Cerebras spent about 18 months spending roughly $8 million a month while unable to build the thing it had raised money to build. Every six weeks there was a board meeting, and the only item on it was that they still could not make it.
The standard startup answer to a dominant incumbent is to be cheaper, or somewhat faster, and take share at the margin. Andrew Feldman's view is that against a company with Nvidia's margins that is not a strategy at all, because the incumbent can simply cut price — so the only opening is a product so different that giving the competing part away for free would not close the gap.
"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."
Feldman is on his fifth chip company. Three were sold, one was taken public before this, and Cerebras is now public too; he was Cerebras' first employee alongside four co-founders in 2016, before the transformer existed. Sitting beside him is Eric Vishria of Benchmark, who co-led the Series A that year and has been on the board since.
The full interview is covered here so you can skip it. 63 minutes of audio, 25 minutes of reading.
Here are the 18 takeaways that matter.
👤 Guest: Andrew Feldman, Co-founder and CEO of Cerebras Systems, on his fifth chip company after selling three and taking one public
🎙️ Host: Jack Altman, General Partner at Benchmark and founder of Lattice, who runs Uncapped
👥 Also on: Eric Vishria, General Partner at Benchmark, who co-led Cerebras' Series A in 2016 and sits on its board
📰 Published: 15 September 2026 on YouTube (Uncapped with Jack Altman)
🔴 YouTube | 🟢 Spotify | 🟣 Apple Podcasts | ⏱️ 1 hr 3 min | ✅ Time saved: 38 min
Key Takeaways
A new chip has to target 100x the incumbent, because the incumbent doubles every year for the five years it takes you to reach scale
Two to the fifth is 32x, and Vishria says you then need roughly 3x on top of that
Cerebras spent about 18 months unable to build the part, at roughly $8M a month
The rule the team set was that every failure had to be a new one
A fab costing $40B to $50B, running for four or five years, produces a chip that sells for $22
A burrito in the Mission District costs $17
One company in the world makes the lithography machine, and it is not a monopoly built by withholding supply
Data centers in the US use less water than California's almond growers, by a factor of four to seven
Everyone in the supply chain underestimated AI demand except Sam Altman
Feldman includes Cerebras, TSMC, Nvidia and the memory makers in that list
What made Nvidia great was a decade of trading badly as a public company, not CUDA
Speed creates markets rather than improving them: fast internet did not make Netflix better at mailing DVDs, it made Netflix a studio
Cerebras is selling throughput by splitting inference work with rival chips, and claims 5x more tokens at the same speed with AMD
1. Can I, And Should I
Jack Altman opened on 2016, when Cerebras was founded and AI looked nothing like it does now. Feldman's answer was that a computer architect who sees a new workload appear asks two questions in order.
The first is whether the chip can be built, and the second is whether there is enough of the work to justify building it. "So you have two questions. Can I and should I?"
What made AI the right answer to both was that it is computationally intensive rather than power-constrained. He contrasted it with the rise of ARM processors in cell phones, which were a power problem and not a compute problem.
The other half of the read was that the industry was running AI on an architecture that had been remodeled for it rather than designed for it, which is what left room for something better.
Altman then turned to Vishria, who had not previously invested in chips. Vishria's reply was about scar tissue: the more companies you work on, the more you see how hard it is, and a new hard thing needs a degree of naivety to attempt at all.
"And probably never will be again." was how he answered whether he would back a chip company a second time.
Feldman's joke on that was advice to founders: don't tell your investors how hard it is. "Keep it inside. Push it down. Push it down."
2. You Need 100x, Not 50%
This is the strategic core of the conversation, and Vishria put a number on it that founders can check themselves against.
Against a giant with high margins, being two-thirds the price is not a plan. "Even if you're two-thirds the price, they can just cut cost. They can just charge less. They can bundle it. They can do a hundred other things."
So the target has to be a product that survives the incumbent giving its own part away. "What that means is you have to go out with something way better, 10, 100, 500 times faster."
Vishria's arithmetic is the reason the multiple has to be that large. Assume the incumbent gets twice as good every year, and assume a new company needs at least five years to reach scale with everything going right — first tape-out works, first bring-up works. That is two to the fifth, or 32x, before you have any advantage at all, and then you need a further multiple on top.
"So that puts you at 32x and then you need at least a multiple advantage of that." Call the extra 3x, and the target is around 100x.
