OpenAI’s ChatGPT is no longer tethered to Microsoft Azure alone. An updated sub-processor list confirms that Google Cloud now handles enterprise and API workloads in the United States, Japan, the Netherlands, Norway, and the United Kingdom—ending an exclusivity era that defined the AI startup’s infrastructure since 2019. The decision, driven by insatiable GPU demand and a need for hardware diversity, adds Google’s Tensor Processing Units (TPUs) to a roster that already includes Azure, CoreWeave, and Oracle. For Microsoft, it’s a strategic blow but not a severance; the company retains a right of first refusal for additional capacity and deep IP access. For Google, it’s the most high-profile validation yet of its custom silicon story.

The move was telegraphed months ago when Microsoft and OpenAI renegotiated their partnership in January 2025, replacing blanket exclusivity with a more nuanced framework. At the time, OpenAI gained explicit permission to seek outside compute, while Microsoft secured a guarantee that it would get first dibs on any new capacity. Now, the sub-processor list makes it official: Google Cloud Platform is listed alongside Microsoft Azure, CoreWeave, and Oracle for ChatGPT Enterprise, Edu, Team, and the API. That’s more than a paperwork update—it’s a blueprint for how leading AI labs will source compute in an era of scarcity.

The Compute Crunch That Forced the Shift

Large language models consume staggering amounts of compute. Training a frontier model can require tens of thousands of GPUs running for months; inference at scale adds a continuous, global load. OpenAI publicly acknowledged compute constraints as a brake on product velocity throughout 2024. Sora’s launch delays, rolling feature gating, and occasional degraded service modes all pointed to a single bottleneck: not enough GPUs, fast enough, in the right places.

Azure struggled to keep pace. Despite Microsoft’s massive data center investments, the explosive growth of ChatGPT Enterprise and new consumer features strained capacity. Adding Google Cloud and CoreWeave distributes that load across multiple hyperscalers and specialty providers, so a capacity hiccup in one region no longer cripples the entire fleet. The strategy mirrors what the biggest cloud-native companies already do: treat compute as a commodity and optimize for cost, latency, and hardware fit.

Google’s TPU Advantage: More Than Just Capacity

Google Cloud’s most distinctive asset is the TPU, a custom accelerator designed specifically for deep learning workloads. TPUs have long powered Google’s own models—Bard, ImageNet training, and internal recommendation systems—but OpenAI’s adoption signals a broader industry shift. TPUs can offer cost and energy-efficiency benefits for dense transformer architectures, particularly during large-matrix multiplications that dominate training and inference. OpenAI will likely target specific workloads to TPUs: pre-training runs, large-batch inference, or fine-tuning where the compiler can optimize the kernel graph.

This isn’t a wholesale GPU replacement. Nvidia’s H100s and upcoming B200s still dominate for many mixed-precision and sparsity patterns. But by mapping jobs to the best accelerator, OpenAI can shave millions in compute costs and reduce time-to-result for critical experiments. Early cross-cloud testing between TPU v5p and H100 clusters reportedly showed promising parity on throughput per dollar for certain model sizes, though official benchmarks remain scarce. The engineering effort—rewriting kernels, adapting numerical formats, and building a scheduling layer that routes to the optimal silicon—is substantial, but the payoff in hardware agility is hard to overstate.

Geographic Reach and Data Sovereignty

The sub-processor list reveals a geographic logic behind the Google Cloud deployment. The five countries—US, Japan, Netherlands, Norway, and UK—map onto regions with strong enterprise demand and strict data-residency requirements. Japan and the UK, for instance, have large financial services and government sectors that insist on local processing. Norway’s inclusion hints at a Nordic play, possibly linked to renewable energy credits and growing AI regulation. By standing up inference endpoints in these regions, OpenAI reduces latency for local users and checks compliance boxes that were previously hard to satisfy with a single cloud.

This multi-region posture also de-risks the business from geopolitical shocks. If a single cloud’s data center in a contentious region faces policy changes, OpenAI can shift traffic to an alternate provider in a friendly jurisdiction. The days of a monolithic, US-centric cloud footprint are over for any global AI service.

Microsoft’s New Reality: From Exclusive to Preferred

Microsoft remains OpenAI’s most important commercial partner. The $13 billion investment, the deep technology integration across Copilot and Azure OpenAI Service, and the revenue-sharing agreement all endure. What changed is exclusivity. In January, the companies formalized a right of first refusal: whenever OpenAI seeks additional compute capacity, it must offer Microsoft the chance to supply that capacity on competitive terms before turning to others. Microsoft also retains privileged access to OpenAI’s intellectual property and can use it in its own products without fear of competition from the API.

Still, the optics are painful. For years, Satya Nadella touted OpenAI as Azure’s crown jewel. Now, Google Cloud—an archrival in AI and cloud—is powering the very service millions use daily. Microsoft’s response will likely be twofold: accelerate its own custom silicon (Maia AI accelerators) to match TPU economics, and sweeten commercial terms to keep new workloads on Azure. The company has already broken ground on dedicated AI data center campuses, and its capital expenditures continue to soar. The loss of exclusivity is a wake-up call that even the deepest partnership can’t guarantee sole-provider status when demand outstrips supply.

