Reuters says Anthropic is in talks to rent servers powered by Microsoft-designed chips to help handle rising demand for its AI services. Reuters is also clear that these are still early discussions and may not lead to a final deal. There is no agreement yet, but the fact that Maia 200 is even in the conversation is enough to get the market's attention.

If Anthropic does end up using Maia 200, it would be a meaningful win for Microsoft's in-house chip push and another sign that frontier labs are starting to open new inference paths outside Nvidia's expensive and constrained lane. The AI infrastructure race looks a little less like a pure benchmark contest from here. It increasingly looks like a fight over who can pull real demand onto their own chip stack.

OpenAI says Codex is now inside the ChatGPT mobile app so users can monitor and steer work running on connected machines. In the same rollout cycle, OpenAI also introduced Appshots and remote computer use on Mac even when the screen is locked. If this works reliably in real workflows, Codex stops looking like a coding window that opens and closes on a laptop.

It starts looking more like a job layer that keeps running, waits for approval, and can be pulled back from a phone. That is where computer-use agents start touching real work: not in how impressive the demo looks, but in whether they can survive inside a workflow after the user walks away from the screen.

The Information, via Reuters pickup, says the White House Office of the National Cyber Director briefed AI companies on a planned executive-order path that could let intelligence agencies and other parts of government review advanced models before release. No rule has been signed and no review regime is operating yet, but frontier rollout no longer looks fully confined to internal lab decisions.

Next to Maia 200 and the Codex update, this completes the third control layer in today's issue. First comes compute, then execution surface, and now release governance. If this direction holds, frontier models will face more layers of review before they reach the public.

NVIDIA is trying to attach provenance and risk metadata to the skills layer: NVIDIA AI says it shipped Verified Agent Skills, along with skill cards and the agentskills.io spec, to run across Claude Code, Codex, and Cursor.

Anthropic is turning compute into a giant, visible line item: TechCrunch and Axios say Anthropic is committing about $1.25B a month for compute capacity on SpaceX/xAI infrastructure through 2029. Frontier AI shows up here as procurement contracts and real invoices, not as some vague backend assumption sitting outside the market's view.

Anthropic is also pushing model values into a more public trust surface: Anthropic says it is widening conversations with clergy, philosophers, ethicists, lawyers, writers, and civic leaders to inform Claude's constitution, model values, and evaluations. Alignment is moving further out of the back room and into a trust narrative labs increasingly have to present in public.

xAI is carrying its subscription entitlements into OpenCode: xAI says SuperGrok and X Premium now work inside OpenCode, a coding agent with both terminal and desktop surfaces. Vendor entitlements are starting to flow into third-party agent shells faster instead of staying locked inside first-party UI.

Windows 365 for Agents is being pitched as a governed execution environment: Microsoft says Windows 365 for Agents is expanding in public preview as an environment for running agents with identity and policy governance attached. The pitch is straightforward: enterprise runtime for agents is being sold as its own substrate, not just as a workaround for computer use.

Google Gemma is showing that local models want a lane into software control too: Google Gemma demoed Gemma 4 E4B driving an iOS simulator directly with Argent, framed as local models handling software navigation autonomously. Local-first execution is moving another step beyond chat demos and into visible UI control loops.

Boris Cherny says the next version of Claude Code will add a /usage command that breaks token usage down by Skills, Agents, MCPs, and Plugins, with replies clarifying that downstream tokens are attributed back to the right component.

As agent workflows get longer, the key question stops being whether the model answered well and starts becoming which layer in the stack is silently inflating the session.

If you spend 20 minutes on this today, use the same mental model to audit any agent stack you run, even outside Claude Code:

MINI SCORECARD:

HOW TO INTERPRET: If Cost visibility is still low, do not blame the model first. Very often the thing making a session expensive is not the capability layer, but the orchestration layer expanding without anyone noticing.

Read yesterday’s One Practical Use of AI: Use Google Managed Agents to package an agent workflow into a callable runtime

Last week, Viktor wrote a campaign brief, built a landing page, opened a pull request, generated a board-ready PDF from live Stripe data, and sent a follow-up email to a churned customer. All from Slack. Same colleague that also pulls your reports and monitors your dashboards. 5,700+ teams. 3,000+ integrations.

Start free. $100 in credits →

This Axios piece is worth reading because it compresses capability, compute economics, and policy pressure into one tight cycle instead of leaving each story isolated. If you want the clearest map of why AI suddenly looks like a multi-layer control stack, start here.

TechCrunch asks the other necessary question: even if Big Tech can assemble the full stack of agents, surfaces, and integrations, do mainstream users actually want to live inside that system? That makes this a useful counterweight to the infrastructure and policy-heavy pieces above.

The Week reads this White House move as a real policy pivot, from relatively hands-off rhetoric toward a willingness to scrutinize advanced models before release. It helps move the White House item out of bureaucratic noise and into the politics of release control.

Axios keeps one lens alive that today's issue needs: even with policy and runtime grabbing headlines, the AI race is still a race around distribution, scale, and cash flow. It is a useful reminder that the control stack only matters in the long run if it stays tied to real distribution power.