The error he says he hears constantly is companies benchmarking against today. "But what exists today isn't the target." Nvidia, Google and Trainium are all moving the ball materially every year.
"You can't incremental your way to vastly better." — Eric Vishria
The second consequence is that everything hard has to be in-house, because that is the part you control: chip, board, system, software, all the way up to the API.
Radical innovation means no vendor is waiting for you. "When you build a chip the size of a dinner plate, you can't go to a catalog and find a heat sink."
The compensation is expertise nobody else has. "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."
Feldman added the cost side of attacking an incumbent: the larger competitor buys silicon for less, buys manufacturing capacity for less, and probably pays less for its chip-design software.
3. 18 Months at $8M a Month
Asked about the near-death experiences, Feldman described chest pain and the specific shape of the problem.
He said Cerebras always had two axes — can you make it, and can you sell it — and it was confident on the second from the start. Nobody had ever done wafer scale, plenty of people said it could never work, and the company itself was unsure.
Then came the stretch with nothing to report. "I mean we believed we could do it but there was no evidence and there was a period of time about 18 months and we couldn't make it and we're spending about 8 million a month."
"You got board meetings every six weeks and all you've got to say is still can't make it right that's the board meeting."
The board asked what it could do to help. Vishria's answer, from the investor's side of the table: "There's nothing that can be done. Nothing to be done."
The rule that got them through was a rule about the kind of failure that was allowed. Every failed build got a full failure analysis, and the same failure was not permitted twice. "Only new mistakes, right? Only new failures."
The progress was measurable in how long a wafer survived. They shattered wafers in seconds, then it took minutes, then one ran for an hour, then they shattered some in minutes again.
The day it worked was July 2019, in a converted office with a hole drilled in the wall to extract the air. The temperature went flat and stayed flat.
"Holy crap, we've solved this problem that nobody in 75 years of compute had ever solved." He called it one of the great minutes of his life.
4. The Next Wafer-Scale Bet
Altman asked what the next five years of research look like. Feldman reduced a computer to three jobs and said a chip company in AI has to be working on all three.
The three are calculation, storage and movement — the core that computes, the memory that holds the result, and the input/output that ships it somewhere useful.
On memory, Cerebras has US government programs to stack HBM onto an SRAM-based wafer, which he said would give the capacity of HBM with the speed of SRAM.
On movement, the work is optical wafer stacking — putting an optical switch onto a wafer, also under large government contracts. He said it would change the world.
He credited the framing to Jack Dongarra, one of the pioneers of large-scale computing, who observed that the industry has been better at making flops than at moving them.
The stated goal on I/O is 10x, 100x or 1,000x faster movement of results, not a percentage improvement.
5. Sand to a ChatGPT Answer
Altman asked for the supply chain explained simply, from sand to a ChatGPT answer. Feldman's first answer was one line: "Dude, it's just TSMC."
Vishria's version: "There's a black box. You call that TSMC by another black box called ASML."
A leading-edge fab is a factory measured in football fields. "It's one of the greatest things humans make." He put the build cost at $40 billion to $50 billion and the useful life at four or five years.
His illustration was Samsung's plant in Texas, where the company built a power plant first — to make the concrete. "They ran concrete trucks, hundreds of concrete trucks, 7 by 24 for years to pour enough concrete to build the foundation."
ASML makes the photolithography machine, each one roughly 50 to 60 feet long and 20 feet high. Altman guessed half a billion dollars apiece; Feldman said only that they are expensive.
The same machines go to different fabs and produce different results, which is how TSMC achieves things its competitors cannot with identical equipment.
Nobody else in the world makes them. Vishria's gloss: this is a monopoly that was not created the way De Beers created one, by controlling supply. "They have technology that others haven't been able to replicate."
Lithography machines are probably not the binding constraint. The constraint is that demand moves exponentially and a fab takes five years to build.
"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."
The process itself: an ingot of silicon is sliced into wafers, the wafers run through a lithographic process that etches transistors into them, and the wafers are then diced into chips — except at Cerebras, where they are not cut.
6. A $22 Chip, a $17 Burrito
The price comparison is Feldman's way of making the economics of the fab land, and the hosts ran with it.
Some of the smartest people you have ever met work on a design for two years, it runs through a factory that cost $40 billion to $50 billion, and out comes a $22 chip. "It blows your mind when you think about it."
Altman's response: "And a burrito is $23."
Feldman took the joke and localized it. "And you go down the street here in the Mission District and you pay $17 for a burrito."
