Andrew Feldman | Uncapped with Jack Altman Summary

1. Key Themes

The "100x Rule" for Hardware Startups Attacking Incumbents

Andrew Feldman's core strategy for Cerebras was built on the idea that incremental improvement is a losing strategy against an entrenched giant like NVIDIA, because incumbents can simply cut prices or bundle to defend share. Eric Vishria quantified this: "if you think of whoever the incumbent is in whatever market... they're getting say twice as good every year... For a new company to get to scale... it's going to take five years at least... you're basically at two to the fifth. So that puts you at 32x. And then you need at least a multiple advantage of that... you're basically at 100x." 00:07:19 Feldman echoed this: "You can't get there with a collection of modest improvements... you got to aim at 100, 500,000 times better if you're going to arrive... ahead of them." 00:09:19

Radical Vertical Integration Is the Only Path to Radical Speed

Cerebras deliberately took on the full stack — chip, board, system, software, API — because true 10-500x advantages can't be bought off a shelf. "When you build a chip the size of a dinner plate, you can't go to a catalog and find a heat sink... nobody has stuff ready for you." 00:06:10 The payoff of this painful process was deep, defensible expertise: "We didn't start world leaders in packaging. We're right now the best in the world of packaging. We earned it failure after failure, year after year until we got it." 00:06:35

Decision-Making Under Radical Uncertainty: "Hop to Hop" Not "Fixed Circuit"

Feldman described navigating strategy in deep tech like an internet router rather than an old dedicated telephone circuit — moving forward step by step and reassessing at each new vantage point rather than committing to one long-range fixed vision. "We knew there was a pot of gold out there, but the path to it we knew was unknowable. So each time you get over a new mountain, you look around and you re-decide." 00:30:11 This is how Cerebras evolved from a training-focused chip company (built pre-transformer, when TensorFlow and ResNet dominated) into an inference powerhouse once the trajectory of AI demand became clear.

The Supply Chain Behind AI Is a Series of Underappreciated, Nearly-Irreplaceable Monopolies

Feldman walks through how sand becomes a ChatGPT answer, highlighting that ASML is the sole global maker of extreme-UV lithography machines: "Right now it's only ASML. In the world? In the world... this is a true monopoly... They have technology that others haven't been able to replicate." 00:18:38 TSMC then converts these machines into functioning fabs at a scale almost nobody else can replicate: "the actual making of the fab is skills that only TSMC has. And even there, they can't do 12 at once." 00:21:57

Data Centers Are Now the Binding Constraint, and the Buildout Is Historically Unprecedented

Jack Altman cites a striking statistic comparing AI infrastructure spend to past infrastructure booms: "1% of GDP was spent per year on highway and telecom... 2% for railroads... this AI build out is like 3.5% of GDP." 00:42:23 Feldman frames the mismatch simply: "AI is moving at the speed of software and data centers are moving at the speed of real estate. And that's why we're behind." 00:19:35

Tech's Communications Failure Around AI Infrastructure

Feldman is blunt that the industry mishandled community relations around data center buildout, particularly on water usage. "All the data centers in the U.S. use less than the California almond growers. Not by 1X or 2X or 4X, but between four and seven times the almond growers use more." 00:46:05 Yet: "as a community, we didn't do a good job of communicating with local communities... and now we're paying the price." 00:46:31 Altman adds: "it kind of goes with the whole theme of tech doing a terrible job communicating about AI in general." 00:47:00

Speed Creates New Markets Rather Than Just Serving Existing Demand Better

Feldman argues that Cerebras's focus on speed isn't just a competitive edge on existing use cases — historically, speed creates entirely new categories of product. "There's no market for slow search... When the internet was slow, Netflix delivered DVDs in envelopes. And when the internet got fast, they became a movie studio." 00:53:01 He connects this directly to the launch of GPT-5.1/"Sol" and expects a new wave of applications to emerge purely because instant frontier intelligence becomes available.

