Ramp Data Shows AI Token Spending Grew Just 1.5% in August-September as Anthropic IPO Valuation Talk Slips to $1.2 Trillion

Ramp's spending data shows that business spending on AI model tokens grew just 1.5% across August and September 2026, even as usage kept climbing, and the finding arrived on 6 November as prospective investors in Anthropic's initial public offering trimmed their valuation talk to $1.2 trillion to $1.3 trillion. The pairing matters because the revenue growth that justifies those valuations depends on customers paying more for AI, and Ramp's economist says customers are instead learning to "use more AI, but use it for less".

TechCrunch reported the Ramp analysis on Friday. The same day, Investing.com relayed a Bloomberg report that some prospective investors are now discussing a range of roughly $1.2 trillion to $1.3 trillion, down from the $1.4 trillion to $1.5 trillion investors had previously communicated to bankers and from figures as high as $2 trillion in early October.

What does the Ramp data show?

Ramp, the corporate-card and spend-management company, tracks what its customers pay AI providers. Its lead economist, Ara Kharazian, told TechCrunch that token spend rose 42% in July from the prior month, then grew only 1.5% across August and September combined. It re-accelerated in October, with token spend up 8.9% between 23 September and 23 October. Token spend means money paid for the units of text that models process.

Ramp's own AI Index gives the broader picture. In September, 55.9% of US businesses in its data paid for AI, up 0.6 points from August and 12.1 points from a year earlier. Anthropic was used by 43.8% of businesses and OpenAI by 40.6%. The median paying firm spent $15.70 per employee per month, up from $14.74 in August, while AI's share of non-payroll business spending slipped to 5.77% from 5.96%.

The token figures in the index come from a sample of Ramp customers that use its Token Spend Management product, so they may not represent all Ramp spending. For the week ending 23 October, the index shows total token volume up 37.1% from the 25 September comparison point while total token spending rose 8.9%. The blended price of OpenAI and Anthropic tokens fell 26.8% to $0.56 per million. OpenAI's own blended price dropped 42.4% to $0.46 and Anthropic's fell 15.7% to $0.61. Anthropic's share of token spend rose to 60.5% from 51.3%, and OpenAI's fell to 35.1% from 44.2%.

Why is spending flat while usage rises?

Kharazian's explanation is price, not appetite. He said the slowdown is not a demand problem and called the market a buyer's market in which the subsidise-then-raise-prices playbook of Big Tech will not work, because competition is too intense. TechCrunch reports that in September OpenAI cut API prices in half for GPT-6 Sol and GPT-6 Luna, and that Anthropic cut prices significantly for Fable 5.1 and Opus 5.5. Anthropic said in a recent filing that its cuts reduced revenue growth by about $600 million, according to TechCrunch.

Buyers are also getting better at cost control. Kharazian pointed to default settings that block expensive models on long tasks, per-employee budgets, and routers that send simple work to cheaper models. Meta and Nvidia are launching enterprise offerings built on cheap or open-source models, TechCrunch adds, alongside open-model companies in China and the United States.

Concentration is the second risk. "Almost the entire spend on AI is from venture-backed, high-growth tech companies," Kharazian said, and a slowdown in the top 1% of firms alone is, in his words, "sufficient to slow down revenue entirely" even with volumes rising. The median firm spends roughly $12 to $15 per employee per month, according to TechCrunch's account of his analysis, while the top 10% spend $600 to $1,000. The most intensive 1% had reached about $10,000 per employee at their summer peak and dipped slightly in August and September. Kharazian doubts that sectors such as doctors' offices or car dealerships will approach those levels.

How does this connect to the IPO?

The Bloomberg report, as summarised by Investing.com, ties the lower valuation talk to Thursday's technology selloff, in which the Nasdaq 100 fell 6.4%, erasing about $2.3 trillion of market value in what the outlet describes as its largest one-day drop on record. It also cites investor concern about the scale of AI infrastructure spending and its returns, possible political and regulatory opposition to new data centres after Democratic midterm gains, and the Cboe Volatility Index at its highest since April 2025. The Wall Street Journal reported on Thursday, per the same summary, that investors doubted Anthropic could reach $1.5 trillion, that one suggested a figure nearer $1 trillion, and that a smaller raise or a delay was possible.

A person familiar with the matter told Bloomberg that "no final decisions have been made, and discussions remain ongoing," and an Anthropic spokesperson declined to comment. Investing.com's own reading is that the lower range may indicate a repricing rather than a postponement.

Debt adds a second pressure. TechCrunch reported on 30 October, citing the Financial Times, that Blackstone, the anchor investor in a $63 billion deal to finance Anthropic's chips, told investors to expect a "high" junk rating after the IPO, having earlier expected investment grade before 2027. TechCrunch's Ramp article adds, citing the Financial Times, that OpenAI lowered its proposed valuation to $1.25 trillion from $1.4 trillion and cut its latest fundraising target by about $10 billion; that figure was not independently confirmed for this article.

What we don't know yet

Ramp's sample skews toward venture-backed technology companies, which cuts both ways: it captures the heaviest spenders but may say little about large enterprises. TechCrunch notes that the data comes from Ramp's own dataset and does not detail its composition or methodology. October's 8.9% rebound suggests the August and September lull may be a pause, and Kharazian's conclusion is conditional: "based on what we can observe today", AI spending has started to slow.

The open question is what replaces coding agents, the most commercially viable product of the past year and a half, as the next growth engine. OpenAI, SpaceXAI and Meta have released personal agents, but Kharazian is sceptical they will prove as economically productive. If they do not, falling prices and rising volumes would leave model makers with more usage to serve and less revenue per unit, which is the arithmetic IPO investors are now repricing.