Vishria's assessment of the whole process: "It's the most impressive thing humans have ever done, right? It's among the most impressive things."
Having the chip is not the finish line. "Once you have the chip, you've solved about a fourth your problem."
7. How the US Lost Its Fabs
Altman asked how much of the bottleneck sits at TSMC. Feldman's answer was that it is not a construction problem but a skills problem.
Building an apartment building has hundreds of qualified builders; building a fab does not. "The actual making of the fab is skills that only TSMC has." Even TSMC cannot build twelve at once.
He attributed the US position to three decades of bad policy that pushed the fabs offshore, and said the ecosystem left with them — the tool vendors and the packaging firms, naming Amkor and ASE.
The remedy put forward on the episode was a 20-year window in which local ordinances are waived so GlobalFoundries, TSMC and Samsung can build domestic capacity.
Vishria's framing was blunter: the US punted a strategic industry and has to do better.
His evidence that the exposure is real was the pandemic-era chip shortage, when a container ship stuck in a canal was followed by Americans being unable to buy washing machines.
"I mean if we lost our chip capacity it would be catastrophic for our industry. Not just for our industry but for the country for the economy."
8. What a Board Can Do
Altman asked what a board is for at a hardware company, where there is far less to contribute on product. Both men answered, and Vishria answered about himself.
Vishria's starting point was a disclaimer. "I have a very small circle of competence." Asked to explain packaging, he said he could not do it, certainly not adequately.
His conclusion is that during long technical stretches the board's job is to stay out of the way and keep the company financed.
Feldman's version of a good board member is someone who knows the edge of their own expertise. Contribute in the domains where you are a real expert, and do not talk in the others.
He said the Cerebras board never tried to solve technical problems, and that the useful questions were about financing and about the long game.
Vishria's argument for why investors see something founders cannot: the company is going deep, the venture firm sees wide across an industry.
The questions he says are worth asking on a long project are structural, not technical. Are you hiring the right people? Have you made the right trade between specialization and flexibility?
Feldman's warning about hardware product management is that customer research is nearly worthless. You spend two or three years building before anyone can use it, and a survey costs the customer nothing, so everyone says it sounds great. "Nobody wants to sort of put their foot on the throat of somebody else's idea."
The corollary investors have to accept is that a first chip is rarely good. His example: Google's TPU team, where "the fourth one was good."
Vishria said the investment profile is entirely different from software — much longer time to revenue, revenue arriving in large chunks, far heavier customer concentration and far less spending on go-to-market. What is the same is needing excellent people from several domains to stay motivated for a very long time.
9. Measure Three, Cut Once
Cerebras was founded before the transformer, when TensorFlow was dominant and ResNet was still current, and the architecture choices made then had to survive what came after.
Feldman said the networks of that era were very small and very simple compared with today's, and the AI itself looked nothing like it does now.
The company was built as a training system, and the shift toward inference came later. Feldman resisted the word pivot.
His account of how the direction changed is a routing metaphor rather than a vision one. One way to plan is a dedicated telephone circuit all the way to the destination. The other is the way the internet works: hop to hop, deciding at each node.
They were aiming at something they were sure existed — "a pot of gold out there" — while treating the route to it as unknowable. Each time they got over a mountain they looked around and re-decided.
The new information that changed the destination was watching AI get good enough that everyone would want it — and everyone using it means inference, not training.
In hardware you cannot take software's approach of changing your mind the next day. Feldman's contrast is that in chips you measure three times and cut once, because a big mistake has to be lived with for years.
"it is really unforgiving."
The architecture decision that paid was a decision not to specialize. Cerebras chose not to embed technology that would accelerate convolutional networks, and to accelerate the underlying algebra instead.
"That was a really good decision because when Transformers came out, we were the fastest at those two, even though we'd never seen them and never heard of them and they hadn't been invented when we set the architecture."
Vishria noted that hardware is now fashionable to fund on the back of SpaceX, Anduril, Cerebras and Palantir — and that walking through the details makes clear how hard it actually is.
10. Young Product Leaders
Altman observed that the founding engineering team had worked together for years while the product and go-to-market leaders are notably young, and asked whether that was deliberate.
The split is between where experience predicts performance and where it does not. Feldman said the product organization is unusually smart, young, and promoted from within.
"We don't have big company rules." If someone is extraordinary they get more responsibility, with no minimum time in a job.