NVIDIA's Greatness Comes from Grit, Not Technology

Feldman offers a contrarian read on what actually made NVIDIA dominant — not CUDA, not chip architecture, but a decade of enduring failure as a public company. "For a decade... they traded horribly. And you're a public company and you're fighting tooth and nail and no one's listening to you... The relentlessness and the grit that that takes is awesome." 00:55:51

Hardware Founders Need Deep Domain Experience, Unlike Software/AI Founders

Feldman contrasts hardware with software-era founders (AI, social networking), who often succeeded by building for themselves without prior domain expertise. In chip-building, this doesn't work: "success has historically been predicted by some experience in the field... there's just not a lot of chance in college or in graduate school to actually build silicon." 00:33:29

2. Contrarian Perspectives

"We All Got AI Data Center Demand Completely Wrong — Except Sam Altman"

Feldman admits that virtually the entire industry — including Cerebras, TSMC, NVIDIA, and memory makers — underestimated the scale of compute demand, and that only Sam Altman correctly modeled the exponential. "What Sam's really good at and I think is so hard is he saw an exponential and wasn't afraid... Everybody else was afraid... I think we got it wrong. I think TSMC got it wrong. I think Nvidia got it wrong. Everybody, the memory guys got it wrong. We all got it wrong except Sam." 00:42:55 This is a striking admission from an industry insider that the "obvious in hindsight" AI infrastructure boom was in fact a near-universal forecasting failure among the most sophisticated players in the space.

More Government Involvement, But Ideally From People Who "Don't Understand"

When asked about growing government interest in AI, Feldman gives a surprising, almost sarcastic take that current political engagement is worse than none: "What we need is more people who don't understand making decisions... I mean, I think more people with absolutely no clue, making important decisions." He clarifies the real critique is a lack of substantive policy thinking: "there's really thoughtful discussions to be had about what's good for the US... and those aren't being had by the politicians. They're knee jerk." 00:51:29

Board Members Should Deliberately Stay Out of Technical Decisions

Contrary to the instinct that active, hands-on boards add more value, Feldman argues the best board behavior during Cerebras's most difficult technical stretch was near-total technical silence: "there weren't efforts by the board to try and solve technical problems. They didn't have that... that wasn't their expertise." 00:25:16 Vishria confirms this from the investor side: "there are periods of time in any company... where there's just... you just have to let the engineers engineer and the scientists do their thing and stay out of the way and keep them financed." 00:24:08

First-Generation Chips Are Supposed to Fail — Even for the Best Teams

Feldman states as near-doctrine that the first (and often second) version of any hardware product will be bad, using Google's TPU as evidence: "even really strong teams like Google's TPU team, the fourth one was good. The first two were... the fourth or fifth were really good parts. It takes years." 00:26:55 This runs counter to how outsiders (and possibly investors) often judge early hardware iteration as a signal of team quality.

CUDA and Chip Architecture Are Overrated Explanations for NVIDIA's Moat

Most market commentary attributes NVIDIA's dominance to its software moat (CUDA) or superior chip design. Feldman explicitly rejects both as the primary driver: "I think most people are wrong about what makes NVIDIA great... I don't think it's CUDA... I don't think it's their chip architecture." 00:55:28 Instead he attributes it to organizational grit forged during a decade of stock market underperformance — a much harder-to-replicate, culture-based advantage that investors may be underweighting when assessing competitive threats to NVIDIA.

3. Companies Identified

Cerebras — Wafer-scale AI chip company founded by Andrew Feldman in 2016 (Feldman's fifth chip startup). Built the first working wafer-scale processor after industry-wide skepticism it was possible. Mentioned as the central subject of the episode, now serving government, sovereign cloud (G42), OpenAI, and hyperscaler customers. "We're only interested in solving problems that other people can't solve." 00:59:23

NVIDIA — Dominant AI compute company. Cited as the incumbent Goliath Cerebras was built to challenge, and later analyzed as one of the greatest companies in history due to cultural grit rather than technology. "The most valuable company in the world... that sort of intensity and grit for a company of their size, in my view, makes them one of the great companies in history." 00:56:20

TSMC — The dominant chip fab, described as operating skills "only TSMC has," central bottleneck in global chip supply. "TSMC able to achieve things that others can't." 00:18:08

ASML — Sole global manufacturer of advanced photolithography machines used by all leading-edge fabs. "Right now it's only ASML. In the world." 00:18:38

OpenAI — Cited as a landmark customer win for Cerebras ("we won OpenAI") 00:58:32 and praised via Sam Altman's foresight on AI infrastructure demand.

