Cerebras Systems CEO Andrew Feldman to discuss AI scaling at Disrupt 2026

Andrew Feldman, CEO and co-founder of Cerebras Systems, will take the stage at TechCrunch Disrupt 2026 in San Francisco to tackle a question that keeps the entire AI industry up at night: can AI actually keep scaling?

The session, scheduled for 2:30 p.m. on the conference’s main stage during the October 13-15 event, comes at a pivotal moment for Cerebras. The company went public in May 2026, raised $5.55 billion, and disclosed a backlog worth $25 billion.

The chip that ate the whole wafer

Cerebras builds something called a wafer-scale engine. Most chipmakers cut a silicon wafer into hundreds of individual chips. Cerebras looked at that process and said, “What if we just used the whole thing?”

The result is the WSE-3, a single chip packed with roughly 4 trillion transistors and 900,000 cores. For context, a high-end consumer GPU has a few tens of billions of transistors. Cerebras fits more than a hundred times that onto one piece of silicon.

The practical upside is speed, particularly for inference, which is the process of running a trained AI model to generate outputs. Every time you ask ChatGPT a question or an autonomous agent writes code, that’s inference at work. Cerebras claims its architecture delivers faster inference than traditional GPU setups, a claim that matters more every quarter as the industry shifts from training-centric workloads to inference-heavy deployment.

Feldman has been vocal about this architectural divergence. While Nvidia’s GPU clusters dominate the training side of AI, Cerebras is carving out a lane where latency and throughput per watt are the metrics that matter most.

Big contracts, bigger questions

The company’s commercial traction has been hard to ignore. Cerebras holds a multi-year contract with OpenAI focused specifically on inference workloads.

Then, on September 29, 2026, General Compute announced its own multi-year deployment partnership with Cerebras. That deal targets agentic coding workloads, with systems set to come online in Q1 2027.

The $25 billion backlog number is eye-catching, but it also invites scrutiny. A backlog is a promise, not revenue. Customer concentration adds another wrinkle. When a significant share of your revenue pipeline flows through a handful of mega-contracts, any single cancellation or renegotiation can ripple through the entire business.

The scaling wall

Data centers are bumping up against hard physical limits. Power grids in key markets are oversubscribed. Cooling infrastructure is being pushed to its thermal boundaries. Land with adequate utility connections is getting scarce and expensive.

This is where the inference-versus-training distinction becomes strategically important. Training a model is a one-time (or periodic) event. Inference runs continuously, at scale, for every user query. As AI gets embedded into more products and workflows, inference workloads are growing faster than training workloads.

What investors and the industry should watch

Feldman’s TIME 100 Most Influential People in AI recognition for 2026 reflects the growing influence Cerebras has on the conversation around AI infrastructure.

The competitive landscape is also getting more crowded. Google continues to develop its TPU line for internal and cloud workloads. Amazon’s Trainium chips are scaling up inside AWS. Startups like Groq are targeting inference speed with their own novel architectures.

The data center power constraint alone could reshape capital allocation across the tech sector for the rest of the decade. Cerebras is one of a small number of companies arguing that the answer isn’t just “build more of the same infrastructure, but bigger.” Whether the market agrees will likely become clearer as contracts like the General Compute deal move from announcement to deployment in early 2027.