Who Pays the AI Bill? Few Customers, Huge Bets, and How to Protect Yourself

In July I worked out what a token really costs. The result: cheap tokens don't make the bill smaller, they make bigger jobs affordable. In October I shared a number from my own project on LinkedIn. An AI research run cost me $218 on DeepSeek. The same token volume at OpenAI list prices would have cost $7,926. The post drew 123 comments, which I analyzed one by one (in German). The result: 26 positions, no consensus. Nobody knew whether AI would get cheaper or more expensive.

This time I'm not looking at the price but at the money behind it. Who pays for AI, how much, and how loyal are they? Who builds the data centers, and with what? And what does that mean for a company that uses AI day to day? The short version: the market rests on a few big spenders who can leave at any time, and on investments that are now larger than what the providers earn from operations. That is no reason to avoid AI. It is a reason to hedge.

Where the numbers come from. It started with a video making the rounds. I traced every number in it back to the original source. Some held, some were out of date, some attributed to the wrong place. Then I read the originals: Anthropic's IPO filing of 26 October, the quarterly SEC filings of the five largest cloud providers and of Nvidia, the Brookings paper on financing the buildout and its revised version, the NBER firm survey of almost 9,000 executives, Ramp's monthly payment data, and Eurostat and Bitkom for Germany. Every number below has a link. Where a number only comes from the press, I say so.

Part 1: The buyers

Wide but shallow

The best public source on what companies really spend on AI is the Ramp AI Index. Ramp sells corporate cards and bill pay in the US and tracks what more than 85,000 US businesses pay for AI through it. It's not a survey, it's actual payments. The limits: US only, skewed toward tech, and large enterprises with direct contracts are missing.

In September 2026, 55.9% of US businesses on Ramp paid for AI, up from 24.5% in July 2023. Lately the share has grown by only about half a point a month. Ramp says it is "at the top of the S-curve" (Ramp, October 2026). The new customers have largely been won.

Which leaves the question of how much each one pays.

$15.70 per employee per month is roughly one basic subscription. Adoption is wide but shallow. It is growing, though: a year earlier the median was $5 and the top 1% $2,515. AI now makes up 5.8% of all non-payroll business spend that runs through Ramp, up from 0.8% a year earlier (Ramp AI Index). Ramp calls the top 1% figure volatile and revises it often. The order of magnitude holds.

A few pay for almost everything

A second, independent source shows the same pattern. Economists around Nicholas Bloom and Steven Davis survey executives in eight countries about AI for the NBER, most recently almost 9,000 of them, including about 2,300 in Germany (NBER Working Paper 34836, October 2026 version).

1.8% of US firms account for half of all AI spending, 1.3% in the UK. The top 10% of US firms account for 90%. If you read about "average AI spend" in a press release, know that the average is made by very few firms.

A good story, not yet true

This is where the viral number comes in: 80% of OpenAI and Anthropic revenue comes from 1% of customers. It isn't quite true.

The number comes from Ramp's lead economist Ara Kharazian. Ramp phrases it more carefully: 80% of the spend on OpenAI and Anthropic that Ramp sees comes from 1% of firms (Ramp, October 2026). That covers Ramp's US business customers, not the providers' revenue. What a provider reports itself has been in Anthropic's IPO filing since this week.

The answer has two parts. Yes, revenue is concentrated: about 6,000 customers spending at least $100k a year brought 74% of revenue in the first half of 2026, out of about 40 million customers (Anthropic, Form S-1, pp. 142, 148). No, Anthropic doesn't depend on a few giants: no customer exceeded 10%, and the top five fell from 36% of revenue in Q1 to 21% in Q3 2026 (pp. 56, 143). The real dependence is elsewhere: 48% of revenue runs through the marketplaces of Amazon, Google and Microsoft (p. 56).

OpenAI looks different. More than a billion people use ChatGPT, about 50 million pay for it, roughly one in twenty. According to OpenAI, business revenue now exceeds consumer revenue. OpenAI publishes no audited figures; these numbers come from its own statements and the press.

I mention this because the stronger story here is the wrong one. If you rely on it, you make bad decisions, whichever way you lean.

Little loyalty

The most important point for anyone buying AI is already in the first chart: Anthropic had 16% of US businesses as customers in September 2025 and 43.8% a year later, more than OpenAI. Market share only moves that fast when customers are barely locked in.

