For three years, the answer to “where do I run OpenAI in production?” had exactly one syllable: Azure. That ended Monday. On April 27, 2026, OpenAI and Microsoft announced a renegotiated partnership that strips Azure of its exclusive right to host OpenAI models, and within hours GPT-5.4 was live in limited preview on AWS Bedrock, with GPT-5.5 announced for preview within weeks.

This is the single biggest shift in the enterprise AI buying map since OpenAI signed the original Microsoft deal in 2019. Azure’s most durable competitive advantage — being the only cloud where you could call OpenAI natively — is gone. AWS Bedrock just absorbed the frontier model that Azure built its entire AI strategy around.

Quick Summary: What Changed on April 27

Detail	Info
Date	April 27, 2026
What ended	Azure’s exclusive right to host OpenAI models
What launched	GPT-5.4 in limited preview on AWS Bedrock; GPT-5.5 announced for preview within weeks; Bedrock Managed Agents powered by OpenAI
What Microsoft kept	Non-exclusive IP license through 2032; primary cloud partner status; revenue share capped through 2030
What OpenAI gained	Right to sell across any cloud — AWS, Google Cloud, others
Official sources	OpenAI blog · Microsoft blog

Bottom line: Azure stays Microsoft’s home court but loses its monopoly on the highest-volume frontier model in enterprise. The buyers who win are the ones whose stack is multi-cloud or AWS-native — they get OpenAI without renting Azure to do it.

---

What Actually Happened on April 27

Both companies posted near-simultaneous announcements on their official blogs — OpenAI’s titled “Evolving our partnership with Microsoft” and Microsoft’s framing it as a “next phase”. The two posts were nearly identical in scope, which is the thing you notice if you read both: this was a coordinated release, not a leaked breakup.

The terms, as both companies described them:

• Azure exclusivity is over. OpenAI can now sell its models (including GPT-5.5, GPT-5.4, and future releases) to any cloud provider. AWS and Google Cloud are explicitly named as eligible distribution partners.
• Microsoft retains a non-exclusive IP license through 2032. Microsoft can keep using OpenAI’s model weights, research, and API across its products (Copilot, Azure AI Foundry, GitHub) without paying per-call rent.
• Revenue share is capped through 2030. OpenAI continues to share a percentage of revenue with Microsoft, but that obligation has a hard end date and a defined dollar cap. Coverage from Bloomberg and CNBC put the cap in the tens of billions, but not unbounded.
• Microsoft remains OpenAI’s primary cloud partner. This is the line both blog posts repeated. Most OpenAI training and serving still runs on Azure. Most net-new Azure capacity OpenAI commissions over the next decade is still earmarked Azure-first. The two companies are still very married — they just stopped being monogamous on the customer-facing side.

Within hours, AWS confirmed GPT-5.4 was live in limited preview on Bedrock, with GPT-5.5 announced for preview within the following weeks and general availability for both targeted for late Q2. AWS also announced Amazon Bedrock Managed Agents powered by OpenAI: a managed agent runtime that uses GPT-5.4 as its initial default reasoning model, sitting inside the existing AgentCore framework.

Google Cloud has not yet shipped OpenAI models in Vertex AI Model Garden, but the partnership terms now allow it. The question for Google is whether it wants to host a competitor’s flagship the way it already hosts Claude as a first-class option. My read: yes, eventually, but Google has its own Gemini story to protect first.

Why This Matters More Than the Cash

The headline writers wanted to make this a Microsoft loss story. Microsoft’s stock dipped briefly on the news before closing up on Tuesday. That’s the wrong frame.

What actually changed is the buying motion for every enterprise that wasn’t already on Azure. Three years of “we’d love to use GPT-4 but we’re an AWS shop and the latency from Azure is killing us” just dissolved. Every multi-cloud enterprise AI committee that paused a deployment because OpenAI lock-in was a procurement red flag now has a green light.

