The AI race is no longer just happening in model laboratories. More and more, it’s being decided in infrastructure contracts. Within just a few months, OpenAI has shifted from relying almost entirely on Microsoft Azure to deploying an unprecedented multicloud strategy, with major agreements with Amazon Web Services (AWS), Google Cloud, and Oracle. The result is a shifting power map, and rapidly: the company behind ChatGPT guarantees computing capacity on multiple fronts, forcing a rebalancing among the big hyperscalers.
The latest move, announced on November 3, 2025, was the most striking: a seven-year contract with AWS valued at $38 billion to run and scale training workloads, inference, and — crucially — agentic AI. It’s not just “more cloud”: the deal includes hundreds of thousands of NVIDIA GPUs (series GB200 and GB300) interconnected in Amazon EC2 UltraServers, and the option to scale to tens of millions of CPUs for the logical and sequential parts that agents require. Capacity will begin deploying immediately, aiming to have most of it up and running before the end of 2026, with room to grow into 2027.
But OpenAI’s move isn’t possible without considering the other pillars of this multicloud approach. Microsoft remains a strategic partner — with investment and rights —; Google joined in June to expand capacity; and Oracle has been collaborating since 2024 and reportedly closed a monumental deal for the latter half of the decade, according to financial press. Here’s a breakdown of these pieces and what they mean for the AI war.
Microsoft: From Nearly Exclusive Provider to Key Shareholder… with Multicloud Allowed
For years, Azure was OpenAI’s home. In October 2025, following a recapitalization and legal restructuring, Microsoft acquired about 27% of OpenAI (through an investment in its new PBC), solidifying a financial link that coexists — and this is new — with openness to other cloud providers. The same framework also included an additional committed Azure spend of $250 billion, while eliminating the preemptive right that gave Microsoft priority over new workloads. Translated: Microsoft remains involved and strong, but OpenAI reserves the freedom to diversify.
This diversification addresses physical needs (energy, space, chip supply, intercontinental latencies) and financial needs: with models growing and a business plan aiming to scale services and agents rapidly, relying on a single provider increases risk.
AWS: $38 Billion to Put Infrastructure at the Core
The agreement with AWS marked a turning point: OpenAI will immediately use Amazon’s infrastructure for training, inference, and agents. The technical components are crucial:
- Large-scale GPUs: GB200 and GB300 NVIDIA GPUs in UltraServer clusters, equipped with internal networks designed for low latency and sustained performance.
- Massive CPUs: capacity for tens of millions of CPUs, not just marketing language — a sign that OpenAI will orchestrate swarm agents with states, queues, and policies.
- Timeline: deployment starts now, with goal by end of 2026, and expansion in 2027+.
For AWS, this move is strategic: for months, it was perceived as “lagging behind” in AI discussion compared to Microsoft and Google; now with Anthropic integrated and OpenAI onboard, it becomes the “Switzerland” of frontier computing: the highway for many of the world’s top models. For NVIDIA, the message is clear: it remains a cornerstone of the sector; OpenAI’s GPU needs are met directly by NVIDIA, not vendor-specific alternatives.
Google Cloud: Additional Capacity for a Non-Stop Demand
In June 2025, other media reported that OpenAI would incorporate Google Cloud to expand capacity. No large public figures or headlines were announced, but the gesture was significant: two direct competitors in models (Google with Gemini) and in products (ChatGPT vs. others) collaborate because physics demand it. If frontier AI requires megawatts and silicon at record pace, adding data centers worldwide reduces latencies, increases redundancy, and mitigates supply risks.
From Google’s perspective, gaining OpenAI’s workload is a validation of its AI infrastructure (networks, high-performance storage, security) and a further step in its narrative of “AI-ready cloud”. From OpenAI’s perspective, it’s ensuring no region or provider becomes a single point of failure when demand spikes.
Oracle: From “Azure on OCI” to the Big Contract of the Second Half of the Decade
Oracle was the first third-party involved alongside Microsoft to extend Azure AI over Oracle Cloud Infrastructure (OCI), announced in June 2024: more capacity for Azure hosted on OCI, with low latency thanks to proprietary interconnections between both clouds. A year later, in September 2025, The Wall Street Journal and other outlets mentioned a historic deal between Oracle and OpenAI for the second half of the decade: around $300 billion over five years, with 4.5 GW of data center capacity starting from 2027. Oracle didn’t disclose details; OpenAI didn’t deny it, and though unconfirmed, the figure aligns with the race to secure energy, land, and networks at large scale.
