When Clouds Start Selling Their Rivals' Models

2026-10-21·6 min read·OKCodex Editorial

When Clouds Start Selling Their Rivals' Models

Read together, this week's signals point at one thing: the deciding factor in AI competition is moving down from "whose model is better" into three lower layers — distribution, compute delivery, and whether agents are allowed into the transaction itself.

1. Distribution: the only cloud selling all three

Google Cloud now offers OpenAI models alongside its own Gemini and Anthropic's Claude. According to The Decoder, it is the first major cloud provider to offer all three families on a single platform; the initial lineup is GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, available as a private preview for enterprise customers.

The implication for model vendors is blunt: once three frontier models can be called under one bill and one permission system, switching cost drops from "change your cloud" to "change a model name." Model comparison collapses into an A/B test, and the cloud provider occupies the genuinely scarce position — the purchasing entry point and the billing account.

2. Compute: whoever can't buy capacity can't close deals

TechCrunch reports that Anthropic is struggling to secure enough compute from cloud providers that increasingly see it as a competitor, while its infrastructure spending balloons ahead of a planned November IPO. The report is framed as "reportedly" and is unconfirmed by either side, but the logic is not hard to follow: when your cloud provider also sells models, selling you compute means funding a rival.

On the supply side, delivery is speeding up. Nvidia's DSX POD pitch is to ship AI data centers preassembled in containers, which The Decoder says could cut deployment from six months to as little as 14 days. If that holds in real projects, the gap between "can get a facility" and "cannot" shrinks from years to weeks — and the bottleneck moves faster toward power, networking and operations talent rather than civil construction.

In the same week, Mistral and Nvidia announced a partnership to accelerate engineering simulation and industrial product design with AI (so-called physics AI). Compute vendors are binding upward into concrete industry workflows, not just selling cards.

3. Access: whether an agent can spend is the dividing line

The retail story is arguably the one most worth watching for teams in China. TechCrunch reports that Amazon blocks AI agents from shopping on its site, while Walmart takes the opposite approach: agents can make purchases on customers' behalf, with controls over spending and transaction approvals.

This is not a morality tale about open versus closed. It is a question of where to place the control point. Blocking preserves certainty in traffic and ad monetization; opening buys an early position as the default interface agents call. For agent builders the implication is practical: what determines whether your loop closes is not model capability but whether the counterparty offers an authorization interface with limits and approvals. When choosing platforms, weigh "is there programmable payment authorization" as heavily as the API docs.

4. Organization: tools outrunning management

Another survey cited by TechCrunch finds more than 50% of managers say evaluating employees' work has become harder since adopting AI. That sits in direct tension with the infrastructure narrative: deployment is accelerating at the bottom while the evaluation layer degrades at the top.

On the model side, some signals deserve caution. Inception claims its diffusion language model Mercury 3 is the most capable diffusion LM released so far, runs four to eight times faster than GPT-6 Luna and Claude Haiku 5.5, and beats both on eight of 12 benchmarks — but the report states plainly that no independent tests exist yet. Until someone reproduces the numbers, treat them as vendor claims. Likewise, Demis Hassabis shortening his AGI timeline from three-to-five years to two-to-three is a personal judgment, not an industry consensus.

Conclusion: look away from the leaderboard

For AI application and agent teams in China, this week suggests three concrete moves:

Treat multi-cloud, multi-model as the default architecture. When one cloud can serve three model families, substitutability is itself negotiating power. Abstract a model-calling layer; don't weld business logic to a single SDK.
Audit your platform dependencies. Check each key channel: do agents have access, and are there spending limits and approval controls? The Walmart–Amazon split will replay across more industries soon.
Fix the evaluation system before claiming productivity. When more than half of managers can't articulate how to measure output, AI-driven efficiency gains are hard to prove. Defining measurable output metrics before rolling out tools is a better investment than swapping in another model.

When Clouds Start Selling Their Rivals' Models