to China, I didn't think it was right.”
Chinese Government Reimburses Venture Funds for Losses on AI Investments
“They backstop their venture groups, right? So if you lost money in an AI company, the government would make you whole.”
Local fire ordinances delayed Samsung's Texas fab by up to 10 months
“Samsung built a fab in Texas, and they had to change the design of a fab because they had local fire ordinance. And the U.S. Government worked for years to get deployment of billions of dollars in taxes, and a local fire ordinance set them back eight months, 1…”
Feldman: US universities are starved of compute for AI research
“I think we have starved our universities of compute. If you want to do interesting training work at a university, very hard to get enough compute to do that.”
UAE Accounted for 75% to 80% of Cerebras' H1 2024 Revenue
“I think in the S-one it says and that was for maybe the first half of 24, we haven't published others, but a lot. I think 75, 80%.”
Cerebras Burned $7M Monthly During 15-Month Period of Chip Failures
“We had a period of about 15 months. Between about 2017 and early 2019 where we couldn't make one. And we were running a burn of about six million a month, seven million a month. And we stayed with it and our board stayed with it.”
The Silicon Industry Is Not a Place for 25-Year-Old CEOs
“The silicon industry is not a place for twenty-five-year-old CEOs, no matter how smart you are. The returns to having built parts before in what we do are enormous.”
Feldman: Many AI projects fail due to bad data, not AI
“I think most, many AI projects fail on those fronts. Have nothing to do with the AI. They fail because the data was a disaster.”
Feldman: Data cleaning and pipeline infrastructure are profoundly under-invested in AI
“I think that's an area that, that's profoundly under-invested in.”
Feldman: AI adoption will diffuse slowly rather than causing immediate labor disruption
“Economic dislocation. Isn't resolved in very short periods of time. That might be true in 15 years. But I think in the three to five year time frame I certainly don't believe that will be the case. I think AI will the adoption of AI or the diffusion of AI into…”
DeepMind's AlphaFold Has Not Yet Produced a Single Drug
“AlphaFold solved one of the hardest problems in in chemistry. A problem that had been open for years. Name a drug that's resulted from it. Not one.”
Feldman: AlphaFold increased demand for X-ray crystallographers rather than replacing them
“Sort of where are the X-ray crystallographers who are displaced because of it, right? That was what X-ray crystallographers were doing only physically. They're not out of work. They're, in fact, there's more demand for them.”
Feldman: AI will surpass human entry-level workers at spreadsheets and research summaries
“I think what it means to be entry level in a company will change a great deal because what entry level has generally meant in consulting firms and at investment banks has been doing shit work in particular being really good at spreadsheets and writing summarie…”
Feldman: Cerebras sales would default to NVIDIA without 10x superiority
“Every dollar we sell is a dollar that if we didn't work at it, if we didn't think, if we didn't invent, if we weren't 10 times better would default to NVIDIA.”
Feldman: Building something great requires every waking minute, not a 40-hour week
“This notion that, that somehow you can achieve greatness. You can build something extraordinary by working 38 hours a week and having work-life balance. That, that is mind-boggling to me. It's not true in any part of life. And your willingness to jump in and s…”
Shorts cut from this episode
▶ The Real AI Bottleneck · 20VC with Harry Stebbings (@40:25)
▶ Where is AI Going? · 20VC with Harry Stebbings (@0:00)
▶ You’re Using AI WRONG · 20VC with Harry Stebbings (@29:55)
▶ Competing with Nvidia · 20VC with Harry Stebbings (@1:16:44)
▶ The CHEAPEST Power on Earth? 🌎🔋 · 20VC with Harry Stebbing (@35:28)
▶ Is Regulation Holding back the USA? 🇺🇸 · 20VC with Harry S (@58:48)