Andrew Feldman
Frontier Insights
Frontier Thesis: AI inference is bound by memory bandwidth, not raw compute. By leveraging wafer-scale on-chip SRAM, Cerebras bypasses GPU data-movement bottlenecks (HBM/CoWoS packaging), drastically slashing token costs and targeting Nvidia’s vulnerable pricing power.
Strategic Moves: Exploiting hardware scarcity to capture sovereign Gulf capital ($1B+ raised; G42 alliance) and validating extreme workloads while bypassing immediate IPO scrutiny and Nvidia’s balance-sheet lock-in.
Risks & Warnings: Severe revenue concentration (G42 alone accounts for ~87%), geopolitical exposure from export controls, enterprise adoption friction via legal/security gatekeepers, and macro demand volatility as Nvidia aggressively defends its moat.
Key Views & Dialogues
Open Source Wins, AGI Is Here, and Scorsese’s AI Toolkit with CEOs of Cerebras & Black Forest Labs
Cerebras reports a $25 billion backlog as faster inference turns token capacity and data-center execution into scarce product advantages, while open models take routine and sovereign workloads. Feldman expects its architecture to improve well over 2x in 18 months; staged releases, red-teaming and recursive model improvement remain catalysts—and risks—as multimodal AI reaches film and robotics.
View summary (7)
Cerebras says AI infrastructure is serving booked demand, not betting that customers will eventually appear. Andrew Feldman disclosed a $25 billion backlog and described customers as “trying to capture yesterday’s demand,” with football-field-sized facilities drawing more power than midsize cities. For investors, the operational challenge is keeping customers while demand outstrips the ability to build and fill data centers.
Reasoning models make inference speed and token availability part of product capability. Rombach says models such as “Fable” and “5/6” increasingly infer intent without a “prompt whisperer,” while his Hermes-agent experiment used a Bittensor project running Z.ai’s GLM-5.2 and debated its own research strategy before acting. Long-running jobs can already produce “amazing things”; Rombach’s hypothetical was that 15-times-faster Cerebras inference could compress “weeks or months’ worth of thinking” into a day.
Cerebras expects its young architecture to improve substantially faster than conventional Moore’s law in the next 18 months. Feldman’s view is “way over 2x,” because a new architecture still offers workload-specific optimization unavailable to a 20-year-old GPU design reliant on smaller fabrication geometries. Hyperscalers’ custom chips do not invalidate merchant silicon demand: their primary objective is avoiding total dependence on any supplier.
Open models are becoming the cost, control, and sovereignty layer beneath frontier AI. Feldman’s analogy is that users should not “take your Ferrari to the grocery store”: frontier models handle the hardest problems, while open models absorb routine G&A and the “cutting and pasting economy.” Regulated enterprises also want domestic, on-prem deployment; Cerebras therefore runs gpt-oss-120B, GLM, Kimmy, the “Quincy” set of models, proprietary OpenAI models, and customer-built models.
Staged release and red-teaming become defensible when a model can expose serious software flaws in hours. Rombach relayed that Palo Alto Networks found previously unknown bugs and paused other work for six weeks of patching; Feldman said giving government defenders “2 or 3 weeks to patch any obvious holes” is reasonable. He simultaneously warned against reflexive regulation and treated a future massive data breach as inevitable: “Something will happen.”
By every AGI definition used 20 years ago, Feldman believes the threshold has already been crossed; the open question is where recursive improvement stops. Repeatedly asking models to learn, retry, and broaden their search yields “not a little bit better answers, but vastly better answers,” compressing thousands of human-style generations into machine-speed iteration. His ledger includes real labor dislocation, but also a chance that children and their loved ones do not die of cancer and that every child receives an adaptive tutor.
Black Forest Labs sees generative media as the foundation of multimodal world models, not merely an image-and-video tool business. Robin Rombach traced FLUX back to latent diffusion, then toward models pretrained on images, video, and audio and combined with action prediction—the same model could help make a movie or become “a brain on a robot.” Near-term value remains human-directed: Martin Scorsese used the models to externalize a scene in his head, while IP owners can combine controlled generation, customized models, and potentially licensed fan creativity.
Cerebras CEO on the Future of Data Centres, Token Costs & Memory | Should US Companies Sell to China
AI demand is running ahead of infrastructure, with Cerebras carrying a $25B backlog after models became genuinely useful somewhere in 2025. Structural HBM shortages could persist for years as new fabs require $40B and five years, while Cerebras’s SRAM architecture avoids HBM and CoWoS constraints. Cerebras reports Kimi K2.6 running 6.7x faster than the next-fastest GPU cloud, while its $20B+ OpenAI deal keeps customer concentration in focus.
