SoftBank announced plans to invest up to €75 billion to build data centers in France, a move that underscores the deepening infrastructure arms race fueling AI development.

The Japanese conglomerate said it aims to develop and operate up to 5 gigawatts of additional data center capacity in the country. That’s enough to power facilities equivalent to roughly a dozen hyperscale campuses, the kind that tech giants and cloud providers use to train large language models and serve AI applications at scale.

The investment scale is notable even by recent standards. While Microsoft, Google, and Amazon have been steadily expanding their global footprint, SoftBank’s commitment represents one of the largest single announcements focused on a European market. It’s also a signal that infrastructure spending isn’t cooling despite concerns about AI returns and questions around when model capabilities will translate to profit.

France has been courting these investments aggressively. The country offers competitive energy pricing, a relatively stable regulatory environment, and proximity to major European markets. SoftBank’s bet suggests it sees French infrastructure as a strategic wedge into broader European AI demand, where capacity constraints are already pricing out smaller players.

The timing is also revealing. As model training runs grow more expensive and inference workloads multiply, the companies best positioned to capture value are those who control the picks and shovels. SoftBank, which has stumbled on high-profile bets like WeWork and suffered from volatility in its Vision Fund portfolio, appears to be pivoting toward hard infrastructure plays where demand is clear and contracts are long-term.

Meanwhile, GitHub Copilot’s Token Billing Sparks Developer Backlash

In a separate development that highlights the growing pains of AI commercialization, Microsoft’s GitHub Copilot introduced token-based billing, and developers are not happy.

The change replaces Copilot’s flat subscription model with usage-based pricing tied to the number of tokens consumed by code completions and chat interactions. For many developers, this feels like a bait-and-switch. The flat rate made Copilot predictable and easy to expense. Token metering introduces uncertainty, especially for teams working in verbose languages or large codebases where completions can rack up quickly.

The backlash has been swift. Developers on social media called the move “a joke” and questioned whether the value Copilot provides justifies variable costs that could spiral. Some pointed out that the change comes after GitHub spent years evangelizing Copilot as a productivity multiplier, only to introduce a pricing structure that penalizes heavy use.

This isn’t just about affordability. It’s about trust. Developers adopted Copilot under one set of assumptions. Changing the deal midstream, particularly as competitors like Cursor and Codeium offer alternative models, risks alienating the user base GitHub spent years building.

The move also reflects broader industry tension. AI tools are expensive to run, and companies are under pressure to show that these products can be profitable, not just popular. But squeezing users too early or too aggressively can backfire, especially in developer tools where switching costs are low and sentiment spreads fast.

Meta Doubles Down on AI Hardware

Meta is reportedly working on an AI pendant, according to recent reports. The device would join the company’s Ray-Ban smart glasses and Quest headsets as part of a broader hardware strategy centered on ambient AI.

Details are scarce, but the pendant form factor suggests Meta is exploring wearables that can capture context throughout the day, audio snippets, location, activity, and feed that into AI models to provide proactive assistance. It’s a bet that the next wave of AI interaction won’t happen on screens but through devices you wear and forget about.

It’s also a risky play. Wearable AI has a graveyard of failures, from Google Glass to Humane’s AI Pin. The challenge isn’t just technical. It’s social. People are uncomfortable with always-on recording devices, especially when the value proposition is unclear.

Meta’s advantage is distribution. It already sells millions of Ray-Ban smart glasses, and it has the infrastructure to integrate AI features across its product line. But hardware is unforgiving. If the pendant doesn’t solve a real problem, or if it creeps people out, no amount of marketing will save it.

What It All Means

Taken together, these stories illustrate the same dynamic playing out across the AI industry. Infrastructure is expensive, monetization is hard, and the gap between hype and sustainable business models is widening.

SoftBank’s data center bet is a vote of confidence in long-term demand, but it’s also a reminder that AI’s winners may not be the model builders. They might be the companies that own the power, the racks, and the connectivity.

GitHub’s pricing controversy shows what happens when companies try to turn AI tools into profit centers before the value is fully proven. Developers will pay for tools that make them measurably better. They won’t pay unpredictable fees for something that feels like a tax.

And Meta’s pendant is another experiment in a long line of attempts to figure out what AI hardware should actually be. The company is placing multiple bets because it doesn’t know which one will land. That’s rational, but it’s also a sign that we’re still early. The form factor that makes AI indispensable hasn’t been invented yet.

The infrastructure is being built. The business models are being tested. The products are still being figured out. What’s clear is that the easy money phase is over. Now comes the hard part.