May 26, 2026 · 1h 7m · news
Cerebras CEO on the Future of Data Centres, Token Costs & Memory | Should US Companies Sell to China · 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 interview, Cerebras Systems CEO Andrew Feldman joins Harry Stebbings to discuss the future of AI hardware, data center infrastructure, and geopolitical semiconductor policies, while sharing intimate leadership insights from Cerebras' historic IPO journey.
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 17.8% of the talking time here. How this is scored →
Harry as informed peer 3.3 Guest teaching 4.2 Guest disagreement 1.9 Harry pushing back 2.6
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 44:00 Firm refusal to sell frontier chips to China
Andrew emphatically rejects selling cutting-edge semiconductors to China, rejecting commercial interests in favor of national security concerns against an industrial adversary.
Hardest push from Harry ▶ 46:30 Challenging China chip export restrictionsHarry directly challenges Andrew's hardline stance, arguing that withholding chip sales simply incentivizes China to build superior self-reliant capabilities.
Biggest teaching moment ▶ 38:09 Lawyers and CISOs as enterprise adoption bottlenecksAndrew flatly rejects Harry's premise that messy enterprise data limits AI adoption, re-educating him on how risk-averse legal teams and security officials block technology implementation.
Harry holds his own ▶ 51:44 Defending European AI ecosystem depthHarry uses his deep market expertise in European tech to push back against Andrew's broad critique of European innovation, explicitly listing category leaders like DeepMind, 11Labs, and Synthesia.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| Welcoming Andrew Feldman & Congratulations on the Cerebras IPO | 3 | 4 | 1 | 2 | Harry introduces a macro framing comparing the AI infrastructure bubble to historical booms, citing investor Gavin Baker. Andrew clarifies that unlike fiber optics or railroads, current AI infrastructure lags existing demand rather than building on speculative future demand. | |
| OpenAI's Exponential Growth Insight & Elon Musk's Hardware Deal | 2 | 6 | 2 | 3 | Harry questions whether early compute commitments matter if companies can buy hardware on demand from Elon Musk. Andrew corrects him, explaining Elon sold older-generation H100 hardware rather than cutting-edge B200s, before detailing memory supply chain pressures. | |
| Real Utility and the Rise of Daily AI Use Cases | 3 | 3 | 1 | 1 | Harry references commentary from Sarah Fry to ask about model commoditization. Andrew explains that inference demand is surging because models recently became smart enough for daily utility across diverse demographic groups. | |
| Hyperscalers vs. Neoclouds and the Value of Full-Stack Credibility | 2 | 4 | 1 | 1 | Andrew outlines Nvidia's strategy of backstopping neoclouds to compete with hyperscalers. He uses a truck seat analogy to contrast raw cheap compute with full-stack enterprise cloud value. | |
| Semiconductor COGS, SRAM Advantage, and Token Cost Reduction | 3 | 4 | 1 | 2 | Harry asks how semiconductor COGS will evolve given 5x memory price increases. Andrew details how Cerebras relies on TSMC-etched SRAM rather than HBM memory or CoWoS packaging, insulating them from market bottlenecks. | |
| Google's TPU Strategy, Full-Stack Ownership, and the NeoCloud Valuations | 5 | 4 | 2 | 4 | Harry presents a detailed thesis that Google's full-stack ownership guarantees lowest token costs and asks if neoclouds are dramatically overvalued. Andrew points out Google's single-customer volume constraint while defending CoreWeave's financial engineering. | |
| Speed as a Moat: Running Kimi Faster and the Value of Rapid Inference | 3 | 5 | 3 | 3 | Harry pushes on whether marginal speed improvements still matter past a certain threshold. Andrew forcefully rejects the premise, comparing slow AI inference to dial-up internet and arguing speed is essential for complex workflows. | |
| Fulfilling Large Deals, Customer Concentration, and the UAE Stargate | 3 | 4 | 2 | 3 | Harry questions customer concentration risks and operational pressure from massive multi-billion dollar deals. Andrew reframes concentration as a necessary milestone to build operational muscle before scaling further. | |
| Energy as the Ultimate AI Bottleneck | 4 | 5 | 3 | 3 | Harry raises concerns about data center energy bottlenecks and local permitting friction. Andrew laughs off permit delays as routine construction reality using a home renovation analogy, while admitting the tech industry handled community engagement poorly. | |
| Job Disruption: "AI-Washed" Layoffs and Engineering Productivity | 4 | 5 | 2 | 3 | Harry cites Marc Benioff's figures on developer token spend to question whether spending will scale to justify AI valuations. Andrew compares software token budgets to hardware EDA engineering tool costs to prove the economics work. | |