View summary (8)
- Feldman’s core anti-bubble argument inverts the historical analogy: rail in the 1880s and fiber in the late ’90s were “if we built it, they would come” — infrastructure ahead of demand — while AI is the exact opposite. “We can’t build data centers fast enough to keep up with demand”: Cerebras carries a $25B backlog, and Nvidia, AMD and others have their own. “We’re not building ahead of demand. We’re building behind demand.”
- The demand inflection has a date: “somewhere in 2025 the models got smart enough to be really useful” — before that, AI “was like cool and then nobody used it.” Now it’s sweeping demographics from his 85-year-old father to his 11-year-old niece, and demand does not peak if AI continues to improve in usefulness — the one bear case he concedes against the “electricity into intelligence” thesis.
- On the compute deal involving Elon: “They bought down rev gear” — H100s, not B200s, “a generation and a half, maybe two generations behind.” “This was not a great deal. It was a good deal for Elon” — forced action in an exponential, versus Sam Altman’s superpower of believing exponential demand data years out and contracting for power, data centers, and hardware ahead of everyone.
- Memory is the No. 2 shortage after fab space, and it’s structural: HBM comes only from Samsung, Micron and Hynix, a new fab costs $40B and takes 5 years, so “we’re going to continue to see memory shortages for at least the next several years” if demand holds. Micron is printing 80–85% gross margins — “software gross margins on making memory.” Cerebras sidesteps it entirely: SRAM etched by TSMC, no HBM, no CoWoS, 5nm while 3nm is the oversubscribed node.
- The speed thesis is absolute: “How big is the market for slow search? It’s zero”— you wouldn’t take $1,000/month to keep slow internet, so there will be zero market for slow inference. Cerebras posted Kimi K2.6 running 6.7x faster than the next-fastest GPU cloud “while one bozo at an analyst firm was on TV saying we couldn’t do it,” and is now digesting a $20B+ OpenAI deal — same concentration critique investors gave him at $1B with G42, one customer-size rung earlier.
- Nvidia has “funded and backstopped and overallocated to the neo clouds” to create hyperscaler competitors — “a dependence, which is probably not healthy.” Neo clouds buy hardware carrying Nvidia’s 70–80% gross margin before adding their own; Google and Cerebras putting their own silicon in their own data centers don’t, though Google’s full-stack edge is capped by having only one TPU customer: itself.
- On jobs: “most of the layoffs were AI-washed” — 90–95% of terminations are COVID over-hiring and old productivity gains finally harvested, “none of this is AI yet.” Meanwhile the real enterprise adoption blocker isn’t data cleanliness but lawyers and security (“no credit, no credit, failure, blame”), and Feldman sees no problem with software engineers using $50–100K/year in tokens — hardware engineers’ EDA tools are already closer to 15–20% of salary. At 47M software engineers, that is “$5 trillion just in software engineering token use.”
- On China, arguing against his own book: leading-edge chips sold to China will be used by its military and its state-backed industry — “there is no debate on that point” — so don’t sell; “I’d like to keep my industrial adversaries more than down rev.” His one free policy wish: give TSMC and Samsung 20 years free of all local ordinances to build US fabs — “fabs are modern pyramids.” The IPO itself was unblocked when “we got a new government” and likely-CFIUS concerns over large customers “disappeared” — the largest semiconductor IPO ever, $185 to $311.
Cerebras CEO, Andrew Feldman on Why Raise $1BN and Delay the IPO & Why NVIDIA’s Worried About Growth
Cerebras raised $1bn at a record valuation because it had margins while rivals had negative ones, despite still intending to go public; UAE orders may have represented 75-80% of revenue and consumed manufacturing capacity. Feldman argues two-year chip depreciation is empirically wrong and Nvidia’s pre-announcements and OpenAI investment signal defensive demand capture, while power sits in the wrong places; investors should watch utilization, customer concentration, permitting, and whether AI adoption delivers productivity.
View summary (7)
- Cerebras raised $1bn — the largest round ever in its category, at the highest valuation — led by Fidelity (“the Oxford or Cambridge of investing”) with Tiger Global, Valor and 1789, and Feldman says Cerebras still intends to go public: “We still have every intention of going public.” Feldman says the round was possible because Cerebras had margins while rivals shopping for money had negative ones — and he thinks the S-1 said the UAE accounted for about 75-80% of revenue, perhaps in H1 2024, with orders so large they “consumed all our manufacturing capacity.”