The quiet number to watch is GPT-5’s enterprise penetration outside Azure. Roughly 60% of Fortune 500 cloud spend runs through AWS. If even a third of that base wanted OpenAI but wouldn’t move workloads to Azure to get it, that’s a meaningful chunk of demand that was structurally suppressed. It just got unsuppressed.

The deeper shift: Azure stops being a model platform and goes back to being a cloud. For two years, the answer to “why Azure?” included “because OpenAI.” Strip that out and Azure has to compete on the same dimensions as AWS and Google Cloud — pricing, region coverage, services, network performance. Those are dimensions where AWS has historically led and Google has been catching up.

What Microsoft Actually Lost (and Kept)

Microsoft did not get hollowed out by this deal. The press is overplaying that angle.

What Microsoft kept that matters:

• The IP license through 2032 means Microsoft can keep training Copilot on OpenAI weights, keep running GPT-5 inside Azure AI Foundry, and keep shipping OpenAI-powered products without paying per-call rent. This is the underrated half of the deal.
• The primary cloud partner status means most new OpenAI capacity gets built on Azure first. AWS Bedrock gets the served version; Azure gets the trained version and the priority lane.
• The revenue share through 2030 is still real money. Capped, but real.
• Microsoft’s MAI hedge — its own internal model line — was already in motion before this deal. The exclusivity ending is not a surprise to Microsoft; it’s a milestone they planned for.

What Microsoft lost that matters:

• The single biggest reason for an enterprise to pick Azure over AWS for AI workloads is gone. That reason was structural, not feature-based. You can’t replace it with a UI improvement.
• Pricing leverage on GPT pass-through. Azure used to charge a premium to host OpenAI calls because Azure was the only place to host them. That premium evaporates the moment AWS Bedrock and (eventually) Google’s Vertex AI offer the same models.
• The narrative grip on “the OpenAI cloud.” Azure spent three years being narrated as the place where the frontier lives. As of Monday, the frontier rents space across town too.

The cleanest analogy is what happened to BlackBerry Enterprise Server when Exchange ActiveSync became universal. BlackBerry kept selling devices and software for years — but the structural advantage of being the only way IT could deploy enterprise email to mobile was gone, and the rest of the company’s positioning had to absorb that loss in slow motion.

What This Looks Like on AWS Bedrock

If you’ve used Bedrock before, the ergonomics are exactly what you’d expect. GPT-5.4 (with GPT-5.5 to follow within weeks) shows up as a model choice in the Bedrock console alongside Claude Opus 4.7, Claude Sonnet, Anthropic Haiku, Meta’s Llama, AI21, Cohere, and the AWS-native Nova models. You pick the model, you pick the version, you wire it up. IAM and VPC endpoints behave the same way they do for Claude.

What’s different: Bedrock Managed Agents now have an OpenAI-default mode. The framework that AWS launched at re:Invent 2025 — which previously assumed Claude as the default reasoning engine — can now spin up agents using GPT-5.4 (or GPT-5.5 once it reaches preview) with no code changes other than the model parameter. AWS is positioning this as a head-to-head capability test against Google’s Gemini Enterprise Agent Platform and Microsoft’s Azure AI Foundry agent runtime. The pitch: same agent framework, your choice of model.

What’s not different: Bedrock pricing tracks the underlying model provider’s pricing. GPT-5.5 input/output rates on Bedrock match OpenAI’s direct API rates plus AWS’s standard infrastructure margin. Don’t expect AWS to undercut OpenAI’s own pricing.

For teams that have been running multi-cloud Claude on Bedrock for cost or compliance reasons, this changes nothing about Claude. It just means the same Bedrock setup now lets you A/B GPT against Claude inside the same console, the same IAM, the same audit log. That’s the operational improvement most engineering leads will care about.

How to Decide Where to Run OpenAI Now

The right answer depends on what your stack already looks like. Three quick decision frames:

If you’re already on Azure and OpenAI works: Stay. The IP license through 2032 means Azure continues to ship OpenAI models inside Azure AI Foundry with full feature parity and priority access to new releases. There is no strategic upside to migrating mature workloads off Azure just because the option exists.

