Google aims for 12–15 million TPUs by 2028 — supply chain says that outpaces NVIDIA AI GPU units

Google is treating NVIDIA scarcity as a design constraint, not a fate. Supply-chain reporting published August 2, 2026 claims Google plans to deploy roughly 12 to 15 million TPU chips by 2028 — a unit count that would exceed NVIDIA’s projected ~12.4 million AI GPU sales in the same window. The point is not a public GPU rival SKU; it is vertical integration for Google’s own training and inference fleet.

TPU v8 split, TPU v9 packaging

Current-gen talk centers on TPU v8 in two flavors: a training-leaning v8T configuration (huge HBM pools, dense pods) and an inference-leaning v8i node design. Next up, TPU v9 is described as a multi-chip package path using Intel’s EMIB technology, with Intel reportedly in line for on the order of ~3 million related units across that generation. Take the exact flop counts as vendor-circuit rumor until Google publishes a Cloud paper.

What it means outside Mountain View

If the volume claims hold, hyperscalers keep rewriting the AI hardware leaderboard in units owned, not GPUs sold at Best Buy. NVIDIA still sets the merchant-silicon price; Google is betting captive TPUs win on cost-per-token inside Search, Gemini, and Cloud. Watch Google Cloud capacity announcements and Intel advanced-packaging commentary for confirmation breadcrumbs.

Custom silicon is no longer a side quest. For Google, it is the main campaign map.