TL;DR
On October 28, 2026, Google announced it will sell its custom TPUs externally through partners such as Oracle, with Wells Fargo estimating Q3 TPU sales near $9B: versus NVIDIA's direct GPU model with CUDA lock-in, Google is taking a lower-margin (~20%) external-sales route, and the differences lie in channel, revenue recognition, margin and total cost of ownership.
Comparison Overview
On October 28, 2026, per Cailian Press, Google reshaped its cloud strategy: instead of offering its chips mainly through its own cloud, it will expand external sales of its custom TPUs through outside partners. Oracle will become the first hyperscaler outside Google to offer TPUs, with the associated compute expected to deploy starting in 2027. Meanwhile, NVIDIA's latest FY2027 Q2 results show data-center revenue of $89.023B (as reported by Shanghai Securities News). This article compares Google's external TPU sales against NVIDIA's GPU business across channel, revenue scale and recognition cadence, margins, total cost of ownership and ecosystem lock-in.
Comparison Table
| Dimension | Google external TPU sales | NVIDIA GPU |
|---|---|---|
| Sales channel and model | Served mainly through Google Cloud, with added external hyperscaler partners such as Oracle; Oracle-channel compute expected to deploy from 2027 | Direct chip sales plus systems (GB200 servers) sold via global cloud providers and system integrators |
| Latest-quarter chip-related revenue | Per Wells Fargo estimate, Google's Q3 TPU sales were about $9B | Official results: FY2027 Q2 (ended July 26, 2026) total revenue $96.221B, including $89.023B from data centers, 92% of total |
| Revenue recognition cadence | Alphabet CFO Anat Ashkenazi said on the July Q2 earnings call that only a small portion of existing TPU system sales agreement revenue is recognized in 2026, with the vast majority expected in 2027 | Revenue recognized quarterly; FY2027 Q2 data-center revenue up 117% YoY |
| Operating margin | Per TD Cowen estimate, Google's external TPU business runs at about 20% operating margin, below Google Cloud's actual 35.6% operating margin in Q2 2026 | NVIDIA's operating margin for the quarter not disclosed within this research scope |
| System TCO / payback period | Per SemiAnalysis estimate, the latest Ironwood TPU costs about 44% less to own and operate than NVIDIA GB200 servers; Google Cloud chief Thomas Kurian said custom-chip server payback is roughly half that of purchased AI servers (under two years) | Benchmark: GB200 servers are the baseline of SemiAnalysis' TCO comparison |
| Ecosystem lock-in | TPUs tied to the Google Cloud software stack; Oracle channel broadens reach | CUDA ecosystem lock-in |
| Anchor customers | Anthropic is Google's largest TPU customer (agreement last October for up to 1 million TPUs worth tens of billions of dollars); per National Business Daily on October 28, OpenAI also plans to buy Google TPUs | Specific anchor customer list not given within this research scope |
| Dependence on partners | External sales depend on hyperscaler partners such as Oracle for facilities, power and networking; Oracle channel deploys only from 2027 | No dependence on a single cloud partner; NVIDIA itself is the chip and platform supplier |
Evidence and Positions
Google's official line: reshaping cloud strategy from offering chips mainly through its own cloud to expanding external TPU sales via outside partners; Oracle will be the first hyperscaler outside Google to offer TPUs, with compute deployment expected to begin in 2027 (per Cailian Press, October 28, 2026). Alphabet CFO Anat Ashkenazi said on the July Q2 earnings call that only a small portion of existing TPU system sales agreement revenue is recognized in 2026, with the vast majority expected in 2027 (per Cailian Press). Google Cloud chief Thomas Kurian said custom-chip server payback is roughly half that of purchased AI servers (under two years).
NVIDIA's official line (as reported by Shanghai Securities News): FY2027 Q2 (ended July 26, 2026) total revenue of $96.221B, up 106% YoY; data-center revenue of $89.023B, up 117% YoY, 92% of the total.
Third-party estimates: per Cailian Press citing Wells Fargo, Google's Q3 TPU sales were about $9B, rising to an estimated $99B in 2027; TD Cowen estimates H2 2026 TPU sales of $9.4B growing to $62.5B in 2027; TD Cowen also puts Google's external TPU business operating margin at about 20%. Per Cailian Press citing SemiAnalysis, the latest Ironwood TPU costs about 44% less to own and operate than NVIDIA GB200 servers. Per Cailian Press, Anthropic is Google's largest TPU customer, under a last-October agreement for up to 1 million TPUs worth tens of billions of dollars; per National Business Daily on October 28, OpenAI also plans to procure Google TPUs.
Interpreting the Disagreement
First, TPU revenue forecasts diverge widely: Wells Fargo projects 2027 TPU sales of $99B while TD Cowen projects $62.5B — a gap of about $36.5B. Both are analyst estimates, not official disclosure; Google has only confirmed that the vast majority of agreement revenue is recognized in 2027 and gives no annual sales guidance.
Second, margins are not directly comparable: TD Cowen's 20% is the operating margin of the external TPU business, while the 35.6% benchmark is Google Cloud's overall operating margin in Q2 2026 — different business scopes. NVIDIA's quarterly operating margin is not disclosed within this research scope, so the three margins cannot be put on the same ruler.
Third, TCO baselines and attribution: SemiAnalysis' 44% cost advantage compares owning and operating Ironwood TPU versus NVIDIA GB200 servers, a third-party estimate; Kurian's payback claim is a Google statement with purchased AI servers (under two years) as the implied baseline, and purchase scale and power-price assumptions are not disclosed.