The data confirm it from several sides:

• Switching: the share of firms switching AI provider hit a record 8% a month (three-month average) in September 2026. Ramp: "It could be an early sign that AI models are becoming commodities" (Ramp, 6 Oct 2026).
• Several providers: 79% of Anthropic's customers on Ramp also pay OpenAI. Among large companies, 81% use three or more model families, according to a16z (a16z, 2026, n = 100).
• Contracts: Anthropic writes in its own filing: "Many of our customer contracts allow customers to reduce usage or terminate with limited or no notice. Customers are not obligated to renew" (p. 42). And open-weight models "could exert pricing pressure on our offerings" (p. 41).

I'm one of those customers. My big research run went to DeepSeek, $218 instead of $7,926. The result was good enough for that task, and a factor of 36 isn't something you save every day.

It isn't just hobbyists. According to Reuters, the Federal Register website, run by the US National Archives, used a search tool for public comments built on Alibaba's Qwen model. It was taken down on 16 September 2026 (Reuters, 17 Sep 2026). The case shows both sides: how quickly even a government agency reaches for a cheap model, and how quickly that can turn into a risk.

For providers that's a problem. They sell something that works more and more like gasoline. You fill up wherever it's cheap and good enough.

Part 2: Price and payoff

More tokens, barely more money

In October Ramp reported "two straight months of virtually no growth in token spend" while volume kept rising. The blended price for OpenAI and Anthropic was $1.15 per million tokens in March 2026 and $0.56 at the end of October. Good for customers. For providers it means delivering more and more to earn the same. Your phone plan works the same way: ten years ago a few gigabytes, today unlimited data for the same price.

It also explains what providers brag about. In May 2026 Google reported more than 3.2 quadrillion tokens processed a month, up from 9.7 trillion two years earlier (Google, 19 May 2026). They report the volume. What a token earns on average is nowhere to be found.

A token is not a unit of measure

Tokens aren't comparable. A token is a piece of a word, about four letters. How many a task takes depends heavily on the model.

Artificial Analysis has every model solve the same test tasks. GPT-6.1 Sol and Claude Sonnet 5.5 have the same list price and similar scores. Yet per task, Sonnet costs almost eight times as much because it burns many times the tokens (Artificial Analysis, as of 30 Oct 2026). Conversely, Anthropic writes in its filing that Opus 5.5 cut cost per task by about 40% compared to its predecessor (p. 10).

So the only number that matters is: what does a finished task cost? Including failed attempts, re-prompts, agent steps and the time a person spends checking the result. That is where hedging starts.

The payoff is slow

Why does the median pay so little? One obvious answer: many companies don't see the payoff yet.

• NBER: over three years, 83% of executives see no effect of AI on productivity, 91% in Germany. The measured effect so far: plus 0.46% overall, plus 0.24% in Germany, cumulative over three years. For the next three years German firms expect plus 0.87%, US firms plus 2.33% (NBER Working Paper 34836).
• McKinsey: 37% of respondents say AI has had at least some effect on operating profit (EBIT), about the same as last year. Only about 6% get at least 5% of EBIT from AI (McKinsey, The State of AI in 2026, n = 1,719).
• Deloitte: 25% of AI-using companies have moved 40% or more of their pilots into production. 20% achieved revenue gains from AI, 74% had hoped to (Deloitte, State of AI in the Enterprise 2026, n = 3,235).

That doesn't mean AI does nothing. In software development, for example, many people have long seen the benefit. It means the extra output hasn't shown up in most books yet. A CFO reading this renews the contract. They don't raise it.

And Germany?

According to Eurostat, 26% of German firms with ten or more employees used AI in 2025, against an EU average of 20% (Eurostat). Other surveys find higher shares because they ask differently: Bitkom 57% among firms with 20 or more employees, ifo 54.5%. All agree on the gap between small and large.

More interesting is what German firms pay for. According to Bitkom, the big cost items are infrastructure (51%) and data preparation (50%). Licences are named by 21%, tokens by only 8% (Bitkom 2026, n = 603). The Bundesbank puts generative-AI spending among using firms at 1.2% of turnover (Bundesbank, Monthly Report March 2026). For mid-sized firms, the token price is rarely the biggest item. The cost sits in everything around it, and that's also where the dependence builds up.

Part 3: The bill

Spending more than comes in

Now to the other side. Who builds the data centers, and with what?

The five largest cloud providers spent $230bn combined in Q3 2026. Their operating cash flow was $199bn. On a trailing twelve-month basis the ratio rose from 41% at the end of 2023 to 95.5% (quarterly SEC filings of Alphabet, Microsoft, Amazon, Meta, Oracle). Amazon's free cash flow over twelve months is minus $42bn, which Amazon says "primarily" reflects investment in AI.

