Skip to content
Go back

Every Cloud Now Sells Claude, GPT, and Grok. Who Runs Them Is Where They Differ.

Take Maya, the ML platform lead at a US health insurer. Her shortlist of three models clears every cloud’s frontier catalog in an afternoon, then breaks on the cheapest model on the list and on a footnote about who actually runs Claude.

In her brief, engineering wants Claude Fable 5.2 for hard coding work. The claims team has built on OpenAI’s GPT-6.2 Sol. A document-triage pipeline needs a cheap open-weight model, Kimi or Qwen, at high volume. Compliance wants every prompt processed in the US. The tension sits in a contract detail: every cloud can put Anthropic-run Claude on its own bill, and Anthropic-run Claude is processed outside that cloud’s boundary.

Models are one layer of a larger map of AI and data services across the three clouds, and companion posts cover agent platforms, agent identity, and cloud marketplaces.

The frontier rows converged

Claude Opus 5.5, OpenAI’s GPT-6 models, xAI’s Grok 4.7, and Mistral models appear in the Amazon Bedrock model list, Microsoft Foundry’s models sold by Azure, and the model versions on Google’s Gemini Enterprise Agent Platform. Grok 4.7 is GA on Bedrock since September 28 and GA on Google, and Foundry lists it as generally available.

OpenAI is on all three, at different stages. GPT-6.2 Astra and Sol went GA on Bedrock and Foundry on November 17. Google’s OpenAI page lists GPT 6.2 Astra and Sol alongside four older GPT-6 models, and states that OpenAI partner models are in private preview, with access through your Google Cloud account team. The spread follows from a contract: in April, Microsoft’s license to OpenAI’s models became non-exclusive, and OpenAI can now serve its products on any cloud.

Gemini is the exception. Neither Bedrock nor Foundry lists a Gemini model, and Google launched Gemini 4 Neon on its own platform today. For a closed model, what remains is launch stage and who operates the endpoint.

One open-weight model, three ways to serve it

Think of a concert venue. One uses its house crew, another books a touring crew that arrives with its own rig, and a third hands you the stage and expects you to bring your own sound system. The analogy breaks on paperwork: each crew comes with its own launch stage, quota, and data terms, which a venue crew never does.

ModelAWS (Amazon Bedrock)Azure (Microsoft Foundry)Google (Agent Platform)
Moonshot KimiK3 and K3.1 Pro (Highspeed), serverless, US-only us. profileKimi-K2.7-Code, K2.6, K2.5 sold by Azure (preview); K3 via Fireworks and BasetenSelf-deploy only
QwenQwen3 serverless; Qwen3.5 and 3.6 on JumpStartQwen3-32B; Qwen3.6 and Qwen3.8-27B in preview, listed mainly for fine-tuningSelf-deploy only
DeepSeekV3.2 serverless; V4 and V4.1 on JumpStartV4-Pro, V4-Flash, V3.2-Speciale sold by Azure; V4.1-Flash via BasetenV4.1-Flash serverless
Z.ai GLM5.3, cross-Region only, eligible enterprise customers5.3 via Baseten (preview)5.2 serverless
OpenAI gpt-oss120b and 20b serverless120b and 20b (preview)120b serverless
Gemma 4ServerlessVia Fireworks26B serverless
Llama 4ServerlessMaverick sold by AzureServerless

Bedrock runs its open-weight catalog itself. The model list includes open-weight models from ten publishers (DeepSeek, Google, Meta, MiniMax, Mistral, Moonshot, NVIDIA, OpenAI, Qwen, and Z.ai), billed per token beside Claude. DeepSeek V4 and V4.1 and Qwen3.5 and 3.6 come through SageMaker JumpStart instead, where you pick an instance and pay while it runs. Custom Model Import accepts Mistral, Mixtral, Flan, Llama 2 through 3.3 and Mllama, GPTBigCode, Qwen2 through Qwen3, and GPT-OSS architectures, with text weights under 200 GB. DeepSeek, Kimi, GLM, Gemma 4, and Llama 4 are not on that list.

Microsoft adds partner-run catalogs to Foundry. Fireworks on Foundry has been generally available since June 2, 2026, with Kimi K3 and DeepSeek V4; Microsoft’s Kimi K3 post says “Fireworks AI provides the optimized inference infrastructure, while Microsoft Foundry handles enterprise deployment and governance,” and prices a Data Zone deployment. Baseten entered public preview on November 17 “as a Microsoft first-party service,” offering Kimi K3, GLM 5.3, and DeepSeek V4.1-Flash “globally or in the US Data Zone.” Both bill through the Azure agreement.

Google narrowed its serverless list to seven models. On October 21, 2026, it retired 16 serverless open-model endpoints: four Qwen3 models, Kimi K2 Thinking, MiniMax M2, gpt-oss-20b, Llama 3.3, four DeepSeek models, two GLM models, and two embedding models, each with a “self-deploy alternative” in Model Garden. The current list has two DeepSeek V4.1-Flash versions, Gemma 4 26B, GLM 5.2, gpt-oss-120b, and two Llama 4 models.

