The Captive Silicon Wave — In Five Days, Five Buyers Became Chip Designers

The Pattern — Five Days, Five Buyers Turning Into Designers

The most distinctive thread in this week's US semiconductor news flow was not packaging, nor Huawei's new scaling architecture, nor the EU's €120B Chips Act 2.0, nor SK hynix crossing $1 trillion in market cap. It was the common denominator beneath all of those stories: the largest chip buyers are all simultaneously declaring themselves chip designers.

On May 28–29, BYD unveiled what it called "China's first domestically designed 4nm automotive SoC," locking in an in-house silicon path for ADAS and assisted driving. The same week, ByteDance was reported to be designing its own AI CPUs and inference chips modeled on the LPU architecture used by Groq (an AI inference chip startup). SpaceX's IPO filing explicitly stated that its orbital AI ambitions require "significantly more" chips than are available, and named its in-house TeraFab program as a material risk. Nvidia, posting a record quarter, simultaneously cancelled one GPU SKU and elevated its in-house CPU strategy on the roadmap. And Beijing placed nine domestic AI accelerators on its "secure and reliable" procurement whitelist for the first time, institutionalizing captive demand.

Five different industries. Five different motivations. One conclusion: depending on merchant silicon will not survive the next cycle.

Why Now — Memory Tightening Meets GPU Rationing

DDR5 16Gb spot closed May 30 at $41.9. AI-driven HBM demand has dragged commodity DRAM up with it; SK hynix's market cap crossed $1 trillion, and the combined Samsung + SK hynix value is approaching Bitcoin's. Bloomberg wrote about an Asia-wide memory chip "frenzy." Goldman tagged Nvidia and Micron as the biggest AI winners.

But the demand side tells a sharper story — rationing. SpaceX writing in its IPO that it cannot get enough chips is the most honest statement in the entire news cycle. Bismarck Brief warned TSMC's advanced node and CoWoS capacity will be the binding constraint on 2026 AI buildouts. Asia Times reported a sharp escalation of H200 smuggling routes through Japan. Translation: even with the money, you cannot buy enough chips, and every large buyer has internalized that fact at the same time.

The response is twofold — (1) design captive silicon for the workloads that matter most, so you have direct claims on TSMC / SK hynix / Micron lines, and (2) keep buying merchant Nvidia/AMD for the general-purpose tail, but use the captive program as negotiating leverage.

Huawei Is a **Consequence** of This Pattern, Not the Cause