The head of TSMC confirmed that over the next few years, the industry will not be able to fully meet the growing demand for AI chips. The shortage of production lines is forcing leading tech companies to look for alternative suppliers. Intel has taken advantage of this, managing to attract new major customers despite recent financial losses and missed deadlines.
New Intel Contracts
Why capacity cannot simply be added is a fact about the process, not about willingness to spend: a leading-edge wafer takes roughly a hundred days to run, a new fab takes years to build, and a design is written against one foundry's toolchain, so moving it means quarters of requalification — see semiconductor manufacturing. The step that has most often been the true constraint is not the wafer at all but advanced packaging, the assembly line where the compute die and its memory stacks are bonded into one accelerator. The full map of who owns which link is in the AI chip supply chain.
Due to TSMC's capacity constraints, orders are being shifted:
- Google has already ordered the production of over 3 million Tensor Processing Units (TPUs) from Intel, with delivery scheduled for 2028.
- Nvidia has not yet signed an official contract but is actively testing Intel's manufacturing processes for its future Feynman GPU architecture.
Memory Partnership
In parallel, South Korean manufacturer SK Hynix has begun testing the compatibility of its memory with Intel's new chip packaging technologies, further strengthening the US company's position as a backup manufacturer in the AI hardware market.