The gap is filled with outside money. Amazon raised $115bn of new debt in the first nine months of 2026, Alphabet $85bn of debt plus $43bn through new shares. In October Meta announced preferred stock and a share sale of about $37bn combined.

Bigger than the railroads

How big is this historically? Stijn Van Nieuwerburgh of Columbia University calculated it for the Brookings Papers on Economic Activity (Financing the AI Buildout, September 2026).

His scenario: 182.7 gigawatts of new data centers between 2025 and 2032, $10.3 trillion of investment, an average of 3.63% of US GDP a year. The railroad boom came to 2.24%, the fiber boom around 2000 to 1.10%. One difference from the railroads: tracks last for decades, graphics chips for a few years. The paper itself says GPUs may be obsolete "within three to six years" (p. 24).

The debt you don't see

A large part of the obligations isn't on the balance sheets.

• Leases: Microsoft has signed $365bn of leases that haven't started yet, up from $106bn a year earlier. Meta has $356bn, up from $104bn at the end of 2025. With Amazon and Alphabet, that is almost $990bn that only shows up on the balance sheet once the buildings stand (lease notes of the quarterly filings).
• Special-purpose vehicles: Meta's Hyperion data center in Louisiana sits in a separate entity. According to the Brookings paper, about $27bn of its $30bn is debt-financed, and the bonds yield 6.58%, at least one percentage point above Meta's own bonds. The reason, per the paper: "not cheaper financing, but preservation of an asset-light corporate profile" (pp. 8 to 10). Meta itself puts its maximum loss exposure from such entities at $68.4bn.
• Guarantees: Nvidia guarantees up to $105bn of data-center leases for an affiliate of OpenAI (Nvidia, 10-Q). Nvidia writes that if OpenAI defaults, "we may assume the applicable lease".
• Money in circles: Microsoft holds about 24% of OpenAI and booked $10bn of revenue from OpenAI in the quarter to September 2026 (Microsoft, 10-Q). Nvidia put $42bn into equity stakes in the first half of its fiscal year, often in its own customers. Morgan Stanley expects about $800bn of the financing gap through 2028 to come from private credit.

The AI labs: rising revenue, bigger commitments

How are the companies that build the models doing? Anthropic's filing is the first audited answer, and it's better than its reputation.

Anthropic's revenue grew from $4.6bn for all of 2025 to $16.3bn in the first half of 2026 alone. In Q2 it posted its first operating profit, $974m. Compute costs fell from 251% of revenue in early 2025 to 48% (pp. 22, 157 f.). The huge net losses in the reports are mostly revaluations of convertible notes, not cash going out the door.

The catch lies in the future.

By the end of October Anthropic had committed to about $855bn of compute and infrastructure, my sum from the filing's notes (F-35, F-58, F-65). That's about 18 times annualized revenue. The largest items: TPU chip leases via Broadcom ($161bn), compute from SpaceXAI (up to $120bn), Google (at least $114bn), Amazon (at least $110bn). These commitments must be met "regardless of our revenue or utilization" (p. 141). Anthropic warns that margins "may compress" (p. 133) and that it may be required "to reduce our prices" (p. 60). For OpenAI the press reports a compute plan of about $856bn through 2030 against recent annual revenue of about $50bn, and an operating loss of $12.3bn in Q2 2026 (Financial Times, Wall Street Journal, Reuters; not regulatory filings).

The buildout bill

What does all of this have to earn? Van Nieuwerburgh calculates it like a mortgage: you don't just pay back the price of the house, you pay the interest too. With a 10% required return, a 50% operating margin and six-year server life, the data centers built through 2032 need about $3.7 trillion of revenue a year from then on. That's 9.2% of expected US GDP in 2032 (p. 26).

For comparison: all of US manufacturing came to 9.5% of GDP in 2025, finance and insurance to 8.7% (BEA via FRED). Revenue and value added aren't the same thing; the comparison shows the scale.

The paper sets this against today's revenue at OpenAI and Anthropic, about $100bn. To reach $3.7 trillion, revenue would need to grow about 80% a year until 2032 (p. 27). The comparison is imperfect, and that should be said: the $3.7 trillion is what data centers must earn, and part of the compute serves Google's, Meta's or Microsoft's own businesses such as ads and search. Still, it shows how far there is to go.