| New Tech Roles, AI Governance, and Enterprise Barriers (Lawyers & CISOs) | 3 | 5 | 4 | 2 | Harry asks if messy data structures are the primary barrier to enterprise AI adoption. Andrew flatly rejects the premise, identifying risk-averse legal teams and security officers as the true bottleneck. | |
| Legal Industry Tipping Points & Open-Source Complexity | 3 | 4 | 1 | 2 | Harry inquires about legal tipping points and enterprise anxiety around open-source AI. Andrew notes that open-source legal compliance is uniquely complex, especially with frontier models coming from Chinese research labs. | |
| Should US Companies Sell Semiconductor Chips to China? | 5 | 4 | 5 | 6 | Harry asks if US companies should sell chips to China, challenging Andrew with the counter-argument that restricting sales forces China to develop self-reliance. Andrew firmly rejects selling cutting-edge chips to industrial adversaries. | |
| Onshoring TSMC & Solving US Industrial Infrastructure Policies | 3 | 4 | 2 | 2 | Harry asks about domestic manufacturing policy regarding TSMC. Andrew critiques US regulatory fragmentation and proposes giving top chipmakers 20-year exemptions from local ordinances to rapidly build domestic fabs. | |
| Europe's Innovation Deficit: "Regulate, Tax, and Fail" | 5 | 4 | 3 | 5 | Harry asks if Europe should be worried about lagging in AI infrastructure. Andrew criticizes Europe's regulatory mindset, prompting Harry to push back by highlighting European champions like DeepMind, 11Labs, and Synthesia. | |
| IPO Strategy: Luck, Grit, and the Relentlessness of Going Public | 4 | 4 | 2 | 4 | Harry presses on whether Cerebras deliberately timed its IPO to beat peers like SpaceX or OpenAI. Andrew corrects him, insisting the timing was entirely driven by luck and persistence after long regulatory delays with CFIUS. | |
| Grit, Controlling Your Own Destiny, and the Realities of an IPO | 2 | 4 | 1 | 1 | Andrew shares reflections on maintaining focus on controllable execution through macro crises like the 2008 crash. Harry humorously points out he was only 11 years old during that period. | |
| Intellectual Horsepower, Stanford Roots, and Creating 800 Millionaires | 3 | 3 | 0 | 1 | Harry asks how wealth changes entrepreneurs. Andrew describes growing up on the Stanford campus where intellect was the only currency, expressing pride in creating 800 millionaires through the Cerebras IPO. | |
| Empathy, Board Support, and the Brutality of Hard Technical Challenges | 2 | 3 | 0 | 1 | Andrew highlights board empathy during an 18-month period where Cerebras burned $8M monthly struggling to solve a core hardware challenge. Harry wraps up the interview with warm mutual appreciation. |
Statements from this episode (33)
Feldman claims Nvidia funds neoclouds to create competitors for traditional hyperscalers
“I think it has been NVIDIA's strategy to try and create competitors for the traditional hyperscalers. They have funded and backstopped and over allocated to the neoclouds. They have created a dependence, which is probably not healthy.”
Feldman says Elon Musk sold outdated H100 hardware in suboptimal deals
“They got H 100. They didn't get the B 200. They didn't get the most current. They are a generation and a half, maybe two generations behind. So the, this was not a great deal. It was a good deal for Elon. He had them sitting around, but they were forced to tak…”
Feldman claims Micron achieves software-like 80% to 85% margins on memory
“Micron is producing numbers where they have 80, 85% gross margins. I mean, they're getting software gross margins on making memory.”
Feldman claims HBM memory shortages limit traditional GPUs but not Cerebras
“That is a limitation for all GPUs, but not us. We don't use it.”
Feldman notes SRAM costs remain stable with no supply shortages
“We use SRAM, and there's no shortage of SRAM. The cost of SRAM hasn't changed.”
Feldman says Cerebras avoids TSMC CoWoS supply bottlenecks by not using it
“We are advantaged by the fact that there are constraints on COOS at TSMC. We don't use COOS.”
Feldman claims Cerebras chips are 15 times faster than competing GPUs
“Now, we are 15 x faster because of architectural reasons.”
Feldman claims NeoClouds face massive cost disadvantages buying Nvidia hardware
“When Google or when Cerebris puts our equipment in our own data center, right, we have a significant advantage Over a NeoCloud, because NeoClouds are buying hardware with gross margins of 70, 80% for NVIDIA.”
Feldman praises CoreWeave for pioneering innovative debt financing in AI infrastructure
“I think CoreWeave has been an extraordinarily innovative company. I think they've solved a series of financial challenges with really innovative sort of financial engineering. They were the first to use debt very innovative way.”
