Model prices fall again, while AI capacity turns into fixed infrastructure

September 23, 2026

Alibaba introduced its Zhenwu V900 AI chip and said it plans to expand Alibaba Cloud’s global data-centre capacity to more than 20 gigawatts by 2032. NVIDIA also launched a qualification programme for battery storage and cooling gear against its AI factory reference designs. This suggests the infrastructure layer is being organised less like a short-lived rush and more like a long build-out with standard parts, which matters beyond Alibaba and NVIDIA because anyone buying cloud capacity or planning private AI estates is being pulled into a market with longer commitments in power, cooling and site design.

OpenAI released GPT-6 Sol and GPT-6 Luna at API prices 50% below its GPT-5.6 promotional levels, while Anthropic launched Claude Opus 5.5 saying it reaches Claude Fable 5.1 on most tasks at 40% lower running cost than Opus 5. OpenAI also upgraded prompt caching for GPT-6, that is, a way to avoid paying in full when repeated parts of a prompt can be reused for a short period instead of recomputed each time. We read these moves together as a shift from selling prestige models toward selling cheaper usable throughput, because the change a customer feels is not only a new model name but lower unit economics and more tooling to control repeat spending. That is likely to weaken any provider, Microsoft included, that still relies on packaging AI as a premium add-on without matching the fall in underlying inference costs.

The United States president told the UN General Assembly that the country rejects any global scheme to control AI, even as US and Chinese officials formalised ongoing talks on AI alongside trade and investment ahead of a planned leaders’ summit. This appears to mean AI governance is becoming less a single rules project than a mix of national control and bilateral bargaining, which changes the operating environment for providers and deployers because access, compliance and acceptable safeguards may now depend as much on political relationships as on one common standard.

Qualcomm unveiled new 2 nm Snapdragon mobile platforms designed for more on-device AI in premium Android phones. Hugging Face hired the creator of oMLX to improve conversion of open models to Apple’s MLX stack for local use on Apple Silicon. Cisco Talos disclosed Windows malware that uses large language models to choose post-compromise actions autonomously.