Most of it hinges on an assumption you won't find in any press release: how long the servers last.

If servers last three years instead of six, required revenue rises to about $6 trillion, 14.8% of GDP. On their books, Microsoft, Alphabet and Amazon use six years, Meta five and a half. All of them lengthened server lives in recent years, which raises profits. Only Amazon shortened them again in 2025 for part of its servers, to five years, "due to the increased pace of technology development, particularly in the area of artificial intelligence" (Amazon, 10-K 2025). Nvidia ships a new chip generation about every year.

Two caveats belong here. First, on 2 October Van Nieuwerburgh published a revised version with newer data: $3.55 trillion, or 8.8% of GDP (NBER w35865). Second, he writes himself that his numbers "do not imply that financial distress is imminent" (p. 13). At the conference he nonetheless called the buildout a source of systematic risk to the US economy (Brookings, transcript 25 Sep 2026).

What this means for prices

Which brings us back to the opening question: will AI get more expensive or cheaper? The numbers point both ways.

Cheaper, because competition is fierce, customers without lock-in move to the cheapest offer, and open models are catching up. According to Epoch AI, the price for a fixed level of performance fell 9 to 900 times a year, depending on the task (Epoch AI). Anthropic itself writes that some competitors sell "at zero or negative margins" (p. 40).

More expensive, because at some point someone has to earn back the trillions. That can happen through higher list prices, tighter usage limits in subscriptions, cheap models being retired, or surcharges for features you get for free today. There are first signs: OpenAI's new top model GPT-6 Astra costs $50 per million output tokens; GPT-5 cost $10 in 2025 (OpenAI, pricing). DeepSeek also raised prices in August 2026.

I don't know which way it goes. In the debate under my post, none of the 26 people knew, and neither did I. That's why I think "more expensive or cheaper?" is the wrong question. The right one is: is my company ready for both?

How to protect yourself

Hedging doesn't mean doing without AI. It means buying it so that a price jump, a retired model or a provider in trouble isn't an emergency. These nine steps cost little and work at any size.

1. Count in tasks, not tokens. Build a test set: 20 to 50 real tasks from your business, each with the result it should produce. Run it through three or four models once a quarter. Measure how many tasks were solved correctly, what each one cost, and how long checking took. Chart 5 shows why: same list price, a factor of eight in cost.

2. One socket, many plugs. Build your workflows so the model plugs in at exactly one point, through your own thin layer or a gateway. Prompts, test set and data live with you, not with the provider. Features only one provider offers are fine, as long as you use them on purpose and write down where.

3. Expensive models only for hard steps. Most workflows have many easy steps and a few hard ones. Sorting, extracting and summarizing are fine on small, cheap models. The most expensive model is often needed only for the one step where it really matters. My research run shows the scale: a factor of 36 between two providers for the same volume.

4. The shadow run. Send a small share of your real requests, say one in twenty, to a second model as well and compare the results. It costs a few percent more. In return you always know whether a switch would work. A project becomes a decision. The record switching rate shows many are already doing it.

5. A backup generator. Test an open-weight model you could run yourself or with a European host if needed. Not as a replacement for everything, but for the two or three workflows that must never stop. Check origin and license. The Federal Register case shows that a cheap model can also become a political problem.

6. Contracts like electricity, not like a lease. The providers are locking themselves in with hundreds of billions for years. You don't have to. Prefer monthly terms over prepaying a year of volume. Watch three things: how early is a model deprecated? What happens when prices change? How do you get your data, prompts and logs out?

7. Use the discounts that already exist. Many providers give a discount if you don't need the answer right away. At Anthropic, batch processing costs half, and repeated context gets much cheaper through prompt caching (Anthropic, pricing). Overnight reports, analyses, translations: much of it doesn't need to be done in seconds.

8. A budget with brakes. Set a hard limit per team and workflow, with alerts at 50 and 80%. And one metric everyone understands: cost per finished task, month over month. If it rises, that's a signal, whether it comes from the price or from a new model that uses more tokens.

9. Two notes. Write down what you'll do if the price of your main model triples. On a second note, write what you'll do if it drops to a third. The first is your emergency plan. The second is often the more interesting one: which tasks would suddenly be worth it? With both notes in hand, you don't have to answer the price question.

