HBM / CoWoS Bottleneck

last updated 2026-05-27
Mature Foundry PositioningChipletsBackside Power Delivery (BPD)HBM / CoWoS…

The two simultaneous supply-side constraints that define AI accelerator economics through 2027. TSMC CoWoS-S advanced packaging and HBM3e high-bandwidth memory are both running at full utilisation, both allocated almost entirely to Nvidia, AMD, Broadcom, and the hyperscaler custom-silicon programmes. Any new entrant requiring either is queueing.

The bottleneck is not just a cost / lead-time problem; it’s a structural moat against new AI accelerator companies, which is why bypass-the-bottleneck approaches like Physics-Native Compute and Mature Foundry Positioning have a window.

CoWoS allocation

TSMC’s Chip-on-Wafer-on-Substrate advanced packaging family. Three variants:

Industry-tracked capacity numbers (verify with Yole or TrendForce before citing):

New entrants typically wait 18–24 months for first allocation. Cerebras’s WSE generation requires CoWoS-L specifically, smaller production runs.

HBM allocation

HBM3e (8-stack, 24Gb-die) supply allocated:

HBM4 (12-stack, 32Gb-die, base-die customisation) target H2 2026 general availability. Likely slip to early 2027 per industry tracking.

Why this matters as a positioning thesis

Two responses available to a new AI hardware startup:

Bypass approaches have a window through 2027. Once CoWoS capacity catches demand and HBM4 ramps, the bottleneck eases and the bypass edge erodes. The window is what makes pre-seed deals in this category time-sensitive.

Bypass approaches across the cohort

DD signals

Strong signal. Founder names “no CoWoS” or “no HBM” as a positioning choice. Maps the bypass mechanism specifically (on-die SRAM, in-substrate weights, photonic interconnect). Acknowledges the 2027 bottleneck window honestly.

Weak signal. “We’ll figure out packaging later.” “HBM is just one option.” No supply-chain awareness.

Sources

Related concepts

Frontier questions