📌 Google Must Resolve Delivery Bottlenecks Before Expanding AI Compute Capacity
Google’s order backlog of nearly USD 514 billion means that the focus of AI capital expenditures is no longer only whether it dares to invest, but whether it can complete data centers and supporting infrastructure on time. According to relevant claims, this backlog is about five times the current level; only if supply constraints ease can the accumulated customer demand be converted into revenue more quickly.
This also changes the perspective on capital expenditures. Expanding server rooms, securing power, and deploying chips first bring cash outlays, but what they correspond to is not only a judgment about distant demand, but also orders that have not yet been delivered. Google’s record profits and investment returns provide a financial buffer, but an order backlog does not equal recognized revenue, and the delivery pace will still affect when revenue and cash flow are realized.
If TPUs are provided to external customers as a service, they could turn infrastructure that primarily serves internal operations into a new source of revenue. Barclays’ projection is that externally sold TPUs could generate about USD 22 billion in revenue per GW, with a gross margin of about 25%; deployments of slightly more than 2 GW could potentially reach USD 50 billion in annual revenue. This is an estimate based on the conditions that deployment scale, revenue per unit, and profit margins can all be achieved, and it cannot be directly equated with existing cloud revenue or backlog orders.
What remains to be watched next is the supply side: whether power, server room, and chip deliveries can remove the constraints; and whether external customers are willing to continue taking on TPU compute capacity. The former determines the pace at which backlog orders are released, while the latter determines whether the external sales business can become a verifiable new revenue pool.
▌ Sources