Cerebras CEO to Address AI Scaling Limits at TechCrunch Disrupt

The Compute Bottleneck Takes Centre Stage

Andrew Feldman, CEO and co-founder of Cerebras Systems, is set to appear at TechCrunch Disrupt 2026 to discuss a question that now dominates boardrooms from Palo Alto to Shenzhen: whether artificial intelligence can maintain its current trajectory when compute, energy and infrastructure are all under strain.

At Opentechwire, we have tracked the infrastructure crisis unfolding across the AI supply chain over the past eighteen months. Training runs for frontier models now consume gigawatt-hours of electricity, data centre waiting lists stretch into 2027, and chip allocations remain the most closely guarded numbers in quarterly earnings calls. Feldman's session arrives at a moment when the industry is openly debating whether the scaling laws that have driven progress since 2017 are approaching a hard ceiling.

Cerebras' Bet on Wafer-Scale Architecture

Cerebras has staked its position on a fundamentally different approach to AI silicon. Where Nvidia and AMD build processors from discrete chips assembled into clusters, Cerebras manufactures single wafer-scale engines that span an entire silicon wafer, delivering what the company describes as orders of magnitude more on-chip memory and interconnect bandwidth.

The architecture is designed to eliminate bottlenecks that emerge when model parameters and activations must move between chips over comparatively slow interconnects. For training runs that involve trillions of parameters, that data movement becomes a dominant cost in both time and energy. Cerebras argues that keeping computation on a single die reduces latency and power consumption per inference, though the trade-off is a higher upfront cost per unit and complex thermal management.

The approach has attracted attention from national laboratories and research institutions running large-scale simulations, but it remains a minority architecture in a market where Nvidia's H100 and forthcoming B-series chips set the de facto standard. Feldman's appearance will likely offer a window into whether Cerebras sees an opening as hyperscalers confront power and cooling limits in their existing GPU clusters.

Energy and Infrastructure Under Pressure

The energy question is no longer theoretical. Utilities in Northern Virginia, home to the world's densest concentration of data centres, have warned that new AI facilities may face multi-year delays for grid connections. In Singapore, the government imposed a moratorium on new data centre construction in 2019 that has only partially lifted, citing land and power constraints. Across the Asia-Pacific region, AI infrastructure is competing with industrial electrification and residential demand for finite generation capacity.

Feldman is expected to address how Cerebras positions its technology in this environment. If wafer-scale engines deliver materially better performance per watt, they could appeal to operators facing hard power caps. But the technology must also prove itself on total cost of ownership, a metric that includes not just electricity but also capital expenditure, cooling infrastructure, and the operational complexity of running non-standard hardware at scale.

What Happens When Scaling Stalls

The more provocative part of Feldman's session will likely be his view on what the industry does if current hardware cannot support the next generation of models. The assumption underpinning much of the venture capital and corporate investment in AI over the past four years has been that doubling compute budgets will continue to yield predictable improvements in capability. That assumption is now being tested.

Research labs have begun to explore alternative paths: sparse models that activate only subsets of parameters, multimodal architectures that compress information more efficiently, and techniques like distillation that transfer knowledge from large models into smaller, faster ones. If the era of brute-force scaling is ending, the competitive landscape shifts toward companies that can extract more capability from constrained resources.

Cerebras, as a challenger in a market dominated by established GPU vendors, has an interest in accelerating that shift. Feldman's remarks will be scrutinised for signals about whether the company sees the current constraints as a temporary supply-side crunch or a structural inflection that reshapes how AI systems are built.

The Regional Dimension

For readers tracking the AI hardware race from Seoul, Taipei, or Bengaluru, Feldman's session also carries implications for how compute capacity is distributed globally. Export controls on advanced chips have created a bifurcated market, with cutting-edge silicon concentrated in the United States and a handful of allied jurisdictions, whilst other regions pursue domestic alternatives or work with de-rated hardware.

Cerebras has not publicly detailed its export footprint, but any discussion of scaling constraints will inevitably touch on questions of where the next wave of AI infrastructure gets built and who controls access to it. The company's wafer-scale approach, if it proves viable at scale, could represent a point of differentiation in markets seeking alternatives to the incumbent GPU supply chain.

Open Questions

Feldman's session will unfold against a backdrop of earnings calls in which hyperscalers have acknowledged that their capital expenditure on AI infrastructure is growing faster than revenue from AI services. The gap between investment and return is widening, and the market is waiting to see whether that gap closes through new applications or whether the current cycle resembles earlier waves of overbuilding.

The technical question of whether AI can keep scaling is inseparable from the economic question of whether it makes financial sense to keep scaling. Cerebras, as a private company that filed for an initial public offering in 2024, will eventually need to demonstrate that its technology can command sustainable margins in a market where Nvidia enjoys pricing power and ecosystem lock-in.

Feldman's appearance at TechCrunch Disrupt 2026 will not resolve those questions, but it will offer a view into how one of the industry's most distinctive architecture bets is positioning itself as the constraints become impossible to ignore.