Advanced packaging, not wafer supply, is what limits accelerator shipments into 2027

The question everyone asks about accelerator supply is how many wafers the leading edge foundries can produce. It is the wrong question and it has been the wrong question for about two years. Logic wafer capacity at the current node is adequate for the volumes being shipped, and utilisation across that segment sits in the low seventies. The constraint sits one step later, in the sequence of operations that takes a finished logic die, bonds several stacks of memory beside it, connects the whole assembly on a silicon interposer and mounts that on an organic substrate with twenty or more layers.

Packaging used to be the part of the process nobody wrote about. It was the back end, it was cheap, it was located wherever labour was cheapest, and capacity could be added in a year by anyone with a clean room and a wire bonder. High bandwidth memory ended that. The stacking step requires equipment with placement accuracy measured in single digit microns, the interposer is itself a silicon product built in a front end fab, and the substrate is a printed circuit board problem that has been pushed far past where printed circuit board suppliers are comfortable.

Where the capacity actually binds

Breaking an accelerator package into its steps shows the problem clearly. Three of the six operations have spare capacity and three do not, and a package cannot be built faster than its slowest step.

Two rows sit above 95%. Memory stacking is the one the market discusses, because it is visible in the memory makers' own disclosures and because it is where the yield losses are largest. Substrate is the one that decides schedules, and almost nobody outside the supply chain tracks it. A high layer count organic substrate for an accelerator package is a laminated structure with more than twenty build-up layers, very tight registration tolerances and a defect rate that punishes any supplier attempting to scale quickly.

Capacity has more than tripled in three years and demand has grown faster. The 2028 and 2029 bars are where the constraint eases, and both depend on equipment that has been ordered but not delivered. Anything a customer wants in 2027 is already allocated, which is why the negotiation between accelerator designers and packaging suppliers has moved from price to prepayment. Several designers have funded packaging capacity directly in exchange for guaranteed allocation, a structure that was almost unheard of in this part of the chain before 2024.

Prepaying a packaging house to build a line for you is an admission that the market clearing mechanism has stopped working. When the customer has to finance the supplier's balance sheet to get allocation, price is no longer doing its job, and that situation resolves in one of two ways. Either capacity catches up or somebody walks away from a prepayment.

Hannah Ilves, semiconductor analyst at Quillon Data

The substrate problem is a concentration problem

Substrate supply concentrates for reasons that are hard to unwind. The manufacturing know-how is tacit and lives in process engineers rather than in documents. Yields on the highest layer counts are materially below those on mainstream products, which means a new entrant loses money for years while learning. And the customers, having been burned by qualification failures, are reluctant to dual source in a shortage.

The commercial consequence is a transfer of margin. When the scarce step in a chain is owned by four private companies and the abundant step is owned by three public ones, pricing power moves toward the private four. Substrate prices for accelerator grade product have risen an estimated 34% since the start of 2025 while leading edge wafer pricing has been roughly flat, and that gap is showing up in the gross margins of the firms that have to buy both.

Sanborn Micro's August update, which pushed roughly $1.9bn of front end spending from 2027 into 2028, should be read alongside its packaging commitments rather than instead of them. The company is not cutting capital spending. It is moving it from the step where it has spare capacity to the step where it does not, which is the correct decision and the one the aggregate capital expenditure line will entirely fail to communicate.

The useful indicator through the rest of this year is not bookings at the large equipment suppliers. It is lead times quoted for substrate drilling and lamination equipment, which currently run past a hundred weeks. When those come in, the constraint is a year from easing. While they keep extending, every accelerator shipment forecast for 2027 is a forecast about four private companies nobody names.