If you’re AWS-native and were running Claude as a forced compromise: Re-evaluate. The mix between Claude and GPT-5.5 is now a genuine model-fit decision rather than a procurement constraint. The GPT-5.5 vs Claude Opus 4.7 split shows the two models genuinely diverge on coding accuracy, instruction following, and long-context behavior. Pick on capability, not on which cloud hosts what.

If you’re building new and have no incumbent cloud: AWS Bedrock just became the most flexible AI cloud for procurement. Claude as primary, GPT-5.5 as secondary, Llama and Nova as fallbacks, all under one IAM domain and one billing account. Microsoft’s pitch for net-new workloads narrows to “Microsoft 365 deep integration” — which is real but doesn’t outweigh the procurement flexibility AWS now offers.

The tradeoff worth naming: Azure still has the deepest agent integration with Microsoft 365, Office, and the Copilot suite. If your end users live in Word, Excel, Outlook, and Teams, the Microsoft Copilot deployment story keeps Azure first. If your end users live in your own applications, the Bedrock or Vertex AI path is now genuinely competitive.

Why Did OpenAI Push for This?

A reasonable question: why would OpenAI break a deal that delivered three years of revenue stability?

Three reasons line up.

Distribution math. OpenAI’s ARR roughly doubled in 2025 and is on pace to grow again in 2026. The Azure-only constraint means every dollar of net-new enterprise revenue had to come through Microsoft’s sales motion. That’s a ceiling. AWS sells AI to a different buyer set with different timelines, and Google Cloud sells to a third. Three distribution channels beat one — even if Microsoft was OpenAI’s biggest single channel by far.

Microsoft’s hedge made the exclusivity expensive. Microsoft has been publicly building MAI as an internal model line and shifting Copilot workloads to it where economics favor it. From OpenAI’s side, exclusivity was always a two-sided deal — Microsoft promised to be the cloud of choice, OpenAI promised to be the model of choice. When Microsoft started favoring its own models for Copilot, the math behind exclusivity wobbled. Renegotiating let OpenAI capture the upside Microsoft was no longer guaranteeing.

The Anthropic mirror. Anthropic structurally locked in distribution across two hyperscalers over the last two weeks — Amazon and Google. That gave Claude pricing leverage and a narrative advantage in any procurement conversation that mentioned “model availability.” OpenAI watching that and concluding “we need the same option” is the obvious move. Frontier labs are converging on multi-cloud as the default, not the exception.

What’s striking is that the OpenAI–Microsoft revenue share through 2030 is the kind of clause that, ten years ago, you’d assume guaranteed monogamy. In 2026 it apparently doesn’t. The cap and the expiration are what make the deal restructurable.

What Enterprise Buyers Should Do This Week

Specific, actionable, not generic.

1. Re-open any procurement decision you froze on OpenAI lock-in concerns. The single biggest blocker is gone. If you tabled a GPT-5 deployment in Q1 because your CISO didn’t want to add Azure as a fourth cloud, that conversation just got simpler. Pull it back off the shelf.

2. Audit your Copilot vs ChatGPT Enterprise spend. A lot of mid-market companies have been buying both Microsoft 365 Copilot and ChatGPT Enterprise because the Copilot integration didn’t quite cover their use cases. Now that GPT-5.5 is on Bedrock, some of those workloads can consolidate to a single model accessed through your existing AWS contract — instead of two separate per-seat licenses.

3. Stop overpaying for “Azure premium” on OpenAI inference. AWS Bedrock pricing for GPT-5.5 and GPT-5.4 is published and matches OpenAI’s direct rates. If your Azure contract embeds GPT calls at a rate higher than that — and many do, for legacy reasons — your renewal conversation has new leverage.

4. Plan for model-portable architectures. The lesson of the last 18 months is that frontier model availability changes faster than enterprise procurement cycles. Build agent and pipeline code that abstracts the model choice — same prompt structure, swappable model identifier. This is what the enterprise AI deployment guide recommends as the default architectural posture, and it’s even more relevant after this week.

