Europe’s €30 Billion AI Compute Bet

Europe’s Gigafactory Gamble

Europe wants greater control over AI computing, but its planned gigafactories will still depend heavily on imported processors — and none will provide capacity any time soon.

That gap between ambition and delivery is what makes this proposal worth a closer look. On July 30, the European Commission opened a formal tender for up to seven AI gigafactories. It’s offering as much as €10 billion in EU and national public funding, intended to attract at least €20 billion in private investment.

Each facility is expected to contain at least 100,000 advanced AI processors and substantially exceed the capacity of Europe’s existing AI factories. That could eventually give European companies and researchers much larger regional infrastructure for training models and running sensitive workloads.

Eventually is the key word. Applications close on November 12. Awards are expected in early 2027, with construction beginning after that. The sites, energy arrangements, winning consortia, final financing and delivery schedules are all undecided. Part of the EU contribution also depends on a future long-term budget that hasn’t been finalised.

The Commission has letters of intent with AMD, Nvidia and Qualcomm, reinforcing Europe’s reliance on predominantly US-designed chips.

For organisations, the likely gain is more control over where important workloads run. My reading, though, is that operational sovereignty is more realistic than technological independence. The test will be whether Europe can finance, power and deliver these facilities while depending on a hardware supply chain it doesn’t control.

Chrome’s Security Surge

That European build-out is a long game. On systems already in widespread use, Google says AI has sharply accelerated Chrome security work.

The company reported that AI-assisted vulnerability discovery, triage and patch generation contributed to 1,072 security fixes across Chrome versions 149 and 150. Google says those two releases fixed more security bugs than the preceding 23 Chrome milestones combined.

It also says its systems prevented more than 20 vulnerabilities from reaching production in May, including at least one in its highest severity category. The security harnesses run on locked-down machines with network allowlists and restricted file access, rather than being given unrestricted internet connectivity.

There are limits to the headline figure. Google hasn’t published enough detail to compare the practical severity of all 1,072 fixes or establish how many flaws were found solely by AI. These are company-reported totals, not an independent measurement of risk removed.

Even so, the operational effect is credible. AI can help scale defensive work, but that pushes pressure downstream into validation and deployment. For users and IT teams, a discovered flaw isn’t fixed until the corrected build reaches the device and Chrome is relaunched. Faster patch production makes prompt, reliable updating more important, not less.

Anthropic’s Day in Court

Here’s a useful test of what happens when government procurement power runs into disputed technical claims.

At a federal court hearing on July 30, Judge Rita Lin said the US government’s evidentiary position had not improved in its attempt to designate Anthropic a supply-chain risk. She said she hadn’t seen evidence that Anthropic could alter a model after delivery or activate a remote kill switch.

A preliminary injunction issued in March continues temporarily to block the designation. Anthropic and the government are both seeking summary judgment, so the hearing didn’t produce a final ruling. There’s no announced decision date, and an appeal is expected whichever side loses.

For AI suppliers and government contractors, the case leaves a basic question unresolved: how much technical evidence must a procurement agency produce before it imposes a label with serious commercial consequences?

My assessment is that the label becomes vulnerable to political use if an agency can’t identify the supplier control it fears, show how that control could operate and connect it to an auditable risk. That doesn’t prevent governments from acting against genuine supply-chain threats. It does suggest they may need a stronger technical record before excluding a provider from public-sector work.

Nscale Moves Up the Stack

From government contracts, let’s shift to the companies trying to own more of the AI infrastructure stack.

Nscale has signed a definitive agreement to acquire Anyscale, the company behind Ray, an open-source platform for coordinating computing jobs across clusters. Anyscale also sells a managed service built around Ray, while Nscale operates GPU infrastructure.

After the transaction closes, Anyscale says its customers will gain access to Nscale’s GPU capacity. The potential benefit is a tighter connection between the software orchestrating distributed AI workloads and the hardware underneath them.

Anyscale says its managed platform will remain multi-cloud, and that open-source Ray will continue to be developed under the PyTorch Foundation. Those commitments matter, but the acquisition hasn’t closed and the companies haven’t published detailed integration, migration or pricing plans.

For developers, fewer seams between orchestration software and large GPU clusters could make demanding workloads easier to operate. The counterweight is commercial gravity. A platform can remain technically portable while its managed version becomes increasingly optimised around one infrastructure owner.

Ray’s foundation governance gives the open-source project some separation, but customers will need the eventual product and pricing details before they can judge whether Anyscale’s promised multi-cloud neutrality survives in practice.

