April 2026 · AI News · GoCodeLab

Microsoft-OpenAI Exclusivity Is Over — AWS Picked Up OpenAI in 24 Hours

On April 27, 2026, a quiet announcement went up. Microsoft was giving up its status as OpenAI's "sole cloud provider." A seven-year exclusive partnership was effectively over. The AI industry started buzzing the next day.

On April 28, AWS moved. It added OpenAI's GPT-4o, GPT-4.1, and o3 to Bedrock. The speed of the announcement was proof that preparations had already been finished. OpenAI models that had only been available on Azure were running on AWS within a single day.

That same week, Google finalized a $40B investment in Anthropic. SoftBank announced it would pour trillions of yen into Japan's AI infrastructure. In one week, the AI infrastructure alliance map changed completely. This article is my attempt to lay out what actually changed for developers and how to respond.

The Day Seven Years of Exclusivity Ended — What Happened

In July 2019, Microsoft invested $1B in OpenAI. It wasn't a simple investment. A contract bundling all of OpenAI's commercial APIs exclusively through Azure came with it. With additional investments of $2B in 2021 and $10B+ in 2023, the relationship deepened further. The result was that anyone wanting to use GPT-series models in the cloud had exactly one option: Azure.

Microsoft reaped enormous benefits from this structure. Azure OpenAI Service revenue grew at double-digit rates every year. After GPT-4 launched in 2023, Azure revenue rode the AI wave and hit record highs quarter after quarter. Developers had no choice. If you wanted OpenAI models, you had to use Azure. Using the OpenAI API directly was an option, but enterprises needing SLAs, compliance, and formal contracts were effectively forced through Azure.

On April 27, 2026, the structure changed. Microsoft and OpenAI restructured the partnership. Microsoft remains OpenAI's largest shareholder, but the "sole cloud provider" exclusivity clause was removed. From Microsoft's perspective, it wasn't a bad deal either. With OpenAI's valuation having risen high enough, lifting the exclusivity to let OpenAI reach more customers could grow the overall pie. From an investment return standpoint, growth could outperform exclusivity.

The relationship with Azure itself continues. Microsoft still operates Azure as OpenAI's primary infrastructure partner. What changed is the exclusivity. OpenAI can now enter direct supply agreements with AWS, Google Cloud, and other clouds. The single-cloud pipeline that had held for seven years ended that day. Turning points in the AI industry usually start with exactly this kind of quiet announcement.

AWS Moved in 24 Hours

On April 28 — the day after the Microsoft-OpenAI announcement — AWS published a Bedrock update. The timing was not a coincidence. AWS had already been in behind-the-scenes negotiations with OpenAI and had been waiting for the exclusivity announcement. GPT-4o, GPT-4.1, and o3 appeared in the Bedrock model catalog. The launch region was us-east-1.

The speed of the announcement was itself the message. Normally, adding a new model to Bedrock takes weeks of evaluation, safety review, and infrastructure preparation. Going live in a single day meant preparations had been completed well before. AWS was waiting for this moment. Amazon CEO Andy Jassy stated in an internal interview that his goal was for every major AI model to run on Bedrock. The last blank in that goal was OpenAI.

The integration method is the same as existing Bedrock. You call invoke_model with boto3's bedrock-runtime client. Auth is IAM-based. No API key management needed. Anyone with an AWS account can start using it immediately after granting IAM permissions. Existing code that called Claude or Llama through Bedrock can be swapped over to OpenAI models with the same structure.

AWS's calculation is clear too. Bedrock had been Claude (Anthropic), Llama (Meta), and Titan (Amazon) heavy — adding OpenAI reduces the chance of customer churn. Customers who had gone to Azure specifically to use GPT-4o can now get what they need within AWS. It's a defensive strategy to capture demand that was leaking to competing clouds, and an offensive strategy to attract new customers at the same time.

Three OpenAI Models Added to Bedrock

GPT-4o is the general-purpose model. Text generation, image understanding, and audio processing all in one. It's the best fit for services that need chatbots, document summarization, and image analysis together. Vision works the same on Bedrock — pass images as base64-encoded data. It has the best cost efficiency in the GPT-4 series, making it the default choice for high-traffic services. The straightforward structure also makes it a good entry point for anyone new to Bedrock.

