Google Cloud doubles revenue and turns the TPU into a product in its own right

28 October 2026. Alphabet reports third-quarter results: $127.6 billion in revenue, up 25% year over year, propelled by a division that spent years dragging on the books. Google Cloud has effectively doubled: $30.7 billion, up 103%, and more tellingly $11.1 billion in operating income against $3.6 billion a year earlier. Why it matters: the cloud is no longer a promise, it is the group’s profit engine — and the heaviest signal is not the growth rate but the way Google now sells its TPU silicon as a product in its own right.

The cloud becomes Alphabet’s profit engine

Google Cloud now accounts for roughly 24% of consolidated revenue, up from 15% a year earlier, with a segment operating margin of 36.2%. Of the $9.3 billion increase in consolidated operating income, $7.5 billion came from the cloud alone. Sundar Pichai and Thomas Kurian lead with one number: first-party Gemini APIs now process nearly 32 billion tokens per minute, four times the prior-year level, and the Gemini app claims one billion users.

The flip side is a financial strain the release does not hide. Net income falls to $26.2 billion from $35.0 billion, with diluted EPS of $2.12 versus $2.87. The cause is not operational: other income and expense swung from a $12.8 billion gain to an $8.2 billion charge, mostly unrealised investment losses. Wall Street, which expected about $128 billion in revenue and EPS near $3.02, punished the stock — but most of the miss is finance, not cloud.

The TPU steps out of the cloud and becomes a product

The most interesting structural change in the 28 October investor call is a distinction Alphabet makes explicit for the first time: building capacity for its own services, versus manufacturing complete TPU systems for external customers.

Concretely, Google is no longer only selling compute hours on its own machines. It is distributing integrated TPU systems — compute, storage, optical networking and software — into customer or third-party facilities, naming Oracle, Crux AI, and also Anthropic, Apple, Citadel Securities, Jump Trading, Caltech and Physical Intelligence. Thomas Kurian’s logic: deliver AI compute to the customer where data cannot move to the cloud, and reduce the need to build data centres for those deployments yourself.

CFO Anat Ashkenazi spells out the economics. Components are purchased and held in inventory against signed contracts; cash flows out as inventory builds, and revenue is recognised on delivery. That is a different cash-flow cycle from a classic hyperscaler, and it explains part of the negative $9.1 billion free cash flow for the quarter — alongside contracted TPU inventory build-up and a $5.2 billion European Commission fine.

On performance, Google claims a 2.7× training gain for the TPU 8t generation and 80% better inference price-performance for the 8i. TorchTPU, developed with Meta, aims to make PyTorch portable to TPUs. For a buyer the consequence is heavy: the TPU stops being an infrastructure footnote and becomes a catalogue line with its own lead times and delivery cycle.

Read together, these moves reposition Google as a competitor to NVIDIA in AI accelerators, not just a hyperscaler customer. Selling TPU systems into third-party facilities is a hardware business with different unit economics — lower per-unit margin than software, but a direct claim on workloads that would otherwise run on someone else’s silicon. The practical upshot for a buyer is that “Google Cloud” and “Google TPU” are now two distinct purchasing decisions.

An order book that spells scarcity

The capacity numbers tell the story better than any statement. The cloud backlog reaches $561 billion, of which just over 55% is expected to convert to revenue within 24 months; on top of that sit $924 billion in future purchase commitments and $152 billion in uncommenced leases. Capital expenditure guidance for 2026 stays at $195–205 billion, with a “significant” increase expected in 2027.

Two points should hold a CIO or platform lead’s attention. First, Alphabet says it will lean on third-party compute starting to ramp in the fourth quarter, while warning this will pressure near-term cloud margins. Second, the infrastructure split — roughly 65% servers and networking, 35% data centres — confirms the bottleneck is moving to the supply of chips, memory and networking, not concrete.

For the busy reader, the summary fits in one table.

What this changes for a cloud buyer

Three practical consequences follow from these numbers.

Capacity is the real bottleneck, not price. With a $561 billion backlog and an explicit shift to third-party compute, part of the capacity ordered today will not be delivered for several quarters. Enterprises with predictable AI workloads should lock in long-term commitments rather than bargain over spot pricing.

The TPU becomes an architecture choice, not just a billing choice. For a sizeable PyTorch workload, the alternative — “rent TPUs from Google, buy TPU systems to run on-premises, or stay on NVIDIA” — is now a genuine engineering decision, with TorchTPU as the bridge. This is not a detail: the accelerator choice locks in software dependencies for years.

Margins will compress, and with them discounts. Third-party compute ramps in Q4 2026 with an explicit margin warning. For buyers that may mean less generous terms in the near term, but also rising competitive pressure if AWS and Azure answer on the same ground — AWS reporting 39% growth and up to $220 billion in AI capex, and Microsoft a quarterly capex of $45.2 billion.

Verdict

If you run large-scale AI workloads, secure long-term capacity commitments now: the $561 billion backlog and the move to third-party compute indicate that capacity, not price, is the constraint that will bite first. If you are evaluating accelerators, treat the TPU as a real option: benchmark TorchTPU on your models before committing, because today’s accelerator choice freezes your software dependencies for years. If you are simply a cloud customer, expect a stretch of margin pressure: the third-party compute ramp in Q4 2026 and heavy 2027 spending will not produce softer buying terms in the short run. Google Cloud is no longer telling a catch-up story — it is telling a story of profit and scarcity, and both are written in the same order book.

References

• Alphabet — Q3 2026 earnings release (PDF)
• Converge Digest — Google Cloud Profit Tops $11B as Gemini Usage Soars (29 October 2026)
• CRN — Google Cloud Sales Explode 103 Percent As CEO Touts TPU Demand (29 October 2026)
• TechTimes — Google Cloud Hits 103% Growth as Alphabet Misses Revenue (30 October 2026)