How Andrew Feldman and Cerebras Built the World's Largest Chip

So picture a single piece of silicon, right? But well, not your standard tiny chip. Picture something roughly the size of a large dinner plate.

Right. Which is already just conceptually wild.

Yeah, exactly. And now imagine you're trying to pump, uh, like 700 watts of electricity straight through it.

Oh, and knowing that if the temperature across that plate fluctuates by even a few degrees, the whole thing just physically shatters.

Literally cracks in half. And if there's even a single microscopic speck of dust on that plate during manufacturing, the conventional wisdom of, you know, the last 40 years says you have to throw the entire thing in the trash.

Right. It's garbage. I mean, it's, it's a piece of hardware that fundamentally breaks the established laws of computer science and materials physics, really.

Yeah. And that's exactly why you've tuned into the deep dive today, because we are taking all the best insights from our stack of sources to look at how this, this absolute white whale of Silicon Valley was actually captured.

Because for decades, the rule was just, you build small chips, you print a bunch of them on a wafer, cut them up, wire them together. Building one giant continuous chip was an idea that historically just set hundreds of millions of venture capital dollars on fire.

It totally did. But we're looking at this now in May of 2026, and the tech world is frankly reeling from the largest IPO of the year.

Right. The Cerebras Systems debut.

Exactly. A $95 billion $1 billion market cap debut. But the thing you really need to understand as you're listening today, this isn't just some boring Wall Street financial story.

No, not at all.

It's a 10-year white-knuckle engineering epic. Our mission today is to unpack the story of Andrew Feldman, the contrarian bet that created the wafer-scale engine, and how solving just the most punishing physical problems imaginable basically challenged NVIDIA's absolute dominance in AI.

And to really get why this matters, yeah, you have to look at how the tech industry actually moves forward. Like every massive leap, the internet, mobile, AI, it always hits a wall, right? The initial software magic happens, everybody gets super hyped, and then progress is abruptly bottlenecked by something completely boring and physical.

Yes, the foundational plumbing.

Exactly. The people who rule the world aren't the ones making flashy apps. They're the ones who go down into the basement, realize the pipes are bursting, and fix the plumbing. Whoever fixes that bottleneck rules the ecosystem.

I always wonder like how someone even develops the mindset to spot those invisible bottlenecks, because you don't just wake up and look at the most sophisticated silicon architecture on earth and go, oh yeah, everyone's doing this wrong.

No, you definitely don't.

You have to be conditioned to see inefficiencies. And that leads us straight into Andrew Feldman's origins, which our sources highlight as the super crucial detail because he grew up literally living on the Stanford University campus. In the 1980s.

Yeah, his dad was a craftsman and his mom was a project manager.

Right. So he's just a kid in the absolute epicenter of the digital revolution, but he's raised by people whose whole mindset is grounded in physical reality, building things, managing resources.

He wasn't just visiting the campus, he was watching Silicon Valley literally construct itself. From mainframes to workstations, the actual cables being laid, the servers going in.

Right, the unglamorous stuff. And it seems like that environment permanently wired his brain because if you track his career, it's almost weirdly consistent. Look at 2002.

Yeah, Force10 Networks.

Exactly. He's VP of marketing and product management. Force10 is building these huge heavy-duty network switches. Cisco was the king of that space, but their stuff was bloated. It used too much power. It was expensive.

And Force10 just engineers a cleaner, way more efficient solution.

Yep.

And Dell acquires them for $800 million, which is really the first clear look at what we could call the Feldman pattern.

The Feldman pattern. I like that.

Right. Because then between like 2004 and 2006, he's VP of corporate development at Riverstone Networks, another infrastructure play. He helps take them all the way to a NASDAQ IPO.

He just loves the basement plumbing.

He really does. But the true crystallization of this, The thing that perfectly previews Cerebras is a company called C-Micro. He co-founded it and was CEO from 2006 to 2012.

Oh man, C-Micro is such a fascinating prelude to all this AI chip stuff because think about the internet in 2006. Google, Amazon, the early cloud, they were all hitting a massive physical wall.

The data centers were just eating unbelievable amounts of power.

Exactly. Everyone was buying these giant monolithic servers that were basically descendants of old mainframes. And they generated so much heat that the air conditioning bill was bigger than the hardware budget.

Right. Power and cooling literally became the physical bottlenecks of the entire internet.

And the whole industry is just like, ah, maybe we can tweak the software.

Right.

Or get better fans. But Feldman looks at it and realizes the architecture itself is fundamentally wrong.