5. Don’t overreact to Microsoft’s stock move. Microsoft is fine. The IP license through 2032 is the load-bearing piece of the new deal, and Microsoft retained it. Read the Microsoft blog post all the way to the bottom — the part where they emphasize they’re still the primary cloud partner and that Copilot continues to ship with OpenAI underneath. Anyone telling you Microsoft is the loser of this deal is selling a hot take.

How the Three-Hyperscaler AI Map Looks Now

	Azure	AWS Bedrock	Google Cloud (Vertex)
Default frontier model	GPT-5.5 (OpenAI)	Claude Opus 4.7 (Anthropic)	Gemini 3.1 Pro (Google)
Other frontier models	Anthropic via Azure AI; GPT family	OpenAI GPT-5.5/5.4 (preview); Anthropic Claude family; Meta Llama; Cohere; AI21; Nova	Anthropic Claude family first-class; 200+ models; OpenAI not yet shipped (terms now allow it)
Native agent runtime	Azure AI Foundry + Copilot Studio	Bedrock Managed Agents (now OpenAI-powered)	Gemini Enterprise Agent Platform
Office/productivity tie-in	Strongest (Microsoft 365 Copilot)	None native	Workspace, weaker than Microsoft 365
Primary advantage	Microsoft 365 integration, IP license access	Most model-flexible cloud	Most model-neutral platform brand
Primary disadvantage	Lost OpenAI exclusivity	No native productivity suite	Gemini-vs-Claude internal tension

The cleanest split that wasn’t true two weeks ago: AWS is now the most flexible cloud for frontier model choice. Google is the most model-neutral platform. Azure is the most Microsoft-integrated. None of them is dominant — and the procurement question for the next 18 months is which of those three positions matches your stack.

For the broader picture on how the frontier labs are landing across these clouds, the Anthropic vs OpenAI 2026 comparison walks through model-level differences. The enterprise AI procurement guide and Amazon’s $200B AI infrastructure spend cover the cloud-vendor competitive context that this week’s news lands on top of.

Our Take

This is the deal that was always coming, and the only surprise is the timing.

The OpenAI–Microsoft exclusivity was a 2019 arrangement built for a world where neither company knew if generative AI would be a real business. It was a real business by 2023. By 2025, the exclusivity was a structural drag on both sides — capping OpenAI’s distribution and forcing Microsoft to fund a hedge it hadn’t planned. Renegotiating in April 2026 was the orderly version of an unwind that would have happened messier in 2027.

The enterprise buyer is the unambiguous winner. For three years, “where do I deploy OpenAI?” had one answer. Now it has three. That doesn’t make any one cloud worse. It makes all three better, because the procurement question is finally a real procurement question.

The harder question is what it does to the model layer. If Anthropic, OpenAI, and Google are all available across all three hyperscalers within 18 months — and the trajectory says they will be — then the model becomes a commodity choice and the cloud becomes the platform. That’s a worse business for OpenAI than what they had on Monday. It’s a much better business for AWS, who is best positioned to be the model-neutral platform of choice.

Microsoft’s response will not be to fight the unbundling. They’ve already started (MAI, Copilot’s deeper integration, the Office moat). The interesting question is whether OpenAI’s response is to lean harder into being a consumer product company — where ChatGPT, not the API, is the thing — and treat the enterprise API business as commodity infrastructure. That’s a real possibility, and it’s the version of OpenAI that would scare Anthropic the most.

For now: if you’ve been waiting for OpenAI on AWS, it’s there. If you’ve been on Azure and it works, stay. If you’re new and you want optionality, Bedrock just became the easiest place to start.

Frequently Asked Questions

When did OpenAI’s Azure exclusivity end?

OpenAI and Microsoft announced the renegotiated partnership on April 27, 2026, in coordinated posts on the OpenAI and Microsoft official blogs. The exclusivity ended the same day the agreement was announced.

