TSMC has sold out its manufacturing capacity through 2028, according to Piper Sandler, with artificial intelligence demand the driving force behind a backlog that now stretches four years out.

The call comes from the brokerage's semiconductor research team and lands amid the sharpest capacity squeeze the foundry industry has seen in a decade. AI accelerators, high-bandwidth memory interposers, and networking silicon are consuming advanced-node wafers faster than TSMC can add cleanroom space. The result, in Piper's reading, is a foundry whose leading-edge capacity is effectively committed before the fabs that will host it are fully built.

A sold-out position through 2028 matters commercially in three ways.

First, pricing. When customers pre-commit years of volume, the foundry gains leverage over wafer pricing, and the usual annual discount negotiations tilt in TSMC's favor. Piper's framing implies that customers wanting guaranteed allocation at advanced nodes will pay for certainty — whether through upfront commitments, co-investment in dedicated capacity, or accepted price increases.

Second, allocation discipline. AI chip designers are not the only customers in the queue. Smartphone, PC, and automotive SoC vendors compete for the same N-series wafers, and a four-year backlog forces product planners to lock in supply far earlier than the industry's traditional 12-to-18-month planning horizon. Smaller fabless companies without committed allocations face the sharpest risk of being squeezed to mature nodes or second-tier foundries.

Third, capital spending. Sold-out capacity validates aggressive capex, and TSMC has been spending accordingly, with its annual investment program running at historic highs to fund new fabs in Taiwan, Arizona, Japan, and Germany. Each sold-out year strengthens the case that this spending is demand-backed rather than speculative — a distinction investors have watched closely since the memory industry's last overbuild cycle.

The geopolitical backdrop reinforces the dynamic rather than easing it. Export controls have pushed AI chip designers toward a narrower set of manufacturable configurations, concentrating demand on the specific advanced nodes where TSMC holds its strongest position. At the same time, governments in Washington, Tokyo, and Berlin are subsidizing new capacity on their own soil, but those fabs ramp years after they are announced. Piper's 2028 horizon sits squarely inside that gap between subsidized construction and delivered output.

For competitors, the analyst view cuts both ways. Samsung Foundry and Intel Foundry gain a commercial opening: customers who cannot secure TSMC allocation have a concrete reason to qualify alternative processes, however painful the porting effort. But a four-year TSMC backlog also signals that AI demand exceeds what the entire industry can supply, meaning rivals can win share without yet threatening TSMC's revenue trajectory.

Investors should note the epistemics. This is an analyst assessment, not a TSMC disclosure; the company itself has consistently described AI-related demand as insatiable in recent quarters without publishing an allocation calendar. Piper's sold-out-through-2028 claim is best read as a synthesis of customer checks and capacity modeling rather than a leaked booking schedule. TSMC's next earnings call, where management updates utilization, pricing, and capex guidance, will offer the first structured test of the claim.

If the backlog holds, the competitive question for 2026–2028 shifts from who wins AI designs to who can actually manufacture them — and on that question, Piper's note suggests the answer is already booked.