Conclusion

From the outside the AI market looks like a boom. From the inside it looks like a big building on a few pillars. More than half of US businesses pay for AI, but most pay little, and loyalty is low. The labs are growing fast; Anthropic now earns more than its compute costs, yet has locked in compute worth 18 times its annual revenue. The big cloud providers invest more than their business brings in from operations and fund the rest with bonds, shares and contracts off the balance sheet. And within six years the buildout has to support an industry the size of US manufacturing.

Nobody knows whether this ends well. Maybe it balances out, like earlier technologies that went through similar phases. In the railroad boom many investors lost their money, and the tracks stayed anyway. For a company using AI, the answer is simple: build so you don't have to guess. If you count in tasks, stay swappable and have a tested plan B, you benefit from falling prices and survive rising ones.

Sources

Spend, adoption, switching

• Ramp AI Index, revised monthly series July 2023 to September 2026 and weekly values to 23 Oct 2026. ramp.com
• Ramp: "Cracks in the AI Thesis, Part 3", 14 Oct 2026. ramp.com · "Businesses are changing AI model providers more often", 6 Oct 2026. ramp.com · June 2026, sample and method. ramp.com
• Ara Kharazian (Ramp) on the 80/1 figure, LinkedIn, September 2026. linkedin.com
• Yotzov, Barrero, Bloom, Bunn, Davis et al.: "Firm Data on AI", NBER Working Paper 34836, October 2026 version, Table 2, Figs. 6, 10, 11. nber.org
• a16z: "Leaders, gainers and unexpected winners in the enterprise AI arms race", 2026, n = 100. a16z.com · Menlo Ventures: "The State of Generative AI in the Enterprise", December 2025, n = 495. menlovc.com

The labs

• Anthropic PBC, Form S-1, filed 26 Oct 2026, in particular pp. 10, 22, 40 to 42, 56, 60, 133, 141 to 143, 148, 157 f., 170 f., F-35, F-58, F-65. sec.gov
• Microsoft, Form 10-Q, quarter to 30 Sep 2026, disclosures on OpenAI. sec.gov
• OpenAI revenue, users and compute plan: OpenAI statements and Financial Times, Wall Street Journal and Reuters, 2026. Not regulatory filings; treat with caution.

Payoff

• McKinsey/QuantumBlack: "The State of AI in 2026: On the Road to ROI", survey 4 May to 8 June 2026, n = 1,719. mckinsey.com
• Deloitte: "The State of AI in the Enterprise", January 2026, n = 3,235. deloitte.com

Germany

• Eurostat: Use of artificial intelligence in enterprises, 2025, table isoc_eb_ai. ec.europa.eu
• Bitkom: "Künstliche Intelligenz in der Wirtschaft 2026", n = 603. bitkom.org
• ifo: "More than half of companies in Germany use artificial intelligence", 5 Jun 2026. ifo.de
• Deutsche Bundesbank, Monthly Report March 2026. bundesbank.de

Tokens and prices

• Google, Sundar Pichai at I/O 2026, 19 May 2026. blog.google
• Artificial Analysis, Intelligence Index v5 and model pages, as of 30 Oct 2026. artificialanalysis.ai
• Epoch AI: "LLM inference prices have fallen rapidly but unequally across tasks". epoch.ai
• OpenAI, API pricing. developers.openai.com · Anthropic, pricing for batch processing and prompt caching. platform.claude.com
• My own bill, $218 vs $7,926: AI: more expensive or cheaper? (in German)
• Reuters, Courtney Rozen, 17 Sep 2026, on the Federal Register search tool (syndicated copy). wsau.com

Investment, debt, supply chain

• Quarterly filings (10-Q, 8-K) for Q3 2026: Alphabet · Microsoft · Amazon · Meta · Oracle. Earlier quarters via SEC XBRL data.
• Nvidia, Form 10-Q, quarters to 26 Apr 2026 and 26 Jul 2026. sec.gov · sec.gov
• Morgan Stanley, "Thoughts on the Market" on the financing gap. morganstanley.com

Buildout and economy

• Stijn Van Nieuwerburgh: "Financing the AI Buildout", Brookings Papers on Economic Activity, conference draft 4 Sep 2026, Tables 1, 5 to 7, Appendix D. brookings.edu · Revised NBER version w35865, 2 Oct 2026. nber.org · Conference transcript, 25 Sep 2026. brookings.edu
• BEA, GDP by Industry, value added as share of GDP, via FRED: Manufacturing · Finance and insurance · Real estate and rental

Images. Cover photo generated with Bria (FIBO model). Charts my own; data and calculations in scripts/data/ki-rechnung.json and scripts/gen-figs-ki-rechnung.mjs.