Agents Move to Always-On Cloud Computers — and So Do the Bills

2026-09-25·6 min read·OKCodex Editorial

Agents Move to Always-On Cloud Computers — and So Do the Bills

Several signals this week point to the same shift: AI agents are turning from chat boxes into always-on execution environments in the cloud, while the compute, compliance and security costs behind them are becoming explicit.

1. Meta Muse: an Ubuntu cloud computer for every user

According to the-decoder, Meta gives every Muse user a free cloud computer running Ubuntu Linux, where they can install software, write code and browse the web. A process called Sentinel monitors sensitive actions outside the user's workspace, while users can inspect every file in the system. The report says Muse passed 500,000 users early on.

The implication: an agent's capability boundary is now identical to its operating-system permissions. Competition shifts from how well a model answers to what this machine can do for you — and who audits it. Sentinel-style side-channel monitoring plus full file inspectability is a pragmatic compromise: broad permissions, but with traceability.

2. Microsoft splits Copilot into three, and adds usage-based billing

Microsoft is splitting its Copilot app into Home, Code and a new agent called Autopilot. Built on OpenClaw, Autopilot runs continuously in the cloud, where it can monitor Teams channels and complete tasks on its own. Microsoft says both Autopilot and Code will use usage-based billing.

That departs from flat subscription pricing: an always-on agent's cost is tied to how long it runs and how many tools it calls, and vendors tend to move that uncertainty onto the bill. For buyers, the practical change is that agent budgeting shifts from seats to workload.

3. The other side of the bill: compute, cloud contracts and oversight

The cost pressure is not only on users. According to the-decoder, the NSA is already spending billions of dollars testing advanced AI models, mostly on computing power, and lawmakers expect full-scale AI oversight to cost tens of billions of dollars a year. The Pentagon, in a budget request, wants to spend $30.3 million over five years on an improved lie detector called Polygraph+ or Polygraph Next, focused on scoring algorithms that use AI and machine learning.

On infrastructure, Anthropic has committed $11.6 billion over seven years to Akamai's cloud infrastructure — a bet on CPUs that could grow to about $20 billion. In an unusual arrangement, Akamai is giving Anthropic a potential stake of up to 5% of its stock that grows as Anthropic spends more. In the same period, British AI neocloud Nscale secured $3.36 billion in convertible financing ahead of a US IPO, from investors including Third Point and Nvidia, to fuel its data center buildout.

4. Risk: insecure defaults are the main exposure surface

TechCrunch reports that some Supabase customers are publicly exposing reams of user data — a finding that highlights how AI-generated and vibe-coded apps can spill and expose data when not configured or secured properly.

Two countervailing signals appeared the same week. Google DeepMind researcher Robert O'Callahan quit, saying AI's current rate of change is far too high; he worked on chip design tools that helped make AI cheaper and faster, a contribution he says he can no longer justify, and adds that many colleagues share his concerns but rarely speak out. A federal appeals court also upheld the Pentagon's decision to bar Anthropic from military contracts: Defense Secretary Hegseth argues the company's safety restrictions could jeopardize military operations, while Anthropic says the designation has already cost it billions.

One caveat: the framing about model release pace comes from a TechCrunch podcast, which describes Meta Muse stealing the spotlight from OpenAI and Anthropic and mentions Anthropic rolling out Opus 5.5 and OpenAI's GPT-6 updates. These are statements made on the show and are not repeated here as verified product facts.

Conclusion and recommended actions

Manage agents as identities, not features. An always-on cloud computer means long-lived credentials and file access. Before launch, define the workspace boundary, a list of sensitive actions, and an audit-log retention period.
Move budgets to usage-based. Set usage caps, circuit breakers and per-task cost attribution so bills don't drift with autonomous runtime.
Put secure defaults first. The Supabase cases show exposure usually comes from default configuration, not from being breached. Disable public read/write by default on new projects and scan for exposed data before launch.
Watch compliance cost pass-through. Once oversight and defense procurement reach the billions, those costs eventually reach enterprise customers as pricing, regional availability or contract terms.

Agents Move to Always-On Cloud Computers — and So Do the Bills