# Who Pays for AI? Big Tech, Loss-Making Labs and Flat-Rate Subscriptions

Every time you ask an AI model something, someone pays for electricity and GPUs. It’s not a detail: it’s the main difference between generative AI and the software we’re used to, where serving one more user costs practically nothing.

I did a small piece of research on who pays this bill and how. What comes out is two teams playing the same game with opposite economic rules. The figures below come from financial statements, IPO documents and press investigations, with links at the bottom. When they’re estimates or leaked documents, I say so.

At a Glance

Two Teams, Two Wallets

Big Tech (Alphabet, Microsoft, Meta, Amazon) has something the labs don’t: a business that already makes money. Advertising, cloud, software licenses. For them AI is a huge cost center, but covered by profits that arrive anyway.

And the spending really is huge. After the April earnings, the 2026 infrastructure investment (CapEx) guidance looks roughly like this:

For Microsoft, 25 of those billions are just the rise in memory prices. And analysts are already modelling over a trillion in 2027.

The pure labs (OpenAI, Anthropic, xAI) have no advertising or legacy licenses to lean on. They live on subscriptions, APIs and above all funding rounds, and they rent their servers, often from Big Tech itself.

The Labs’ Accounts

OpenAI closed 2025 with about $13 billion in revenue and an operating loss of almost $21 billion; the reported net loss is around $39 billion. In Q1 2026 revenue tripled to $5.7 billion, but the operating loss was $9.3 billion. These are figures shared with shareholders and reported by the press, not audited statements.

Anthropic is the case where the numbers are easiest to misread, and in fact my research had them the wrong way round. According to the draft IPO prospectus leaked at the end of September:

• the 2025 operating loss is about $8 billion, over 91% of it compute costs;
• the net loss is about $42 billion, but around $34 billion of that is an accounting charge tied to investor funding, not money spent;
• compute infrastructure commitments reach $518 billion over the next ten years, about 80% of them non-cancellable.

The Financial Times also reported that Anthropic was on track to close the current quarter with an adjusted operating profit.

xAI, now inside SpaceX, lost about $6.4 billion in 2025 on $3.2 billion of revenue, and spent $12.7 billion on infrastructure. In Q2 2026 AI spending rose to about $15.8 billion, more than all of 2025; the improvement in its accounts comes mostly from renting the GPUs of the Colossus supercomputer to others, including Anthropic and Google.

1.65 Trillion You Don’t See

Then there’s a number that doesn’t appear in the balance sheets, but in their footnotes. According to an analysis by Nikkei Asia, Alphabet, Microsoft, Amazon, Meta and Oracle have about $1.65 trillion in off-balance-sheet AI-related obligations, more than the debt they report on their balance sheets (about $1.35 trillion).

It’s not money hidden in a drawer: it’s mostly long-term data center leases and future purchase commitments for servers and chips, often through special purpose vehicles. The practice is currently allowed by accounting rules, but its critics didn’t hesitate to mention Enron. Note that Nikkei directly verified only the figures for Meta (about $420 billion) and Oracle; the others are estimates.

The Flat Rate Doesn’t Hold

This is where it gets personal. In classic software a flat subscription works because one more user costs nothing. With AI every question costs, and costs more the harder it is: an answer with long reasoning, code or big documents uses far more tokens than a search.

The result is that a $20 or $200 monthly plan makes money on the people who barely use it and loses money on those who use it a lot. It’s not a hypothesis: in January 2025 Sam Altman wrote publicly that OpenAI was losing money on the $200 Pro plan, because people were using it much more than expected.

Someone coding with an AI agent all day is exactly the user a flat plan loses money on. And companies are reacting in two ways that anyone using these tools will already have noticed:

1. Tighter limits, sometimes silent: shorter context windows, shorter answers, lower usage caps at peak hours.
2. Hybrid models: a fixed fee plus metered usage beyond a threshold, like phone plans with data.

Why They Keep Spending

If these are the accounts, why does nobody slow down?

• Big Tech defends its territory. If Google didn’t put Gemini into search, some users would go and search elsewhere, and the advertising would follow them.
• The labs are betting on consolidation. The market won’t have room for everyone; whoever is among the few survivors will be able to raise prices. Meanwhile they need market share, and the stock market: SpaceX is already listed, Anthropic and OpenAI are getting ready.

What It Means for Those Who Use AI

My reading, as a daily user:

• Today’s prices are subsidized. A $20 subscription with access to frontier models is a land-grab price, not an equilibrium one. Sooner or later something changes: price, limits, or both.
• It’s worth knowing how much you actually use. Heavy agent users should compare subscription and API every now and then, and watch the limits, which change more often than the prices.
• Local models become insurance. They don’t replace frontier models, but for many tasks they’re enough, and you decide what they cost.
• Look at the dependencies. Anthropic rents capacity from Amazon, Google, Microsoft and even xAI; OpenAI from Microsoft and Oracle. Those who sell AI and those who sell servers are often the same companies, or partners.

Generative AI is the first big piece of software that costs something every time you use it. As long as someone is willing to pay the difference, we pay less than the cost. Worth remembering, before the difference starts being paid by someone else: us.

Sources

• 2026 CapEx: Fortune, Tom’s Hardware
• OpenAI: The Decoder, Techstrong.ai
• Anthropic: SiliconANGLE, Khaleej Times
• xAI: TechCrunch, Baxtel
• Off-balance-sheet obligations: The Next Web, The Wire
• ChatGPT Pro losing money: TechCrunch, Fortune