The shift was first telegraphed in Q1 2026, when CEO Sundar Pichai announced plans to sell TPUs to a "select group of customers" who wanted to run them in their own facilities . By Q2, those deliveries had commenced and revenue was being booked .
Alphabet is simultaneously prioritizing its own highest-priority effort: artificial general intelligence (AGI) research. Pichai stated on the Q2 earnings call that Alphabet is managing its fleet of AI accelerators to prioritize AGI development rather than renting them all to cloud customers . This creates an inherent tension: the more TPUs Google sells externally, the fewer it retains for internal AGI training and inference workloads. The decision to sell chips at all signals that Alphabet believes the strategic benefits — locking in enterprise AI relationships, monetizing excess supply, competing with Nvidia — outweigh the cost of diverting some compute from its AGI ambitions.
TSMC's CoWoS (Chip-on-Wafer-on-Substrate) advanced packaging capacity is the binding constraint across the entire AI chip industry. TSMC CEO C.C. Wei confirmed CoWoS is sold out through 2026 . Multiple reports indicate Google was forced to cut its 2026 TPU production target by roughly 25% — from an estimated 4 million units down to about 3 million — due to insufficient CoWoS allocation . This means Google's ability to scale TPU sales and internal use is directly gated by how many chips TSMC can package, not by demand or design capability.
All of the financial figures cited here are confirmed by Alphabet's Q2 2026 earnings report and multiple credible sources:
| Factor | How It Interacts |
|---|---|
| TPU sales (new vendor role) | Diversifies revenue, competes with Nvidia, but diverts chips from internal use |
| AGI hoarding priority | Pichai explicitly prioritizes keeping TPUs for AGI research, limiting how many can be sold |
| CoWoS bottleneck | TSMC packaging capacity caps total TPU output (~25% production cut); both internal use and external sales are constrained |
| Cloud revenue surge (82% YoY) | Enterprise AI demand is enormous; backlog of $514B proves it is multi-year committed demand |
| Capex at $195–205B | Alphabet is betting aggressively on AI infrastructure scale to capture that demand |
| Negative FCF (-$5.9B) | The cost of that bet: first negative free cash flow in 22 years as a public company, raising investor questions about near-term ROI |
Bottom line: Google's Q2 2026 TPU sales launch marks its transformation from a cloud-rental-only chipmaker into a direct semiconductor merchant, but that move operates under tight supply constraints (CoWoS), internal competition for chips (AGI priority), and extraordinary but cash-burning financial scale ($514B backlog vs. -$5.9B FCF).