Lifecycle terms differ in detail. Google promises preview open models at least 45 days, and calls that “a minimum, not a scheduled deprecation.” Its own Gemini 4 Neon, Hydrogen, and Carbon sit in its shorter-availability table; only Gemini 4 Argon carries a 12-month commitment. Bedrock’s Kimi K3.1 card promises “at least 45 day EOL Notice.” Microsoft “doesn’t recommend using preview models in production.”

For Maya, “Kimi is available” means three different jobs. Bedrock and Foundry both give her an API call, to a model AWS runs on one and Fireworks or Baseten runs on the other. Google gives her a GPU endpoint her team sizes, patches, and pays for while idle.

Who runs Claude, and where the prompt goes

Every cloud sells Claude two ways, and the difference decides where a prompt is processed.

AWSAzureGoogle
Runs on the cloud’s infrastructureAmazon Bedrock”Hosted on Azure”: Opus, Sonnet, Haiku 5.5Claude on Agent Platform
Runs on Anthropic’s infrastructureClaude Platform on AWS”Hosted on Anthropic”: all models, including Fable and MythosClaude on Google Cloud (Anthropic managed)
Fable 5.2 on the cloud’s infrastructure?Yes, US geographic profileNoYes, US multi-region
Fable 5.2 retentionOpt in to AWS review; kept up to 30 days, not shared with AnthropicAnthropic’s termsKept up to 30 days; prompt-response sharing with Anthropic required

On Azure, the hosting comparison says that for both options “Anthropic is the seller and operator.” Hosted on Azure means prompts are processed on Azure infrastructure and “never leave Azure.” Fable 5.2 and Mythos 5.2 appear only on the Anthropic-hosted side, where “data might be processed outside of Azure including outside of your selected Azure geography,” with Global Standard deployments only and no US Data Zone.

AWS’s Fable 5.2 model card requires setting your data retention mode to aws_review, and offers a us. geographic profile with no single-Region endpoint. The abuse detection page says Bedrock retains all Fable traffic up to 30 days, that “classifier-flagged traffic will be subject to potential human review performed by AWS,” that retained data is “not shared with third-party model providers,” and that under cross-Region inference it is stored in the Region that processed the request. Customers eligible for Enterprise Frontier Safeguards can use zero data retention for internal use through at least June 30, 2027.

On Google, the Fable 5.2 page lists it as GA in us and eu multi-regions and stores prompts and responses up to 30 days. It also says: “As required by Section F of Anthropic’s Service Specific Terms, enable prompt-response sharing with Anthropic for each model location.” The logging page says shared logs go to the model partner “in real time,” and the Advanced AI Safety Addendum allows human review by authorized personnel.

Fable 5.2 can stay in a US geography on AWS and Google, both with a 30-day retention window and with different rules on who else sees the prompts, while on Azure it may be processed outside Azure.

Maya’s three options

RequirementAWSAzureGoogle
Fable 5.2 processed in the USYes, us. profile, retention opt-inNo: Anthropic-hosted, Global Standard onlyYes, us multi-region, sharing with Anthropic
GPT-6.2 Sol, GAYesYesListed, private preview only
Kimi or Qwen, pay-per-token, USYes, Kimi K3.1 Pro us. profileYes, Kimi K3 through Fireworks (Data Zone) or Baseten (preview)No: self-deploy

My reading: AWS clears all three rows, at the price of a retention opt-in for Fable. On Azure, Fable 5.2 is the blocker, and Opus 5.5 hosted on Azure is the substitute that keeps prompts inside Azure. On Google, Maya’s compliance team has to accept prompt sharing with Anthropic, wait on GPT-6.2’s preview, and run Kimi on her own endpoint. Swap Gemini 4 in for GPT-6.2 and Google becomes the only cloud with all of her frontier models, though Kimi there still means self-deploying.

What’s missing

Anyone can download open weights, and that portability stops at the serving endpoint. The same Kimi model is a Bedrock model ID on one cloud, a partner deployment on another, and a GPU cluster on the third, each with its own launch stage, quota, retention terms, and retirement clock. Google’s retirement table shows the clock: deprecation notice on July 21, shutdown on October 21.

Where this comparison stops:

So what

If you are choosing a cloud for models, skip the frontier comparison. Write down the open-weight model that will carry most of your tokens, and ask each cloud who runs its endpoint: the cloud, a partner, or you. Then ask whose infrastructure runs the Claude model you need, and who else sees the prompts.

The question I have not answered: does a cloud-run open-weight catalog hold up as releases speed up? The newest DeepSeek and Qwen releases on AWS sit on SageMaker JumpStart, outside Bedrock serverless. Google narrowed its serverless list to seven models. Microsoft adds partner-run catalogs from Fireworks and Baseten. If partner-run inference becomes the usual way to get a new open-weight model in its first months, then “which inference provider, billed through which cloud” becomes the question for that layer. I don’t know yet whether Maya should plan her triage pipeline around the cloud’s catalog or around the partner behind it.


Share this post on:


Previous Post
Three Cloud Marketplaces, Three Commit Rules, Three Front Doors
Next Post
Who May Call What: Per-User Authorization on AgentCore With Cedar