Google Compute Deals: Inside the Multi-Billion AI Infrastructure Shift

On Friday, June 5, 2026, SpaceX disclosed a contract under which Alphabet's Google would pay the rocket-and-satellite company $920 million per month for access to approximately 110,000 Nvidia GPUs, along with associated CPUs, memory, and infrastructure at SpaceX's Colossus data centre complex. The deal, running from October 2026 through June 2029, had a total contract value north of $30 billion. It was the second mammoth compute agreement SpaceX had signed in a matter of weeks, and it landed just seven days before the company's initial public offering, AFP reported on the day of the announcement.

The Google-SpaceX deal was not an isolated event. It arrived at the tail end of a six-week period during which the largest AI labs and their cloud providers had effectively renegotiated the terms of who owns compute capacity, who rents it, and who is prepared to write a cheque measured in hundreds of billions of dollars. On May 31, Anthropic had expanded its partnership with Google Cloud to secure access to up to one million chips in a compute commitment that Morning Overview reported exceeded $200 billion. On June 9, at its Worldwide Developers Conference, Apple confirmed that its most advanced AI workloads would run on Nvidia GPUs inside Google Cloud, 24/7 Wall St reported. And on June 22, SpaceX signed yet another deal, this time with open-source AI startup Reflection, worth up to $6.3 billion, CNBC reported.

Taken together, these agreements do more than add zeros to balance sheets. They reveal a structural shift in how frontier AI compute is provisioned. The old model, dominant from roughly 2020 through 2025, was bilateral: each major lab struck a primary cloud deal with one hyperscaler, and that relationship defined the lab's infrastructure strategy. OpenAI had Microsoft Azure. Anthropic had AWS and, later, Google Cloud. Google DeepMind ran on Google's own infrastructure. The new model, visible in the contracts signed between May and June 2026, is multilateral, multi-vendor, and organised less like a procurement function than a commodities trading desk.

Consider the positions of the three largest AI labs as of mid-2026. OpenAI, still anchored to Microsoft Azure, has disclosed no equivalent to the Anthropic-Google Cloud expansion but saw Microsoft commit $80 billion to AI infrastructure in its 2025 fiscal year alone. Anthropic, now the most valuable AI company by private market metrics, has compute relationships with Google Cloud, AWS, Akamai, and SpaceX. Apple, historically the most vertically integrated company in consumer technology, has ceded its most demanding inference workloads to Google Cloud running Nvidia silicon. The map of who provides compute to whom is no longer a set of exclusive lanes. It is a mesh.

The SpaceX dimension of this story is the most peculiar and, for the hyperscaler incumbents, the most disruptive. A company whose primary revenue driver remains launch services and satellite internet has become, in the space of a few months, one of the largest commercial compute providers on the planet. SpaceX acquired xAI earlier in 2026, merging Elon Musk's AI venture into the parent company and gaining control of the Colossus data centre in Memphis, Tennessee. The facility, already among the largest single-site GPU clusters in the world, was repositioned almost immediately as a neutral compute marketplace, selling capacity to any lab willing to pay.

Google is the most conspicuous buyer. The $920 million monthly payment to SpaceX, Yahoo Finance reported, secures capacity Google needs to train and serve its Gemini family of models at a time when internal data centre buildout cannot keep pace with demand. Anthropic is also renting SpaceX capacity, according to CNBC, alongside its primary Google Cloud commitment and a separate $1.8 billion seven-year deal with Akamai that Forbes reported in early May. The Akamai agreement is particularly telling: a content delivery network, known for moving video bits, has retooled its edge infrastructure into what it calls an Inference Cloud, and Anthropic is its anchor tenant.

What changed? The bottleneck. Throughout 2024 and into 2025, the binding constraint on AI progress was access to the most advanced Nvidia GPUs. Labs competed for allocation slots on the H100 and then the B200. By mid-2026, Nvidia had ramped production of its GB300 series and the supply of advanced silicon, while still tight, was no longer the only game in town. The new bottleneck is energy, land, and the permitting timelines required to build and power data centres at gigawatt scale. The hyperscalers have committed enormous capital to data centre construction. Amazon projected over $100 billion in infrastructure spending for 2026. Microsoft and Google have made similarly vast pledges. But a financial commitment and a powered, operational data centre are separated by years of regulatory approvals, grid interconnection queues, and physical construction.

SpaceX, with Colossus already operational and scaled, offered something the hyperscalers could not match in the near term: capacity available now. The trade-off, for Google, was straightforward. Pay a premium to a third party for immediate access to 110,000 GPUs, or wait for internal data centres to come online and risk ceding model performance ground to competitors who had already solved their compute equations. Google chose the premium. The contract structure, at $920 million monthly, implies a daily rate of roughly $30 million. That number is large enough to suggest that what Google is buying is not merely GPU hours. It is buying time.

Anthropic's $200-billion-plus commitment to Google Cloud represents the other end of the spectrum: a long-duration, deeply integrated partnership that locks the lab and the cloud provider together for years, perhaps a decade. The scale of the commitment has no real precedent in the technology industry. By comparison, the largest enterprise cloud deals of the 2010s, such as the CIA's $600 million AWS contract or the Pentagon's $9 billion JEDI cloud procurement, are rounding errors. The Anthropic-Google Cloud deal, as Morning Overview noted, secures access to up to one million of Google's custom TPUs alongside Nvidia GPUs, making Anthropic the largest single tenant in Google's AI infrastructure by a wide margin.