Feldman predicts there will be zero market for slow AI inference
“Why do we believe that inference will be any different? There'll be zero marking for slowing them.”
Feldman notes AI data center power requirements rapidly escalated to multi-gigawatt scale
“And it used to be the case that 20 megawatts was a lot. And then a hundred megawatts was a lot. And then a gigawatt was a lot. And now we're running around looking for multi-gigawatt facilities. And that's, in any other time, 750 megawatts would have been a mi…”
Feldman says Sam Altman and Elon Musk uniquely envision 500-gigawatt AI infrastructure
“And by the way, that, that is exactly where, where I think Sam is the best in the world. Maybe Elon is where everybody else's brain shuts down, right? When you're trying to think about a hundred gigawatts or 500 gigawatts, those guys, they have sort of this ab…”
Feldman admits the tech industry did a poor job engaging local communities
“Yeah, I think our industry did a shitty job of engaging the community properly.”
Feldman claims modern AI data centers do not require massive water consumption
“Our data centers don't need to use a ton of water. They can recycle it. You can have a closed loop.”
Feldman asserts 90% to 95% of recent tech layoffs are AI-washed
“I think to date, most of the layoffs were AI washed. They were, because we did boneheaded hiring, During COVID. It is actually because a great deal of productivity gains has been, have occurred over the years that we're just now harvesting. The ability to gath…”
Feldman predicts AI productivity gains will increase engineering hiring, not reduce it
“If we get, as we get more productive, we do more things, we're gonna hire more engineers. We're not gonna hire less engineers.”
Feldman claims global software engineers represent a $5 trillion AI token market
“There are forty-seven million software engineers in the world. I mean, that's five trillion dollars just in software engineering token use.”
Feldman predicts AI will eliminate routine informational human resources jobs
“I think the role of HR changes fundamentally. The part of AR, HR that just, that answered questions, that provided information about benefits, that, that disappears.”
Feldman asserts legal and security teams bottleneck enterprise AI adoption
“The wide scale adoption and use of AI in organizations is today limited by security and legal.”
Feldman says Jensen Huang overrode Nvidia lawyers to adopt AI tool Cursor
“I think what's happening is the leaders are tipping. I think even Jensen told a story that, that he was battling with his own internal lawyers around the use of, I think it was cursor, and finally he just decreed. We're gonna do it. And I think I got that righ…”
Feldman says top Chinese open-source AI models lag closed-source models slightly
“This is made doubly worse by some of the best open source models were made by Chinese companies. And they are exceptionally good models. Kimi Ketu, Deep Seek. When the GLM, these are extraordinarily good models. They're not quite as good as the closed source m…”
Feldman asserts China will inevitably use advanced US chips for military applications
“If we sell leading edge technology to China, will their military use it? Everybody says yes. There is no debate on that point. Their military will use it. You ask a second question, which is, if you sell our leading edge technology, will they, will their gover…”
Feldman argues TSMC, ASML, and Samsung are effective semiconductor supply chain chokepoints
“I think the chip industry Requires you to go through TSMC, and TSMC requires you to go through ASML or Samsung. I think there are reasonable choke points to manage those challenges.”
Feldman claims the US power grid relies on outdated 1950s technology
“Their power infrastructure is extraordinary. And in the US, we are a patchwork of 19 fifties technology, if we're lucky.”
Feldman proposes granting TSMC and Samsung 20-year exemptions from local US ordinances
“I would allow TSMC and Samsung both to a 20 year period free from all local and legal, local ordinances, all of them, to build fabs in their desired location in the US.”
Feldman calls semiconductor fabs the modern pyramids of human manufacturing achievement
“Fabs are modern pyramids, Harry. I mean, they are the greatest things humans make in manufacturing, in the manufacturing world, by far.”
Feldman argues Europe's regulate-and-tax culture stifles tech entrepreneurship
“There has emerged in Europe a sort of be afraid of it, then regulate it, tax it, or sort of mentality that works against entrepreneurship.”
Feldman reveals CFIUS delayed Cerebras' IPO by 18 months
“We tried to go public a year and a half earlier, and we couldn't get it done because we bumped into CFIUS.”
Feldman claims Cerebras is the only pure-play AI public company
“But what we did know was that we had a chance to be the first and only AI peer play in the entire market. There's only one, and that's us.”
Feldman states the Trump administration is unwaveringly better for business
“Unwaveringly better for business. You know, there are things I agree with, there are things I disagree with in this administration, but unwaveringly better for business.”
Feldman asserts vendors inflate prices tenfold for companies preparing for an IPO
“As you prepare to go public, the number of people who call you and try and sell you stuff is insane. Just suddenly developing a presentation, Which should cost 20,000 dollars is a 200,000 dollar project.”