Andrew Feldman maps the three bottlenecks choking AI accelerator shipments and explains how

Cerebras sidesteps them all Andrew Feldman, co-founder and chief executive of Cerebras, laid out the three supply-chain chokepoints that are stalling deliveries across the AI accelerator industry: HBM memory, CoWoS advanced-packaging capacity, and access to TSMC’s 3-nanometer line. He argues that Cerebras’ architectural choices let it bypass all three simultaneously — a rare position in a market where every major vendor is constrained by at least one.

HBM: not in stock, not in the plan

High-bandwidth memory has become the tightest resource in the supply chain; leading GPU makers depend on it entirely to feed their compute cores. Cerebras eliminated that dependency from the start. Its Wafer Scale Engine embeds massive on-chip SRAM, removing the need for external HBM stacks and their attendant memory controllers. The result: fewer components to procure, fewer suppliers to manage, and fewer points of failure in the supply chain.

CoWoS: a bottleneck avoided entirely

TSMC’s Chip-on-Wafer-on-Substrate process is the standard way to bond GPU dies to HBM stacks, and its slots are booked months ahead. Because Cerebras’ architecture is monolithic — a single wafer-scale die — it does not need multi-chip advanced packaging and therefore does not compete for scarce CoWoS capacity. The logistical advantage is twofold: lower packaging cost and independence from the packager’s schedule.

3-nanometer: staying on mature 5-nanometer

The scramble for TSMC N3 capacity leaves little room for new customers or rapid volume ramps. Cerebras chose to remain on the previous node, N5, a mature process with significantly higher availability. That decision sacrifices peak transistor density, but it guarantees stable, immediate production — a critical asset when customers need delivery now, not next quarter.

A fundamentally different supply profile

The bottom line, Feldman says, is that Cerebras enjoys a “materially different supply profile” compared with standard GPU systems. While rivals fight for HBM allocations, CoWoS slots, and 3-nanometer wafers, Cerebras relies on internal memory, simple packaging, and an available process node. In a market where shipping ability determines revenue as much as silicon performance, that operational edge is hard to ignore.

The remarks were made on the MAD Podcast with Matt Turck

The full discussion appeared in an episode of The MAD Podcast hosted by Matt Turck, also published on Cerebras’ YouTube channel. The conversation took place against the backdrop of continuing tension in the global AI accelerator supply chain, as of October 2026.