BREAKING: Cerebras CEO… | Sourcery Newsletter Summary

1. Key Themes

Inference Is the New Battleground in AI Compute

The center of gravity in AI has shifted from training models to running them — and speed is the differentiator.

"We make AI with training, & we use AI with inference. And so as the models and as the AI we made becomes useful, everybody wants to use it, & inference is the mechanism through which we use it. And so now we have smart AI, & people wanna use it. When they wanna use it, they want it to be fast, & that's where we come in, & where the fast is not by a little bit but by 20x."

Cerebras's $20B+ OpenAI deal — 750MW of inference compute over three years — is the clearest validation of this shift, representing the majority of Cerebras's ~$24.6B backlog.

The AI Infrastructure Deficit Is Real and Severe

Against the popular narrative that AI infrastructure spending is ahead of demand, Feldman argues the opposite: the build-out is structurally behind because real estate moves slowly and AI demand arrived suddenly.

"Data centers had historically moved at the speed of real estate. Somebody would decide to build a building, and two years later, after permits, they poured concrete, & the building would go up. Two years ago, nobody cared about AI. The AI explosion's happened so quickly. We've outstripped the demand for compute, for memory, and where do these things go? They go to these buildings, & we haven't made them fast enough."

The energy crunch is so acute that operators are "borrowing jet engines for power" — and Cerebras itself announced a 7x CS-3 production expansion alongside 200MW of European compute capacity by late 2027.

Hardware-Software Co-Design Is a Durable Moat — But Harder Than It Looks

The next competitive edge in AI chips isn't raw specs alone; it's integrating hardware and software development from the start.

"Historically you made chips & you ran software on them. AI has gotten so large & speed is so important that what they're doing is they're thinking about the design together. What changes could we make in software that would advantage the hardware, or as we're designing the hardware, what changes could we make that would make the software easier to run?"

Cerebras's visibility into OpenAI's roadmap, and Google's TPU co-development with Gemini/DeepMind, are cited as the clearest examples. The catch: "The misconception is that it's easy and all you need to do is get in a room. The software guys think one way. The hardware guys think a slightly different way. Anything you do to make it easier to write the software makes it harder to do the hardware. And these are really hard trade-offs."

Supplier Concentration Risk Is the Defining Strategic Threat for AI Buyers

Feldman's sovereign AI thesis centers not on ownership, but on optionality — the danger of locking into a single vendor for compute or models.

"You shouldn't be dependent on Nvidia. You shouldn't be dependent on one model maker. What you'd like to be as a nation or as a big company is you like to have choices."

The corollary on proprietary data: "If you have unique data, be sure you don't give that away, & you get the credit for your data, & it doesn't help make somebody else's model better."

AI's Biggest Long-Term Value Creation Is in Medicine and Education

Feldman argues that AI's true economic and social return dwarfs productivity gains in coding — and that critics systematically ignore the upside ledger.

"We have a chance for our children or the next generation not only to not die from cancer, but to not know anybody who died from cancer. I think that is a real, achievable goal in 25 years."

On education: "We've known for 2,000 years the right way to educate children, and we never do it. Ever... With AI, we can do that. We can get you a tutor that's right for you."

2. Contrarian Perspectives

The AI Infrastructure Build-Out Is Behind, Not Ahead

The consensus among critics is that massive data center capex is a bubble — that supply is racing well ahead of monetizable demand. Feldman directly inverts this: "We are behind." His reasoning is structural: physical infrastructure (permitting, concrete, power) compounds on multi-year timelines, while AI demand materialized almost overnight. The evidence: Cerebras itself is executing a 7x production expansion and building 200MW of European capacity just to keep pace with existing contracted demand from OpenAI — suggesting the backlog is real, not speculative.

"Free" Compute Credits from Nvidia Are a Predatory Lock-In Strategy, Not Generosity

The market largely treats Nvidia's investment-and-credit ecosystem as a sign of the company's confidence and partner-friendliness. Feldman frames it as deliberate competitive foreclosure:

"I'm concerned that they are ways frequently for Nvidia to exercise market strength. They're ways for Nvidia to use their balance sheet to limit competition. If they invest in a Neo cloud, the Neo cloud is less likely to use a non-Nvidia chip. If they invest in a model builder, there's pressure not to use other people's chips."