Minnesota’s AI Image Law

The next dispute asks where legal responsibility should sit when an image generator can be used for intimate abuse.

On July 27, xAI filed a federal lawsuit seeking to block Minnesota’s prohibition on providing or operating AI nudification technology. The first-of-its-kind state law was scheduled to commence on August 1 and creates civil exposure for entities that provide access to, or operate, covered technology. Reported penalties can reach $500,000 per violation.

xAI says it doesn’t oppose banning the distribution of non-consensual intimate images. Its argument is that Minnesota’s provider-level prohibition is overbroad and lacks an adequate safe harbour for lawful tools and conduct. The company has also raised the possibility of restricting its service in Minnesota.

The court hasn’t ruled on the request, so the law could be upheld, narrowed or delayed. Its final scope remains uncertain.

For generative-image providers, the distinction is substantial. A rule focused on harmful outputs supports enforcement against abuse; one focused on the underlying tool may force state-specific controls, service withdrawal or acceptance of considerable liability.

The policy challenge is to place meaningful responsibility on providers without leaving loopholes that undermine victim protection or defining the technology so broadly that lawful image systems become unworkable. Minnesota’s case may become an early judicial guide to that boundary.

Siri’s Possible Subscription Cost

Consumer AI keeps encountering a quieter constraint: cloud inference has an ongoing cost, even when the device was expensive.

Apple chief executive Tim Cook said on July 30 that the company expects to offer some form of iCloud+ upgrade for customers who want to use its new AI-powered Siri heavily.

This isn’t a charge Apple has introduced. There’s no announced price, usage threshold, feature list or regional availability, and the next-generation Siri remains in beta ahead of wider availability expected in the northern autumn.

Still, the comment changes a reasonable assumption around Apple’s AI strategy. Buying a supported device may provide access to the assistant without guaranteeing that intensive use of its cloud-dependent capabilities is included indefinitely.

For customers considering new Apple hardware, the useful conclusion isn’t that a particular Siri subscription has been confirmed. It’s that headline AI features may increasingly carry recurring service costs after the hardware purchase. Apple’s eventual boundary between included use and paid capacity will determine whether the upgraded Siri feels like a feature of a premium device or another subscription attached to it.

Zoox Clears a Federal Hurdle

Now for an AI system where there’s no steering wheel waiting for a person to take over.

US regulators have approved Zoox to charge passengers for rides in its purpose-built robotaxis, which have no conventional driver controls. The federal approval covers deployment of up to 2,500 vehicles over two years.

Zoox has already offered free rides in Las Vegas and San Francisco. Charging passengers would move the service beyond demonstrations, but federal clearance isn’t permission to launch everywhere. The company still needs the relevant state and local approvals, and it hasn’t established where, when or at what scale commercial rides will begin.

There’s also a significant evidence gap. Independent safety researchers say the available crash and mileage data remain too limited to support reliable comparisons with human-driven vehicles. Regulatory clearance therefore shouldn’t be read as proof that the system is safer.

For the public, paid deployment will be a more demanding test of the purpose-built robotaxi model. Zoox must show that vehicles without human controls can operate reliably, handle unusual situations and earn permission to scale. Regulators are allowing that process to advance before strong comparative safety rates are available, so the commercial fleet may help produce the evidence that passengers and cities would ideally already have.

What Changes for You

One practical change is already reaching enterprise development teams, and its default setting deserves a look.

GitHub has extended enterprise-managed Copilot settings to the GitHub Copilot app and Copilot cloud agent. Administrators can use the same managed-settings.json mechanism to govern plugins, plugin marketplaces and supported prompt-bypass behaviour across more Copilot surfaces. GitHub has also introduced a separate policy controlling access to the app.

The notable detail is that the app-access policy defaults to Enabled everywhere. Organisations shouldn’t assume an existing restrictive posture automatically covers this new surface. Administrators may need to review the policy explicitly, while developers could encounter new access decisions after their employer updates its settings.

Enforcement also differs between clients. The cloud agent doesn’t support the interactive bypass-prompt controls available in the app, command-line interface and Visual Studio Code. One shared configuration file therefore doesn’t guarantee identical behaviour everywhere, and GitHub hasn’t said when the remaining cloud-agent controls will arrive.

These features require an enterprise Copilot plan and administrative access. For organisations already on that plan, central governance now reaches further into agentic coding. Testing each client separately is newly worth doing, because apparent policy consistency isn’t yet enforcement parity.