For years, the frontier model was the product. Enterprises chose their cloud partly by which AI they could access inside it. That logic died quietly on October 21, 2026, when Google Cloud announced that OpenAI put GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna into private preview on Google Cloud’s Model Garden, completing a distribution sweep that now spans Azure, AWS Bedrock, and Google Cloud simultaneously.
The deal marks the first time OpenAI’s models will be sold natively through all three major US cloud providers at once. The question that follows is not about OpenAI. It is about who actually benefits once the model itself is everywhere.
What the Distribution Map Looks Like Now
Azure has sold GPT-6 Astra as generally available since September 3, 2026; AWS Bedrock since September 8, 2026; Google Cloud opened private preview on October 21, 2026, with broader availability described as “coming soon.” For enterprise buyers, the billing terms may be the bigger news: Google said customers can purchase OpenAI models through Google Cloud’s Model Garden and Marketplace, including by using existing cloud spending commitments. Google’s Flexible Savings Plans documentation also says participating third-party foundation models available through Vertex AI and Cloud Marketplace can be eligible for coverage under certain plans.
That billing consolidation is precisely what each of the three clouds is selling. The model is no longer the differentiator. Infrastructure, price, and proprietary data advantages are.
Google’s Structural Edge
Google Cloud enters this commodity model era with one asset neither Microsoft nor Amazon can replicate quickly: its own silicon. Google owns Gemini and TPUs top to bottom, with no licensing fees to OpenAI or Anthropic. That said, the claim that Gemini is served exclusively on TPUs is too absolute. In practice, Google leans heavily on TPUs for Gemini, but Google Cloud also offers GPU infrastructure broadly, and the underlying hardware choices vary by product and deployment.
Google Cloud revenue reached $24.8 billion in Q2 2026, up 82% year-over-year. Evercore ISI is already projecting that TPU sales could drive Google Cloud revenue toward $213 billion in 2027.
The pricing dimension sharpens that structural edge. Google launched Gemini 4 Argon at a fifth of what OpenAI charges for GPT-6 Astra, a move that turns its own-silicon cost advantage directly into a price weapon against hosted rivals.
Amazon’s Data Gravity
AWS arrives at the model-neutral era with a different kind of advantage: the largest installed enterprise base in cloud computing. Andy Jassy has spent the past eighteen months making AWS the default landing pad for every major AI lab, validating a multi-model Bedrock strategy that AWS has described as offering access to more than 100 models. By 2026, Bedrock has become a common choice for enterprises running their workloads on AWS that want managed access to multiple model families while keeping data within an AWS account. The stickiness is not the model catalog. It is the decade of S3 buckets, IAM roles, and compliance workflows that makes leaving painful.
Microsoft’s Difficult Position
Azure is where the model-neutrality transition hurts most. Microsoft remains OpenAI’s primary cloud partner under a license to OpenAI IP for models and products through 2032, but the unilateral lock-in that defined the partnership since 2019 is gone. Microsoft has no proprietary frontier hardware comparable to TPUs, and its own first-party model business is structurally entangled with the partner it no longer exclusively controls. OpenAI’s “ship first on Azure” commitment ensures Microsoft’s infrastructure remains the default path for new deployments for now, but that is a contractual delay, not a competitive moat.
What Could Go Wrong
The commodity model argument assumes the major labs continue releasing through all three clouds at roughly equal quality and latency. That could break. OpenAI and Microsoft have said “Microsoft remains OpenAI’s primary cloud partner,” which means Azure-first rollouts remain a real asymmetry. And Google Cloud is now claiming it is the only hyperscale cloud provider offering access to all three leading closed-source LLM families alongside more than 200 first-party, third-party, and open models, a breadth claim that may not persist if exclusivity deals shift again.
The long-term verdict is not that models stop mattering. It is that owning the infrastructure beneath them, and the data workflows locked into them, is now the compounding advantage. Google’s TPU stack and Amazon’s enterprise data gravity are the two most defensible positions in this new map. Microsoft has the longest runway with OpenAI but the least structural protection if the model layer commoditizes fully. The investor question heading into 2027 is simpler than it looks: who owns the power plant, not who sells the electricity.
Physical infrastructure is now the compounding variable. Alphabet’s €13 billion Finland commitment and a 22-year nuclear power deal illustrate how vertical control of land and energy becomes a durable cost moat.