His example is his co-founder Sean, hired as an individual contributor at his previous company in 2007 at 25 or 26, made a corporate fellow when AMD acquired it four years later, and now co-CTO of a public company.
Where experience does matter is chip-building itself. Feldman said years of previously building chips is the best predictor of whether someone delivers an exceptional chip, and that the same holds on the mechanical and system sides.
The reason is that you cannot practice. Silicon is too expensive and too slow to build, so even a doctoral program rarely lets anyone tape out a chip, bring it up, put it on a board, power it and write the software for it.
Where experience does not matter is knowing what AI customers want, because nobody has a long history of that — so methodology, intelligence and insight beat experience.
On founder count he was direct. "First, five is too many founders without question except that we'd worked together before." Everyone knew what the others were good at, so there was little headbutting.
His test for a new hire is fast. "I think within the six or eight weeks of working with someone, you can tell if they're extraordinary."
The first signal is usually an email. "Every list is in descending order of importance. There's not a lot of fluff. There's high signal." Then you watch them run a meeting, then deliver something.
Vishria's own recruiting heuristic is whether he learned something real and wanted another meeting for that reason alone. He said it is his best predictor.
11. Dozens of Vendors
Altman asked about relationships outside the company, and the answer turned into the clearest contrast in the episode between building software and building hardware.
Feldman said the relationships are carried in from previous companies — Cerebras had been building chips with TSMC and working with its contract manufacturers for decades, which mattered most in periods of contention.
His prescription for new partners is unglamorous. "Be a good person. Write a thank you note. Do what you say you're going to do." Be good to your word in bad times as well as good ones, and get people information early.
Vishria's learning from the COVID era was about how short a software company's critical vendor list is. For a software infrastructure company it is AWS, and maybe a foundation model provider. Two vendors that matter.
Feldman's list is dozens — the specialty manufacturers that build the cooling plate, the water system, every component that goes into the system, each one a dependency.
"They call it a chain for a reason."
The lead times are the part that does not translate. On AWS you add capacity online; adding wafer capacity is a 15-month decision, made against a large capital commitment, and the vendor has to make its own allocation decision and buy into it too.
12. 3.5% of GDP
Altman said that internally they keep finding the AI build-out hard to size, and put a chart to the room.
The host-read figure: about 1% of GDP a year went into highways and telecom, about 2% into railroads, and the AI build-out is running at about 3.5% of GDP. "I actually saw a chart this morning that was like 1% of GDP."
The point made in response was that one person forecast this and it was Jack Altman's brother, Sam Altman. What he did, in Feldman's reading, was see an exponential and not be afraid of it — take it out three or four or five years and accept the answer.
Feldman recalled the reaction to the original Stargate number. "Remember the Stargate like 7 trillion or whatever," and said it was mind-boggling at the time and people laughed.
Vishria's comment now: the initial Stargate looks sadly small.
Feldman put his own company on the list of people who got it wrong. "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."
The same admission covers data centers. Asked what he would have said in 2018 or 2020 about data centers becoming a critical constraint, both said they would have got it wrong.
13. The Grid and the Almonds
Feldman's data center section is partly an engineering account and partly a complaint about how his own industry has behaved.
The data center is one link in the chain that delivers compute, and it binds whether the customer rents cloud capacity or puts equipment on its own premises.
"Our grid is pathetic and sort of built on 1940s or 50s technology."
He said the US stopped working on nuclear, and noted the irony if AI is what brings it back.
Backup power is where the innovation is appearing, after decades of nothing: diesel and gas gensets from GE Vernova and Caterpillar, jet-derived designs being proposed for data center power, battery backup, and fuel cells from Bloom Energy. He credited necessity.
On the industry's own conduct he was harsh. It tried to pass costs onto local communities and to take advantage of municipalities. "There is no reason a data center shouldn't pay its way."
The water objection he rejects with a comparison. Cerebras uses closed-loop systems, and: "All the data centers in the US use less than California almond growers. Not by 1x or 2x or 4x, but between four and seven times, the almond growers use more."
Vishria: "The water thing is just not a thing."
What the industry failed to do was communicate — get local buy-in, and show the thousands, sometimes tens of thousands, of high-paying construction jobs and the ongoing jobs and tax base. "And now we're paying the price."
Vishria widened it to AI generally, and Feldman agreed without qualification. "Horrible job. We're doing a horrible job."
14. 50MW Takes 18 Months
Altman asked how long a data center takes from start to finish, which produced both a definition and a set of lead times.
