streugut.ch
Archive Subscribe

AI & Microsoft.
As we see it.

Up Close

Microsoft turns Copilot resale into metered spend as open-weight Beam lands on rival clouds

Microsoft set an end-of-sale date of 1 December for Microsoft 365 Copilot offers in the Cloud Solution Provider channel and said usage-based billing will become the default for new Microsoft 365 Copilot Business licences there. That changes the purchase from a partner-resold package into metered consumption by default. We read this as Microsoft pushing Copilot further into the budget discipline already common in cloud infrastructure: buyers and resellers now have to manage usage, not just approve seats, and enterprise AI is being sold more like controlled capacity than ordinary software.

AWS added Reflection’s open-weight Beam model to SageMaker JumpStart, where customers can deploy and customise it inside their own AWS accounts, with optimisation for Trainium3. Google Cloud announced day-0 support and deployment recipes for the same model on its AI Hypercomputer and NVIDIA AI infrastructure. An open-weight model is a model whose parameters are published for others to run and adapt themselves. Taken together, this suggests model competition is shifting away from any one provider’s front door and toward the cloud control plane that makes a model governable, deployable and cheap enough to run. That matters beyond AWS and Google because it lowers the cost of backing one model family without surrendering control to its maker, and it makes Microsoft’s own mix of cloud and local AI governance look less exceptional.

Anthropic received a BB+ issuer credit rating for about $4.4 billion of subordinated debt ahead of a planned $35 billion chip-financing package led by Apollo. Apple opened UK pre-orders for the foldable iPhone Duo with Siri AI built on Apple Intelligence. TechCrunch reported that Replit's chief executive said Replit was no longer working with Anthropic's Claude models and now preferred OpenAI's Sol and Astra models, though unnamed sources disputed that the Anthropic partnership had fully ended.

Evidence

Over Time

AI is splitting by delivery mode while control and infrastructure harden around it

Generative AI is no longer settling on one way to be used. It is spreading into distinct delivery modes - local PCs, regulated cloud enclaves, specialised decision models and device-bound consumer interfaces - because cost, latency and trust do not line up in one place. Microsoft’s Azure Enclave reached general availability for regulated workloads, while Hugging Face released Llama Windows, a Windows 11 app that runs quantized models locally and exposes them to other software through a local interface, and Microsoft put its own small decision model into Foundry. Microsoft Foundry is the shop where Microsoft keeps models side by side, its own and other people's. Apple added Siri AI and a full-screen AI discovery layer to Apple TV 4K, and Meta put one frontier model behind its API and another behind a subscription product. This appears to sharpen the line from earlier local-first agents and selective feature retirements: the market is segmenting by what must stay local, what must be governed in cloud isolation, and what is packaged as a device feature rather than a general assistant.

Competition is also moving away from the model as a standalone release and toward the full stack that lets agents and workloads run. Microsoft said Azure Maia will become the primary inference system for its own products and services and plans multi-gigawatt deployments from 2027, while Google made Gemini 4 Argon generally available in Google Cloud and tied it into its enterprise agent platform. AMD, Intel and NVIDIA each used the Open Compute Project summit to publish rack designs, optics or infrastructure specifications built around agentic AI workloads. This likely widens an older shift from model launches alone to integrated systems of chips, compilers, schedulers, cloud services and workflow tooling, because the contest now reaches from the model catalogue down to power delivery and cooling. For anyone buying or building with generative AI, that suggests the practical choice is increasingly between stacks that arrive as an operating environment, not between isolated models.

Local AI on personal devices is moving past the old privacy niche into a real execution path for capable agents and specialised models. NVIDIA detailed the RTX Spark PC platform with up to 128GB of unified memory and associated agent apps ahead of launch, AMD introduced a Ryzen AI Halo developer platform for hosting very large locally run models, LiquidAI released both a local reasoning model for private agentic workflows and open-weight edge decision models, and Microsoft had already opened preorders for a Surface Laptop Ultra built to run large models locally. This appears to strengthen the new line rather than merely repeat it: local execution now has hardware, runtimes and task-specific models arriving together. The effect reaches beyond Microsoft because ordinary Windows and PC estates start to look less like thin clients for cloud AI and more like places where part of the workload can stay on the machine.

Evidence

Our forecast

We expect Microsoft, Apple or Google to extend their operating system by the end of January so that any app can use an AI model directly on the device.

what we hold ourselves toDue 2027-01-31

By 2027-01-31, at least one of Microsoft, Apple or Google must have announced or released a first-party desktop or device operating-system feature that lets third-party applications call a locally running generative AI model through a documented on-device interface.

The settlement

The first expectation came due on October 15, 2026. It is settled in the next weekly text.