OpenAI has completed roughly $11 billion in financing at a $1.25 trillion valuation, while Thinking Machines Lab unveiled a 594-billion-parameter model the same day. AI remains as hot as ever, but the more consequential development is that Microsoft is reportedly reducing its capacity on older chips. The industry has entered a phase in which scale keeps expanding and efficiency must keep pace.
What this is
These developments span models, capital, and infrastructure. OpenAI's financing shows that leading research laboratories can still secure enormous amounts of money. Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released Inkling 2 Small, a mixture-of-experts model (MoE, an architecture that activates only some parameters per inference to reduce costs) with 594 billion total parameters and 17 billion active parameters. Meanwhile, Microsoft is reportedly phasing out older AI chips leased from Oracle because their internal specifications no longer meet current requirements.
AWS SageMaker Inference also added support for NVIDIA GPU physical partitioning, and the vLLM team proposed an asynchronous reinforcement learning training method. Together, these developments point in the same direction: extract more useful computation from a limited number of chips rather than simply adding more devices.
Industry view
We believe model parameter counts and funding rounds will continue setting records, but the marginal return on capital is already being tested. GitHub data shows that weekly code commits on the platform exceeded 1.2 billion in October 2026, up more than 400% year over year. Yet faster code generation has not translated into an equivalent increase in overall R&D productivity. There remains a gap between what models can do and what enterprises can actually deliver, with process, data, and accountability boundaries still in between.
The risks are equally clear. Inkling 2 Small's results are based mainly on official and third-party benchmarks, which are not yet sufficient to establish its stability or cost advantage on long-running tasks. Microsoft's reduction of older compute capacity also means enterprises cannot treat rented external compute as a permanent solution. We do not accept that every technical advance automatically translates into realized commercial value. Model scale, benchmark scores, and actual revenue still require more time to validate.
Impact on regular people
For enterprise IT: Compute utilization, model deployment costs, and system reliability will be closer to purchasing priorities than sheer parameter count. Enterprises holding legacy cloud resources may need to reassess their contracts and the pace at which they replace equipment.
For individual careers: AI assistants will take on more planning, document analysis, and code collaboration work. But GitHub's data is a reminder that increased tool use does not automatically increase job output. Task decomposition, review, and result delivery remain the decisive factors.
For the consumer market: Doubao Work Personal Edition has replaced its five-hour usage limit with weekly quotas, allowing individual users to concentrate AI use around peak business periods. Higher model usage will also intensify competition over subscription pricing, data privacy, and service reliability.