AMD — Partner with Cerebras on inference disaggregation, delivering major throughput gains: "with AMD, we're seeing 5x additional throughput." 00:54:37 Also historically acquired Feldman's prior company, where his cofounder Sean was made a corporate fellow.

AWS — Partner with Cerebras on disaggregation as well, and maker of Trainium chips, seen as one of four major chip players in the category. "We're seeing similar numbers with AWS." 00:54:37

Google (TPU team) — Cited as an example that even elite technical organizations require multiple iterations to build a great chip: "the fourth one was good." 00:26:29

G42 — Sovereign cloud customer of Cerebras, part of the customer trajectory from government to sovereign cloud to frontier labs to hyperscalers.

Cognition and Cursor — Cited as examples of AI/software companies built by elite engineers building tools for themselves, contrasted with hardware's need for deep domain expertise. "If you look at cognition, if you look at cursor, these are some of the best engineers, software engineers in the world, they're building tools for themselves." 00:33:54

GE Vernova and Caterpillar — Traditional generator/genset manufacturers now facing new competitive innovation pressure due to AI data center power demand.

Boom (Boom Supersonic) — Cited for applying jet engine design principles to power data centers: "The guys at Boom want to use what they designed for jets to power data centers." 00:45:11

Bloom Energy — Cited for innovative fuel cell technology being adopted for data center power backup.

Amkor (referenced as "Amcor") and ASE — Packaging companies cited as part of the U.S. semiconductor supply chain that left the country due to policy failures.

Samsung — Mentioned as building a major fab in Texas requiring an entire dedicated power plant just to produce concrete for the foundation.

Global Foundries — Cited as one of the few companies capable of operating advanced fabs.

Microsoft (referenced as "Macro Hard"/"Macro Harder") — Vishria describes flying over their data center construction near Memphis: "the size of buildings... the construction nearby and everything. I was like, whoa... That is tremendous." 00:50:12

SpaceX, Anduril, Palantir — Cited by Altman as fellow examples of hardware/deep-tech companies succeeding in a moment where "hardware is now hotter than hot."

4. People Identified

Andrew Feldman — CEO and founder of Cerebras, on his fifth chip startup, previously sold three chip companies and took one public (acquired by AMD). Renowned for building the world's first commercially viable wafer-scale processor after 18 months of failed attempts. "We drilled a hole in the wall to suck the air out and stared at a server... I said, holy crap, we've solved this problem that nobody in 75 years of compute had ever solved." 00:12:22

Eric Vishria — Benchmark investor and Cerebras board member since the early days, credited by Feldman as an unusually effective board member who stayed within his expertise. "I have a very small circle of confidence... you have to be like, hey, this is where you what you can do. And this is what you can't do." 00:24:08

Sean (Feldman's co-founder) — Hired as an individual contributor at Feldman's prior company at 25-26 years old; by the time of the AMD acquisition four years later, was made a corporate fellow, later became a founder and CEO/CTO of a public company. "There was just no end." 00:36:50

Sam Altman — Singled out as the only person in the industry who correctly forecasted the scale of AI compute demand, including the scale of Stargate. "He saw an exponential and wasn't afraid... we all got it wrong except Sam." 00:42:55

Jensen Huang — NVIDIA's CEO, praised for maintaining an "underdog" fighting mentality even after NVIDIA became a multi-trillion dollar company. "Could you imagine old tech leaders before Jensen doing that? No." 00:56:59

William Shockley — Inventor of the transistor at Bell Labs, was Feldman's childhood neighbor on the Stanford campus; credited by Feldman as the person whose move to the West Coast "created the foundation for Silicon Valley." 01:00:31

Amos Tversky — Nobel Prize-winning behavioral economist (with Daniel Kahneman), was another of Feldman's childhood neighbors, played weekly doubles tennis with Feldman's father.