On free credits to startups: "I think these are drug pushers. Here, little girl, try a little bit. Just a little bit." His supporting evidence: the circular capital structure (invest → create captive customers → lock out competitors) is a documented pattern, not incidental.

Most High-Profile AI Deals Are Structurally Opaque — and Many Are Hollow

The market largely takes headline AI partnership announcements at face value. Feldman suggests many of them are closer to press releases than binding commitments — and that even industry insiders can't easily tell the difference:

"Even from inside the industry, he said, the true shape of many headline deals is hard to read, with some carrying real commitments and others little more than announcements."

He traced the GPU-leasing market back to a specific, somewhat accidental origin: X leasing spare Grok capacity to Anthropic. "X had available capacity because their Grok model wasn't used very much. So they had these GPUs that were sitting around, & so they leased a whole block of them to Anthropic. And they looked up & said, 'Whoa, that's a pretty good idea.'" The implication: a significant portion of the AI infrastructure market was not strategically planned but opportunistically assembled.

3. Companies Identified

Cerebras Systems (Nasdaq: CBRS) AI chip company, wafer-scale hardware Core subject of the article. IPO'd at $56B, briefly reached ~$95B market cap on day one, now trading around $60B. Builds the WSE-3 — a single wafer-scale chip with 4 trillion transistors and 900,000 cores — and the CS-3 system. Benchmarked at ~2,500 tokens/second/user on Llama 4 Maverick, more than double Nvidia's DGX B200.

"We announced in January a huge partnership, one of the biggest deals done in Silicon Valley history. We'll be doing more than $20 billion of hardware for them over the next several years."

OpenAI AI model developer Primary commercial customer for Cerebras inference compute. The $20B+ deal covers 750MW of inference over three years and represents the majority of Cerebras's ~$24.6B backlog.

"We'll be doing more than $20 billion of hardware for them over the next several years."

Nvidia GPU and AI accelerator market leader Cited as the incumbent Cerebras competes against, and as the central actor in a strategy to lock out competitors through investment and free compute credits.

"I'm concerned that they are ways frequently for Nvidia to exercise market strength. They're ways for Nvidia to use their balance sheet to limit competition."

Google / DeepMind AI research and cloud infrastructure Cited alongside Cerebras as a leading example of hardware-software co-design, specifically the TPU developed in tandem with Gemini and DeepMind.

"[Feldman] pointed to Cerebras's visibility into OpenAI's roadmap, and Google's development of the TPU alongside its Gemini and DeepMind teams, as the clearest examples."

X (formerly Twitter) Social media and AI Identified as the inadvertent origin point of the GPU-leasing market, when spare Grok capacity was leased to Anthropic.

"X had available capacity because their Grok model wasn't used very much. So they had these GPUs that were sitting around, & so they leased a whole block of them to Anthropic."

Anthropic AI model developer Named as the first recipient of GPU leasing from X, which sparked the broader Neo-cloud leasing model.

"[X] leased a whole block of them to Anthropic. And they looked up & said, 'Whoa, that's a pretty good idea.'"

SpaceX Aerospace and satellite Referenced in the context of multi-billion dollar AI deals and the broader discussion of data centers potentially moving to space.

(Timestamp at 18:31): "Inside SpaceX's multi-billion dollar AI deals."

New Limit (Brian Armstrong's longevity company) Biotech / longevity Cited by Feldman as an example of AI-driven drug discovery and the broader longevity wave alongside GLP-1 work.

"[Feldman cited] Brian Armstrong's New Limit and the broader wave of longevity and GLP-1 work."

AMD Semiconductor company Named as one of the alternative compute suppliers founders should diversify toward rather than remaining dependent on Nvidia's free credits.

"His solve for founders is to take the offer without becoming reliant on it, spreading demand across suppliers including AMD & Cerebras."

Palantir Enterprise AI and data analytics Referenced via CEO Alex Karp's sovereign AI thesis, which Feldman broadly agreed with but reframed around optionality rather than full ownership.

"Palantir CEO Alex Karp said enterprises want to 'own the means of production' instead of 'transferring their alpha' to OpenAI or Anthropic. Cerebras CEO Andrew Feldman says Karp is right."

Meta Social media and AI Briefly referenced in a sidebar tweet about Meta's Muse Spark 1.1 pricing 75% below OpenAI and Anthropic, signaling the onset of AI model price wars.