The phrase "shovel ready" is a marketing term. Feldman's translation: we want to sell you a pile of dirt, ready for your shovels. Asked whether it at least implies permits, he said sometimes not.
He allowed that it sounds better than raw land or dirt in Oklahoma.
From raw land and permits, a good builder takes about 18 months to stand up 50 megawatts, which is the standard block — even large sites unfold in 50-megawatt blocks.
Brownfield sites are the shortcut, and they are being hunted in the rust belt. Buyers are taking old factories and paper mills because they already have the power and the permits for it, and retrofitting them.
Every chunk of the supply chain has its own supply chain. Concrete is plentiful and labor usually is too — though in Wyoming, where sites were being fought over, electricians were being brought in from as far away as Denver.
The long poles are generators and electrical transmission switches. A cold shell — a concrete tilt-up or metal building — goes up fast. "But then you have to fit it out and turn it into a high-powered hotel for compute." Electricians, cooling and chillers are all long lead time.
Altman described flying over Elon Musk's Memphis site the previous night and being astonished by the scale of the buildings and the construction around them.
Feldman: "Nobody can build like Elon. It's unbelievable."
Asked whether the risk is permanent shortage or an eventual overbuild, Feldman took the shortage side. Ten years out is unforecastable — a decade ago transformers did not exist and models were identifying cats and chairs — but: "But five years we will still be chasing data centers and we'll still be chasing chips. And for those who use HBM, they'll still be chasing HBM."
15. Tokens Per Watt
Asked whether a permanently constrained denominator means the only metric that matters is tokens per watt, Feldman gave two answers: speed, and throughput bought from other people's chips.
His first claim is historical: speed creates markets rather than improving existing ones. "There's no market for dialup." There is no market for slow search either.
His example of what a speed change actually does: "And when the internet got fast they became a movie studio." Netflix had been mailing DVDs in envelopes; it did not get better at that, it became something else.
He offered a domestic version of the same point — to punish your children, do not take the phone away, put it back on dial-up speed.
He said the launch of GPT-5.6 Sol, in limited availability the week before, was already producing new application ideas because frontier intelligence arrives instantly.
The throughput route is disaggregation — splitting the work of inference between Cerebras and someone else's chip. He said the partnerships are live with AMD and with AWS.
The AMD partnership produces "5x additional throughput" with the speed unchanged, and "we're seeing similar numbers with AWS."
He framed it as available across the whole GPU landscape. There are four major chip makers in the category — Nvidia, AMD, Google's TPU and AWS's training parts — Cerebras is working with two, and would like to work with Nvidia.
16. What Made Nvidia Great
Altman asked what explains Nvidia's run. Feldman said most people answer this wrongly.
He calls Nvidia the great company of the first quarter of this century, without qualification — and then rejects the usual explanations.
He rejects the two standard answers, CUDA and the chip architecture. "They look to their chip architecture. I don't think it's a chip architecture."
His explanation is the decade nobody remembers. From roughly 2003 or 2004 to 2013 or 2014, the stock traded badly while the company fought for every sale with nobody listening.
"It's this unbelievable grit and intensity that was born of a decade of not having success as a public company."
Coming out of that as the most valuable company in the world is, in his view, what makes it one of the great companies in history. "These are just things people say."
Vishria's addition was about Jensen Huang personally — asking whether you could imagine an earlier generation of technology leaders still fighting at that level at that size.
Feldman said it is what he measures himself against as a chief executive.
17. A Professional David
The section where Feldman is most explicit about how he runs, and it doubles as the company's commercial history.
He said what is interesting about Huang is that he still sees himself as the underdog, and that he works the same way.
"I'm a professional David in the battle with Goliath and I wake up every day with that mentality" — five startups, three sold, two taken public.
His account of the last decade is a sequence of moved goalposts, each one named: you can't do wafer scale; you did it but you can't yield it in volume; you can't package it in volume; you only have a government customer; then a sovereign cloud; then you don't have a frontier lab, and they won OpenAI; then a hyperscaler; then you can't run big models, and now they are serving GPT.
Asked what the objection is now, his answer was one line: "CEO is a boomer."
The motivation he gives is not financial. "It's because we love building cool things, and we really like building cool things that are so hard that other people can't build them."
18. Shockley's Candy Bars
Altman closed by asking about Feldman's childhood, which Vishria had told him was unusual.
