In 2026, Uber reportedly spent its entire annual AI budget in four months, and 62% of organizations in one survey said unexpected AI cost spikes had changed strategic decisions. Meanwhile, the major AI vendors replaced flat monthly fees with metered, per-token billing.
I have seen this before. I worked at Ericsson through the boom of mobile voice and the first GPRS data services. Back then, the unit was the minute and the kilobyte. Today it is the token. The economics, and the frustration of the people paying, are remarkably similar.
In the early 2000s, a phone that used mobile data was barely usable. Today no one can imagine living without one. AI is on the same path, only faster, and how we pay for it will shape that path as much as the technology does.
Tokens are the new minutes
A token is a unit of work you pay for but cannot see. Few buyers can say how many tokens a task will consume, before or after it runs. The price list has many dimensions. A reasoning model decides how long to think, an agent decides how many steps a task needs, and success pulls usage into automated workflows nobody watches.
A price per token that falls every year does not help when the number of tokens per task rises faster. Budgets are set per year. Consumption is decided per request.
I’ve seen this before
I worked in telecom from 1996 to 2002 and saw this transformation from the inside. A late-1990s mobile bill priced every minute by time of day, destination network and roaming, and every SMS separately. Germans called it the Tarifdschungel, the tariff jungle.
Then data arrived. In September 2002, one German provider charged 29 cents per 10 kB of pay-as-you-go GPRS, roughly €30 per megabyte. Customers did not know how large a web page or an attachment was, just as today’s users do not know how many tokens an agent will spend. Nobody could predict that bill.
The operators were not trying to be unfair. They were trying to pay back an investment on a scale the industry had never seen: Germany’s six 3G licenses alone cost €50 billion in August 2000, before a single 3G antenna was built. The AI industry is in the same position. Google, Amazon, Microsoft and Meta plan about $725 billion of capital expenditure in 2026, and usage-based pricing is how they expect to earn it back.
How similar are the two pricing stories?
| Mobile | AI | |
|---|---|---|
| Unit | Minute, SMS, 10 kB block | Input, output and reasoning tokens |
| Hard to predict because | Users did not know how big a page or attachment was | Users do not know how many tokens an agent will spend |
| First flat rates | Voice flat rates from 2005 | Chat and coding subscriptions, 2023 to 2025 |
| The retreat | 2010: AT&T ends unlimited data, citing heavy users and network capacity | 2026: Claude Enterprise moves to a $20 base fee plus usage; Microsoft adds per-use charges to Copilot Cowork; OpenAI meters Codex; GitHub Copilot moves to token-based credits |
| Where customers pushed it | Flat monthly plans with premium options on top | Still to be determined |
Tariff comparison, early mobile and AI days
Complexity created an industry
Complex tariffs need complex billing, and that created a software category of its own. One of its stars came from near Frankfurt. LHS was founded in Germany in 1990 and became a leading provider of telecom billing software with its BSCS system. Several of my friends worked there. In 2000, Sema Group acquired LHS in a $4.7 billion stock deal, and Schlumberger acquired Sema in 2001. In 2004, the LHS software business was bought back by investors, including founder Hartmut Lademacher and General Atlantic, and it returned to the stock market in 2006. Ericsson acquired a majority stake in 2007, valuing LHS at about €310 million, completed full ownership in 2010 and phased out the LHS brand in 2011.
By then, billing had become infrastructure rather than a differentiator. No one treats BSCS as business-defining software anymore. Complexity does not buy prosperity. It creates a market for managing the complexity, and that market shrinks once buyers force prices to become simple.
AI pricing is in its 2010, not its 2017
Mobile pricing did not move to flat rates in a straight line. In 2005, E-Plus launched what it called Germany’s first voice flat rate. Then came a retreat: in June 2010, AT&T ended unlimited data, arguing that a small group of heavy users was straining the network while 65% of customers used under 200 MB a month. AI pricing is at exactly that point.
GitHub shows how far the retreat has gone. Telecom eventually moved toward flat rates; GitHub moved away from them. On June 1, 2026, Copilot replaced flat-rate premium requests with token-based credits at $0.01 each. Developers protested, with more than 400 comments and nearly 900 downvotes. There was no exodus. Two months later, Microsoft reported “stronger-than-expected GitHub Copilot consumption following the June business model change”, with Copilot revenue up more than 60% quarter over quarter.
Complaining is not sufficient to inflict pain. Customers protested in the comments and kept paying at the checkout. GitHub and Microsoft could make this change because their customers let them, and every other AI vendor watched. The vendors’ argument is the one AT&T made: capacity is scarce, and heavy users consume most of it. In telecom, that argument did not last.
What telecom put in, and what it got back
I was at Ericsson when this investment cycle turned. The public numbers show how it played out.
Operators built the networks and usage grew beyond anything imagined in 2000, but revenue did not follow. Most of the value went to customers, and to the devices and apps on top of the networks. Metering did not protect the investment. Competition and price pressure set the price, not the cost of the network.
What did telecom put in, and what did it get back?
| Period | What happened |
|---|---|
| 2000 | Ericsson’s record year: SEK 273 billion in sales and 105,000 employees, as operators raced to build mobile networks. |
| 2000 to 2002 | Operators cut network investment sharply. Deutsche Telekom’s share price fell from €103.40 in March 2000 to €8.42 in September 2002, and telecom stocks lost about $700 billion of market value in two years. |
| 2002 | Deutsche Telekom posted a €24.6 billion loss, then the largest in German corporate history, mostly from write-downs. |
| 2010 to 2015 | European mobile service revenue fell 14%, from about $205 billion to $175 billion. |
| Ten years to 2022 | European operators invested more than €500 billion in their networks. |
| 2019 to 2026 | Global mobile data traffic grew about sevenfold, from just over 30 to more than 220 exabytes a month. |
| 2024 | European operators invested €64.6 billion in a single year, while mobile revenue per user, adjusted for GDP, was lower than in 2015. |
Telecom industry investment progression
What AI is putting in now
The AI industry is now in its own 2000, at a much larger scale.