GPT-4.1 is optimized for coding. Its 1M token context window is the key feature. Think of it as a reader who can take in ten paperback books at once and process them all. You can feed in an entire large codebase and instruct it to refactor. Since it's priced lower than GPT-4o, GPT-4.1 is the more rational choice for coding agent use cases. Long document processing, legal contract analysis, and large-scale code reviews are particularly well suited.

o3 is a reasoning model. Think of it as a mathematician who doesn't give you an answer right away — it runs through a long internal calculation before delivering a conclusion. On tasks that require "thinking" — complex multi-step logic, math problems, root cause analysis of code bugs — it's noticeably more accurate than GPT-4o. The tradeoff is longer response times and higher cost. It fits batch processing and pipelines where accuracy matters more than speed. Not suitable for real-time user-facing responses.

OpenAI on Bedrock — I Tried the API Myself

I called GPT-4o on Bedrock directly through boto3. Compared to Azure OpenAI, only two things changed: authentication and the modelId format. The request body structure is the same OpenAI Chat Completions format. The prompt and message array I had written with the existing OpenAI SDK worked as-is.

# pip install boto3
import boto3
import json

# IAM role or ~/.aws/credentials-based auth — no API key needed
bedrock = boto3.client(
    service_name="bedrock-runtime",
    region_name="us-east-1"
)

# GPT-4o call — body format is identical to OpenAI Chat Completions schema
response = bedrock.invoke_model(
    modelId="openai.gpt-4o",
    contentType="application/json",
    accept="application/json",
    body=json.dumps({
        "model": "gpt-4o",
        "messages": [
            {
                "role": "user",
                "content": "Confirm GPT-4o is working on AWS Bedrock. Keep it short."
            }
        ],
        "max_tokens": 512,
        "temperature": 0.7
    })
)

result = json.loads(response["body"].read())
print(result["choices"][0]["message"]["content"])
# → "Yes, GPT-4o is running normally on AWS Bedrock."

Running it directly, response quality was identical to Azure OpenAI. That's expected — it's the same model. Response latency out of us-east-1 came in nearly identical to Azure's eastus region. The cost per token was set slightly lower than Azure OpenAI. AWS appears to have offered early-adoption pricing incentives. In real usage, the first token came back in the 200–400ms range.

One thing to watch out for is streaming. To receive streaming responses, you need to use invoke_model_with_response_stream. On Azure, a single stream=True parameter handled it, but on Bedrock the method itself is different. You also need to implement a separate chunk parser. If your service needs a real-time typing effect, expect some extra work there. Function calling worked fine with the same format.

Azure OpenAI vs AWS Bedrock — What's Different

Same OpenAI models, but each cloud has different integration methods, pricing, and ecosystem. There's no absolute winner. The right choice depends on the infrastructure you're already using.

Item	Azure OpenAI	AWS Bedrock OpenAI
Authentication	API Key or Azure AD	IAM role / AWS credentials
Model Access	Deploy a Deployment unit first, then use	Enable in model catalog, use immediately
Python SDK	openai (AzureOpenAI class)	boto3 bedrock-runtime
Network Isolation	VNet, Private Endpoint supported	VPC Endpoint supported
Data Used for Training	Disabled by default. Azure contract guarantees.	Disabled by default. AWS BAA applies.
Number of Regions	20+ regions (varies by model)	Initially us-east-1, expanding sequentially
Fine-tuning	GPT-4o fine-tuning supported	Not supported at launch, on roadmap
AWS Service Integration	Requires additional configuration	S3, Lambda, SageMaker native

What the Google-Anthropic $40B Alliance Means

That same week, Google finalized a $40B investment in Anthropic. The structure mirrors Microsoft-OpenAI like a reflection. The investment terms included a clause requiring Anthropic to use Google Cloud (GCP) as its primary cloud infrastructure. Vertex AI became the exclusive platform for the Claude series. The same strategy Microsoft used to lock OpenAI into Azure.

$40B is not a simple investment. At that scale, it effectively covers Anthropic's entire valuation. From Google's perspective, investing in Anthropic makes sense because it's the only independent AI research lab competing directly with the GPT series. Claude 4 has shown performance that directly competes with GPT-4o on coding, long document processing, and reasoning tasks. Anthropic fills the market space that Google's Gemini lineup alone can't cover. With Gemini and Claude both on GCP, there's less reason for GCP customers to leave.

How you use Claude on Vertex AI is also changing. Previously, you called Claude through the Vertex AI Marketplace. Now Claude has been integrated as a core model in the Vertex AI lineup. You can select Gemini and Claude from the same interface in the GCP console. GCP customers no longer need a separate Anthropic API contract to access Claude. Two models on one bill.

