By Javier Cuervo
Monday, 7:12 AM. An alert goes off on your mobile: OpenAI could be moving part of its operations out of Microsoft Azure . If confirmed, it would not be a sharp turn, but a natural phase of maturity: when your demand for computing grows faster than supply, you open up alternatives. I see it often: first you choose a provider for speed; then you diversify by resilience, cost and bargaining power.
To understand it, you have to look back. OpenAI was born in 2015 as a non-profit organization. In 2019 he created the "capped-profit" structure (OpenAI LP) to attract capital while maintaining control in the non-profit entity. That same year, Microsoft announced an investment of 1,000 million and put Azure as the preferred provider. In 2020, Microsoft obtains the exclusive license for GPT-3. In January 2023, Microsoft announces a new "multi-year and multimillion-dollar" investment and that Azure will continue to be the cloud provider exclusive for your primary uploads. In November 2023 comes the biggest governance crisis: board removes Sam Altman , the team mobilizes, Microsoft offers to incorporate him, and a few days later Altman returns as CEO. The board is recomposed with Bret Taylor (chairman), Larry Summers and Adam D'Angelo, and Microsoft gets a non-voting observer seat. In March 2024, following an external investigation, OpenAI announces governance improvements, Altman's return to the board, and the addition of independent profiles such as Nicole Seligman, Sue Desmond-Hellmann, and Fidji Simo. That is the context: Microsoft capital, cloud exclusivity declared in 2023, and reinforced governance but with tensions between mission, scale and market.
Let's go back to this week of November 2025: Monday, 10:30. If the specialized press is anticipating that OpenAI Evaluates AWS for part of their compute, the first question from large customers is operational: latency, availability, and SLAs. Which operations would you move first? Obviously those that allow greater tolerance to change: peaks of inference, experimentation, peripheral services. Large-scale training usually comes later,... if it goes.
Monday, 1:10 p.m. Technical reasons, no mystery.
1) Capacity: The demand for GPUs continues to outstrip the supply.
2) Cost per token: If you can run some of the inference on alternative hardware (e.g., AI-designed chips like Inferentia/Trainium on AWS) with a lower TCO, your P&L notices it.
3) Resilience: multicloud reduces the risk of concentration.
4) Compliance: More locations, more options for regulatory requirements.
Monday, 5:45 p.m. Strategic motive: bargaining power. A buyer who concentrates all his demand on a single supplier loses the ability to improve his conditions. Conversely, diversifying gives you better prices, better capacity reservations, and sometimes priority access to new generations of hardware. It is not "breaking" with Microsoft; it is managing dependency.
Tuesday, 8:20 a.m. Does this change OpenAI's governance? In principle, no. The structure remains the same: the non-profit entity controls; LP raises capital with a return cap; The management executes. Deciding to move operations between clouds is typically the responsibility of management, not the board, unless it affects material contractual commitments. If there are exclusivity or minimum spend clauses with Azure, any sustained and significant changes would open a renegotiation table. But that is commercial negotiation, not redesign of the mission.
Tuesday, 12:05. And the relationship with Microsoft? The strategic alliance is still there: integration of models into products (Copilot, M365, Bing), co-development of infrastructures and a relevant financial relationship. OpenAI exploring AWS doesn't end there. The adjustment is contractual-operational: capacity reserves, tiered pricing, deployment priorities, and perhaps new limits on what is executed where and with what deadlines.
Wednesday, 11:00. Fine technique. The training of new models is likely to continue on the latest generation NVIDIA GPUs due to ecosystem and maturity. Inference, on the other hand, supports greater diversity: per-model optimizers, specific kernels, model servers that abstract hardware, and paths for specialized chips. The desirable outcome: declining cost per 1,000 tokens, stable latency, and more resilience.
Thursday, 16:00. Effect on the cloud market. If OpenAI breaks ground, it validates AWS's commitment to proprietary chips and reinforces the "true multicloud" thesis for generative AI at scale. Microsoft, which already invests in dedicated supercomputing for AI on Azure, will accelerate capacity commitments, supply agreements, and more integrated platform services. Google is left in an ambivalent situation: a great provider of AI and infra, but a direct competitor in models.
Friday, 9:00. What changes for the business customer? Three things:
1) Portability standards (serving and orchestration) matter more than ever.
2) Cloud contracts with real elasticity clauses, not just credit.
3) Metrics that control business, not just technical: cost per useful conversation/response, P95 latency, drop rate, multi-region availability.
What do I expect to happen, if this is confirmed:
- Pragmatic multicloud: Azure remains the primary partner; AWS gains a significant but minority portion of loads, starting with inference and spikes.
- Training mostly in NVIDIA in the short term; More experimentation on alternative hardware for inference.
- Contractual change between Microsoft and OpenAI: more capacity committed to Azure in exchange for price and priority; limits and usage windows outside of Azure.
- No change in the advice for this reason. Yes, more focus of the risk/security committees on technological dependence and concentration of suppliers.
- The most important thing , the example : More actors AI-first They adopt selective multicloud strategies, prioritizing cost per token, resiliency, and compliance.
What I would do as CEO/CIO today, regardless of headlines (and what we will start doing this week in Proportione):
- Negotiate multi-year capacity reservations with at least one provider and credible options with a second.
- Design the AI platform (serving, observability, feature store, vector DB) with hardware and cloud abstraction from day one.
- Measure business, not just technical: cost per useful result, not per GPU/hour.
- Prepare an AI-specific business continuity plan: functional degradation, queuing, caching, and fallback between vendors.
- Review exclusivity clauses and MFN in contracts; that the technical strategy is not trapped by paper.
This is, in essence, about operational independence without breaking alliances. OpenAI has to secure capacity, lower unit costs and gain flexibility. Microsoft wants to maximize adoption on its platform and protect the relationship. AWS is looking to validate its AI stack at scale. None of these objectives are incompatible. Equilibrium is traded in watts, tokens, and SLAs, not holders.
— Javier Cuervo