- Nvidia is showing a worried giant’s tells: “use your balance sheet more and your technology less” — buying business rather than winning it (as Cisco did from 1992-2001), “predatory pre-announcing” B300s before anyone can get B200s, and silence on field failure rates “which are massive.” The $100bn OpenAI investment “was designed for nobody to understand it… it’s just not an analyzable thing” beyond Nvidia locking up a slice of OpenAI’s demand.
- Two-year chip depreciation is “empirically wrong”: H100s are still earning past two years, A100s are at three-to-four and “could be as long as five or six.” The real depreciation variable is generational improvement — and apples-to-apples (8-bit to 8-bit) Feldman estimates 2-2.5x per meaningful generation, because memory bandwidth, not flops, is the binding constraint for inference on GPU architectures.
- Demand is unknowable even to buyers — customers ask Cerebras for “between five and 40 million queries per second,” with Harry highlighting the uncertainty as an order of magnitude — so read megadeals as options on the future, and read “up to $100 billion over five years” as “the great CYA word in marketing history.” Chance he’s still underestimating demand: “100%. I’ve been wrong.”
- On Mag7 concentration, the risk isn’t valuation (Nvidia at $4.5T: “maybe it’s too low”) — it’s mispriced diversification: the S&P “is not an index of the global economy — it’s 30% or 50% seven companies,” so holders carry sector risk they never signed up for. “Risk comes in financial markets where people fundamentally underestimate risk.”
- On energy, the scarcity narrative is “strictly wrong”: “We have plenty of power. It’s in the wrong places” — West Texas gas and upstate New York hydro sit far from people and fiber, nuclear is reasonable but not unavoidable, and America’s patchwork permitting (a local fire ordinance set Samsung’s Texas fab back 8-10 months) is the real handicap versus China’s central planning.
- Contrarian diffusion calls: Groq’s Jonathan Ross’s five-year AI labor-shortage prediction is “absolutely wrong — that might be true in 15 years”; AlphaFold won Nobels but “name a drug that’s resulting from it. Not one.” Productivity jumps only when society reorganizes around AI — the electricity-and-the-dynamo lesson — not when AI just replaces Google.
20VC: The Daily Deal: Coreweave IPO | Scale Hits $25BN on $2BN EOY Revenue | Sequoia’s 25x Return on Wiz | Tech Stocks Tank with Tariffs | Cursor: Defensible or Dangerous Example of Lost Moats in Tech
AI is creating a golden age of software and extraordinary revenue growth, but thin wrappers, dilution, and uncertain recurring revenue are exposing the limits of the old scale-equals-durability model. Moveworks’ $2.85B sale shows the value of enterprise distribution and workflow integration, while CoreWeave’s IPO, a possible $2B put obligation, and constrained M&A make execution and liquidity the next catalysts.
View summary (8)
- Jason Lemkin’s central call is that AI is opening a “golden age of software” while destroying the old assumption that scaled revenue is durable. If software expands from 2% to 4% of GDP, he sees 100 or more decacorns. Harry’s framing was that every VC has five or ten unicorns that are no longer unicorns, now growing in the single digits or teens; Lemkin expects late-stage slowdowns to become commonplace by 2026 and those stranded assets to become less attractive to acquire.
- Hot AI rounds can look cheap at roughly 20x forward ARR only because investors extrapolate extraordinary slopes while remaining uncertain whether the “R” is recurring and durable. Lemkin says markups still grade VCs and some LPs, while Harry notes that entering at $200M ARR can shorten duration; the counterweights are dilution that Lemkin estimates might approach 10% annually and a venture cash cycle stretching toward 20 years.
- “Triple-triple, double-double” remains elite operating performance but is currently “silver in a gold rush” to perhaps 80% of B2B investors. Lemkin’s prescription is a year-long, relationship-led process with roughly 50 investors and honest monthly updates: “It only takes one.” Founders without Cursor-like heat should avoid artificial Friday deadlines and may be better off accepting a valuation 30% lower if that gets the round closed.
- AI demand is producing real revenue faster than technical substance, while investors remain “addicted to top-line growth.” Lemkin has seen companies reach $2M in 60 days or eight figures inside a year with humans running prompts behind thin wrappers; yet AI efficiency does not necessarily reduce capital needs, because RevenueCat took a 2x productivity gain and “plowed it all into new hiring.”
