Key Takeaways

  • The explosion of AI from 2015-16 created a rare 'aperture' to invent entirely new computing architectures, not just refine existing ones.
  • Cerebras made two bold bets: building silicon dedicated solely to AI and radically departing from the GPU design blueprint.
  • Their 'dinner plate-sized' chip, with memory integrated directly next to compute, addresses the 'fundamental problem in AI' — moving data — far more efficiently than traditional designs.
  • This architectural shift results in 15 to 18 times faster performance for critical OpenAI workloads compared to conventional GPUs.
  • The market for slow AI is zero. Real-time processing is non-negotiable, demanding hardware built for speed from the ground up.

"Dinner Plate" Chips Beat GPUs By 18x

When Andrew Feldman, CEO of Cerebras, talks about AI silicon, he cuts past incremental upgrades. He argues that the “hard part here, the hard part is moving data from memory to compute. This is the fundamental problem in AI.” Most of the industry has been focused on making the compute faster, but the data still has to travel. Cerebras decided to tackle that core bottleneck.

Their solution looks nothing like a GPU. It's a single, massive "dinner plate-sized" chip. This isn't just about sheer size; it's about integration. By placing memory right next to the compute units on the same wafer, Cerebras dramatically reduces the time and energy spent shuttling data back and forth. This radically different architecture isn't a minor tweak; it’s a re-imagining of the entire system for a specific workload.

The results speak for themselves. Feldman reports that when major players like OpenAI use Cerebras hardware, they see a significant leap in speed. “When OpenAI uses us, we're 15 or 18 times faster than a GPU.” This isn't just an engineering flex. It translates directly to user experience. As Feldman puts it, “That means your answers are delivered more quickly. It means your engagement with the AI is more enjoyable.” He makes the case that speed isn't a luxury; it's a necessity. “How big is the market for slow search today? Is zero. You will not wait for AI. We have to deliver it to you in real time.”

The Unseen Opportunity in New Workloads

Feldman's journey with Cerebras started with a core insight: new computing workloads open the door for entirely new architectures. “What AI did starting in about 2015 16 is it opened the door the aperture to say maybe we could use computers on images.” This was the shift. Before, AI was niche; suddenly, it was a tsunami of compute demand. “We saw that AI would be an enormous consumer of compute. And historically for computer architects, new workloads were the opportunity for share to change.”

This wasn't about building a slightly better GPU. It was about seeing that the existing paradigm was fundamentally misaligned with AI's unique demands. Feldman's philosophy is sharp: “Our view as computer architects is if you want to be 20 times better than somebody, your architecture can't look like them. They have enjoyed and eaten all the lowhanging fruit.” Trying to out-iterate NVIDIA on their own turf, with their own design principles, was a losing battle. Instead, Cerebras went to first principles, asking what AI really needed, not just what existing chips could provide. They didn't optimize; they reinvented.

What to Do With This

Stop optimizing where you should be reinventing. Identify a core, "fundamental problem" in your industry that everyone accepts as an inevitable bottleneck. Don't look for a 10% improvement within existing frameworks. Instead, question the framework itself. If you're building a SaaS, ask: What's the equivalent of 'moving data from memory to compute' in my users' workflow? What if I blew up the existing process and designed it from first principles for 10x speed or simplicity? Pick one process, sketch out the current solution, then brainstorm a solution that would make the existing one look laughably slow. That's your Cerebras moment.