What is AI putting in now?
| Company | Investment in hardware and compute |
|---|---|
| Google, Amazon, Microsoft and Meta | About $725 billion of capital expenditure planned for 2026 |
| OpenAI | More than $400 billion of committed compute through 2028, including a five-year $300 billion deal with Oracle, and about $50 billion of compute spend in 2026 |
| Anthropic | A $50 billion US data center plan, 3.5 GW of Google TPU capacity from 2027, and up to $84.5 billion for SpaceX compute through 2029 |
| SpaceX, including xAI | The Colossus cluster in Memphis, approaching 2 GW, with plans for at least 1 million GPUs |
AI industry investment progression
The revenue is real, but not yet on the same scale. According to the Financial Times, Anthropic took in $16.8 billion in the third quarter of 2026, and OpenAI’s revenue is approaching, but not yet exceeding, its compute spend.
Most of that revenue is metered. By one estimate, about 45% of Anthropic’s revenue comes from per-token API usage and about 35% from enterprise contracts with committed spend. Claude Code adds about 10%, priced per seat plus usage. The monthly consumer plans most people associate with Claude bring in only about 8%. The flat rate is the shop window. The business is the meter.
Telecom operators expected their money back in 2000 too. The networks got built, the traffic came, and customers kept most of the value. Whether AI buyers repay this investment token by token will be decided by what they accept.
Customers voted with their money, and telecom fell in line
The retreat from unlimited data did not hold, and the incumbents did not move first. Challengers with good-enough networks sold simplicity: E-Plus in Germany, T-Mobile and Sprint in the US. Customers followed them, and in February 2017 Verizon brought back unlimited data at $80 a month. Pricing settled on a flat monthly fee, a generous allowance and premium services on top. None of it happened because operators wanted it.
AI has its challengers too. Open-weight models such as Kimi, DeepSeek, Qwen, GLM, Mistral and OpenAI’s gpt-oss run on specialist inference providers, European clouds or your own infrastructure. According to Epoch AI, they have trailed the closed frontier by about four months in 2026, at a fraction of the price: GLM 5.3 lists at $4.40 per million output tokens against $25 for Claude Opus 5. For summarizing, classifying, extracting and much of everyday coding, good enough is good enough.
So far, enterprises have barely used this lever. In Menlo Ventures’ survey, the open-source share of enterprise generative AI spending fell from 19% to 11%. Telecom was a giant, capital-intensive industry with very strong incumbents, and market pressure still simplified its pricing. I see no reason why AI should be different.
Vote with your wallet
Telecom pricing changed because buyers had alternatives and used them. AI buyers have more alternatives than mobile customers ever had, and an enterprise contract carries far more weight than a consumer. If you negotiate AI contracts, own the budget or set the strategy, these are the levers I would use:
- Start from the bill you already have. You know what you pay today. Price the same workloads with a competing frontier vendor, an open-weight model on a cheaper provider, or a smaller model for routine tasks. The same for less is either a saving or a stronger negotiating position.
- Buy outcomes, not tokens. Push for pricing per agent, per seat with a defined allowance, or a capped monthly amount, so the vendor shares the risk of overconsumption.
- Insist on caps and transparency. Spending limits and overage alerts should be standard terms. Cloud has FOCUS, an open billing standard; AI has none yet, and Anthropic and OpenAI are not among the initial backers of the Linux Foundation’s effort to create one. Demand standardized usage APIs and reporting in the contract.
- Keep the exit open. Make the model swappable. A credible alternative is your strongest negotiating position.
- Avoid long commitments at today’s prices. Token prices keep falling. A multi-year deal at a 2026 rate card locks in the vendor’s margin, not your savings.
- Reward simple pricing. When offers are close, choose the one you can predict. Vendors watch where the money goes.
The market decides, if you let it
Telecom once looked too big and too entrenched to change its pricing. It changed anyway, because the people paying refused to accept bills they could not predict. AI vendors make the same arguments today. They may be right about their costs, but pricing is set by what buyers accept.
If you are preparing an AI negotiation or rethinking next year’s AI budget, I’d like to talk with you about how to approach it. At CaeliQ, we built our platform to compare what each option will actually cost before you commit, across AI and cloud providers.
Sources
- Forbes: When flat-fee AI pricing ended (October 2026)
- Tom’s Hardware: Skyrocketing component prices push Big Tech capex to record $725 billion
- heise: 25 years of the UMTS auction
- Wikipedia: Telecoms crash
- Wikipedia: LHS Telekommunikation
- Ericsson Mobility Report: Mobile traffic update
- CNBC: Verizon brings back unlimited data (2017)
- Light Reading: Ericsson buys billing vendor LHS (2007)
- Linux Foundation: Tokenomics Foundation launch
- Crypto Briefing: Anthropic could pay SpaceX up to $84.5B for compute
- LinkedIn: Tokens are the new minutes (full source list)
Additional sources cited in the article include Economy.ac, MLQ, Microsoft FY2026 Q4 earnings, GetCoreTech, TeleGeography, Connect Europe, Ericsson Mobility Report, Presenc, Value Add VC, The Next Web, SiliconANGLE, HORIZONT, Channel Insider, DisCo, teltarif.de, and Epoch AI / Menlo Ventures survey figures.