The short version
of every page

The world's largest chip foundries told customers this week that lead times for AI accelerators will not return to normal before 2027, according to three people briefed on the guidance. The constraint is no longer the number of wafers that can be etched, but the capacity to bond finished dies to high-bandwidth memory.

Lead times for advanced packaging have roughly doubled over the past year. Hyperscalers, which now account for well over half of accelerator demand, have responded by pre-booking capacity two to three years ahead, effectively locking smaller buyers out of the queue.

"Every slot that opens up is already spoken for," said one executive at a mid-sized AI lab, who asked not to be named because the negotiations are private. "We're being quoted delivery dates that would have been a joke eighteen months ago."

New fabs, late arrivals

Relief is coming, but slowly. Packaging plants under construction in Arizona and Kumamoto are expected to come online in late 2026, with meaningful output not before the following year. Analysts expect the squeeze to push prices for the newest accelerators up by a further 10 to 15 percent.

For smaller labs and startups, the practical effect is a two-tier market: those with long-term contracts, and everyone else renting whatever spare capacity the cloud providers choose to resell.