Cerebras CEO to Tackle AI Scaling Limits at TechCrunch Disrupt 2026
On September 30, 2026, Cerebras Systems confirmed that CEO and co-founder Andrew Feldman will lead a “Can AI Keep Scaling?” session at TechCrunch Disrupt 2026 (October 13–15, Moscone West, San Francisco). The talk confronts a central industry paradox: ever-improving AI models demand ever-rising compute, energy, and physical infrastructure—and it is uncertain how much longer this trend can continue. Cerebras rejects conventional chip design. Its wafer-scale computing architecture builds a processor directly on an entire silicon wafer—bypassing the need to slice wafers into discrete chips and wire them together—yielding ultra-high interconnect bandwidth ideal for massive AI training workloads. Key commercial milestones include a $5.5 billion May 2026 IPO and a multiyear deal with OpenAI for 750 megawatts of deployed Cerebras systems from 2026 through 2028. The CS-4 platform launched in August 2026.
A Decade-Long Bet Against Conventional Wisdom
Feldman co-founded Cerebras in 2015 after running microserver startup SeaMicro (acquired by AMD in 2012) and network-equipment firm Force10 Networks. The most counterintuitive move? Pursuing wafer-scale integration when industry consensus held it impractical—then delivering commercial production a decade later. Traditional chips cut silicon wafers into many dies, package them separately, and connect them via PCB traces or interposers; this imposes latency and power penalties at scale. Cerebras’s monolithic approach eliminates those bottlenecks by keeping all cores on a single die, maximizing data movement speed while minimizing energy per operation. InfrastructureScale is now matching compute ambition: over 600 megawatts of data center capacity is live or under contract for delivery by end-2027, including 200 megawatts planned for Cerebras’s first European facility, also targeting completion by 2027. Manufacturing capacity is being scaled more than tenfold during 2026 alone. These figures support Feldman’s view that scaling AI requires more than faster chips—it demands parallel expansion of power, cooling, and manufacturing capabilities.
| Area | Confirmed Figure | Timeline |
|---|---|---|
| OpenAI contract | 750 megawatts | 2026–2028 |
| Datacenter capacity | >600 megawatts | Operational by end-2027 |
| European capacity | 200 megawatts | By end-2027 |
| 2026 manufacturing growth | >10× | Calendar year 2026 |
| IPO proceeds | $5.5 billion | May 2026 |
Who Should Tune In?
This session is especially valuable for:
- AI research leads who must coordinate model size with available infrastructure cadence;
- Datacenter operators evaluating rack-level power, cooling, and footprint requirements for wafer-scale systems;
- Seed-to-Series C investors tracking how infrastructure capital intensity reshapes the AI supply chain economics. If you’re waiting for real-world validation of wafer-scale reliability or concrete deployment timetables beyondSigned contracts, the Q&A segment will likely address these concerns directly.
In Closing
The AI scaling bottleneck has migrated from transistor density to megawatt density. Cerebras’s trajectory proves architectural innovation must be paired with industrial-scale execution—or promising chips remain laboratory curiosities. As competition shifts from who reels in faster GPUs to who builds bigger power substations, infrastructure players are stepping into the spotlight.