- Moveworks accepted ServiceNow’s $2.85B offer after years of deep integration and accelerating demand made ServiceNow’s distribution especially compelling. Bhavin Shah said 250 of 350 customers already used ServiceNow, while Moveworks had five million users against ServiceNow’s 150M-plus; he concluded that Moveworks could not match the market’s scale and speed independently. The strongest ROI was not saving an employee three hours, but automating core workflows across systems such as SAP, Workday, Salesforce, Concur and Jira.
- Lemkin expects an IPO and M&A “gold rush” within roughly 18 months, but sees little evidence that PE will rescue slow-growth unicorns. He expects large private companies such as Stripe, Figma, Chime and Canva may list, yet says cash-flow-positive assets at $20M, $50M or even $100M are not attracting the “tire kicking” he once saw; PE instead appears to be combining holdings into “Frankensteins” such as SalesLoft with Drift and Gainsight with Skilljar.
- CoreWeave getting public was a major entrepreneurial achievement, but its real scorecard begins on days 180, 365 and 450. Andrew Feldman called an IPO “the beginning of adulthood”; Lemkin nevertheless worries that, as he understands the structure, last-round investors can put back almost $2B of stock if shares fail to trade 70% above the IPO price within two years. That deadline could invite shorts and sacrifice long-term decision-making to a fixed date.
- Feldman’s warning on fashionable hardware and defense investing is that accumulated experience still matters. A chip tape-out can consume $20M-$30M in non-recurring engineering costs—and a bug means paying again—while defense requires trusted relationships, cleared personnel, specialized facilities and patience with “Bible-sized contracts.” Commercial technology can still reshape warfare, but procurement and cost-plus incentives remain the constraint.
Andrew Feldman, Cerebras Co-Founder and CEO: The AI Chip Wars & The Plan to Break Nvidia’s Dominance
Cerebras argues that GPU off-chip HBM is a fundamental inference bottleneck, while wafer-scale SRAM could challenge NVIDIA’s near-total market share. Feldman sees inference demand exceeding 100 times today’s market within five years, with chip providers potentially worth more than model companies, though G42’s 87% revenue concentration remains a key risk.
View summary (7)
- Feldman’s core attack on NVIDIA: the GPU’s off-chip HBM memory is a fundamental architectural limitation for generative inference — serving one word from a 70B-parameter model means moving ~140GB of weights from memory to compute, again for every next word. Wafer scale let Cerebras use fast SRAM at capacity, and “what used to be their advantage is now weakness… it can be beaten and I think they know it.” His market-structure call: NVIDIA goes from “approximately all” of the market today to 50-60% share in five years — between Uber’s 90/5 and cloud’s even split.
- The CUDA moat is “not real at all” in inference — “you can move from OpenAI on an Nvidia GPU to Cerebras… with 10 keystrokes.” The real, rarely discussed moat is market-share leadership itself: Intel made “nearly a decade of catastrophic decisions” and still holds ~75-80% of x86. “I can make a bunch of bad decisions for a decade and only lose 20% share — the moat was just unbelievable.”
- Scaling-law gains continue — Feldman flatly rejects the “far along on compute, algorithms and data” consensus: “I think they’re wrong. I think we are early in all of them.” A GPU doing inference runs at 5-7% utilization — “95 or 93% wasted” — OpenAI’s o1 shows inference scaling laws “fully functional,” and we won’t be as dependent on Transformers in 3-5 years (“100%”). In five years training data is “almost all synthetic.”
- The inference market equation: users × frequency × compute-per-use — all three growing simultaneously, a rare condition. AI flipped from “novelty” to useful in Q4 2024, the market in five years is “way over 100 times bigger,” and faster/cheaper only expands it: “no examples in compute in 50 years in which by making things cheaper, faster, the market got smaller.”
- Chip providers will be worth more than model providers on a 5-year view. Today’s model valuations are option pricing — “uncertainty is a friend of the value of the option” — but Buffett’s weighing machine eventually kicks in. On models generally: competing on release cadence four months ahead of rivals carries “not a lot of value”; staying top-decile for years does.
- Cerebras is cash-flow positive where peers hemorrhage cash — “gross margins were a measure of your technical differentiation… if you’re running a negative gross margin business, you’re selling commodity.” The flip side: G42 is 87% of revenue (a deal estimated north of $1bn), which Feldman defends as a learned muscle: “the way you catch three large customers is to catch one first,” with several relationships targeted in 24 months.
- On China: Cerebras refused to sell — the deal “wouldn’t be used for good” (facial recognition of minorities, military) and failed his mother test — yet he thinks the US underestimates China “100%,” export controls may not be “a tractable problem,” and the attempt to cut off EDA tools just spawned US-VC-backed EDA startups in Shenzhen.