He grew up on the Stanford campus, in a neighborhood where every neighbor was faculty. One of them was William Shockley, who shared the Nobel Prize for the transistor and whose move west laid the foundation for Silicon Valley.
What the children knew about Shockley was that his wife handed out full-size candy bars at Halloween.
His description of the value system there is the point of the story. "the only currency was intellectual horsepower" — nobody cared who was rich or who had started a company; this was the 1970s, and what counted was whether someone was smart and did good work.
Another neighbor won the Nobel Prize in economics for work done with a collaborator.
His father played weekend doubles with a rotation of six or eight men. Feldman realized in his twenties that three of them had Nobel prizes and one had a Fields Medal.
His verdict on their tennis was less generous: "No, let me tell you, it was some old man tennis. I mean, their serves were grim." Their physics, he added, was good.
He expected to be an academic and was working on a PhD, going to Stanford's business school while finishing his qualifying exams, and got pulled into industry by mistake.
His father still asks when he is going to finish it. His answer: "all my professors are dead."
Bonus Insights
On politics in AI, Feldman's answer was sarcastic and then serious. "No, what we need is more people who don't understand making decisions." He said there are genuinely thoughtful discussions to be had about what is good for the US, and for individual municipalities and rural areas, and that politicians are not having them.
The ideal founder profile differs by sector, in his account. In AI and social networking, many of the best founders were building tools for themselves and their friends and had unique insight from that — he named Cognition and Cursor as current examples of the best software engineers building for themselves. In chips, prior experience in the field is what has historically predicted success.
Feldman's closing remark was about his investors. "Benchmark was an extraordinary partner." His advice to hardware founders: "I think if you do hardware, you're going to be in bed with your backers for a decade and pick good ones and I'm proud we did."
Vishria and Feldman traded a running joke about age, after someone called Vishria a boomer. Feldman's response: when you are low enough you are looking up at everyone, and at eight years old 27 looks ancient.
Feldman's bottom line is that a hardware company only gets to exist by picking a problem that cannot be solved incrementally, then surviving the years in which it cannot be solved at all — and that the current constraint on AI is not intelligence but physical: fabs that take five years, data centers that take eighteen months a block, and a grid built in the 1940s.
Products, Companies & Tools Mentioned
Cerebras Systems (The wafer-scale chip company Feldman co-founded in 2016; now serving government, sovereign cloud, OpenAI and a hyperscaler, and public)
Nvidia (The Goliath the strategy is built against, and the company Feldman calls the great company of the first quarter of this century — for grit rather than CUDA)
TSMC and ASML (The two black boxes: TSMC operates the fabs and is the bottleneck, ASML is the only maker of the lithography machine)
Samsung and GlobalFoundries (The other operators of leading-edge fabs on US soil; Samsung's Texas plant needed its own power plant to make the concrete)
Amkor and ASE (The packaging firms that left the US when the fabs did, part of what he calls three decades of bad policy)
AMD and AWS (Disaggregation partners: Cerebras splits inference work with their chips and claims 5x more throughput at the same speed with AMD)
Google TPU and AWS Trainium (The other two of the four major chip makers in the category; Feldman cited Google's TPU as proof that the fourth chip is the good one)
OpenAI and GPT-5.6 Sol (The frontier-lab customer, and the model whose limited release he says is already generating new application ideas)
G42 (The sovereign cloud customer in the sequence of milestones)
GE Vernova, Caterpillar and Bloom Energy (Named as the backup-power incumbents and the fuel-cell challenger in a category with no innovation for decades)
Netflix (His illustration of what speed does: slow internet meant DVDs in envelopes, fast internet made it a studio)
Cognition and Cursor (His example of software founders building tools for themselves, which he contrasts with hardware)
SpaceX, Anduril and Palantir (Vishria's list of the companies that made hardware fashionable to fund)
Benchmark (Vishria's firm, which co-led the 2016 Series A; Feldman's closing advice is to pick backers you can live with for a decade)
Lattice (Jack Altman's company before he became an investor)
Books & Resources Mentioned
Jack Dongarra's observation on flops (Cited by Feldman: the industry has been better at making flops than at moving them, which is the case for optical wafer stacking)
Stargate (The AI infrastructure program whose original headline number Feldman recalls as roughly $7 trillion, and which Vishria now calls sadly small)
Watch the full episode:
If this was worth your time, send it to someone closer to the industry than you are.
Get the latest market chatter as it happens: