AI token spending from 50 cents to $56 threatens earnings

CEOs and CFOs have little visibility into the scale of AI token spending inside their organizations, and that blind spot will soon show up in earnings reports.

AI inference costs span a 100x range across providers — from 50 cents per million tokens for Chinese models to $56 for Anthropic's latest — creating a hidden expense that finance teams are not tracking.

"CEOs and the CFOs, in my opinion, probably have no idea how much tokenmaxxing is going on inside of their organizations," Chamath Palihapitiya, founder and chief executive officer of Social Capital and CEO of AI company 8090, told CNBC's "Squawk Box" on Tuesday. "I suspect what'll happen is one day you're going to have a miss, and EPS will be off by a few pennies, and the CEO will say to the CFO, 'What happened?'"

Palihapitiya's own company, 8090, is spending more than $10 million a year on AI inference, a figure he described as "very scary" for a small startup. He said many companies are "feeding this revenue ramp without getting any meaningful ROI from it." The warning echoes Palantir Technologies Inc. Chief Executive Officer Alex Karp, who earlier this month criticized OpenAI and Anthropic for their token-based pricing models, saying enterprises are "going to chillax and waste my time with tokens."

The issue is intensifying as inference costs have dropped 280-fold over the last two years, according to Deloitte, but usage has exploded even faster. Research from Ookla's Downdetector found high-signal disruption days on major AI platforms rose to 51 in the first quarter of 2026 from just six in the same period a year earlier, highlighting the operational risk of single-provider dependencies. For companies that have hardcoded integrations with expensive frontier models, the margin impact could be material.

A 100x Gap in the Price of Intelligence

Palihapitiya framed the pricing disparity using an oil barrel analogy: a barrel of intelligence — defined as 1 million tokens — costs $26 from OpenAI's standard model, $56 from Anthropic's latest, roughly $1 from Elon Musk's Grok, $1.50 from Meta Platforms Inc., $1 from Google's Gemini, and about 50 cents from Chinese providers. "If you've made it a bet very early around one of these folks that are selling extremely expensive barrels of intelligence, and you try to pass through the cost, you may run into some downstream difficulty," he said.

The divergence is accelerating. Research from Epoch AI found the price to reach a fixed performance level has fallen between nine times and 900 times per year depending on the benchmark, meaning the cheapest model for a given task in January is rarely the cheapest by April. Yet many enterprises remain locked into single-provider contracts, exposed to both pricing and availability risk.

The Infrastructure Blind Spot

Yi Shi, founder of FlashLabs, an AI agent software company, said most finance teams still model AI as a flat per-user cost — roughly 4 cents per chat, according to EY — when the real cost rises to about $1.20 per interaction once orchestration factors such as knowledge-base updates, agent evaluation and human-collaboration design are included. "If you are not budgeting at the tokens-per-task level, your forecasts may be off by an order of magnitude," Shi wrote.

Palihapitiya said the reckoning will come when a company misses earnings by a few pennies per share and traces the overrun to AI inference costs that were never captured in the budget. "That hasn't happened yet," he said, but the conditions are in place.

For investors, the risk is concentrated in companies that have aggressively adopted expensive frontier models without building the infrastructure to measure, route and cache token usage. Companies that have built multi-provider architectures with semantic caching and token-level observability — or that use lower-cost open-weight models, which Epoch AI estimates now lag the frontier by only four months — are better positioned. The divergence in AI cost management could become a distinguishing factor in margin performance across the technology sector in coming quarters.

This article is for informational purposes only and does not constitute investment advice.