Andrew Feldman

1. Key Themes

The AI Compute Supercycle Has Outrun All Forecasts

Demand for AI infrastructure has exceeded every projection, creating acute shortages across chips, memory, and data centers simultaneously. This is not a temporary gap but a structural undersupply driven by a multi-year delay in data center construction that predates the AI boom.

"The demand for AI has sort of outpaced everybody's expectation, everybody's forecast. And so everybody's chasing. They're chasing chips or they're chasing memory or they're chasing data centers." 00:01:17

Data Centers Are the Binding Constraint on AI Progress

Data centers historically moved at the speed of real estate — two-year build cycles — while AI demand accelerated in months. The physical infrastructure layer is now the primary bottleneck for everyone deploying AI at scale.

"We haven't made them fast enough. And so people are chasing data centers around the world. And that's what's happening. And so it is a major limitation for everybody right now." 00:12:31

Co-Design of Hardware and Software Is a New and Powerful Paradigm

For decades, an OS abstraction layer decoupled chip design from software development. AI's scale and speed requirements have collapsed that separation, forcing hardware and software teams to design simultaneously — and the advantages are enormous for those who can execute it.

"AI has gotten so large and speed is so important that what they're doing is they're thinking about sort of the design together... when you sort of design things together, the advantages are enormous. And so this is something that's really taken shape right now." 00:16:16

Nvidia Is Actively Using Its Balance Sheet to Foreclose Competition

Feldman identifies a specific and non-obvious competitive threat: Nvidia investing in NeoClouds and model builders to create dependency and lock out rival chip vendors — a structural market power play that goes beyond product competition.

"I'm concerned that there are ways frequently for NVIDIA to exercise market strength. There are ways for NVIDIA to use their balance sheet to limit competition. If they invest in a NeoCloud, the NeoCloud 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." 00:00:00

Fast Inference Is the New Battleground — and Speed Differential Is Massive

The industry is shifting from training to inference as the primary use case. Cerebras's value proposition is not marginal speed improvement but a claimed 20x inference speed advantage, which changes the economics of AI deployment fundamentally.

"We make AI with training, right? And we use AI with inference... when they want to use it, they want to use it and they want to be fast. And that's sort of where we come in. And where the fast is not by a little bit, but by 20x." 00:03:36

Sovereign AI and Stack Independence Are Strategic Imperatives

Drawing on Alex Karp's framing, Feldman argues that dependency on any single vendor — whether chip maker or model provider — is strategically dangerous for nations and enterprises alike. The right posture is optionality at every layer of the stack.

"You shouldn't be dependent on NVIDIA. You shouldn't be dependent on one model maker. I think that that rarely works out well. What you'd like to be as a nation or as a big company is you'd like to have choices... you like choices at each layer of the stack." 00:23:20

The Data Center Innovation Wave Is Just Beginning

Infrastructure that was static for 20 years — generators, batteries, chillers, cooling — is now being reinvented from scratch. Fuel cells, jet engines for power generation, and new turbine architectures are all being deployed, signaling a massive opportunity in data center infrastructure technology.

"The interesting about data centers hadn't changed a lot in 20 years... And suddenly there's this sort of intense demand and sort of people are looking to innovate. They're using fuel cells and guys are using jet engines like Boom to generate power for these data centers." 00:13:11

AI's Highest-Order Impact Will Be Biological and Medical

Feldman frames AI's most meaningful long-term contribution not as productivity or code generation, but as eliminating cancer and automobile deaths — two of humanity's largest killers — within a 20-25 year horizon.

"I think that 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." 00:24:34

2. Contrarian Perspectives

"Free Tokens" Offers Are Drug Dealing, Not Generosity

The common industry narrative is that free compute credits from AI giants are a democratizing force for startups. Feldman reframes this as predatory dependency creation — a deliberate lock-in strategy that startups should consciously subvert by taking from multiple vendors simultaneously.

"I think these are drug pushers. Here's a little girl. Try a little bit. Just a little bit. And I think what the startups should do is they should take it and then never be dependent. Take some from AMD and come to us and see if we can get you some as well and avoid dependence. That never ends well." 00:20:59

Many High-Profile AI "Mega-Deals" Lack Real Teeth

Despite the press coverage around billion-dollar AI partnership announcements, Feldman — as an industry insider — suggests many of these deals are either aspirational or were born from desperation (excess GPU capacity from failed model deployments) rather than genuine strategic conviction.

