ByteDance's $29.6bn AI loan: cheap money, offshore chips, and a bill running ahead of the revenue

The week's biggest money move in tech didn't happen on any exchange. ByteDance, the private parent of TikTok, quietly locked in a $29.6 billion loan — the largest it has ever arranged and the second-biggest dollar borrowing in Asia this year. No ticker trades on it, so the question a U.S. retail investor has to ask is the one nobody at a press conference will answer: why does a company this profitable need to borrow, and what does that tell us about the business that is going to spend the money?

Start with the mechanics, because they are informative in themselves. Nearly 30 banks signed up, coordinated by Citigroup and JPMorgan, on an unsecured three-year facility extendable to five. Chinese banks took more than 60% of it. The deal started life at a $20 billion target and was oversubscribed to roughly $30 billion. The pricing tells the same story: an opening margin of 68 basis points over SOFR, seventeen basis points tighter than the 85 the company paid on its 2024 offshore loan. This is not a nervous market rationing credit to a struggling borrower. This is the lending market competing to hand one of the world's most valuable companies a discount to build AI infrastructure.

The stated purpose is "general corporate purposes," which is financing-language for "we're not obligated to tell you." The reported substance is sharper: the proceeds are mainly meant to fund AI expansion outside China, with ByteDance acting as offtaker for data centers being built in Southeast Asia. That geography is the real story, and it is a chip story in disguise.

ByteDance cannot legally buy advanced Nvidia parts for anything it operates in China; U.S. export controls draw the line at the physical location of the silicon. So Chinese frontier labs have learned to shop around the checkpoint. ByteDance has been working with a Southeast Asian firm, Aolani Cloud, on plans to deploy roughly 500 Nvidia Blackwell systems, and it reportedly planned to spend up to $7 billion on Nvidia chips outside China in 2025. Alibaba and Tencent run the same play. The export-control regime's focus on ownership and location, not on where the compute is consumed, is precisely the seam this loan is funded to exploit. ByteDance isn't borrowing to buy Nvidia in China; it's borrowing to buy a workaround — a foreign compute base where the best silicon can legally sit.

Why borrow at all, when the company is one of the most profitable operations in technology? ByteDance generated roughly $186 billion in revenue in 2025, up about 20%, and was on a record pace for around $50 billion in profit late last year — before the AI splurge compressed the full-year result. The answer is that the spending plan keeps outgrowing even that cash machine. It entered 2026 planning about $23 billion of AI infrastructure, raised that by at least 25% to well over 200 billion yuan (roughly $29 billion) as memory prices climbed, and by May was reportedly weighing as much as $70 billion for the year — more than double 2025's spend — with ~$100 billion discussed for 2027. The $29.6 billion facility is, in effect, borrowing against a budget that a single year of cash flow can no longer fully cover.

There is a second reason the money is borrowed rather than repatriated, and it doubles as a domestic political signal. A proportionally larger share of ByteDance's budget is now going to Chinese-made chips — initial talks with Samsung about in-house parts targeted at 100,000+ units in 2026, per reports — both to comply with Beijing's directives and to blunt U.S. leverage. The overseas, dollar-funded buildout and the domestic, yuan-funded buildout are diverging into two different supply chains, and this loan is collateral on the dollar side of that split.

Now for the number that should govern every investor's read of the AI trade: the bill versus the revenue. ByteDance claims roughly $4 billion a year in AI revenue, the largest figure any Chinese firm has publicly stated — which, coming from a marketing-minded parent, is best treated as a ceiling, not a floor. The detail behind it is sobering. Doubao, its chatbot, has over 200 million daily users but reportedly pulls in less than $1 million a day, most of it e-commerce commissions rather than payments for intelligence. Its video model, Seedance, is the healthier leg, near $2 billion of annualized revenue at roughly 70% gross margins. Do the division and the gap is stark: an AI business earning billions against an annual AI capex program that ranges, depending on the report, from roughly $29 billion to $70 billion. The people placing the biggest AI bets are spending fifteen to twenty times what their AI units currently earn.

This is the same awkward arithmetic that hangs over the U.S. hyperscalers, and it is why ByteDance's private loan is worth a public investor's attention. On one side, the debt confirms durable, self-reinforcing demand for AI silicon, memory, and power — a Chinese giant with record cash flow borrowing cheap money to build compute offshore rather than slow the race is a bullish read on the supply chain that sells it accelerators and the energy that runs them. On the other, the gap between the capex and the revenue is a reminder that every one of these buildings is a cost booked today against returns that have not materialized. Lenders will happily fund the construction of the road. Whether the tolls cover the debt is a question no loan agreement can answer.