AI Token Pricing Backfires as Corporate Usage Costs Soar

The artificial intelligence industry's shift to token-based pricing is creating an unexpected crisis: AI is proving more expensive than the human labor it was meant to replace.

Major AI providers including OpenAI, Microsoft, and Anthropic recently abandoned flat subscription fees in favor of per-query charges. The repricing has sent corporate AI bills spiraling out of control, with one company reportedly spending half a billion dollars on Anthropic's Claude in a single month, according to The American Prospect.

The usage cap snapback

Companies that previously encouraged workers to integrate AI into every task are now imposing strict limits. Uber, Tesla, Meta, and Microsoft have all capped employee token usage. Palantir CEO Alex Karp stated earlier this month that "something has gone completely wrong" with the billing model.

The irony is stark: these are among AI's most vocal advocates, with several operating their own AI businesses. Tesla exempted only its in-house product, Grok, from usage caps. If these evangelists are scrambling to control costs, the implications for more skeptical enterprises are significant.

Why it matters

This isn't a technical problem that better models can solve. Unlike AI accuracy issues—which cost Ford billions after the company laid off engineers whose institutional knowledge proved irreplaceable—the cost-benefit equation is fundamental. No amount of improvement makes an inherently expensive tool economical if it costs more than the workers it replaces.

The repricing wasn't optional. OpenAI and Anthropic are losing money, with compute spending estimated at 70 percent of entire industry revenues. They needed better balance sheets before planned IPOs. But if corporate customers rebel against the new pricing, the business model collapses.

Broader economic implications

The four largest tech companies expect to spend more than $750 billion on AI infrastructure this year alone. Market valuations for AI firms and their suppliers—chipmakers and data center operators—depend on assumptions of perpetual growth. Capped usage contradicts that vision.

Warning signs are multiplying. Meta is attempting to launch a cloud computing business to utilize excess capacity, acknowledging it overbuilt infrastructure. SpaceX's recent IPO has flattened, and even its bonds aren't selling well. Data center construction pipelines were cut in half by late last year.

A leaked Treasury Department draft report compared AI investment risks to the dot-com bubble, but with greater potential impact on the broader economy. The report noted that any change in conditions could trigger crisis, and mass corporate rollback of AI usage would certainly qualify. The economics of data center construction, complete with questionable lending practices and debt accumulation, show parallels to the housing bubble.

Some market observers note certain valuation gauges now exceed levels seen before the 1929 collapse. Recent tech stock pullbacks, dismissed as technical corrections, are occurring with increasing frequency.

These details were first reported by The American Prospect.