Cerebras Systems CEO Andrew Feldman will tackle one of artificial intelligence's thorniest questions at TechCrunch Disrupt 2026: whether the hardware industry can sustain the exponential scaling that has powered AI's recent breakthroughs.
The San Jose-based chip startup has positioned itself as a contrarian answer to the conventional GPU dominance. While Nvidia's H100 and H200 accelerators remain the standard for training large language models, Cerebras has built specialized processors that prioritize energy efficiency and compute density over raw parallelization. Feldman's appearance signals that the conversation around AI infrastructure is shifting from "how do we get more chips" to "can we keep doing this sustainably."
The core problem is real. Training modern foundational models demands hundreds of millions of dollars in compute infrastructure. Data centers consuming these systems face mounting electricity costs and grid constraints. The industry has largely thrown hardware at the scaling problem, but that approach hits physical and economic walls. Cerebras' wafer-scale engine takes a different architectural approach, fitting more transistors and memory onto a single processor to reduce data movement between chips, a major source of latency and power waste.
Feldman's position at Disrupt matters because Cerebras raised $250 million at a $5 billion valuation in 2023, backed by investors including Sapphire Ventures and Coatue Management. The company competes in a crowded field against Graphcore, Trazer, Cerebras itself, and a growing roster of startups building specialized silicon for AI. But Cerebras also faces Nvidia, which controls roughly 80-90% of the AI accelerator market and has shown surprising agility in addressing efficiency concerns with newer architectures.
The real question Feldman will likely address is whether specialized hardware can penetrate model training at scale. Cerebras has secured customer wins, but the AI development community remains gravitationally pulled toward the GPU ecosystem. Developers optimize code for CUDA. Open-source frameworks favor GPU acceleration. Switching to alternative hardware demands retraining teams and rewriting software, a friction many organizations won't bear unless the cost savings are transformational.
Feldman's framing around "what comes next if today's AI hardware reaches its limits" hints at a deeper inflection. The scaling laws that powered GPT-3 to GPT-4 to the newest multimodal models may be flattening. Training costs have risen exponentially while benchmark improvements have plateaued in some domains. If raw compute scaling becomes economically untenable, the industry pivots toward inference optimization, mixture-of-experts architectures, and efficient algorithms. That environment would favor companies arguing for fundamentally different hardware approaches.
The timing of this talk at a major tech conference also reflects Cerebras' growth narrative. The company has moved beyond "could this work" to "here's how we're deploying it." If Feldman can articulate a vision where Cerebras hardware becomes the standard for a post-scaling-law era of AI development, it reshapes how the venture and enterprise communities think about infrastructure bets.
Cerebras has positioned itself for a conversation it believes the industry will have to have. Whether Feldman's argument converts skeptics depends on translating theoretical efficiency gains into real customer deployments that undercut GPU economics at scale.