Off-Balance-Sheet AI Arms Race

The New Financiers of Tech

You are a giant, world-dominating technology company with a market cap north of a trillion dollars and tens of billions in cash on your balance sheet. You have decided, out of a sense of "paranoia" and competitive fire, that you must spend even more billions—perhaps $72 billion this year alone—on an AI arms race. You need to build gigantic, power-hungry data centers, so many that you’re now signing deals to buy the entire output of nuclear power plants.

Where do you get the money?

An intuitive answer might be "from your giant pile of money," or maybe "from the public markets, which seem to love you." But that is not always the modern answer. The modern answer is that you go to private credit. Here is the Financial Times on Meta:

Meta is looking to raise $29bn to fund its all-in push into artificial intelligence, turning to private capital firms to finance its build out of data centres in the US.

Talks between the Instagram-owner and private credit investors have advanced, with several large players including Apollo Global Management, KKR, Brookfield, Carlyle and Pimco involved in the discussions, according to people familiar with the matter.

There is a simple reason for this, which is that Meta needs to spend a truly breathtaking amount of money to stay in the AI race. We have talked about this before: its latest Llama 4 model was underwhelming, and now it is on a frantic campaign to catch up by hiring top talent and, most importantly, building the infrastructure to power them.

But there is also a complicated financial reason, which is more fun. If Meta went out and borrowed $26 billion on its own, its public shareholders might get nervous. They might look at the balance sheet, see all that new borrowing, and think "gosh, that's a lot of new debt for a company whose main business is still showing people ads next to pictures of their friends' vacations."

So you do a trick. You don't borrow the money yourself. You create a new, separate company—a special purpose vehicle or joint venture—and have the private credit firms lend the money to that company. That new company then builds the data centers, and you agree to be its main, guaranteed tenant for the next couple of decades. The debt, crucially, is not on your balance sheet. It's on the new company's balance sheet. Your public investors are happy, and you get your AI factories.

This is a great deal for Meta, but why would private credit firms do it? Well, they have their own problem. Asset managers now own or have tie-ups with major insurers and annuity providers, and those insurers are sitting on trillions of dollars of customer money that needs to go somewhere. And that 'somewhere' needs to be a high-quality investment that regulators will approve of, but that also generates better returns than plain old government or corporate bonds.

So they have turned to these bespoke financings. And what could be a better bespoke financing than a brand-new building with a 20-year lease guaranteed by one of the richest companies in human history? The private credit fund isn't betting on the success of Llama 5; it's betting that Meta will be able to pay its rent. Which it probably will.

So you have this perfect symbiosis. Big Tech gets to fund its AI moonshots with off-balance-sheet financing that doesn't spook the public markets. And private credit gets a new supply of high-quality, infrastructure-like assets to invest in. Everybody wins. The only remaining question is whether anyone other than Nvidia will ever actually make a profit from the AI boom itself. But that’s a problem for another day.

The Data Curtain

For a while now, the big AI story has been a hardware story. The US controls the supply of the fanciest chips from Nvidia, which are made by TSMC, which in turn uses unique, bus-sized machines from ASML. The US strategy has been to use these chokepoints to keep the best hardware out of China. This is a simple, if fraught, way to wage a tech war. You can, in theory, count the chips.

But the AI race is not just about who has the best hardware; it’s about who has the best data. And that front of the war just got a lot more explicit. Here is Reuters:

Germany's data protection commissioner has asked Apple and Google to remove Chinese AI startup DeepSeek from their app stores in the country due to concerns about data protection, following a similar crackdown elsewhere.

Commissioner Meike Kamp said in a statement on Friday that she had made the request because DeepSeek illegally transfers users' personal data to China. … "Chinese authorities have far-reaching access rights to personal data within the sphere of influence of Chinese companies," she added.

This is a fascinating development. For one thing, it’s not Germany banning DeepSeek; it’s Germany asking Apple and Google to ban DeepSeek, which turns the App Store and Google Play into the de facto border control for the digital cold war. Being the world’s unwilling tech regulators seems like an uncomfortable, though perhaps profitable, position for Apple and Google to be in.

But more importantly, it shows that the lines of conflict are being redrawn around data. The hardware war was about preventing China from building powerful AI. The data war is about preventing China from using it on you. The concern is that every prompt you type into a Chinese-made AI app is another drop of fuel for its models and, potentially, another data point for its intelligence services.

This is, of course, a great marketing opportunity for Western tech companies. Apple, after arriving spectacularly late to the AI party, has made privacy its core pitch. Its "Apple Intelligence" is, for the most part, designed to run on your iPhone, not in some distant data center, precisely so you don’t have to worry about where your information is going. Anthropic, another leading AI lab, spends a lot of time publishing research on the ethics and "values" of its models to build user trust. The implicit sales pitch is: our AI might be expensive, but at least it won't snitch on you to the government.

The problem, as DeepSeek demonstrated earlier this year, is that the other side’s AI is very good and very, very cheap. The world was briefly excited by the prospect of a powerful, low-cost alternative to the expensive models from OpenAI and Google. But now it seems the price of that efficiency might be your personal data. The splinternet is no longer just about firewalls and blocked websites; it's about which AI you trust to read your emails. It’s a choice between two stacks, and increasingly, you have to pick one.

The Scoreboard

• AI: OpenAI Turns to Google’s AI Chips to Power Its Products, Source Says (Reuters)
• AI: Softbank CEO Says He’s ‘All In’ on OpenAI, Reveals He’s Long Wanted Microsoft’s Spot as Main Backer (CNBC)
• Semiconductor: Intel’s Top Strategy Officer to Depart This Month (Reuters)