Yet even with that commitment, Anthropic is spreading its infrastructure across four providers: Google Cloud for the bulk of training and inference, AWS for legacy workloads and a portion of inference, Akamai for edge inference distribution, and SpaceX for overflow compute during peak training cycles. Diversification is the watchword. The strategic logic is not difficult to decode. A lab that sources all its compute from one provider is exposed to that provider's capacity constraints, pricing power, and strategic priorities. A lab that sources from four providers has negotiating leverage, redundancy, and the ability to route workloads to whoever offers the lowest cost per teraflop at a given moment.

Apple's entry into this multilateral compute market is, in some ways, the most revealing signal of all. The company spent the better part of two decades building an integrated hardware-software stack that, in the consumer electronics business, is a durable competitive advantage. The M-series chips that power MacBooks and iPads are Apple-designed, Apple-commissioned, and Apple-exclusive. For its Private Cloud Compute initiative, Apple initially built its own data centres running M-series silicon, pitching the architecture as a privacy guarantee: user data would never leave an Apple-controlled environment. At WWDC 2026, that pitch was quietly amended.

Craig Federighi, Apple's senior vice president of software engineering, confirmed that the company's most advanced AI model, referred to internally as FM Cloud Pro, would run on Nvidia GPUs hosted inside Google Cloud, according to 24/7 Wall St. The decision, which would have been unthinkable at Apple five years earlier, is a concession to the physical reality that training and serving frontier-scale AI models requires a density of compute that Apple's own data centres, even with custom silicon, could not economically provide. Apple is not exiting the infrastructure business. It is acknowledging that the frontier has moved beyond what any single company, no matter how vertically integrated, can build alone on a reasonable timeline.

Apple's FM Cloud Pro runs on Nvidia GPUs inside Google Cloud, abandoning private infrastructure for its most demanding AI workloads., 24/7 Wall St, reporting on Craig Federighi's WWDC 2026 confirmation

The Apple deal also reveals something about Google's strategy that was less visible in the Anthropic and SpaceX agreements. Google Cloud is not merely selling raw compute. It is selling an ecosystem. By hosting Apple's AI workloads on Nvidia GPUs, Google positions itself as the neutral platform on which even its platform competitors can run. This is the playbook AWS used to dominate enterprise cloud in the 2010s: sell infrastructure to everyone, including direct business competitors, because the scale economics of cloud make neutrality more profitable than exclusivity. Google Cloud, long the third-place hyperscaler behind AWS and Azure, appears to be executing a variant of that strategy for the AI era, and the results are showing up in contract values that would have seemed fantastical three years ago.

The question that hangs over this cascade of deals is what happens when the capacity that Google, Anthropic, and Apple are paying premiums for today becomes commoditised. The SpaceX Colossus facility is, at root, a collection of Nvidia GPUs in a warehouse with a power connection. Nvidia sells GPUs to anyone with the capital to buy them. The know-how required to cluster tens of thousands of GPUs into a coherent training fabric is nontrivial, but it is also increasingly well-documented and replicable. If SpaceX can turn a data centre into a $30 billion revenue stream over three years, other well-capitalised entrants will try. The barriers to entry in AI compute are capital, energy permits, and operational expertise. Capital is abundant. Energy permits are a political bottleneck, not a technological one. Operational expertise is a hiring problem.

The cheapest signal that this strategy is working for the labs will arrive in the form of model releases. Every major lab now trains its frontier models on infrastructure spread across multiple providers. If the next generation of Claude, Gemini, and Apple Intelligence models demonstrates performance gains commensurate with the capital deployed, the multilateral compute model will be validated. If the gains are marginal, or if a lab running primarily on a single provider's infrastructure jumps ahead, the diversification thesis weakens. The first checkpoint will be the next round of post-training write-ups, which labs typically publish alongside major model releases. Watch for the acknowledgements section. The list of infrastructure partners thanked in the footnotes of a model card has become as revealing as the benchmark scores in the body.

One name that earns its own footnote in this story is Nvidia. The chipmaker is the quiet counterparty to nearly every agreement signed in the May-to-June window. Google's deal with SpaceX is, in economic substance, a mechanism for Google to rent Nvidia GPUs it could not procure directly in sufficient quantity on the timeline it needed. Anthropic's deal with Google Cloud includes both Google's custom TPUs and Nvidia silicon. Apple's arrangement routes Nvidia GPUs through Google Cloud. SpaceX's entire compute business is built on Nvidia GB300 chips. As The Motley Fool noted in its analysis of the Google-SpaceX agreement, Nvidia is positioned to book revenue from every layer of this ecosystem regardless of which company's logo appears on the contract.

There is a calendar buried in the SpaceX IPO filing. June 12, 2026, was the date the company went public at a valuation of approximately $1.7 trillion, MSN reported, making it the largest public offering in history. The Google compute deal was announced June 5. The Reflection deal followed on June 22. The sequencing was not accidental. An IPO prospectus that lists $2.17 billion in monthly recurring compute revenue, from Google and Anthropic alone, tells a story to public-market investors that the satellite-launch business could not tell on its own: SpaceX is an AI infrastructure company now, and the revenue is already booked. The calendar was the message.

What the six-week compute deal frenzy of 2026 did, more than anything, was dissolve the idea that AI infrastructure is a two-sided market between labs and their chosen hyperscaler. It is now a multi-party bazaar in which labs shop for capacity across four or five providers, hyperscalers rent from each other and from upstarts, and a rocket company can become a compute landlord practically overnight. The nameplate on the data centre matters less than the chips inside, the power contract that feeds them, and the date by which they come online. For the next round of model releases, the infrastructure acknowledgements will be the lede. For everyone else, the question is simpler: who, by the end of 2026, has not diversified their compute supply, and what does that undiversified position cost them?