SpaceX will sell Google $30 billion of AI compute, and the deal shows who really owns the boom

Information Technology 5 min 80 sources

Google agreed to pay SpaceX $920 million a month for data-center compute, the Trump administration floated taking equity stakes in AI firms, chip stocks had their worst week in over a year, and OpenAI shipped a security mode against prompt-injection attacks.

Key takeaways

Google agreed to pay SpaceX $920 million a month — about $30 billion through 2029 — for AI computing power, showing that the scarce resource in the AI boom is compute, not ideas.
The Trump administration floated taking ownership stakes in AI firms like OpenAI, which would make the government both regulator and shareholder; its AI advisor is also leaving this month.
Chip stocks had their worst week in over a year, and a data-center lawsuit in Utah is a reminder that AI's real limits may be power, water, and local permits.

The biggest technology story this week is not a new chatbot. It is who controls the machines the chatbots run on. A $30 billion deal between SpaceX and Google, an unusual idea from the US government to take ownership in AI companies, and the sharpest drop in chip stocks in more than a year all point at the same thing: the AI boom is now a contest over compute — the raw computing power that trains and runs the models — and the money behind it.

Google rents its rival’s machines

SpaceX said Google will pay it $920 million a month for AI computing power, a contract that runs from October through June 2029 and could total about $30 billion [1]. The compute comes from data centers tied to xAI, the AI lab Elon Musk’s SpaceX owns [1]. So Google — which builds its own chips and runs some of the largest data centers on Earth — is renting capacity from a competitor’s AI operation.

Why would it do that? Because demand for compute is outrunning supply. Training and running large models needs vast banks of specialized chips, and even the biggest players cannot build data centers fast enough. The deal also lands as SpaceX prepares to sell shares to the public for the first time, turning a steady $920 million-a-month contract into a number investors can price [1].

The angle for anyone in the field: the scarce resource right now is not talent or data, it is the chips and the power to run them. A company that controls compute can rent it to its own rivals. Watch where the capacity contracts go — they say more about who is winning than any benchmark does.

Washington wants a cut of AI

President Trump said on Friday that he has spoken to AI companies about deals “where the American people can benefit from the success of AI” [2]. He named no company, but CNBC reported the administration has discussed taking an equity stake — an ownership share — in OpenAI, the maker of ChatGPT [2]. Some of that stake could seed a “Public Wealth Fund,” an idea OpenAI itself recently floated [2].

This is a striking shift. A government taking ownership in the companies it also regulates raises an obvious tension: the same officials would both police AI firms and profit from them. The details — how much, on what terms, with what say over decisions — are not public, and a separate report walks through what is still unknown [3].

The policy plumbing is in flux too. Sriram Krishnan, the venture capitalist serving as the White House’s AI advisor, said he will leave the administration at the end of June [4]. So the equity idea surfaces just as the person who helped shape AI policy heads for the door.

Chip stocks crack

The market felt all this. The Nasdaq fell about 4% on Friday — its worst day in over a year — as traders sold chip stocks hard, and Wall Street’s “fear gauge,” a measure of how much price swings investors expect, jumped back up [5]. After a long run where any company linked to AI compute only went up, the trade finally reversed.

A sell-off is not a verdict on the technology. But it is a reminder that the compute boom is being paid for with enormous bets, and bets can move both ways. If your work or your company’s plans assume cheap, abundant AI capacity, the price of that capacity is now something to track, not assume.

OpenAI hardens against a quiet attack

OpenAI shipped a feature called Lockdown Mode to defend against prompt-injection attacks [6]. A prompt injection is when an attacker hides instructions inside a webpage or document, and the AI — reading that content to help you — follows the hidden instructions instead of yours. It is one of the hardest problems in AI security because the model cannot easily tell a user’s request from a trap buried in the data it reads.

Lockdown Mode’s defense is to take away the AI’s reach: it disables live web browsing so the model can only see cached pages, and limits other ways untrusted content can flow in [6]. The trade-off is plain — a more locked-down assistant is a less capable one. If you build tools on top of these models, this is the security dependency to audit: any feature that lets a model read outside content is also a way in for an attacker.