"BREAKING: Meta just pricemogged OpenAI & Anthropic with Muse Spark 1.1 that's reportedly 75% cheaper, putting new pressure on both open & closed models."

4. People Identified

Andrew Feldman Co-Founder & CEO, Cerebras Systems Primary interview subject. Serial entrepreneur who previously founded a company that produced ~100 millionaires; Cerebras's IPO produced close to 1,000. Architect of the $20B+ OpenAI deal.

"Getting to an IPO is not the end of a journey. It's sort of a plateau. It's the arrival at corporate adulthood. It is the achieving one plateau so that you can climb others, & our opportunity's gotten bigger."

Alex Karp CEO, Palantir Referenced for his sovereign AI thesis — that enterprises should own their AI stack rather than cede "alpha" to foundation model providers. Feldman agreed directionally but narrowed the argument to optionality.

"You shouldn't be dependent on Nvidia. You shouldn't be dependent on one model maker. What you'd like to be as a nation or as a big company is you like to have choices."

Pierre Lamond Venture Capitalist (now in his 90s) Credited by Feldman as a foundational mentor and early Cerebras investor. Invested in Cerebras at age 84.

"There was a venture capitalist named Pierre Lamond, and he's now in his 90s. He invested in us at Cerebras when he was the ripe age of 84. He's forgotten more about making chips than I'll ever know."

Mark Leslie Former CEO, Veritas Cited by Feldman as one of the "wise people" who taught him to hold himself and others to a high standard.

"[Leslie was] one of the 'wise people' he said taught him to demand a great deal of himself and others."

Brian Armstrong CEO, Coinbase; Founder, New Limit Referenced as an example of a tech leader deploying AI toward longevity and drug discovery.

"[Feldman cited] Brian Armstrong's New Limit and the broader wave of longevity and GLP-1 work."

5. Operating Insights

Take Free Compute Credits — Then Immediately Diversify Away From the Provider

Feldman's tactical advice to founders receiving Nvidia's free compute credits is not to refuse them, but to treat them as a one-time resource, not a foundation. Lock-in, in his framing, is the existential risk.

"They should take it and then never be dependent. That never ends well." His prescribed action: spread workloads across AMD, Cerebras, and other suppliers from day one.

Protect Your Proprietary Data Above All Else in Any AI Partnership

When engaging with foundation model providers — whether for training, fine-tuning, or inference — the single non-negotiable is ensuring your data doesn't improve someone else's model without your benefit. This is Feldman's operationalization of the sovereign AI thesis for companies that can't afford to build a full stack.

"If you have unique data, be sure you don't give that away, & you get the credit for your data, & it doesn't help make somebody else's model better."

Building Conviction Over Fashion Is a Talent Acquisition Filter, Not Just a Culture Statement

Feldman uses an operator's lens on this: he explicitly filters for people who built hardware when it was uncool, not when it became lucrative. This creates a team with intrinsic motivation and long time horizons — which maps to the multi-year bets required in semiconductor development.

"For the type of people I love working with, they like building stuff, & they like building hard stuff when it paid a little, when it was out of fashion. When hardware was uncool, they were still building hardware, because that's what they like to build."

6. Overlooked Insights

Data Centers Are Already Borrowing Jet Engines for Power — Infrastructure Hasn't Been Stress-Tested Like This in 20 Years

The energy constraint inside AI data centers is so acute that operators are repurposing jet engine technology for power generation. Feldman flagged this as a sign that parts of the data center stack that had been static for roughly two decades are now being redesigned under duress. This points to a significant secondary investment opportunity in power infrastructure, fuel cells, and non-traditional energy sourcing for compute — a layer largely absent from mainstream AI infrastructure coverage.

"[Feldman noted] the scramble is reshaping the data center itself, with operators reaching for fuel cells and jet engines to power parts of the stack that had gone unchanged for about 20 years."

Space-Based Data Centers Are a Real Engineering Thesis, Not Science Fiction — But Still 5+ Years Out

Feldman offered a concrete technical rationale for why wafer-scale chips like the WSE-3 are uniquely suited for orbital compute: a single large chip eliminates the inter-chip communication problem that makes clustering many small chips in space impractical. This is not a vague futurist claim — it's a product-architecture argument. For investors tracking long-range compute infrastructure bets, this is worth tracking earlier than the timeline suggests.

"I don't think we're in danger in the near term of having a data center in space. I think it's more than five years away."