CoreWeave and Oracle: Specialists Fill the Gaps

CoreWeave, a specialized GPU cloud provider, inked a deal with OpenAI in March 2025 worth up to $11.9 billion over five years, according to its own press release. That contract, which also gave OpenAI an equity stake, provides dedicated H100 clusters with predictable pricing and rapid scaling. Unlike hyperscalers, CoreWeave doesn’t intermix customer workloads on shared infrastructure; its entire value proposition is raw, high-performance GPU compute. For OpenAI, it’s an escape valve from timeshare-like cloud provisioning and a way to lock in capacity amid supply-chain chaos.

Oracle brings sovereign and enterprise-grade regions to the mix, particularly in markets where it has deep government and financial services roots. While Oracle’s AI infrastructure is smaller than the big three’s, its existing relationships help OpenAI land deals that require specific certifications and on-premises connectivity.

Technical Hurdles: Orchestration and Model Portability

Running the same service across four providers is not trivial. OpenAI must now operate a sophisticated orchestration layer that can:
- Route training jobs to the most cost-effective cluster based on real-time pricing and availability.
- Manage inference traffic globally, shifting between regions and accelerators to meet latency SLAs.
- Maintain consistent performance and security posture across different network fabrics and storage systems.

This requires substantial investment in multi-cloud networking, unified logging, and cross-cloud storage abstractions. The team has likely adopted a containerized, Kubernetes-based approach with service meshes and global load balancing—practices honed by hyperscalers themselves. Still, the operational burden is real. Configuration drift, inconsistent monitoring, and cross-provider troubleshooting can eat into the very savings multi-cloud promises.

Model portability adds another layer. TPUs and GPUs have different memory hierarchies and preferred data layouts. Code optimized for CUDA must be re-tooled for Google’s XLA compiler. While frameworks like JAX and PyTorch XLA bridge the gap, production-grade inference often requires hand-tuned kernels. OpenAI’s infrastructure teams are likely building abstraction layers that allow model graphs to be compiled for multiple backends with minimal human intervention. The industry will watch closely: if OpenAI makes TPU-GPU portability look easy, it will accelerate multi-cloud adoption across the board.

Data Governance and Security Risks

With more providers comes more surface area for data exposure. OpenAI handles sensitive enterprise prompts, some containing proprietary code, legal documents, or personal data. Adding Google Cloud means model weights and data may traverse infrastructure managed by a direct AI competitor. The contractual protections must be airtight: encryption at rest and in transit, customer-managed keys, and strict data isolation. OpenAI’s updated Data Processing Agreements (DPAs) likely mandate that Google Cloud processes data only for the specified services and deletes it promptly.

Moreover, regulatory complexity multiplies. The EU’s AI Act, Japan’s APPI, and sector-specific rules like HIPAA in the US all impose overlapping obligations. Each new region means one more set of audits, data-transfer impact assessments, and breach-notification timelines. Enterprise customers evaluating ChatGPT Enterprise should scrutinize the sub-processor list and demand evidence of compliance controls for each provider and jurisdiction.

Enterprise Takeaways for IT Leaders

The OpenAI multi-cloud play offers lessons for any organization building or buying AI infrastructure:
- Audit your sub-processors. If you use ChatGPT Enterprise, check the sub-processor list to know exactly where your data flows. Negotiate DPAs that align with your data classification policies.
- Embrace multi-cloud as a negotiating tool. The mere threat of shifting workloads to another provider can unlock capacity commitments and better pricing. Even if you stay single-cloud, maintain a contingency plan.
- Plan for heterogeneity. Whether you deploy open-source models or consume APIs, assume you’ll eventually need to support multiple accelerator types. Invest in portable training pipelines and inference stacks.
- Double down on governance. Multi-provider architectures demand unified identity, secrets management, and observability. Treat your AI infrastructure as a distributed system from day one.

The Cloud Market Ripple Effect

OpenAI’s decision will reverberate through the cloud industry. Hyperscalers will intensify their custom silicon races: AWS has Trainium and Inferentia; Google has TPU; Microsoft has Maia. Nvidia’s grip remains firm, but no AI company wants to be locked into one chip architecture anymore. The era of homogeneous GPU clouds is giving way to a heterogeneous mesh of accelerators, each optimized for different parts of the AI lifecycle.

We’ll also see more lab-to-lab multi-cloud deals. Anthropic, Cohere, and Mistral already operate across multiple clouds, but OpenAI’s move normalizes it at the largest scale. Cloud providers, in turn, will segment their offerings: premium services with tight model integration, and commodity compute for portable workloads. Microsoft’s competitive moat—Copilot, Azure OpenAI Service, the developer toolchain—now matters more than ever, because raw compute can be sourced elsewhere.

For Google Cloud, the win is psychological. Landing OpenAI validates years of TPU investment and its network’s reliability. It also blurs the line between provider and competitor: Google Cloud serves OpenAI the same month Google DeepMind launches a rival model. Such symbiotic rivalry will become the norm.

A Forward Look

OpenAI’s multi-cloud architecture is not a temporary fix; it’s a structural response to the physics of AI at scale. By weaving together Microsoft Azure, Google Cloud, CoreWeave, and Oracle, the company gains resilience, hardware optionality, and regulatory dexterity—at the cost of heightened complexity. The coming years will reveal whether the operational overhead pays off or whether it births a new breed of multi-cloud management tools that the rest of the enterprise world can adopt. Either way, the message is clear: in AI, no single cloud is big enough anymore.