Top companies now spend Rs 7 lakh per employee on AI

The top 1% of American companies spent a median of $7,400 per employee on AI in July, roughly Rs 7 lakh at current rates. The number is striking. What it did over the past two months is more interesting: almost nothing.

Ramp's June index put the same cohort at $7,449. July came in at $7,400. After a year of steep climbs, the heaviest corporate AI spenders have flattened out, and Ramp's lead economist Ara Kharazian titled his 12 August update "Cracks in the AI Thesis." The section carrying the spend figures is headed "Our latest data shows businesses are hitting their limit on AI spend."

The spending gap is enormous, and mostly static

AI investment remains extraordinarily concentrated. Against the top 1% median of $7,400 per employee per month, the top 10% of firms spent $650, or around Rs 62,000. The median company spent $11.95, roughly Rs 1,140, which is about the cost of a single enterprise seat on ChatGPT or Claude.

That is a gap of more than 600 to one between the most committed adopters and the typical firm. A small group is running AI across coding, support, finance and internal workflows. Most companies are still paying for a handful of subscriptions.

The gap also reflects workload type. Firms running agents, high volume automation and advanced coding tools consume far more compute than firms using AI for occasional drafting. A simple linear workflow that cost a few cents per interaction in 2023 can cost well over a dollar as an orchestrated agentic system in 2026.

The Fable 5 signal

The sharpest finding in the August index concerns Anthropic's newest flagship. One month after launch, Fable 5 accounted for just 6% of the tokens businesses bought from Anthropic and 11.4% of dollars spent on Anthropic models, despite being priced at roughly $10 per million tokens.

For comparison, OpenAI's flagship GPT-5.6 Sol made up 25% of OpenAI tokens and 23% of spend, at around half the price. In July, Fable 5 generated approximately 75% as much model attributed spend as GPT-5.6 Sol.

Kharazian's reading is that this establishes an upper bound on what businesses will pay for performance.

Fable 5 is the most capable model on the market and companies are largely declining to buy it at the asking price. His caveat: the token data comes from Ramp's spend management product, whose sample skews more technical than the broader index, so actual adoption may be lower still.

Why the caps came in

The flattening follows a visible pullback across large enterprises. The Financial Times reported on 19 June that Amazon, Walmart, Cisco, Uber and Meta had all introduced spending caps, discouraged wasteful use, or pushed staff toward cheaper models.

Uber burned through its entire 2026 AI budget by April and now caps employees at $1,500 per month per agentic coding tool, covering products such as Claude Code and Cursor. Walmart capped tokens on Code Puppy, its internal coding platform, after usage surged. Amazon and Meta both removed internal AI usage leaderboards after engineers began deploying agents to climb them, a practice the trade press nicknamed "tokenmaxxing."

The trigger is pricing structure. As Anthropic and OpenAI moved customers from flat subscriptions to token based billing, companies became directly exposed to the cost of every prompt and every automated workflow.

Speaking at OpenAI's enterprise event on 2 June, Sam Altman said cost had become the second most common complaint from enterprise customers, and quoted a line now circulating among them: the company spent its entire 2026 budget in Q1, can you make this more efficient. He noted that at the start of the year nobody raised cost at all, and that it had become a serious issue very suddenly.

Nvidia's Bryan Catanzaro has said that on his own team, compute now costs more than employees do.

Where the money is moving

Anthropic extended its lead in July. 43.5% of US businesses paid for subscriptions or tokens from Anthropic, up 1.1 percentage points month on month, while OpenAI rose 0.23 points to 39.7%. These are not market shares and do not sum to 100, since many firms pay both. xAI posted its fastest growth since July 2025, adding 0.94 points to reach 4%.

Model serving platforms, which provide access to open source and some Chinese developed models, reached 6.1% of AI using businesses, up 0.2 points. Kharazian's argument is that first time AI buyers are still going to the American labs, not to open source. But growth for OpenAI and Anthropic increasingly has to come from existing customers spending more, and those advanced spenders are the ones shifting toward cheaper models.

What this means for buyers

One caveat before drawing conclusions. Ramp's figures come from card and bill pay transactions across more than 70,000 US businesses on its own platform. Those firms skew AI friendly, Ramp sells AI cost monitoring software, and the data captures only paid corporate spend, missing free tools and anything running on personal accounts. This is a US benchmark, not an Indian one.

With that said, the practical implication for companies moving from pilots to deployment is that cost per task now matters as much as capability. Tracking spend against hours saved, quality gains and output volume is what separates a deployment worth scaling from one worth capping. And as Fable 5's reception suggests, the most powerful model available is frequently not the one worth paying for.