Here is a number that should give every fund manager underwriting the AI IPO of the decade a very uncomfortable night. The median U.S. company, tracked by Ramp's corporate card and bill-pay platform, spent $11.95 per employee on AI in July 2026. That is roughly the cost of a streaming subscription, which is not exactly the deep, structurally embedded spend that justifies a $2 trillion valuation. The top 1% of businesses spent a median $7,400 per employee, while the top 10% spent $650. The gap between the whales and the ocean floor is not a sign of a healthy, broad-based market. It is a sign that a tiny cohort of "AI-pilled" firms is carrying almost all the real revenue weight, while the rest of corporate America treats AI as a line-item luxury, not a mission-critical dependency. Ramp's data draws on its corporate card and bill-pay platform linked to Revelio Labs' workforce records across more than 21,000 U.S. firms, which makes it arguably the most grounded demand-side signal available, built from actual transactions rather than executive survey optimism.
Both Anthropic and OpenAI are racing toward public markets with the enterprise story as their anchor. Anthropic's annualized revenue crossed $47 billion in May 2026 and the company expects a 130% revenue surge that would deliver its first operating profit. OpenAI filed its own confidential S-1 with the SEC on June 8, 2026, at an $852 billion valuation set during a $122 billion equity round in March 2026. The pitch from both camps is that enterprise contracts are sticky, that integrating a foundation model deep into a company's stack raises switching costs, and that the revenue multipliers to come from agentic and API-driven usage will dwarf today's seat licenses. The Ramp data, updated this month, quietly demolishes that premise with the cold authority of receipts.
In traditional software, enterprise adoption often creates long-term commitment. AI model providers, by contrast, have minimal switching costs. That can enable surging adoption in a short window, but it is also a serious liability. The August Ramp AI Index, written by lead economist Ara Kharazian, covers the latest on business adoption of cheaper open-source models, a lookback on Anthropic's Fable 5 launch, why business takeup has been slower than expected, and what that means for the AI trade. The most striking data point: Anthropic's Fable 5 is considered the most powerful AI model on the market, but U.S. companies are barely buying it. According to Ramp data, Fable 5 accounts for only six percent of Anthropic tokens sold. Ramp economist Ara Kharazian attributes Fable 5's slow uptake to its price. The model costs about $10 per million input tokens and $50 per million output tokens, making it roughly twice as expensive as GPT-5.6 Sol or other Anthropic flagship models. Kharazian sees this as a new ceiling on what companies are willing to spend on AI, arguing that the extra performance simply is not worth the cost. For a company staking a $2 trillion IPO on the thesis that enterprise customers will pay a premium for frontier capability, that is a structurally painful data point.
The provider-switching pattern compounds the problem. The share of businesses using model-serving platforms, which provide access to open-source models and some Chinese-developed models, rose again in July to 6.1% of AI-using businesses. Growth has yet to meaningfully impact spending on OpenAI and Anthropic, but adoption of OpenAI and, to a lesser extent, Anthropic has slowed in recent months. First-time buyers on AI are still choosing American model companies, but more of their growth will have to come from existing businesses spending on AI, particularly the advanced spenders, and those businesses are increasingly spending on open-source. Ramp even launched its own product to accelerate this dynamic: Ramp launched Router on August 20, 2026, an AI model-routing service that lets users switch between large language models through one API. Router answers a problem that has become common for any company juggling multiple AI vendors: switching between models usually means rewriting code, managing separate contracts, and losing visibility into what is actually being spent. Ramp's pitch is that it removes that friction entirely, effectively institutionalizing the switching behavior that the AI labs most fear.
Skeptics of the bearish read will point to Anthropic's raw revenue trajectory. The run rate reached $65 billion at the end of July per Reuters, and in July, Anthropic extended its gains as the leader in business AI adoption, with 43.5% of U.S. businesses having paid for subscriptions or tokens from Anthropic, up 1.1 percentage points month-over-month. Those are not fake numbers. But there is a difference between aggregate revenue growing because more companies are adding a $12-per-month Anthropic subscription and aggregate revenue growing because enterprises are embedding Claude so deeply into operations that they cannot leave. The Ramp data captures both in the same bucket, and the distribution is wildly skewed. Both companies forecast large cash burn rates as they scale computing resources. Due to increasing price competition and enterprise pressure to reduce AI spending, those cash burn rates are likely to rise. Across the Nasdaq 100 universe, the average company trades at roughly 34 times trailing earnings and 25 times forward earnings. At those multiples, a $2 trillion Anthropic would need to post annual profits in the neighborhood of $59 billion to $79 billion to keep pace. The Fable 5 adoption data suggests that even Anthropic's most powerful product cannot yet command the pricing power that math demands.
When the receipt data from 70,000 companies shows the median enterprise paying less for AI each month than for a gym membership, the sticky-revenue story is not a thesis: it is a hope dressed in a term sheet.