Cheap models and expensive infrastructure are locking AI into operating systems, not one-off tools

September 23, 2026

OpenAI cut API prices for GPT-6 Sol and Luna, added new prompt caching controls, and kept GPT-6 Astra as its top tier. It is increasingly being sold as a full operating stack - the model, the price controls around it, the caching that avoids paying twice for repeated context, and the workflow surface where people actually use it. Microsoft, according to The Information, authorised Copilot discounts for large seat commitments while tying them to additional usage-based payments. Google put Ask Workday into Sheets and Slides for financial work. This appears to sharpen the line that providers now compete on deployment ergonomics and governed use, not only on who posts the strongest base model. For anyone buying or building with generative AI, the practical choice is therefore looking less like "which model is best" and more like "which stack makes cost, routing and work integration manageable."

The cheapening of inference is not slowing the build-out underneath it. Alibaba introduced the Zhenwu V900 chip and said Alibaba Cloud plans more than 20 gigawatts of global data-centre capacity by 2032. NVIDIA started a DSX Ready programme - a qualification scheme for power and cooling parts that fit its AI factory designs. OpenAI lowered model prices again, and Anthropic launched Claude Opus 5.5 at lower running cost than Opus 5. This suggests falling model prices are being absorbed by a market that expects more usage, not less infrastructure: the easier it becomes to call a model, the more the bottleneck shifts to power, cooling and long-lived capacity. That matters beyond Microsoft because cloud buyers and local AI operators alike are being pulled into decisions that look more like infrastructure planning than software procurement.

Control is also hardening into an operating condition rather than a policy add-on. OpenAI published a paper proposing shared international standards for advanced AI and a framework for third-party assessments. Yet President Trump told the UN the United States rejects any global scheme to control AI, while US and Chinese officials formalised ongoing AI-related talks and 19 EU states launched the IPCEI-AI programme with over €10 billion ($10.9 billion) in planned investment. Anthropic, for its part, published a system card for Claude Opus 5.5 and kept Fable-class access controls around its biology and cybersecurity capacities. The picture appears to be widening, not resolving: AI oversight is becoming part of market access and product design even where governments disagree on one common rulebook. For deployers, that means compliance is likely to arrive through several channels at once - provider access terms, national policy, and the technical documentation needed to justify use.