Yeah, because those heavy-duty processors were made for complex tasks.

Mm-hmm.

So most internet traffic back then was just, you know, fetching an image, loading a tiny webpage.

Right. So he found C-Micro to invent the microserver. Instead of massive power-hungry chips, he networks together hundreds of low-power chips, like the ones you'd find in laptops.

It's totally contrarian.

It's like you wouldn't build a fleet of Ferraris to deliver the morning newspaper, right? You'd build a bunch of highly efficient bicycles.

That's a perfect analogy. And C-Micro's architecture was so dense, it ran so cool, it just slashed data center power consumption. It fundamentally changed server design.

And AMD sees this and buys them for $355 million.

Which totally validates his thesis. He doesn't do consumer apps, no social media hype. He finds the most unsexy infrastructure, pinpoints the physical constraint slowing everyone down, and bypasses it.

Okay, so that proves his thesis. Now fast forward to 2015. Deep learning is suddenly taking off. Real, practical AI. And everyone defaults to using NVIDIA GPUs to run these neural networks.

Right, graphics processing units. They were like 100 times better than CPUs for this.

But Feldman turns that exact the exact same critical eye toward AI, and he spots a fundamental flaw.

Well, yeah, because GPUs were built for graphics. They were engineered to take a 3D model and calculate the color of every single pixel on your monitor 60 times a second.

Right.

And it just so happened that the parallel matrix math for video games looks a lot like the math for training a neural network.

It was an accident that they were good at AI they weren't purpose-built for it. And let's dig into that flaw because the NVIDIA chips, I mean, they were doing the math super fast. The bottleneck wasn't the computation. The thing choking AI was data movement.

Exactly. In computer science, we call it the memory wall because standard chips are small. When you process a huge AI model with billions of parameters, you can't fit the data on the chip itself.

So it has to constantly reach out to external memory modules or across the network to other GPUs in the server rack.

To fetch the data it needs to do the math.

And physically, distance is the time is the absolute enemy of speed here. Moving data off a chip and back takes milliseconds, which sounds fast to you and me, but in silicon time—

A millisecond is an eternity.

Yeah. I actually like to think about this using a kitchen analogy.

Oh, let's hear it. So, imagine a standard GPU cluster is a kitchen full of the world's greatest, fastest Michelin-star chefs. These are your compute cores. They can chop onions faster than anyone alive. But the kitchen is built weird. The pantry, the memory with all the ingredients, is located in a warehouse 3 towns over.

Oh wow. Yeah, that perfectly illustrates the inefficiency.

Right, so these elite chefs just stand around their cutting boards, completely idle, waiting for a delivery truck to bring them onions. They chop them in half a second, and then they go back to standing around waiting for tomatoes.

The kitchen is fast, but it's completely starved for ingredients.

Exactly. Data movement is the bottleneck.

And that is exactly what Feldman saw in 2015. Everyone else was just trying to make the checks chop faster, but Feldman was like, let's eliminate the delivery truck.

Right.

His insight was terrifyingly simple. What if you built an AI chip where the cores and the memory were on the exact same piece of silicon?

But I mean, to hold a whole AI model, that chip would have to be absolutely massive. Exactly. So instead of taking a standard 12-inch silicon wafer, printing small chips on it and cutting them up, You just don't cut it.

You leave it whole.

One single gigantic chip, the wafer-scale engine. Hundreds of thousands of cores, gigabytes of memory, all on one piece of silicon. The data movement happens on chip in microseconds.

You completely kill that off-chip latency tax. Back to the kitchen, they decided to build a kitchen the size of a stadium where every chef has their own massive pantry right within arm's reach.

Nobody's ever waiting for the truck. It's a total paradigm shift. It's practically a different species of computer.

But okay, if it's so elegantly simple, just make the chip bigger, why wasn't NVIDIA already doing it? Why was the undisputed king of silicon still printing tiny chips? Well, as the sources show, it's because the last visionary who tried this had just a spectacular failure.

Yeah, Gene Amdahl, his ghost really haunts this whole story.

Oh, absolutely.

He was a legend, right? One of the main architects of IBM's mainframe dominance. But in 1980, he leaves to start Trilogy Systems to do this exact thing. Yeah. Wafer-scale integration. He raised $230 million.

Wait, let's pause. $230 million in 1980.

Right.

Adjusted for inflation, that's like what, three-quarters of a billion dollars today for a startup?

An astronomical war chest.

Yeah.

