On November 13, SMIC delivered a clear supply-side judgment at its third-quarter 2026 earnings call: the supply gap for AI chips keeps widening, and capacity tightness could last into the second half of 2028. Co-CEO Zhao Haijun said at the call that AI demand remains insatiable, and the industry consensus is that 2027 will be a year of severe capacity shortage, expected to continue through mid-to-late 2028; after that, AI-related memory demand may reach basic balance, while compute and other companion chips could remain tight.

Behind that judgment is a structural reshuffle of foundry capacity. Zhao noted that demand for smartphones and mid-to-low-end consumer products is weakening and may keep declining; the company is redirecting freed-up capacity toward high-demand, high-margin products such as dedicated memory and BCD processes. Beneath the "off-season, not off" framing lies a migration of capacity from consumer electronics to the AI chain.

Concept illustration of a wafer fab line
Concept illustration: silicon wafers and robotic arms on an advanced fab line

The numbers back the storyline. SMIC expects full-year revenue to exceed US$12 billion (about 80.2 billion yuan), up nearly 30 percent year on year; with total depreciation up about 30 percent, gross margin is expected at around 26 percent, roughly 5 points higher than a year ago. For the fourth quarter, shipments are expected to stay steady, with product-mix changes and market-based pricing lifting the average selling price, and revenue forecast to grow 4 to 6 percent quarter on quarter. Rising depreciation and material costs, plus more engineering and R&D wafers in the quarter, will nudge utilization down slightly even as new capacity comes online; the gross-margin guidance of 26 to 28 percent is flat with the previous quarter.

By stretching the tightness outlook to 2028, the foundry leader has anchored a time coordinate for the whole AI hardware supply chain: wafers, companion chips and memory will not loosen at the same pace, memory may rebalance first, and the contest for compute-line capacity will continue. For cloud providers and chip designers downstream, judging when to lock in capacity now has a first-hand reference from the supply side itself.