📺 Today’s recommended deep-dive video: https://www.youtube.com/watch?v=jeop9wfb9jU

Wafer-Scale Winning: How Cerebras Bet on Speed to Disrupt the GPU Giants

For nearly a decade, Andrew Feldman was told his “wafer-scale” architecture was a technical impossibility, a weird detour that could never challenge the dominance of the GPU. Today, Cerebras is a $63 billion public company, having secured a massive $20 billion deal with OpenAI by solving the one thing modern AI users refuse to tolerate: latency.

Core Question: Can a radically different hardware architecture unlock the next phase of the AI revolution by prioritizing inference speed above all else?

Highlights

• Why “wafer-scale” dinner-plate-sized chips succeeded where industry legends failed for 70 years.
• The pivotal $20 billion deal with OpenAI and the 2025 turning point where speed became a necessity.
• Andrew Feldman’s philosophy on being a “professional David” against the Goliath of Nvidia.
• The transition from treating fast AI as a novelty to viewing it as a catalyst for new business models.

⏱️ Reading time: approx. 7 minutes · Saves you about 24 minutes vs. watching.

Want to take notes while watching? Click the image below and let AI Notebook capture the key points for you 👇

The Long Game of Radical Hardware Architecture

Breaking the “Postage Stamp” Paradigm

To build something radically better, you cannot simply iterate on existing architectures; you have to be willing to look like you’re doing it wrong. While the rest of the semiconductor industry was busy perfecting chips the size of postage stamps, Cerebras bet everything on “wafer-scale” integration, creating a single chip the size of a dinner plate. This 46,000 square millimeter behemoth was designed to eliminate the communication bottlenecks inherent in traditional clusters.

For years, the industry consensus was that this couldn’t be done.

The history of computing is littered with the failures of those who tried to achieve wafer-scale integration, including Gene Amdahl, one of the founding fathers of modern compute. Between 2017 and 2019, Cerebras faced its own “valley of death,” spending $8 million a month while failing to yield a working product. Feldman describes the agonizing six-week intervals of board meetings where he had to report that the impossible remained, for the moment, impossible. It wasn’t until the summer of 2019, in a makeshift office in Los Altos, that the team finally saw their machine come to life, proving that 70 years of industry skepticism had been misplaced.

💡 Digging Deeper

Q: Why is a single large chip better than many small ones?
A: Small chips have to communicate over wires and circuit boards, which is slow and power-hungry. A wafer-scale chip allows billions of cores to talk to each other at the speed of light on the silicon itself.

Q: How did the team survive the two-year period of constant failure?
A: Feldman attributes it to a rigorous failure analysis process—getting slightly better with each iteration—and the extreme conviction of investors who understood that the reward for solving this specific problem was a monopoly on speed.

From Novelty to Necessity: The Inference Explosion

The Market for Slow is Zero

Speed is a novelty until a technology becomes smart enough to be useful; once it is integrated into daily workflows, speed becomes a requirement. Between 2023 and 2025, Cerebras was “blisteringly fast,” but the market didn’t care because AI wasn’t yet an everyday tool. Just as there is no market for dial-up internet or slow web searches today, there is no future market for slow AI inference once agents start running 24/7.

When models got smart enough to be useful, the demand for Cerebras hardware didn’t just grow; it exploded.

The company’s trajectory shifted from selling a dozen units of their first generation to tens of thousands of their third. This ramp-up was validated by a massive partnership with G42 and eventually the landmark agreement with OpenAI. These deals weren’t just about capital; they allowed Cerebras to battle-test their equipment at a scale that their own internal QA labs could never replicate.

💡 Digging Deeper

Q: What happened in 2025 to change the market dynamics?
A: AI models reached a threshold of intelligence where they moved from research projects to integrated tools for coding, design, and SaaS. At that point, the “attention span” of the user became the primary bottleneck.

Q: How does speed enable new business models?
A: Speed doesn’t just make existing things faster; it changes the nature of the service. Feldman uses the Netflix analogy: they didn’t just get better at mailing DVDs; fast internet allowed them to become a movie studio.

Scaling the “Professional David”

Competing Against Goliath

Operating in a market dominated by Nvidia requires a specific type of psychological makeup that Feldman calls being a “professional David.” It is a recognition that every dollar won is a victory of brains over the muscle of a much larger incumbent. For a hardware startup to reach “corporate adulthood,” it must maintain a fearless engineering culture that values extraordinary failure over ordinary success.

Settling for “good enough” in hiring or design is the primary cause of death for companies scaling from 800 to 3,000 employees.

One of the most significant hurdles in this journey was the software stack. One of the company’s founders initially estimated that building a compiler would take a decade—a timeline Feldman initially dismissed as “big company talk.” It turned out to be exactly right. Building the bridge between radical hardware and existing AI frameworks is a marathon, not a sprint, and it required a decade of disciplined development to make the hardware accessible to the broader ecosystem.

💡 Digging Deeper

Q: How does Cerebras use AI internally to speed up its own work?
A: The company has shifted from engineers writing code to engineers governing agents. They went from spending nearly nothing on tokens to $30,000 per engineer, creating “100x” productivity for those who can manage agents effectively.

Q: Why did Cerebras choose to go public now?
A: To “graduate from corporate adolescence to corporate adulthood.” Being public provides legitimacy for large-scale enterprise deals and offers a pure AI investment play without the baggage of gaming or PC graphics revenue.

Key Takeaways

The success of Cerebras is a testament to the power of contrarian thinking in high-stakes hardware. By betting on wafer-scale integration when the rest of the industry considered it a dead-end, Feldman and his team positioned themselves to capture the massive demand for inference speed that emerged once AI became smart enough to be useful. The company’s journey from a makeshift office in Los Altos to a $63 billion market cap highlights that in the world of deep tech, the “long road” is often the only way to build a sustainable advantage.

Crucially, the next phase of AI will not be about making current tools faster; it will be about the fundamental reorganization of work. As inference costs drop and speeds increase, we will see the emergence of business models that were previously impossible, much like the transition from physical media to streaming. For those building in the space, the lesson is clear: if you aren’t building for the “speed of light,” you’re building for a market that won’t exist in three years.

Q&A

Q1: How much faster is Cerebras than traditional GPUs?
A: Depending on the model and workload, Cerebras is currently 15x to 20x faster than traditional GPUs for inference tasks.

Q2: What was the most difficult period in the company’s history?
A: The two-year stretch between 2017 and 2019 when the team was spending $8 million a month and could not get the wafer-scale chip to yield.

Q3: How did the deal with OpenAI come together?
A: It started in mid-2025 when Sam Altman recognized the necessity of fast inference. The deal was moved from term sheet to signed master agreement in just four weeks, an exceptional speed for a $20 billion contract.

Q4: Why does Andrew Feldman call himself a “professional David”?
A: He has spent his career starting companies that compete directly against giants like Intel and Nvidia, relying on architectural innovation to beat their sheer size and market power.

Q5: What is the significance of the G42 partnership?
A: G42 provided a $1 billion “bridge” order that allowed Cerebras to scale its manufacturing and battle-test its clusters before the broader market was ready for high-speed inference.

Q6: What is the current head count and how does that relate to their market cap?
A: They are at approximately 800 to 850 people, representing an exceptionally high market cap per employee for a $63 billion company.

Q7: How does Feldman view open source AI?
A: He believes open source has “fed the market” and kept creativity alive, forcing closed-source providers to innovate faster and providing a diverse range of workloads for Cerebras hardware.