Uh, and it was completely incinerated. The company just Failed completely. They couldn't overcome the physical realities, yield issue, you know, defects on the wafer and the thermal constraints. They just couldn't cool a piece of silicon that big without it destroying itself.

It was such a high-profile disaster that wafer-scale integration literally became a toxic phrase. People thought it was a fool's errand.

So now it's late 2015. Feldman and his CTO, Sean Lai, who's this hardware design guru from Intel, they form Cerebras. And they know all about AMDOL.

They know exactly what killed Trilogy.

Yeah, but their bet is that 35 years of quiet advancements in materials science and manufacturing might finally make it possible. Which leads to this absolutely incredible moment in our sources.

Oh, pitching the impossible to venture capitalists.

Yes. January 2016, they get a meeting at Benchmark Capital, arguably the most elite VC firm on Sand Hill Road. And they meet with Eric Visria.

And the sources note that Vishri is just actively annoyed to be there.

He really is. He's muttering to himself, wondering why he even took the meeting. Because pitching a wafer-scale chip to a seasoned VC is basically like pitching a perpetual motion machine. It just triggers instant skepticism.

Right. He's practically looking for an excuse to kick them out of the room. But Feldman is smart. He knows he can't walk in with 50 slides of financial projections. He brings the shortest pitch deck ever.

It's legendary. Slide 1. The company title. Slide 2, the team, you know, dropping Intel and C-Micro names so Vishriya knows they aren't totally crazy.

And slide 3, a single sentence: GPUs actually suck for deep learning.

And that is the pivot point. Vishriya reads that and the annoyance just vanishes. He has this massive realization. He knows everyone's throwing billions at Nvidia, but when he sees that sentence, The data movement bottleneck suddenly clicks for him.

Right, he's like, "Of course they suck, they're graphics chips!" Exactly.

But he doesn't just hand over the money. He calls in Bruce Dunlevy.

Ah, the heavy artillery.

Yeah, Dunlevy is a founding partner at Benchmark. He was around in the '90s, he remembers the crater Gene Amdahl left. He sits down and immediately starts grilling them on everything that killed Trilogy. Yield. Thermal. Packaging.

And after this hours-long interrogation, Dunleavy looks at them and just says, "What you're attempting will be hard." Which is such a stark warning. Mm-hmm.

But they invest $15 million anyway. They lead the Series A in 2016.

Which is fascinating psychologically, right? Benchmark wasn't betting that the physics had magically become easy. They were betting that 2016 was the exact right time, and this was the exact right team to crack a 35-year-old cold case. I really want you listening to just think about the audacity that takes to walk into a top-tier VC, look a veteran in the eye and say, we're gonna do the thing that famously burned a quarter of a billion dollars, but don't worry, we've got it.

You really have to back that up. You do.

And backing it up meant surviving what the sources accurately call the engineering gauntlet. Getting that $15 million was genuinely the easiest part of the whole decade.

Yeah. Now they actually had to build the thing and the problems they hit were existential. Any one of them could have killed the company. Lenny.

Let's break this gauntlet down, because this is where it turns into a crazy extreme engineering story. The first monster was the same thing that got Amdahl: yield.

Right. So chips are made from these flawless silicon crystals sliced into round wafers, and using photolithography, which is basically shining lasers through a stencil, they print billions of tiny transistors onto it.

But it's never perfect.

Never. A fluctuation in the light, or literally one rogue speck of dust in the cleanroom will destroy some transistors. On a normal wafer with hundreds of chips, you just throw away the one broken chip and keep the other 99.

Right. But when your product is the entire wafer—

One speck of dust theoretically ruins the whole product.

You can't just cut around it. If you have a dinner plate-sized chip, one dead gate makes the whole thing useless. So how do you get a commercial yield when physical perfection is impossible?

You don't aim for physical perfection. You aim for logical perfection. Cerebra's solution here is architectural genius. They put hundreds of thousands of cores on the wafer, but they overlaid it with a dynamic communication fabric, like a mini internet built into the silicon.

Okay, let me see if I can translate the mechanics of this for you. Imagine the chip is a huge grid of cities, all connected by highways. During manufacturing, they test the wafer and find a dust particle destroyed City B. In a normal chip, the whole map is broken because the hard wiring says, City A has to talk through City B. But the cerebrus fabric is dynamic. It senses City B is dead, and the logic mathematically rewires the highway to bypass it, going through City C instead.

That's exactly it. They built in about 1% redundancy, just extra cores scattered around, and the chip essentially heals itself. It writes around the dead spots.

So they mathematically conquer the yield problem, but solving that immediately triggers a much scarier physics problem. Thermal management.

