<figure class="news-cover"><img src="https://aposhuang.com/uploads/covers/news-221-bbafb130.jpg" alt="" loading="eager"></figure>
<p><strong>TL;DR</strong>: Google will let Oracle sell its tensor processing units, the first time its own silicon is available outside Google's data centres. If you buy AI compute, you now have a third credible supplier and real leverage in price talks; if you are planning 2027 capacity, price a TPU option before your budget locks. Do not assume the TPU is a drop-in Nvidia replacement — the software stack is the switching cost.</p>
<p>Alphabet said on 27 October that Oracle will offer Google's tensor processing units to Oracle's cloud customers, one day before Alphabet reported third-quarter results. It is the first time TPUs — custom chips Google has been building since 2013 — will be sold outside Google's own infrastructure. Google only began selling TPUs to outside customers in the second quarter of this year.</p>
<h2>① Oracle becomes the first place to rent a Google TPU outside Google</h2>
<p>Google spent years arguing that chips, models and cloud bought from one vendor beat the alternatives. Selling silicon through Oracle inverts that pitch: the company is pushing hardware past its own walls to reach customers it cannot host itself. Cloud boss Thomas Kurian, who spent more than 20 years at Oracle before joining Google in 2018, is running the play.</p>
<ul>
<li>TD Cowen models $9.4 billion of TPU sales in the second half of this year and $62.5 billion in 2027</li>
<li>Wells Fargo expects about $9 billion in the third quarter and $99 billion next year</li>
<li>SemiAnalysis puts Ironwood, Google's latest TPU, at roughly 44% lower cost to own and operate than an Nvidia GB200 server</li>
<li>TD Cowen estimates a 20% operating margin on TPU sales, against about 35% for Google Cloud last quarter</li>
</ul>
<p>Google's own demand is not the constraint on supply — construction is. Chief financial officer Anat Ashkenazi said in July that only a small portion of TPU revenue from existing agreements would be recognised this year, with the "vast majority" landing in 2027, and that Google planned to use more outside capacity in the third quarter while it builds its own infrastructure. Alphabet's cloud unit posted $25 billion of revenue in the second quarter against $42 billion at AWS and $29 billion at Azure, but analysts expect growth of about 90% year on year, after 82% in the prior quarter.</p>
<p>The biggest buyer is also a competitor. Anthropic agreed last October to use up to 1 million TPUs, a deal worth tens of billions of dollars that Mizuho says brings nearly a gigawatt of capacity online this year. This week Broadcom detailed a $63 billion vehicle, backed by Blackstone and three large banks, that buys TPUs and networking gear and leases them to Anthropic for five years; analysts at William Blair estimate the financing covers about half of Anthropic's 2027 needs. OpenAI is also planning to buy TPUs from Google, CNBC confirmed after The Information reported it.</p>
<h2>② Our take: is Google becoming a chip merchant, or defending a cloud it cannot build fast enough?</h2>
<p>It is a supply trade, not a strategy pivot, and the tell is the margin. Google is monetising silicon it would otherwise have to house, at a 20% operating margin instead of the roughly 35% its cloud business earned last quarter. That is a worse business per dollar — but a dollar you cannot earn at all if your data-centre build is the bottleneck is worth more than a margin point.</p>
<p>The uncomfortable part is the direction of travel. A company that sells its differentiator to a rival cloud is admitting that distribution, not silicon, is now the scarce asset. Kurian's own framing is about efficiency rather than generosity: "It's because we can optimize that whole stack," he said at a Goldman Sachs conference in September, describing why Google's servers pay for themselves.</p>
<p>Two readings are defensible. The bullish one: Chris Ballard of Check Capital Management argues Google is "a very strong competitor, and it's very possible that they could eventually take market share from Nvidia." The cautious one: TPU revenue is lumpy, contract-driven and concentrated in two model labs that are also Google's competitors, so a single procurement decision in 2027 can move the number more than any product improvement. We lean to the second. Uday Cheruvu of Harding Loevner put the strategic shift plainly: "Google was great at doing its own thing, but it was never able to serve customers. That mindset has changed." Serving customers through Oracle is what that change looks like when you cannot serve them yourself.</p>
<h2>③ Enterprise AI teams get the leverage; model labs get the bill</h2>
<p>For a company buying accelerators, the practical change is a third quote on the table. Procurement teams that have priced only Nvidia capacity and one hyperscaler now have a credible alternative with a real cost advantage on paper — 44% lower cost to own and operate, per SemiAnalysis — and a financing market that will lease it to you, as the Anthropic structure shows.</p>
<p>For Anthropic and OpenAI, the same shift cuts the other way. Their compute plans now depend on a competitor's roadmap, and on lease vehicles intermediated by Broadcom. Anthropic disclosed in its IPO filing that it uses silicon from Google, Amazon, Nvidia and AMD, choosing "based on the best price-performance and efficiency characteristics for each workload" — a sensible hedge that also describes how exposed the company is to any one supplier.</p>
<h2>④ What AI infrastructure buyers should do now</h2>
<ul>
<li>This week: get a written TPU price and availability quote from Oracle alongside your existing Nvidia and hyperscaler quotes, so the comparison exists before renewal.</li>
<li>This month: cost the migration, not the chip. Ask what your training and inference code needs to port, and who pays for that work.</li>
<li>This quarter: model 2027 capacity against the TD Cowen ($62.5 billion) and Wells Fargo ($99 billion) TPU supply estimates, not against this year's deliveries.</li>
<li>Before signing: check whether your contract lets the reseller change the underlying hardware generation mid-term.</li>
</ul>
<p><strong>What not to do:</strong> treat a 44% cost-of-ownership claim as a 44% saving on your bill — it excludes porting, retraining and staff. And do not sign a multi-year single-supplier deal this quarter on the assumption that TPU supply will stay tight; Google's own guidance puts most revenue recognition in 2027, which is when capacity arrives.</p>
<h2>⑤ Watch Alphabet's cloud growth, 2027 TPU recognition, and the Anthropic lease</h2>
<p>Four numbers settle the argument. Alphabet's third-quarter cloud revenue growth, guided near 90%, shows whether TPUs are already paying. The split of TPU revenue between 2026 and 2027 reveals how much is contract talk and how much is delivery. The first payments under Anthropic's five-year Broadcom lease test whether the financing structure works at scale. And Oracle's TPU region count tells you whether this is a real channel or a single announcement.</p>
<p class="src">Source: <a href="https://www.cnbc.com/2026/10/28/google-alters-cloud-playbook-turning-to-partners-to-sell-more-ai-chips.html">CNBC, October 28, 2026</a></p>