A World Economic Forum commentary published on 10 April returned attention to the physical scale of the AI buildout. The underlying McKinsey analysispublished this month estimated that data centres would require $6.7 trillion in capital investment worldwide by 2030 to keep pace with demand. It attributed $5.2 trillion of that requirement to AI workloads and $1.5 trillion to traditional IT.

Those figures are a forecast built from assumptions about demand, supply and technology. They are not committed spending. Their usefulness lies elsewhere: they show how much physical and commercial machinery may sit beneath services that appear to be just software.

Two separate developments arrived alongside that forecast. A compromised maintainer account put malicious Axios packages on npm. Anthropic introduced a cyber model whose most striking vulnerability claims will take time to verify publicly.

The Forecast Begins below the Model

McKinsey's $6.7 trillion estimate covers land, buildings, power, cooling and computing hardware. These are long-lived commitments, often made before a model is selected or an application reaches production.

They also determine operating choices. A facility and its power arrangements shape availability. Hardware supply affects capacity and replacement. A region, provider, or specialist component can become a point of concentration even when the application is designed to use more than one model.

That does not make concentration a mistake. Some workloads justify it. But the trade should be visible when the capital and contracts are approved, while the organisation can still change the design or price the dependency honestly.

Axios Published Two Malicious Versions for about Three Hours

On 31 March, two malicious versions of the Axios HTTP client appeared on npm. The project's maintainer postmortem identified them as 1.14.1 and 0.30.4 and said they remained available for roughly three hours before removal.

The versions added a malicious dependency, [email protected], designed to install remote-access malware on Windows, macOS and Linux. The maintainers said an attacker published the releases through a compromised maintainer account. They responded by wiping the affected device, resetting credentials and beginning changes to the release process, including plans for short-lived, identity-based publishing credentials.

The publication window establishes exposure. It does not establish that every Axios user downloaded either version, that every download executed the malicious dependency, or that every execution produced a further compromise. Lockfiles, package-manager records, build logs and cached artefacts would be needed to distinguish those states for a particular environment.

The familiar package name offered no protection once publication authority had been taken. Nor would a general instruction to use a trusted registry have answered the question. The evidence belongs to the precise version and bytes that crossed the build path.

Anthropic Made Large Claims and Preserved a Route to Check Them

On 7 April, Anthropic introduced Claude Mythos Preview and Project Glasswing, an initiative involving security organisations including CrowdStrike. Anthropic said the model had found previously unknown vulnerabilities and demonstrated exploit paths during testing across open-source targets.

The announcement is provider evidence. Anthropic acknowledged that many of the vulnerabilities were not yet patched, which made the claims difficult for outsiders to verify immediately. It published cryptographic commitments intended to show the findings existed before later disclosure, and said professional human triagers reviewed the highest-severity cases before sending them to maintainers.

The public examples and described testing method do not independently validate the model's wider production performance.

The operational question is still promising. A model may widen the search for vulnerabilities and help specialists investigate them. But its output becomes an organisational decision only when someone can see the supporting evidence, apply an agreed threshold and own the action that follows. Human review is valuable when it is an accountable judgement, not a decorative click after an opaque result.

Final Perspective

Focus on the point where a forecast becomes a commitment. The headline number may change, but facilities, power arrangements and capacity contracts can narrow the recovery choices before an AI service has proved its value.

That is when you want a credible substitute demonstrated. If the data, capacity or commercial terms cannot move in the time the business needs, the second provider is still an intention rather than a recovery path.