"I think there have been announced some sort of deals that didn't have teeth. Deals that could have teeth later... X had available capacity. And you got to ask why they had available capacity. They had available capacity because the Grok model wasn't used very much." 00:18:58

Five Founders Is Almost Always Too Many — But It Worked Here

Conventional wisdom sometimes celebrates large founding teams; Feldman flatly disagrees, calling five founders "in almost every case too many." The fact that it worked for Cerebras is a notable exception tied to a pre-existing working relationship from his prior company.

"My co-founders, we, there are five of us. Five is in almost every case too many founders. They all work, we all work together at my last company." 00:31:09

Humans Are Categorically Bad Drivers, and That's an Underappreciated Fact

Feldman makes a stronger claim than most AV advocates: it's not just drunk or distracted or elderly drivers who are dangerous — all humans are structurally poor drivers, and the sooner self-driving replaces them, the better. This is a more aggressive stance than the typical "complement human judgment" framing.

"Humans are terrible drivers, horrible, horrible drivers. It's not just when we're 16 or 18 or when we're not paying attention... we don't pay attention. We're on the phone. We're talking to our wife. We're worried about work. Well, that's even before people drink." 00:25:59

3. Companies Identified

Cerebras Systems

AI chip company founded by Andrew Feldman. Makes a wafer-scale chip 58x larger than any other chip on the market, targeting fast inference at 20x speed advantage over competitors. Recently IPO'd and signed a $20B+ hardware deal with OpenAI.

"While we build this sort of super big chip, the chip that's like 58 times larger than any other chip, it goes in a server. It goes in a metal enclosure that's about the size of a fridge for a dorm room." 00:13:39

OpenAI

AI model company. Named as Cerebras's largest announced customer in what Feldman describes as one of the biggest deals in Silicon Valley history.

"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." 00:00:00

Google / DeepMind

Named as an example of successful hardware-software co-design, where the TPU team collaborates directly with Gemini and DeepMind teams to mutually inform chip and model design decisions.

"One of the advantages Google has is that their TPU can be designed in collaboration with the team building Gemini or the team of DeepMind... they can inform their choices back and forth." 00:17:06

Anthropic

Named as a recipient of xAI's excess GPU capacity — an example of the compute leasing dynamic and how GPU constraint is reshaping cloud compute markets.

"They had these GPUs that were sitting around and that's a bad idea. And so they sold a whole block of them or leased a whole block of them to Anthropic." 00:19:11

Boom Supersonic

Noted as a non-obvious entrant into data center power generation — using jet engine technology to generate power for data centers, representing the breadth of industrial innovation now converging on AI infrastructure.

"Guys are using jet engines like Boom to generate power for these data centers." 00:13:39

New Limit

Brian Armstrong's longevity/biotech company, cited as an example of AI-enabled ambition in drug discovery — specifically the goal to eradicate all disease — that would have been considered pure hubris a decade ago.

"Brian Armstrong just came out with this company called New Limit. I think their goal is to eradicate all diseases." 00:27:25

Veritas

Enterprise software company, cited as the company Mark Leslie ran, which invented the file system. Named as context for Leslie's operating credibility as a mentor to Feldman.

"There's a CEO, former CEO named Mark Leslie. He was CEO of Veritas. They invented the file system." 00:30:16

Brex

Intelligent finance platform combining cards, expenses, and banking with agentic finance built in. Cited as the financial platform used by Sorcery and AI-native companies including Vercel, OpenAI, Anthropic, Granola, and Deepgram. [Sponsor]

MongoDB

Database platform used by 75% of the Fortune 100 and leading AI-native startups. Features vector search and embeddings from Voyage AI for real-time AI agent workloads. [Sponsor]

Assembly AI

Voice AI infrastructure company. Builds speech-to-text and speech understanding models used by Granola, HeyGen, Ashby, and ClickUp. [Sponsor]

4. People Identified

CEO and co-founder of Cerebras Systems. Serial entrepreneur who previously created ~100 millionaires at his last company; created approximately 1,000 employee millionaires at Cerebras's IPO. Led the $20B+ OpenAI hardware partnership.

"At IPO, it was more. It was about a thousand. Current and former." 00:06:48

Sachin (from OpenAI)

Named as Feldman's on-stage partner at the Raise conference to discuss the Cerebras-OpenAI partnership and the importance of fast inference. No last name provided.

