Data from the Ramp spending platform revealed a counter-intuitive phenomenon in the artificial intelligence market: companies are consuming growing quantities of AI tokens, but total spending on the model has stalled. Between August and September 2026, token spend grew just 1.5% — a chasm compared to the 42% growth recorded in July. And this happens just as Anthropic prepares for its IPO, at a moment when investors are pricing in the sustainability of AI companies' hyper-growth thesis.
The finding is particularly relevant because almost all corporate AI spend comes from a narrow group of firms: the top 1% of AI-using companies, those analysts call "AI-pilled," spend an average of $10,000 per employee per month. It's this group that drives the growth curve. When it stutters — even slightly, dipping below the $10,000-per-employee peak — the leverage effect is enough to halt revenue growth across the entire industry, according to Ara Kharazian, lead economist at Ramp.
But what's causing this slowdown? It's not a demand problem. On the contrary: usage volumes are rising. What's changing is the economics beneath the model. In September, OpenAI cut its API prices in half for the GPT-6 Sol and GPT-6 Luna models. Anthropic followed the same path, cutting prices on Fable 5.1 and Opus 5.5 — a reduction the company itself acknowledged in its IPO prospectus as responsible for a $600 million hit to revenue.
The result is a situation that Akshay Kothari, Notion co-founder, summed up with surgical precision in internal data published in October: "tokens go brrr, costs go ahhh." Tokens flow in growing volumes, but costs for companies decline.
This phenomenon has profound implications for frontier AI labs' valuation thesis. Anthropic, for instance, reported an operating profit of $974 million in Q2 2026, but the sequential sales growth rate fell from 144% in Q2 to just 47% in Q3 — a deceleration that concerns investors pricing a $1.5 to $2 trillion valuation. If prices keep falling and companies are getting progressively better at managing AI costs, revenue may not track the speed at which volumes grow.
Corporate maturity is also changing how spending is distributed. According to Cloudera, 89% of organizations scaled back or redesigned their AI plans in 2026 due to cost volatility. 45% migrated workloads between different providers to find better prices. And the LLM Token Expenditure Index shows that the price per million tokens fell from over $2 in June to $0.95 in October.
This is, in many ways, a healthy sign of market maturity. When prices fall and efficiency increases, customers win. But for an ecosystem built on the narrative of exponential hyper-growth, the message is uncomfortable: you can use more AI, but for less. And the capital markets won't ignore that distinction for much longer.
The data also reveals a deeper structural shift in how AI value is being captured across the stack. Menlo Ventures' 2026 State of Enterprise AI report, published just days ago, found that enterprise AI deployment spending surged to $177 billion by September 2026 — up from $33 billion the previous year, a 5.4x increase. Every dollar invested in model training generated $1.70 in deployment revenue, a striking return that has fueled years of massive infrastructure bets. Agents, which now account for 80% of AI usage on platforms like OpenRouter, have fundamentally changed the economics: instead of paying per user, companies pay per task performed, creating a direct link between usage and cost that puts downward pressure on per-token pricing.
Yet the very forces that make AI more efficient are also creating the spending plateau Ramp has documented. As agents become more capable and use cache reads — which cost just $0.20 per million tokens versus $4 for fresh input — the average blended price drops. Combined with the competitive pricing wars between OpenAI and Anthropic, this creates a scenario where usage scales while revenue stalls. For a market built on the promise of exponential growth, that distinction may well define the next chapter of the AI investment cycle.
Sources: TechCrunch, Menlo Ventures, IT Pro
✓ Independent sources cross-checked and verified before publishing