Is OpenAI on AWS Bedrock now?

Yes, as of April 27, 2026. GPT-5.4 is available in limited preview on AWS Bedrock immediately, with GPT-5.5 announced for preview within weeks and general availability for both targeted for late Q2. Amazon Bedrock Managed Agents powered by OpenAI launched alongside the model preview as a managed agent runtime that uses GPT-5.4 as its initial default reasoning model.

What did Microsoft keep in the new deal?

Microsoft retained a non-exclusive IP license to OpenAI’s models, research, and technology through 2032; primary cloud partner status; and a capped revenue-share arrangement through 2030. Microsoft can continue to use OpenAI’s models inside Copilot, Azure AI Foundry, and other products without per-call licensing, and most net-new OpenAI capacity is still expected to be built on Azure first.

Will GPT-5.5 launch on Google Cloud Vertex AI?

The new partnership terms allow it, but Google has not yet shipped OpenAI models in Vertex AI Model Garden. Google currently ships Anthropic Claude as a first-class option in Model Garden but has its own Gemini family to protect, so an OpenAI launch on Vertex is plausible but not announced as of April 29, 2026.

Does this lower OpenAI API pricing?

Probably not directly. AWS Bedrock pricing for GPT-5.5 and GPT-5.4 matches OpenAI’s published direct API rates plus AWS’s standard infrastructure margin. The bigger pricing impact is indirect: Azure can no longer charge a premium for being the only cloud where OpenAI runs, which removes structural pricing leverage Microsoft had on enterprise OpenAI deployments.

Does Microsoft Copilot still use OpenAI models?

Yes. Microsoft’s IP license through 2032 covers using OpenAI’s models in Microsoft products, including Copilot. The renegotiation does not unwind Copilot from OpenAI; it removes Azure’s exclusive right to host OpenAI as a third-party API for other companies. Microsoft has also been building its own MAI model family as a hedge, and Copilot increasingly routes between OpenAI and MAI depending on workload.

Is this worse for Microsoft or OpenAI?

Neither company comes out clearly ahead, which is a sign the deal was negotiated honestly. Microsoft loses a structural Azure advantage but keeps deep IP rights and primary partner status. OpenAI gains distribution flexibility but gives up a guaranteed revenue floor. The clearest winner is the enterprise buyer — model availability across multiple clouds is now a real procurement option for the first time.

Should I migrate my OpenAI workloads from Azure to AWS Bedrock?

Probably not yet, if your Azure deployment is working. The IP license through 2032 means Azure continues to ship OpenAI models with full feature parity. The Bedrock launch is a procurement option for new workloads or for teams that wanted OpenAI but were blocked from adopting Azure. Migration of existing workloads is a heavier lift than the savings usually justify in the short term.

How does this compare to Anthropic’s multi-cloud strategy?

Anthropic deliberately structured itself across two hyperscalers from early on — Amazon and Google. The recent $40B Google investment and the parallel $25B Amazon commitment locked in 10 GW of dedicated compute across both clouds. OpenAI is now pursuing a similar multi-hyperscaler distribution model, but with Microsoft still as the dominant primary partner. Anthropic’s split is more even; OpenAI’s is more weighted toward one partner.

---

Last updated: April 29, 2026. Sources: OpenAI — Evolving our partnership with Microsoft · Microsoft — Microsoft and OpenAI evolve partnership · AWS — OpenAI models now available in preview on Amazon Bedrock · Bloomberg — OpenAI and Microsoft restructure landmark partnership · CNBC — OpenAI, Microsoft partnership revenue cap and restructure.

Related reading: Google’s $40B Anthropic Bet · Anthropic vs OpenAI 2026 · Microsoft MAI: The OpenAI Hedge · GPT-5.5 vs Claude Opus 4.7: Coding Showdown · Google Cloud Next 2026: Agents Are the New OS · Enterprise AI Deployment Guide · Microsoft Copilot Cowork Review