The AI Benefactor Making $13 Billion a Year Doesn't Train a Single Model

Deep Thoughts on AI and Aspirations —— ByteDance Deep Thinking Circle

Let’s start with some numbers. Tether made $13 billion in 2024 with a team of 150 people. That same year, OpenAI earned $3.7 billion in revenue but lost $5 billion, with over 3,000 employees; Anthropic earned $1 billion and also lost $5 billion. The combined losses of these two serious AI labs don’t even reach half of this stablecoin company’s annual profit. By per-capita output, the gap is 60-fold.

This company then began pouring money into AI, committing nearly $1 billion over two years. Its position is peculiar: one of the wealthiest backers in the entire AI industry, yet it doesn’t train a single model itself.

How the Money Printer Works: An Interest-Free Deposit

Tether’s business is remarkably simple when broken down. You exchange $1 for 1 USDT, it takes your dollar and buys U.S. Treasury bonds, keeps the interest, and you get nothing. Banks do the same thing but pay depositors interest; Tether doesn’t. With 150 people managing over $130 billion in Treasury bonds, it collected about $7 billion in interest alone in 2024.

This structure isn’t new in financial history—it’s float. Insurance premium reserves, credit card transaction float—essentially all “collect money first, deliver later, use it for free in between.” Float businesses share the same characteristics: they don’t require innovation, just scale and not screwing up. Originally, their relationship to AI was zero.

But float has an Achilles heel: returns track interest rates. When the Federal Reserve cuts rates, Treasury yields drop, and this money printer slows down. Lying back and earning $7 billion in 2024 won’t necessarily continue. The money’s still there, but the growth story stops working. This is the first reason for its pivot to AI.

The second reason is more straightforward: narrative positioning. Call yourself a stablecoin company and investors, media, and politicians barely glance your way; call yourself an AI, brain-computer interface, and humanoid robotics company, and you’re a tech leader. Even money printers need new stories, and AI is currently the only story with enough weight.

How It’s Placing Its Bets

Looking at the portfolio, this isn’t scattershot—it’s a vertical chain.

One move deserves special mention: open-sourcing the dataset. OpenAI and Anthropic’s training data are core secrets; Tether does the opposite, releasing what it claims is the world’s largest open-source training dataset. This isn’t charity—it’s trading zero marginal cost for ecosystem position, claiming the “open” persona in data while adding to its own narrative.

Using Cash Flow to Fund Frontier Tech: An Old Recipe

Before treating “a company unrelated to AI investing its profits in AI” as a curiosity, consider: this structure has appeared repeatedly in tech history. Bell Labs was funded for decades by telephone monopoly profits, producing discoveries that shouldn’t belong to any single company; Xerox PARC, funded by copier profits, invented the graphical interface and Ethernet but never profited from them. More recently, Tencent uses gaming profits to sustain research institutes and unprofitable foundational products.

This leads to my assessment: evaluating who can sustain long-term bets on a long-cycle technology has never been about funding amounts—it’s about cash flow structure. Whether there’s a form of cash inflow unrelated to R&D cycles with near-zero marginal cost determines a company’s upper limit of patience.

Fundraising has cycles; when markets cool, you tell stories and cut projects. Float has no cycle, only interest rates. This is why the most relaxed bettors in the AI race are those with money printers: OpenAI tells investors it won’t burn cash until around 2029, Anthropic around 2028—both still pitching to capital. Tether doesn’t need to pitch; the money’s already in pocket. If they’re right, it’s foresight; if wrong, it’s tuition. Either way, the core business remains unharmed.

The actionable insight for entrepreneurs and investors: when examining a company’s AI investment, first ask where the money comes from. Profit-funded, fundraising-funded, and debt-funded are three entirely different species, with vastly different patience and degrees of strategic distortion. Teams burning fundraising money must re-convince investors every 18 months—their AI strategy is inherently short-sighted. This isn’t a team problem; it’s the mechanics of capital structure.

Two Types of Money, and How Each One Dies

Money for doing AI and money for investing in AI are two different things. The former demands closed-loop business models, bringing loss pressure, fundraising rhythms, and valuation games; the latter only requires principal preservation, allowing purchases of compute, data, and options that won’t pay off for a decade. Tether occupies the latter position—its composure comes from structure, not superior vision.

But even good structures have boundaries. This money printer faces at least three risks. First, it has operated for ten years without a complete audit, with reserve transparency relying on self-disclosure—a vulnerability regulators could close at any time. Second, falling rates erode its core returns. Third, its nearly $1 billion AI investment is just a fraction of its $13 billion annual profit—“chalk it up to tuition if wrong” only works while the printer keeps running. If stablecoin regulation lands and reserve rules tighten, the tuition budget will immediately recalculate.

Finally, the unavoidable irony: the most centralized company in crypto—controlling issuance, self-reporting reserves, unaudited for a decade—now champions “decentralized AI, local execution, returning intelligence to individuals.” It’s somewhat like a casino owner teaching gambling addiction recovery. But irony doesn’t constitute negation. For the invested technical assets, money has no smell; for observers, don’t learn its slogans—learn its structure.

The real takeaway is this sequence: first build a cash flow machine unrelated to trending topics, then use the machine’s spare change to buy the most expensive options. Reverse the sequence—do AI first then find a business model—and you get today’s top labs losing $10 billion a year. AI’s real business model challenge hasn’t disappeared; it’s just been reframed by a company that isn’t in a hurry. It hasn’t solved “how to make money doing AI”—it’s solved “how to make AI losses not hurt.” These are two questions, and everyone entering the field should be clear about which one they’re solving.