Microsoft says Maia chips will run its own AI services

Microsoft wants custom chips for Copilot, Epic is reportedly changing game production, and a Dot agent tested a subscriber complaint.

Today’s issue covers Maia chips, AI-first game work, and a Dot agent testing a subscription bug.

Microsoft is trying to make its AI bill less dependent on Nvidia. On 12th October, the company said its Azure Maia chips will become the main system for running AI inside Microsoft products and services.

The new Maia 300 is due for large deployments in 2027. Microsoft says it should deliver 40% more AI output per megawatt than leading GPU-based systems, and up to seven times the rack-scale output of Maia 200.

That matters because the AI bottleneck is becoming power as much as chip supply. Microsoft wants chips tuned for long documents, live conversations and agents that work through multistep jobs.

The bigger signal is where the large AI platforms are spending money. The companies with enough demand are designing the hardware around their own workloads, not waiting for a general-purpose chip market to solve the bill.

Founders context: When renewing AI-heavy tools, ask what happens when usage rises: whether seats, message caps, response speed or premium features are tied to scarce model capacity.

Epic has reportedly pushed an unannounced Star Wars and Marvel shooter toward an AI-first production model, turning a delayed game into a live test of AI production work.

The quoted internal message said there would be “no more hand-written code”, while art would be generated through full AI workflows. The risk is whether the studio can keep code quality and visual consistency while changing how work gets made.

The review job gets harder when both code and art start from AI output. A game needs code that compiles, art that feels consistent and someone senior enough to reject weak work.

Epic is a useful test case because games expose weak AI work quickly. If the art feels generic, or the code creates bugs, players will notice in the product itself.

Founders context: If you push AI into production work, separate three things: where AI may draft, who approves the result, and which jobs still need a human specialist from the start.

Ethan Mollick, a Wharton professor who tests AI tools in public, posted a small but useful customer-service example on 12th October. A subscriber emailed him about being signed up to other Substacks after subscribing to his.

Without being asked, his Dot agent tried to replicate the complaint. It figured out the problem by testing the signup path, then brought the evidence back for a person to judge.

The workflow is narrow, which is why it is interesting. The agent did not “run support”. It took one complaint, recreated the path, and turned a vague customer note into a testable bug report.

That is a copyable pattern for low-risk support, billing, onboarding and account issues. Let an agent reproduce the problem, collect evidence and hand it back before anyone touches customer data or sends a reply.

Founders context: Start with side-inbox jobs where the agent can investigate but not act. Good candidates are failed signups, missing invoices, broken links, duplicate notifications and routine account-setting confusion.

💸 Techpresso tracked Microsoft’s Maia launch as the day’s big AI infrastructure story. The newsletter framed the move as Microsoft’s biggest bet yet on its own silicon, reducing reliance on Nvidia over time.

📱 A phone-model post said 19 local open-weight models now beat GPT-4. The claim points to models that can run on a modern smartphone, rather than closed-source/API systems such as ChatGPT, Claude or Gemini.

🧰 Microsoft’s MxC sandboxing work drew fresh attention. The open-source project is agent safety infrastructure for limiting processes across Windows, macOS and Linux.

🧠 GeekWire reported that Maia 300 targets long-context agents and live conversations. The specialist report adds context on Microsoft’s push to reduce Nvidia dependence while still using Nvidia systems in its data centres.

🎓 Certificate automation kept the skills-proof debate alive. The demo makes course-completion badges less convincing unless they are paired with proof of applied work.