Training a frontier model now costs hundreds of millions of dollars per run and billions per year in infrastructure. No lab can self-fund that trajectory for long — so the labs sold equity, compute commitments, and distribution rights to the three companies that already own the world's largest clouds. Microsoft has OpenAI. Amazon and Google have both taken large stakes in Anthropic. This is the alliance map of the frontier era: three labs, three clouds, and a flow of money that keeps circling back to the clouds themselves.

Microsoft and OpenAI: the deal that wrote the template#

The original partnership began in 2019 with Microsoft's $1 billion investment into OpenAI. It deepened into roughly $13 billion in cumulative commitments by 2023, in exchange for exclusivity: Azure became OpenAI's sole cloud provider and sole host of the OpenAI API, Microsoft gained preferred rights to commercialize OpenAI models, and an unusual "AGI clause" gave OpenAI's board a tripwire that could terminate Microsoft's commercial rights upon a declaration of artificial general intelligence.

The economics mattered as much as the access. Microsoft integrated GPT models into Bing, Office, and GitHub Copilot, and Azure's AI business became one of the fastest-growing lines in the company. By late 2025, OpenAI restructured into a public benefit corporation in a deal that valued it at $500 billion — with Microsoft's stake, reported at around 27%, revalued at roughly $135 billion, an order-of-magnitude return on its $13 billion outlay.

Then exclusivity ended. In spring 2026, Microsoft and OpenAI restructured the commercial relationship. Key terms of the revised deal:

  • Cloud exclusivity is gone. OpenAI can now host and sell products like ChatGPT on AWS and Google Cloud. Azure remains the primary cloud partner, with new products launching on Azure first unless Microsoft cannot support the required capabilities.
  • IP rights through 2032. Microsoft keeps a license to OpenAI's models and intellectual property through 2032, but the license is no longer exclusive.
  • Revenue share through 2030. OpenAI continues to pay Microsoft a 20% revenue share through 2030, now subject to an undisclosed cap. Microsoft no longer pays a share back to OpenAI for offering the models on Azure.
  • A $250 billion Azure commitment. OpenAI committed to at least $250 billion in Azure services through 2032.
  • The AGI clause is effectively removed. The renegotiated terms eliminate the tripwire that once threatened Microsoft's rights.

The lesson of the template: the partnership was never really an investment in a startup. It was a forward purchase of compute demand and model supply, priced in equity. Microsoft got first access to the best models and a $250 billion revenue commitment back to its own cloud; OpenAI got the capital to scale. The 2026 restructuring keeps the economics while acknowledging that a multi-hundred-billion-dollar OpenAI can't be a single-cloud company anymore.

Amazon and Anthropic: the Trainium bet#

Amazon's relationship with Anthropic followed the template with its own twist — a heavier emphasis on custom silicon. The deal began in September 2023 with a $1.25 billion investment, grew with a $2.75 billion tranche in March 2024, and reached $8 billion total with a $4 billion add-on in November 2024. Throughout, Amazon has stayed a minority investor with no board seat — a structure designed, in part, to keep competition regulators at bay (the UK's competition watchdog declined a deeper probe in 2024).

The compute side is where Amazon's interests show. AWS is Anthropic's primary cloud and primary training partner, and Anthropic trains and deploys its future foundation models on AWS's Trainium and Inferentia chips. The centerpiece is Project Rainier: a cluster of nearly 500,000 Trainium2 chips across multiple US data centers, anchored by a roughly $11 billion campus in New Carlisle, Indiana — described as one of the world's largest AI compute clusters. Amazon reports more than 100,000 customers already run Claude models through AWS, and Claude powers Amazon's next-generation Alexa.

In April 2026 the partnership escalated again: Amazon announced a $5 billion investment with up to $20 billion more tied to milestones, expanding the prior $8 billion base toward a potential total near $33 billion. Alongside it, Anthropic committed to what was reported as more than $100 billion in AWS technology spending over ten years, with planned capacity scaling to five gigawatts and over a million Trainium2 chips.

That last number is the tell. Amazon's capital flows out to Anthropic and flows back as decade-long purchases of AWS compute — a circular arrangement that books cloud revenue for Amazon while giving Anthropic the chips to train the next generation of Claude.

Google and Anthropic: the chip-and-cloud hedge#

Google was Anthropic's first cloud backer. It invested roughly $300 million in early 2023 (reportedly for a 10% stake), then committed $2 billion in October 2023 — $500 million up front plus $1.5 billion over time. In return, Anthropic agreed to a multiyear Google Cloud deal reportedly worth over $3 billion, including use of Google's TPU v5e processors for scaling Claude. Claude became available through Google Cloud's Vertex AI, alongside AWS's Bedrock.

Anthropic sits in the unusual position of being funded by two cloud giants that compete with each other. Amazon gets the "primary cloud and training partner" title and the Trainium workload; Google gets a major TPU workload and a distribution channel. Anthropic, for its part, gets diversified compute — at the cost of operational dependence on vendors whose interests diverge.

In April 2026, Google reportedly went further still with a $40 billion investment in Anthropic — $10 billion up front, $30 billion tied to milestones — which if accurate would be among the largest single private AI investments ever made. The strategic logic is defensive as much as offensive: Microsoft is deeply embedded in OpenAI, and Google needed a frontier-lab relationship of comparable scale to anchor Google Cloud's AI story and give TPUs a marquee training workload.

The circularity problem#

Step back and the pattern is hard to miss. The clouds invest billions in the labs; the labs commit to spending tens or hundreds of billions on the clouds' compute. The money leaves the hyperscaler as an investment and returns as booked revenue — sometimes with a guaranteed margin baked in. Three observations follow:

  1. Compute is the real currency. Equity stakes make the headlines, but the binding terms are compute commitments: $250 billion to Azure, $100 billion-plus to AWS, multibillion-dollar TPU deals. The labs' burn goes almost entirely to the clouds.
  2. The labs are multi-sourcing. The era of one lab on one cloud is ending. OpenAI broke exclusivity and is courting AWS and Google Cloud; Anthropic plays Amazon, Google, and Nvidia off each other. Multi-cloud is leverage, not loyalty.
  3. Regulators are watching. UK and US competition authorities have already scrutinized these arrangements. The deals are carefully structured as minority investments with no board seats — which is also why they keep looking like acquisitions that aren't.

The takeaway#

The frontier of AI is not really three labs competing — it is three clouds financing model progress through the labs, each trying to lock in the demand for its own chips and data centers. For anyone building on these models, the practical takeaway is simple: watch the compute commitments, not just the funding headlines. The money tells you where the models will be trained, on whose hardware, and which cloud will have the cheapest and fastest access to them when they ship.