A decade of keeping TPUs in-house, reversed in one announcement
Google has developed and used tensor processing units internally since 2013, treating them as a proprietary advantage rather than a product to sell outside its own cloud. The company's announcement at Oracle AI World in Las Vegas changes that: starting in 2027, Oracle will rent out Google TPUs from its own cloud infrastructure, marking the first time Google's custom silicon will run on another hyperscaler's platform rather than exclusively inside Google's own data centers.
The deal carries a notable personal thread. Google Cloud head Thomas Kurian spent more than two decades at Oracle before joining Google in 2018, and multiple reports describe him as the executive pushing this particular partnership. A twenty-year Oracle veteran now running the Google Cloud division that just agreed to let Oracle resell Google's most strategically guarded hardware is the kind of detail that makes a purely financial deal read a little more like a personal full-circle moment as well.
The financing structure doing the heavy lifting underneath
Separately, Broadcom organized a large financing facility, reported at roughly $63 billion, involving major banks and investment firms, to fund Broadcom-developed TPUs and networking equipment that get leased to Anthropic over a five-year term, with deliveries expected to begin in 2027. This structure lets Google's TPU ecosystem scale to a customer like Anthropic without Google itself carrying the full balance sheet exposure of building and financing that capacity directly, spreading the financial risk across banks and investment firms with their own appetite for infrastructure-backed lending.
This is not Google's first attempt at this kind of structure. Earlier in the year, a related arrangement used a special purpose vehicle to buy roughly a gigawatt of TPU hardware for Anthropic, with Broadcom backstopping a large share of the exposure if Anthropic's payments stopped. The scale has clearly grown since then, and the pattern is consistent: Google wants the revenue and ecosystem lock-in that comes from supplying Anthropic's compute, without recording the full liability on its own books the way it would if it financed the buildout directly.
Why analysts are calling this a shift from loyalty to supply
TD Cowen's projections, $9.4 billion in TPU sales for the second half of 2026 alone and $62.5 billion for all of 2027, describe a genuinely new revenue category for Google, not a marginal side business. More importantly, the analysis argues that wider TPU distribution through Oracle and the Broadcom-financed Anthropic leases turns chip access into a supply question rather than a vendor-loyalty question: a buyer no longer has to be a Google Cloud customer specifically to run workloads on Google silicon, which weakens one of the structural advantages that has historically locked enterprises into a single hyperscaler's hardware ecosystem.
The practical test, as the analysis frames it, is whether Oracle's TPU capacity actually comes online on schedule in 2027 and whether TD Cowen's projected revenue becomes real booked revenue rather than optimistic modeling. If both hold, Google has a credible path to eroding Nvidia's hold on hyperscale frontier training workloads specifically, by making its custom silicon available through a channel, Oracle's cloud, that a meaningfully different set of enterprise customers already trusts and uses.
The open questions that determine whether this actually works
Several structural details remain genuinely unresolved and worth tracking rather than assuming resolved in Google's favor by default. It is not public whether Oracle receives a wholesale discount on TPU capacity or whether Google retains the full margin on chips sold through Oracle's channel, a detail that materially affects how profitable this arrangement actually is for Google versus how much it is simply a market-share and ecosystem play funded by thinner margins.
It is similarly unclear how much of the $63 billion Broadcom-organized facility is already contractually committed to 2027 deliveries versus how much remains speculative capacity that depends on Anthropic's continued growth to actually get built and leased. Given that Anthropic's compute commitments to Google Cloud are reported to represent a substantial share of Google's committed future cloud revenue, the entire structure carries real concentration risk if Anthropic's own growth trajectory falters for any reason over the life of these leases.
How Anthropic ends up at the center of the entire structure
Anthropic sits at the middle of nearly every thread in this story: it is the customer whose chip purchases the Broadcom-organized financing facility funds, it is reportedly committed to a cloud spending figure with Google that represents a substantial share of Google's own committed future cloud revenue, and it lacks the independent credit rating that would let it finance this scale of hardware purchasing on its own balance sheet. Every financing workaround described here exists specifically because a company without an established credit history needs gigawatts of compute faster than conventional lending would normally allow.
That concentration cuts in two directions at once. For Google and Broadcom, Anthropic's continued rapid growth is effectively a load-bearing assumption underneath a financing structure now reaching tens of billions of dollars. For Anthropic, the arrangement secures compute capacity that would otherwise be difficult to access at this scale on any realistic timeline, at the cost of its growth trajectory becoming a variable other companies' balance sheets are now quietly depending on as well.
What this means for any enterprise currently locked into Nvidia-based planning
Technology leaders who have been building 2027 and 2028 infrastructure roadmaps around assumed Nvidia GPU scarcity now have a credible reason to revisit those assumptions. If Google's TPU ecosystem genuinely becomes accessible through a mainstream channel like Oracle's cloud, rather than remaining a Google-exclusive advantage, the effective AI compute supply available to the broader market expands in a way that could meaningfully soften the pricing power Nvidia has enjoyed through the current scarcity cycle.
The appropriate level of confidence here is cautious rather than certain. Google has announced the intent and the financing structure exists; actual delivered capacity running real enterprise workloads through Oracle in 2027 is still a forward-looking claim, not a completed fact. Enterprises should treat this as a reason to ask their current compute vendors direct questions about TPU access and pricing over the next two to three quarters, not as a reason to assume the chip scarcity problem is already solved.



