AI Notes: Every Layer's Lock-In Is Loosening

The thing most worth noticing this week was a bill.

On October 21, OpenAI and Google announced that GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are coming to Google Cloud's Model Garden, opening first as a preview for some enterprise customers. Billing runs through Google Cloud Marketplace, lands on the same Google Cloud invoice, and can be drawn down against cloud spending commitments customers have already signed.

Read on its own, the headline is two rivals cooperating. Put next to a few other things from the same week, it looks like part of a bigger change: the things that lock a buyer into one supplier along the AI supply chain are loosening, one layer at a time.

Models start following the budget

When a company buys a model, it often doesn't start by asking which one is strongest. It starts by asking which cloud its budget is already committed to. A signed cloud commitment is money that has to be spent sooner or later; any model that can be paid for out of it skips a round of procurement.

Google says it is now the only hyperscaler offering Gemini, Claude and GPT together, with more than 200 models in Model Garden; GPT will also come to Antigravity, Google's own coding tool. OpenAI, for its part, stresses that Microsoft is still its primary cloud partner.

For OpenAI, an exclusive channel used to be a bargaining chip; now it is trading exclusivity for reach. For Google, putting a rival's model inside its own tools looks like a concession, but what it really runs is a shelf: the customer stays on its cloud, and whichever model they use gets settled on its bill. As models look more like goods on a shelf, whoever holds the entrance to the budget carries more weight.

Half the chip moat is software

On October 19, AWS released TorchNeuron in public beta: change one line of configuration and PyTorch code runs on Trainium. The next day Google open-sourced TorchTPU. Before that, PyTorch users who wanted TPUs either rewrote in JAX or used the older torch_xla.

Moving off Nvidia GPUs used to mean rewriting code, re-tuning, and taking on the risk of things breaking. That switching cost is a big part of GPU pricing power. It isn't written in the chip's spec sheet; it's written in developers' time. The two largest cloud providers pushed it down within two days of each other, which amounts to saying: from now on, choosing a chip will be more about how much compute a dollar buys, and less about where your code is locked in.

That said, there are no public comparisons yet of the same model across different chips. The barrier is lower, not gone.

The inference stack laid open

On October 20, DeepSeek open-sourced DeepInfer, the inference framework behind its own service, under Apache 2.0. It supports Nvidia and Huawei Ascend chips, but holds back the hardest parts: KV caching, prefill/decode disaggregation and load balancing. The same day, Google Cloud, Red Hat and IBM released llm-d 1.0, which fills in exactly those pieces and runs on Nvidia, AMD and Intel GPUs as well as Google's TPUs.

Running models cheaply has always been a craft that model companies and inference clouds don't talk about much; a large share of the difference in cost per token comes from this layer. Once that craft becomes public code, the profit earned simply by running things more efficiently than others gets thinner.

Buyers start doing the maths

These changes are arriving together because buyers' attitudes have changed. On IBM's October 21 earnings call, CEO Arvind Krishna said large companies care more and more about the real return on AI spending, and worry that some of it is "out of control." By his estimate, running open-weight models on your own racks costs roughly 10% to 25% per token of what frontier models cost.

Once switching costs fall, pricing power slowly moves from the side being chosen to the side doing the choosing. Models can move between clouds, code can move between chips, inference frameworks can be taken off the shelf. Companies have more options to compare, and their negotiating stance naturally hardens.

Where value lands once the locks loosen

None of this means everyone makes more money. Broadcom's Hock Tan offered some estimates on October 12: frontier labs spend about $100 billion to bring in about $120 billion of revenue, while the open-weight camp spends about as much and recovers only around $30 billion. For buyers, substitutes make the bill lighter; for sellers, the layer that gets substituted struggles to hold on to profit.

So what really changed this week is the way companies compete. Customers used to be kept through lock-in; from here, you have to keep getting chosen when the customer could swap you out at any time. Roughly two kinds of players can manage that: those who hold the entrance to the customer's budget, and those who still come out as the best value every time the comparison is rerun.

This is the other side of the same week's mood in the capital markets: the bigger the infrastructure spend, the more it needs users to keep paying for a long time, while the choices in users' hands keep multiplying.