For three years the AI story has been a story about models — who has the smartest one, who shipped it first, what it can do this month that it couldn't last. This week the story moved house. The action was in the physical and financial world the models depend on: the capital, the chips, the power grids and the fight over who pays for all of it. Three separate developments, read together, sketch the next phase.

1. The money is heading for the concrete

On Friday, Andreessen Horowitz — a firm that made its name on software — closed a $1.1 billion Machine Age Fund aimed at AI's physical layer: chips, memory, data centres, robots, and the cooling, power and real estate that keep them running. Its stated goal is to "open the throttle and accelerate the physical buildout of AI." When one of the most software-native investors on earth raises a ten-figure fund for cooling systems and electrical infrastructure, that is a signal about where the scarce resource now lives. The bottleneck in AI is no longer a clever idea; it's a gigawatt of power and a building to spend it in.

2. The software layer is picking sides

At the other end of the stack, Salesforce and Anthropic announced Claudeforce — embedding Claude across Salesforce as the default reasoning engine, with a "Salesforce in Claude" plugin and 37 prebuilt sales skills. The interesting part isn't the product; it's the strategic turn. Enterprise AI spent two years prizing model-agnostic flexibility — never get locked in, always be able to switch. Claudeforce goes the other way: one model, wired deep. And Marc Benioff used it to swat down the "SaaSpocalypse" thesis that AI agents will hollow out software firms, arguing frontier models depend on CRM data rather than replacing it. The buildout, in other words, isn't just hardware — it's commercial territory being claimed.

3. The backlash is now a battlefield

All that construction has a cost, and it is becoming a political fight. This week X's Safety team said it found roughly 200,000 suspected inauthentic Chinese accounts, about 200 of them pushing AI-generated cartoons against data centres — higher power bills, strained grids, greedy operators. X called it foreign interference in "American AI and energy policy." But the picture is tangled: the underlying grievances are real (US electricity prices rose about 6.9% year over year, per a Goldman Sachs analysis in early 2026, and towns have spent years fighting hyperscale projects), and X's owner, Elon Musk, builds data centres himself — so critics read the disclosure as conveniently self-serving. A real influence operation, a conflicted messenger, and public anger that is entirely home-grown, all in the same comment threads.

The through-line

The money (a16z), the territory (Claudeforce) and the politics (the datacentre backlash) are three faces of one shift: AI is becoming an infrastructure industry, with the costs, the land fights and the lobbying that implies. Even the week's biggest unconfirmed rumour fits the pattern — Nvidia is reported to be closing on a ~$12.9 billion deal for Hugging Face, the "GitHub of AI," though neither company has confirmed it. The models still make headlines. But the fight that will shape the next few years is over the concrete, the capital and the kilowatts — and this week you could see it starting.