Google Sells TPU v8 to Third-Party Data Centers

IO Fund's Beth Kindig reports that Google will begin selling its TPUs to select third-party data center operators, which IO Fund frames as Google entering the merchant AI accelerator market. IO Fund notes Google's latest earnings included a cloud backlog it describes as up about 400% year over year to roughly $462 billion, and argues the company is capitalizing on shifting workload economics toward inference. Alphabet's Q1 2026 filing independently reported total revenue backlog of about $467.6 billion, largely attributed to Google Cloud.

Google's eighth-generation TPUs, announced at Google Cloud Next 2026, split into two designs: the TPU 8t for training and the TPU 8i for inference, the first time Google has bifurcated the line. Per IO Fund, the new TPU pods emphasize a large coherent shared memory as a key architectural differentiator relative to Nvidia GPU clusters, which IO Fund presents as enabling different scaling and latency trade-offs for high-throughput inference serving.

As AI workloads shift from expensive training runs toward high-volume inference, total cost of ownership and cost per token become primary purchasing considerations for operators. A vendor that sells custom accelerators into third-party data centers creates an alternative procurement path to incumbent GPU suppliers, which can reshape price-performance negotiations and long-term capacity planning for inference.

IO Fund frames the merchant-sales push as coinciding with three converging dynamics: rising inference share, pressure on hyperscalers to monetize models, and what IO Fund calls a potential 'Rubin delay' for Nvidia. IO Fund argues these could together open a window for custom silicon to gain inference share, and points to signals of merchant traction such as a Google-Blackstone AI cloud joint venture targeting an initial 0.5 GW of TPU capacity in 2027 and a multi-gigawatt Anthropic TPU commitment. For practitioners, hardware diversity at the data-center level influences deployment architectures, latency envelopes, model partitioning, and inference cost optimization.

• •Reported uptake by third-party operators and any published performance or power-efficiency comparisons between TPU 8i pods and Nvidia GPU clusters.
• •Benchmarks for inference cost per token and latency at scale on coherent-shared-memory TPU topologies versus GPU-based sharded approaches.
• •Integration and orchestration constraints operators describe when fitting TPUs into existing inference stacks.

Google's merchant-sales debut is best read as an inflection in inference hardware procurement options rather than a guaranteed displacement of GPU incumbents; adoption at scale will depend on measured cost per inference, software maturity, and operator integration work.