Google will prioritize cutting-edge AGI (artificial general intelligence) development above all else in allocating its proprietary AI accelerator, the TPU. CEO Sundar Pichai revealed this during the Q2 2026 earnings call on July 22. In the same quarter, Google delivered TPU systems to customer data centers for the first time and began recognizing the associated revenue. Google is advancing internal use and external sales simultaneously. While securing the capacity needed for AGI research and its own services first, it has also opened a supply channel that draws on customer-side facilities and outside capital.
An Order of Compute Resources Anchored in AGI Development
Regarding TPU allocation, Pichai explained: "The first priority is allocating the amount needed to compete at the frontier of AGI development." When further asked about the allocation of compute resources overall, he reiterated that the amount needed for cutting-edge AGI development serves as the baseline. Alphabet has not disclosed the number of TPUs allocated to AGI or the ratio of internal-to-external allocation. Even so, the policy of estimating model development's compute demand ahead of other businesses is clear.
For the remaining capacity, Google Cloud is prioritized alongside Search and YouTube. Within Cloud, in addition to compute resources for providing proprietary models through Vertex AI and Gemini Enterprise, allocations go to core services such as data analytics and cybersecurity. When providing Google's models to many Cloud customers, both TPUs and NVIDIA GPUs are used—operations are not run on proprietary chips alone to meet demand.
This ordering does not mean Google will stop selling TPUs externally. When outside customers want the infrastructure itself, Google installs TPUs at the customer's data center. In other words, before diverting Google's own data center capacity to external sales, the company is expanding the range of installation sites to the customer side.
Splitting Processing Between the 8t for Training and 8i for Inference
The 8th-generation TPU that Google announced in April 2026 splits into the "TPU 8t" for training and the "TPU 8i" for inference and reinforcement learning. The 8t packs 9,600 chips into a single Superpod, delivering 121 exaflops of compute performance and 2PB of shared memory. The design allows over 1 million TPUs to be connected as a single training cluster across multiple sites.
The 8i, meanwhile, is built to more easily hold the KV cache used during inference on-chip, offering 80% higher inference performance per dollar than the previous generation. The training workloads for building cutting-edge AGI models and the inference workloads for returning answers in Search and Gemini Enterprise call for different suited systems. As a result, internal research and customer-facing services are not always competing for the same TPU model number.
However, even with chips split between training and inference, the supply constraints Alphabet has indicated still remain. In Q2, roughly 60% of technical infrastructure investment went to servers, and roughly 40% to data centers and networking equipment. Alphabet has not disclosed what constitutes the biggest bottleneck, but the investment spans both the servers that house the chips and the facilities that run those servers.
External Sales Head to Customer-Side Data Centers
As of Q1, Alphabet had already outlined plans to deliver TPUs to selected customers' data centers. In Q2, it completed its first deliveries and recognized revenue from TPU system sales for the first time. The plan moved into actual shipment. However, the portion of existing contracts that will turn into revenue in 2026 is relatively small; the bulk is expected to be recognized in 2027.
TPU sales contracts also flow into Google Cloud's backlog. The backlog grew by more than $50 billion from the prior quarter to reach $514 billion. The majority consists of standard GCP contracts, and the amount attributable to TPUs is undisclosed. Q2 Cloud revenue was $24.768 billion, up 82% year-over-year, with operating income of $8.814 billion and an operating margin of 35.6%. According to Alphabet, Cloud growth accelerated significantly even excluding the impact of TPU sales.
Supply channels using outside capital are also starting to move. In May, Google and Blackstone announced a joint venture to operate a TPU cloud. Blackstone will make an initial investment of $5 billion, with 500MW of capacity expected to come online in 2027. Google will supply the TPUs, software, and services. Combined with direct sales to customer facilities, this lets Google expand TPU installation sites without shouldering all the data center investment itself.
$195–205 Billion and the Cost of Short-Term Procurement
Alphabet raised its 2026 capital expenditure guidance from the previous range of $180–190 billion to $195–205 billion. In Q2 alone, capital expenditures reached $44.924 billion, exceeding operating cash flow of $39.069 billion. Free cash flow came in at negative $5.855 billion. The pace of capacity expansion is outrunning current cash generation.
Even so, the completion of internal facilities is not keeping pace with demand. In Q3, Alphabet will increase the compute capacity it procures from external operators as a bridge. This cost is expected to temporarily lower Google Cloud's profit margin. Pichai explained that while bearing higher costs upfront for a few months, the company expects to earn high returns on investment through large multi-year customer contracts.
Google's TPU strategy has entered a phase of restructuring even where compute resources are physically located, in order to protect AGI development while expanding external sales. By 2027, most of the existing TPU sales contracts are expected to convert into revenue. In the next earnings report, it will be necessary to separately verify the amount recognized and the pressure that external capacity places on Cloud's profit margin.