A €20 billion telecom deal, and the water under the data centers

Two stories away from the AI headlines are worth carrying. In France, a consortium led by Bouygues Telecom, with Orange and Free, agreed to buy the operator SFR from Altice France for €20.35 billion, about $23.44 billion including debt [7]. It would shrink France’s mobile market from four big carriers to three — fewer competitors usually means less downward pressure on what customers pay.

And in Utah, residents and a non-profit sued to block Stratos, a planned AI data center backed by investor Kevin O’Leary [8]. The project was already cut down in size, with a commitment of water to the Great Salt Lake [8]. The fight is a preview of a constraint the compute boom keeps running into: data centers need land, power, and water, and the communities nearby get a vote. The bottleneck on AI may end up being a county permit, not a chip.

02 · Lesson · why it matters

Who owns the bottleneck owns the boom

When everyone needs the same scarce thing, the company that controls it can charge its own rivals — and control follows scarcity, not invention.

The strange part of the Google deal

Google is one of the most capable computing companies in the world. It designs its own chips. It runs data centers across the planet. And this week it agreed to pay SpaceX $920 million a month for AI computing power — renting from a competitor’s lab.

That sounds backwards until you ask what is actually scarce. The thing in short supply is not engineers, and not ideas. It is compute: the banks of specialized chips and the electricity to run them. When the thing everyone needs runs short, having it matters more than being clever.

Scarcity moves the power

Here is the pattern. In any system, when a lot of people need the same input and there is not enough of it, power shifts to whoever holds that input. Not to whoever is smartest. Not to whoever got there first. To whoever has the bottleneck — the narrow point everything has to pass through.

A bottleneck is the choke point in a chain. Oil is a bottleneck for transport. A single bridge is a bottleneck for a city’s traffic. Right now, compute is the bottleneck for AI. Every model, every company, every demo has to pass through it. So the holder of compute can name a price — even to its rivals — and they pay, because the alternative is to stop.

Why even rivals line up

You might expect a company to refuse to help a competitor. But when you control a bottleneck, your rival is also your customer. SpaceX selling compute to Google is not friendship. It is a toll. The same machines that power Musk’s own AI lab can be rented out by the hour, and a competitor’s need becomes a steady $920 million-a-month payment.

This is why a bottleneck is so valuable. It does not require you to win the race. It lets you charge everyone who is running it.

The government noticed too

The Trump administration’s idea this week — to take an ownership stake in AI companies — is the same instinct, one level up. If a handful of firms control the bottleneck of a whole new industry, then owning a slice of them is a way to sit at the choke point without building anything.

That is the appeal, and also the danger. The government would regulate these companies and profit from them at the same time. Whoever sits at a bottleneck has power over everyone downstream — and a regulator who is also an owner has two reasons to keep the choke point exactly where it is.

What scarcity makes visible

The chip-stock sell-off and the Utah data-center lawsuit are the same story from the other side. The market suddenly priced in a worry that compute is too expensive to keep buying at any cost. Utah residents pushed back because the data centers need their water and power. Both are reminders that a bottleneck is only as strong as the resources feeding it.

When a thing is abundant, no one thinks about who controls it. When it goes scarce, control becomes the whole game. The way to read this week is not “Google made a deal” or “the government wants a cut.” It is: find what is scarce, and you will find where the power went.

The pattern you can carry

This is not only about chips. It repeats everywhere a chain has a narrow point. The skill is not memorizing this week’s bottleneck — it is learning to look for the choke point first.

Ask of any system: what does everyone here need, and is there enough of it? When the answer is “not enough,” stop watching the people competing and watch the one holding the scarce thing. They are quieter, they rarely make the headline, and they are usually the ones being paid by everybody else.

03 · Lab · your turn

You Hold the Bottleneck

Rehearse how a scarce resource turns even your rivals into customers — and how that power vanishes when the scarcity does.