Oh, the heat?

Yeah. The later generation wafer-scale engines have over 4 trillion transistors. This single piece of silicon generates over 700 watts of heat. You can't just slap a heat sink on it and blow air across it like a normal GPU.

No, if you blow air across a baking hot wafer that big, it'll warp.

Yeah.

Or literally catch fire.

Air cooling was totally impossible, so they had to use liquid cooling. Cerebras invented this custom copper heat exchanger. It sits flush against the silicon, and cold water pumps through these tiny microchannels carved in the copper.

But that brings up a massive materials science nightmare: the coefficient of thermal expansion.

Right. Different materials expand at different rates when they get hot. It's like pouring boiling water into a freezing cold glass. It shatters.

Exactly. Copper and silicon have totally wildly different expansion coefficients. If you firmly glue a giant copper block to a giant silicon wafer and turn the power on, The copper expands faster. The shear stress will physically rip the silicon to pieces.

It's a total engineering trap. You need them touching to cool the chip, but if they touch while heating up, they destroy each other.

And Cerebra's fix for this is just one of the most elegant mechanical engineering feats I've ever seen. They didn't rigidly glue them together. They engineered these microscopic structures for the thermal interface that actually flex.

That detail blew my mind. They made micromaterials that act like microscopic shock absorbers. As it powers up and gets hot, the copper and silicon expand at different rates, sliding against each other at a microscopic level.

Right. And these flexible structures stretch and contort. They maintain perfect thermal contact to pull the heat away, but they absorb all the shear stress so the wafer doesn't crack.

They literally engineered a chip that physically breathes as it heats up. It's stunning.

But, I mean, you don't generate 700 watts of heat without pumping 700 watts of electricity in. Which is gauntlet number 3, power delivery.

Right. Standard chips get power from little gold pins around the edge, but you can't push massive amperage from the edge of a dinner plate all the way to the center. The silicon's electrical resistance is too high.

The outer edges would turn into plasma and the center would just shut down from lack of power.

So lateral power delivery was a non-starter. They had to invent a vertical power delivery system. They made this complex power fabric that delivers electricity straight down perpendicular to the chip. Hitting every square millimeter uniformly at the same time.

Which, again, easy to say, agonizing to actually manufacture. Which brings us to the fourth challenge. You can't just draw this up and email it to a factory in Taiwan.

Yeah, they partnered with TSMC, the undisputed kings of chip manufacturing. But TSMC's whole multi-billion-dollar assembly line is optimized for small chips.

Specifically, they're constrained by the reticle limit. The machines that print chips use a stencil called a reticle, and the max size is usually like 800 square millimeters. The machine stamps that stencil over and over across the wafer.

But Cerebras needed TSMC to print one continuous, highly coordinated design across the entire 46,000 square millimeter wafer.

They had to rewrite the rules of TSMC's own equipment, making sure these intricate communication highways lined up flawlessly across the boundaries of every single reticle stamp. Cerebras basically had to invent the manufacturing process just to get the chip built.

And even after TSMC prints it, you still have to physically move the thing, which is the final boss of the gauntlet: packaging.

Oh yeah, a bare silicon wafer that big is so thin and brittle. How do you pick it up, align it with a flexible copper cooler, align it with a vertical power board, and sandwich it all together without snapping the silicon?

You can't just have a guy with a screwdriver tighten down a heatsink.

No. Because if you tighten one corner even a fraction of a millimeter more than the opposite corner, the uneven pressure bends the wafer and it shatters instantly.

So what did they do? The tools to assemble this didn't exist anywhere on Earth, so they built the tools.

They engineered a custom robotic assembly machine. Its only purpose is to drill over 40 retaining screws into the packaging simultaneously.

Just the imagery of that. A custom robot applying perfectly uniform pressure across 40 points at the exact same millisecond, keeping the wafer perfectly flat wallet seals. They had to engineer an entire industrial supply chain just to support the chip.

But here is the brutal reality of deep tech startups. Doing all this consumes unbelievable amounts of time and capital. Physics doesn't care about your quarterly financial targets.

It really doesn't.

And while the hardware team is inventing all this, the business side of Cerebras is fighting a totally different war: surviving the VC desert.

The funding timeline is wild. It really shows how Feldman kept investors focused on the long game. They had that $15 million Series A in 2016. Then in 2017, they raised $40 million for Series B.

Right.

2019, they grab $155 million for Series C. And remember, they're just in R&D this whole time. They're a money furnace, paying TSMC, paying engineers with no commercial product to sell yet.