For developers, this means more options. If you wanted to use Claude in a GCP environment but found a direct API contract inconvenient, Vertex AI now covers it in one place. Using GPT-4o on Bedrock for AWS workloads and Claude on Vertex AI for GCP workloads becomes a natural positioning. If you want both models in AWS, Claude is available on Bedrock too, so a single-cloud consolidation strategy is also viable.

SoftBank and the Reshaping of Japan's AI Market

SoftBank is another variable in this reshuffling. That same week in April, SoftBank chairman Masayoshi Son announced plans to pour trillions of yen into building AI infrastructure inside Japan — covering AI servers, data centers, and power infrastructure. It's a declaration that Japan intends to become not just a consumer of AI services, but a builder of AI infrastructure. The direction is to claim AI sovereignty even at the cost of energy expenses.

SoftBank also has deep ties with OpenAI. SoftBank made a large equity investment in OpenAI and is pursuing the "Stargate Japan" project — a joint effort by OpenAI, SoftBank, and Oracle to build AI data centers in the US and Japan. With the Microsoft-OpenAI exclusivity gone, how OpenAI models get delivered through SoftBank's infrastructure — and through which cloud pipeline — has become an open question. They now have the option to run their own pipeline that bypasses Azure.

In the Japanese market, SB Technology had been a major reseller of Azure OpenAI Service. After the exclusivity ended, reports surfaced that SoftBank affiliates were exploring loading OpenAI models onto their own infrastructure instead of Azure. Asia's largest telecom-and-financial conglomerate is moving toward directly owning its AI cloud supply chain. If this trend continues, it could ripple into Korea, Taiwan, and Southeast Asian markets.

There's a Korea connection too. Samsung Electronics is in negotiations to supply HBM memory to the Stargate project. SoftBank's AI infrastructure expansion ties directly into demand from domestic chipmakers like Samsung and SK Hynix. The reshuffling of alliances at the AI software layer is shaking the hardware supply chain. A monopoly breaking at the top of the stack is affecting chip markets at the bottom.

The Multi-Cloud AI Era Has Started

Before this reshuffling, AI models and clouds were paired one-to-one. GPT series meant Azure. Claude meant a direct Anthropic API contract (or AWS Bedrock). Gemini meant Google Cloud. Choosing a model was the same as choosing a cloud. For developers, it was effectively vendor lock-in — you had to use whatever cloud had the model you wanted.

Now it's different. GPT-4o runs on both Azure and AWS Bedrock. Claude runs on both AWS Bedrock and GCP Vertex AI. Instead of picking a cloud and then choosing a model, the structure is shifting toward choosing a model first and then deciding which cloud to run it on. The coupling between models and infrastructure is starting to come apart — a real decoupling.

What this means for developers is failover. Before, if Azure OpenAI went down, there was no fallback. Now you can implement code that automatically switches to Bedrock when Azure stops responding. You can build higher availability than a single 99.9% SLA by combining clouds. Patterns for multi-cloud AI failover in production are going to spread fast.

On the cost negotiation side, enterprises have more leverage now too. Before, it was "Azure or nothing." Now you have the card of "if AWS gives us a better deal, we'll go to Bedrock." That negotiating card will actually push invoices down. Competition between clouds will drive token prices lower. It always works that way when a monopoly ends.

Cost Comparison — Azure vs AWS vs GCP

Bottom line first: at launch, AWS Bedrock's OpenAI model pricing is slightly lower than Azure OpenAI. GCP Vertex AI's Claude pricing is nearly identical to the direct Anthropic API. The table below reflects prices published as of late April 2026. Cloud promotional policies may change.

Model	Azure OpenAI	AWS Bedrock	Direct API
GPT-4o Input (1M tokens)	$2.50	$2.30	$2.50
GPT-4o Output (1M tokens)	$10.00	$9.20	$10.00
GPT-4.1 Input (1M tokens)	$2.00	$1.85	$2.00
Claude 4 Sonnet Input — Bedrock	Not supported	$3.00	$3.00
Claude 4 Sonnet Input — Vertex AI	Not supported	—	$3.00 (GCP billed)

How to Migrate from Azure to Bedrock

Here's how to move code from Azure OpenAI to Bedrock. The key is changing only the authentication and client initialization. Business logic, prompts, and response parsing don't need to be touched. First, grant bedrock:InvokeModel permissions in IAM, then enable OpenAI models in the Bedrock console. Once those two steps are done, you can test immediately with the code below.

#