Jack Dongarra — Cited as a pioneer of high-performance/supercomputing, quoted by Feldman on the industry's historical focus on flops production over flops movement: "We've been better at making flops and moving flops." 00:15:06

5. Operating Insights

Only New Mistakes — Institutionalize Failure Analysis as Culture

During the 18-month period when Cerebras couldn't get wafer-scale chips to work, the operating discipline was rigorous post-mortem analysis after every failure so the same mistake never recurred. "Each time we built it and it failed, we'd go through sort of good engineering practice. We'd do a full failure analysis. We'd understand it... we sort of had a mantra, only new mistakes, right? Only new failures." 00:11:28

Promote on Demonstrated Excellence, Not Tenure — Especially in Product/Go-to-Market

Cerebras deliberately staffed young, high-talent people in product and go-to-market roles while reserving deep experience requirements for chip engineering. "We don't have big company rules. You got to be in a job for this amount of time before you can get promoted... if you're extraordinary, we're going to give you more and more responsibility." 00:35:55

Identify Extraordinary People Within Weeks, Not Months

Feldman describes a fast, pattern-based method for spotting exceptional hires and collaborators almost immediately. "I think within the six or eight weeks of working with someone, you can tell if they're extraordinary... in the first email they send you, you go, whoa, that's exactly what I needed... Every list is in descending order of importance. There's not a lot of fluff. There's high signal." 00:38:34

Legacy Relationships Are a Compounding Asset in Hardware — Earn Them the Boring Way

Long-standing supplier relationships (TSMC, contract manufacturers) proved critical during periods of supply contention, and Feldman ties this explicitly to unglamorous, old-fashioned reliability. "You're good to your word, not once, not twice, not just in good times, but... over good and bad times... write a thank you note. Do what you say you're going to do." 00:39:39

Architectural Decisions Should Bet on Durable Math, Not Point-in-Time Workloads

Cerebras's first chip architecture deliberately avoided hard-baking acceleration for convolutional neural networks (the dominant workload at the time) and instead optimized for the underlying algebra common to all AI. This decision paid off when transformers emerged. "We decided not to sort of embed technology that would accelerate convolutional networks. Instead, we said... if we work underneath that and accelerate the algebra that underpins all AI we knew about, that was a really good decision because when transformers came out, we were the fastest at those." 00:32:13

6. Overlooked Insights

Disaggregated Inference Partnerships Could Let Cerebras Sell Speed Across the Entire GPU Landscape

Buried in the discussion of AMD and AWS partnerships is a potentially major strategic insight: Cerebras isn't necessarily trying to replace GPUs outright, but instead is building a model where its wafer-scale architecture "disaggregates" parts of the inference workload from any chip maker's hardware to boost throughput industry-wide. "We have an opportunity because of our architecture to do that with the entire GPU landscape. We could do it across the board... there are four major chip makers right now in our category... And we're already working with two." 00:54:37 This reframes Cerebras less as a pure NVIDIA competitor and more as a horizontal throughput layer that could plausibly partner with (rather than fight) most of the industry, including potentially NVIDIA itself — a very different (and much larger) TAM story than "faster chip vendor."

The Real Bottleneck in AI Buildout May Be Generators and Electrical Switches, Not Chips

While the popular narrative fixates on GPU/HBM scarcity, Feldman notes almost in passing that basic industrial hardware — generators and electrical transmission switches — has become a critical long-lead bottleneck, more so even than construction or labor: "Generators and electrical transmission switches are long lead time items right now. And those are hard to come by... those are the long poles usually." 00:49:00 This suggests underappreciated investment opportunity in unglamorous industrial equipment categories (gensets, switchgear, fuel cells) that feed data center buildouts — a theme reinforced by his mention of innovation from Boom (jet-engine-derived generators) and Bloom Energy fuel cells, but never developed into its own discussion by the hosts.