And then the market cools down. But in 2021, they raise a $250 million Series D. By now, early versions of the chip work, but enterprise sales are just agonizingly slow.

Because their original model was CAPEX, capital expenditure. They were trying to sell these million-dollar refrigerator-sized supercomputers for companies to install in their own data centers.

And that's a tough sell. Hey, buy this totally novel architecture that requires custom power and cooling. It takes a year to close a sale like that.

But the investors don't blink. 2022, they raise another $500 million. Still highly unprofitable, no huge recurring revenue, but the VCs keep writing checks because they know if Cerebras operationalizes this, the payoff is massive.

Which brings us to a huge crossroads in 2023 and 2024. They've got the WSE-3 chip. The hardware is flawless. But the CapEx sales model is too slow. They need cash and public legitimacy. So they quietly file for an IPO in late 2023.

And they immediately hit a massive unexpected geopolitical roadblock.

According to the sources, the IPO was derailed by intense national security scrutiny from the US government. At the time, Cerebras' biggest commercial customer was G42, a massive tech holding company in Abu Dhabi.

Right. And the US government raised serious national security and supply chain risk concerns about Middle Eastern investment and involvement in advanced AI hardware.

Yeah.

And I want to pause here. For you listening, we need to be exceptionally clear. We are not political analysts, and we are absolutely not taking a side on the geopolitics here.

Exactly. We are strictly reporting the historical realities from the source material.

The US government had stated concerns about this sensitive technology, and for Cerebras, the political debate was kind of irrelevant compared to the structural impact. Their IPO was completely dead in the water.

They had to withdraw their paperwork and stay private. And beyond the geopolitics, This showed a classic startup vulnerability. Relying on a single massive customer for your revenue is so dangerous.

It really is.

If that relationship fractures for whatever reason, your whole revenue model collapses.

At the time, withdrawing an IPO when you're burning that much cash looks like a disaster. But in hindsight, it forced Cerebras into probably the most brilliant business pivot of the decade.

A total blessing in disguise.

Yeah.

Blocked from going public. They realized the CapEx model was flawed. Selling physical supercomputers was too hard. So between 2024 and 2025, they inverted the whole business.

They shifted from selling hardware to keeping the hardware, racking it in their own data centers and renting out the computing power over the internet.

They became a SaaS company—compute as a service—just like AWS.

And the genius of this is it instantly diversified their clients. They didn't need just G42 anymore. They landed AWS themselves. And crucially, the biggest AI lab on the planet, OpenAI.

Yes, which leads directly to the nuclear announcement in January 2026, a $20 billion multi-year infrastructure deal with OpenAI, 750 megawatts of raw computing capacity through 2028.

The largest dedicated infrastructure deal in AI history. And think about what this means for you, the listener. The SaaS pivot democratized access to the fastest chip on Earth. Before, an AI researcher needed millions in grants and a custom facility. Now, just a cloud account and a credit card. You rent it by the hour.

And we have to acknowledge the poetic justice of OpenAI signing this deal. OpenAI is the lab that popularized NVIDIA GPUs for language models. They ignited the GPU boom. Right. But for their future infrastructure, they write a $20 billion check to Cerebras. OpenAI basically told the world that Andrew Feldman's 2015 thesis was right. GPUs are bottlenecked by data movement, And wafer-scale computing is the answer.

Okay, but I'm gonna inject some skepticism here. Sam Altman doesn't hand over $20 billion just because you have a neat architectural theory. The hardware has to be mathematically untouchable. You need raw performance. So let's look at the specs of the WSE-3, because these numbers just defy logic. They really do.

A single WSE-3 has 900,000 AI-optimized cores. That's 4 trillion transistors on one piece of silicon. 4 trillion! For perspective, NVIDIA's flagship H100 GPU has roughly 50,000 cores. One WSE-3 is theoretically equivalent to 62 H100s combined.

But the core count isn't even the craziest part. It's the memory. 44 gigabytes of on-chip SRAM. And we should explain why SRAM is so important compared to standard DRAM that NVIDIA uses.

Please do. It's the heart of the whole thing.

Think of standard DRAM like a massive filing cabinet down the hall from your desk. It's cheap, it holds a lot of data, but every time you need a file, you have to walk down the hall, find it, and walk back. It takes time. Right. SRAM is like having a sticky note slapped directly on your computer monitor. It's super expensive, holds less data, but access is basically instantaneous. By putting 44 gigs of SRAM directly next to the cores, Cerebras gave every single core an infinite supply of sticky notes.