"I'm on stage with Sachin from OpenAI." 00:03:05

Pierre Lamond

Veteran venture capitalist and semiconductor expert. Invested in Cerebras at age 84 and served on its board. Described as having forgotten more about making chips than Feldman will ever know.

"There was a venture capitalist named Pierre Lamont and he's now in his 90s. He invested in us at Cerebras when he was in the ripe age of 84... he's forgotten more about making chips than I'll ever know." 00:30:16

Mark Leslie

Former CEO of Veritas Software (inventor of the file system). Named as a key mentor to Feldman for his exceptionally high standards and leadership teachings.

"There's a CEO, former CEO named Mark Leslie. He was CEO of Veritas. They invented the file system. And these were sort of wise people. And they had extraordinarily high standards." 00:30:16

Alex Karp

CEO of Palantir. Cited for his public position on sovereign AI — that companies should own their full stack including data and models — which Feldman partially endorses and partially nuances.

"Karp said not too long ago on CNBC... talking about sovereign AI and how every company should own the full stack." 00:22:51

Brian Armstrong

CEO of Coinbase. Named as the founder of New Limit, a longevity/drug discovery company with the goal of eradicating all disease.

"Brian Armstrong just came out with this company called New Limit. I think their goal is to eradicate all diseases." 00:27:25

Tony Kim

Referenced from a prior Sorcery episode discussing new data center architecture. Described as an authority on how data centers are being rethought as AI factories. (Tony Kim is a Managing Director at BlackRock focused on technology investing.)

"We just had Tony Kim on and we were talking about the new architecture of data centers." 00:12:56

5. Operating Insights

Build Dependency Resistance Into Your Vendor Strategy From Day One

Feldman's warning about free compute credits as a lock-in trap contains a concrete operating prescription: deliberately diversify across compute vendors even when one offers free or subsidized access. The cost of switching later is far higher than the cost of distributing usage across AMD, Cerebras, and Nvidia from the start.

"What the startups should do is they should take it and then never be dependent. Take some from AMD and come to us and see if we can get you some as well and avoid dependence. That never ends well." 00:20:59

Protect Your Proprietary Data as a Distinct Strategic Asset

In the sovereign AI framing, Feldman offers a specific operational principle: your unique data is your moat, and you must ensure the value of training on it accrues to you — not to the model vendor whose product improves as a result.

"If you have unique data, be sure you don't give that away. Be sure you have multiple choices in different parts of the stack, but you get the credit for your data. And it doesn't help make somebody else's model better." 00:23:44

Post-IPO Attention Management Is a Distinct and Underappreciated Leadership Challenge

The volume of inbound requests after going public is a non-obvious operational burden. Feldman describes it as essentially a second job requiring dedicated triage infrastructure (executive assistant plus chief of staff working in tandem), and distinguishes sharply between this noise and actual work.

"The number of people who want something has exploded. If you get 80 emails a day of different people asking you for something, that's not work. That's just this range of people who want something of one form or another." 00:05:54

6. Overlooked Insights

xAI's GPU Leasing Business Was Born From Model Failure — Not Strategic Vision

This was mentioned in a single throwaway paragraph but is highly significant for investors evaluating xAI and the NeoClouds broadly. The implication is that several NeoClouds offering GPU capacity may be doing so not because they have a superior cloud strategy, but because their AI model failed to generate demand, leaving them with stranded assets they are monetizing opportunistically. This is a red flag framework for evaluating any NeoCloud whose primary model product underperforms.

"X had available capacity. And you got to ask why they had available capacity. They had available capacity because the Grok model wasn't used very much. So they had these GPUs that were sitting around and that's a bad idea. And so they sold a whole block of them or leased a whole block of them to Anthropic... then they looked around and said, who else can we lease to?" 00:19:11

Cerebras's Wafer-Scale Architecture Is Natively Advantaged for Space Deployment

Feldman mentioned in a single sentence — almost as an aside — that Cerebras's chip is structurally better suited for space-based computing than conventional multi-chip GPU clusters, because the hardest problem in space compute is inter-chip communication latency, which Cerebras's monolithic wafer design eliminates entirely. Given the accelerating interest in orbital data centers and defense-related space compute, this is a non-obvious long-term positioning advantage worth tracking.

"We are really good for space because one of the hardest problems in space is getting all these little chips to talk to each other. And because we're a big chip, we don't have that problem." 00:14:47