Here’s a key often overlooked: multicloud contracts are no longer just “another region” or “better price”. They are lots of energy, permits, and construction. They secure gigawatts, substations, cables, cooling systems, and local contracting. Those who arrive late won’t be able to train or infer, no matter how good their models.
Why OpenAI Is Moving to Multicloud (and Why This Changes the Sector)
- Operational risk: if a provider hits a bottleneck (chips, permits, energy), the program won’t stop. It will be rerouted.
- Coverage and latency: inferences and agents with global users demand proximity; dispersing capacity reduces queues and errors.
- Cost negotiation: multi-year commitments with several providers refine pricing and priority.
- Governance: with regulators eyeing the AI and cloud oligopoly, sharing load reduces concentration risks.
For hyperscalers, the message is twofold. Microsoft maintains a key client and investor — but is no longer alone. AWS regains prominence and accelerates its narrative as the “backbone of AI”. Google enters the capacity picture, and Oracle positions itself for the next phase (2027+). For NVIDIA, all roads currently lead to its GPUs: demand will grow wherever chips and energy are available.
What Do the Numbers Say? Physics and Accounting
Those only looking at OpenAI’s revenue — around $13 billion annually — versus a committed capex that, adding in all agreements, some analysts estimate exceeds one trillion dollars, might see a discrepancy. But it’s not so much if you understand the financial model popularized by general cloud providers in recent years:
- Guaranteed consumption contracts over several years.
- Provider investments (data centers, energy, networks, chips) backed by those commitments.
- Pricing and priority for customers, who monetize their products (ChatGPT, APIs, agents, vertical solutions).
The risk isn’t eliminated — for neither provider nor customer — but it’s spread across a timeline and multiple geographies.
Implications for Companies and Developers
More capacity available in AWS, Azure, Google, and OCI for “ready-to-scale” AI workloads. Expect new instances and managed services built on the improvements in network and storage deployed for OpenAI. Agentic AI will gain momentum because, finally, CPUs and memory will be close to data, enabling the organization of complex workflows with SLOs. And cost per token should stabilize as competition for hosting these large models eases tensions.
For technical teams, the tactical lesson is: design portable architectures, multimodel, and multi-agent; prioritize governance (who asks, with what data, what’s returned, how it’s audited); and measure from day one TTFT, p95/p99 latencies, and cost per interaction. Vendor lock-in won’t disappear, but it can be negotiated more effectively with two or three options on the table.
What to Expect in 2026
- Energy and Permits: Will providers be able to ramp up gigawatt-scale capacity on schedule?
- Networks and Latency: How far can internal networks for multimillion-token clusters go without bottlenecks?
- Regulation: The FTC and European Commission are already watching the AI and cloud oligopoly; remedies and conditions are likely.
- Agent Economy: If agentic AI lives up to its promise, we’ll see refashioned value chains; otherwise, it’ll just be another layer of “assistants” costing more.
Frequently Asked Questions
Does Microsoft remain OpenAI’s main provider?
It remains a key partner: investing (~27% post-recapitalization), and OpenAI has expanded its Azure consumption commitment (≈ $250 billion). But OpenAI is no longer exclusively tied to Azure: it can transfer workloads to AWS, Google Cloud, or Oracle.
What does the $38 billion deal with AWS offer that Azure doesn’t?
Immediate deployment of large GPU clusters (GB200/GB300) in UltraServers, internal networks for low latency, and capacity to scale CPUs at large scale for agents. Essentially, it’s diversification: more capacity, more geographies, and negotiated pricing.
Does OpenAI also use Google Cloud?
Yes. In June 2025, OpenAI incorporated Google Cloud to expand capacity, as reported by Reuters. No specific figures were announced, but the collaboration fits into their multicloud strategy.
What about Oracle? Fact or rumor?
OpenAI has been collaborating with Oracle since 2024 within the Azure on OCI scheme (more Azure capacity hosted on Oracle). Additionally, in September 2025, WSJ and others reported a $300 billion five-year deal starting from 2027. Oracle didn’t confirm details; if true, it would be one of the largest cloud contracts in history.