Which perfectly resolves the memory wall. And the ultimate metric here is memory bandwidth, the speed data moves between memory and cores. The WSE-3 hits 21 petabytes per second. Let me contextualize that.

NVIDIA's ultra-premium B200 chip has a bandwidth of about 8 terabytes per second. The Cerebra's WSE-3 is moving data 2,625 times faster.

It's just a completely different magnitude.

But how does that translate to real-world AI performance?

Well, you have to separate AI into training and inference. Training is feeding the model mountains of data for months. Inference is when you're actually asking it questions in real time. Cerebras really dominates in inference where moving weights in and out of memory rapidly is crucial.

The sources use a phrase here, WSE-3 is compute-bound at batch 1 while GPUs are memory-bound. What does that actually mean?

Batch 1 means one user asking the AI one question. On an NVIDIA cluster, the cores calculate the first part of the answer instantly, But then they sit idle, waiting for the system to fetch the next parameters from memory across the network. The chef's waiting for onions. Exactly. They spend more time waiting than calculating, so they are memory-bound. But the WSE-3, with that 21 petabytes of SRAM right there, never waits. The data is instant, so it's compute-bound. Its speed is limited only by the cores themselves.

And with 900,000 cores, it's astonishingly fast. Running inference on Meta's Llama 370B, it's up to 21 times faster than the B200. Plus, for businesses renting it, the total cost of ownership is 32% lower.

Faster, cheaper, no network bottlenecks.

Okay, I'm gonna push back aggressively here because it sounds like game over for NVIDIA, but NVIDIA isn't just a hardware company. They have CUD, their software ecosystem. Yep. The moat. Right. Millions of developers only know how to write code for NVIDIA. Does a faster chip even matter if developers can't program it?

That is the most valid counterargument. Nvidia's software moat is way deeper than their hardware moat. They control 90% of the market. Because CUDA is the frictionless standard. Cerebras is building their own software stack and compilers to translate PyTorch to run on the wafer. But it's undeniably less mature. It's David versus Goliath.

David has a physics-breaking slingshot, but Goliath has millions of developers. Exactly.

And Nvidia isn't trying to build a giant wafer. They're just building increasingly massive clusters of small GPUs and relying on software to to overcome the hardware limits.

But armed with that $20 billion OpenAI validation, Cerebras finally went to Wall Street.

The May 2026 IPO. May 14th, 2026, they, they debut on the NASDAQ, priced between $115 and $125 a share, aiming to raise $3.5 billion.

And the market reaction was explosive. First day of trading, the stock popped 72%, closed around $300 a share, pushing their market cap to $95 billion, the biggest IPO of the year. But it wasn't just hype.

They released their 2025 financials. The SaaS pivot worked. They posted $510 million in revenue, up 76%, and they were profitable—$87.9 million in net income.

They proved you can conquer physics and make a profit. And the wealth creation. Feldman's 10.3 million shares jumped to $3.2 billion on day one. Sean Lai, the hardware genius, hit $1. $7 billion.

And Benchmark Capital, the firm that invested $15 million into an idea that had previously failed, their day one value was $5.3 billion.

Even early angel investors like Sam Altman made massive returns. I keep thinking about Bruce Dunleavy's warning, what you're attempting will be hard. 10 years of white knuckle moments, surviving yield issues, inventing flexible cooling, the VC winter, a blocked IPO, pivoting the whole business, and it birthed $95 billion in value.

It represents a seismic shift. Cerebras proved NVIDIA's monopoly can be broken, but you can't just copy them; you have to rethink the foundational architecture.

So, as we wrap up this deep dive, what's the ultimate takeaway here?

The real lesson is that this isn't just about AI chips; it's a testament to deep physical engineering. We're obsessed with quick software iterations and hype cycles, but Cerebras proves That real generational value comes from having the patience to solve the hardest physical problems in the universe over a 10-year game.

Which leads me to a final provocative thought for you to chew on. We just spent an hour talking about how Feldman solved the data movement bottleneck by putting everything on one giant chip. But look at that OpenAI deal. 750 megawatts of power. An incomprehensible amount of energy. The equivalent of powering a mid-sized American city for one customer's AI. So if the silicon architecture is no longer the bottleneck holding AI back, is the global power grid itself the next great bottleneck? And if we're hitting an energy wall, who's gonna be the Andrew Feldman of the global power grid? Who's gonna engineer the radical physics-breaking solution for how we generate and distribute power? Thank you so much for joining us on this deep dive.