← Back to feed

Google recognized its first revenue from TPU system sales to customer data centers in Q2 2026. CFO Anat Ashkenazi confirmed the milestone on Alphabet’s July 22 earnings call. It ends a ten-year policy of keeping custom AI silicon entirely internal.

The Numbers

• TPU inventory: Rose from $628M (June 2025) to $7.74B (June 2026). A 12x increase in 12 months. Google’s 10-Q filing identifies the inventory as “primarily hardware related to TPU systems for sale to enterprise customers and devices.”
• First named customer: Blackstone. The joint venture was announced in May 2026.
• Revenue scope: The Q2 recognition is a minority portion of the $514B Google Cloud backlog. The “vast majority” of contracted TPU system sale revenue is expected to land in 2027.

Alphabet has not disclosed unit volumes, per-chip margins, or the total number of external customers. Blackstone is the only named party.

The Break With a Decade of Internal-Only Policy

Tensor Processing Units have been a strictly internal asset since Google built the first generation for inference in 2015. The chips trained and served Google’s models; no external party bought or deployed them. The rationale was compound: vertical integration over third-party competitors, no leaking chip architecture to rivals, and maintaining cost advantages that came from not commoditizing the silicon.

Google telegraphed the shift in April 2026, saying it would sell TPUs to a “select group of customers.” The Blackstone JV announcement in May 2026 provided the first concrete transaction. The Q2 earnings call is the first confirmation that revenue has actually been recognized.

Why It Matters for NVIDIA

Google selling TPUs externally is an attempt to attack Nvidia’s distribution moat. Nvidia’s ecosystem advantage is not just chip performance — it is the training tooling, CUDA libraries, and relationships that hyperscaler neoclouds like CoreWeave, Nebius, and Lambda have refused to abandon. Google has been pushing TPU adoption into that market and found resistance.

Selling TPU systems directly to enterprise customers and infrastructure investors (Blackstone) is a different go-to-market. It bypasses the neocloud relationship and targets buyers building their own data centers who are not already locked into Nvidia’s stack.

Current TPU Lineup

Google’s 8th-generation TPUs are split into two variants: TPU 8t for training workloads and TPU 8i for inference. The split is a deliberate optimization choice — inference and training have different memory and compute requirements. Both variants are part of the Virgo network architecture that Google has described as uniting over one million TPUs across its infrastructure.

The Blackstone deal was described in May as targeting 500MW of capacity by 2027, anchored around the new chip generation. The $514B Cloud backlog — Google’s single largest figure from the Q2 call — includes TPU system sales as a minority component.