AI tokens got nearly 27% cheaper in a month while usage soared, inference startup fal doubled to a $16 billion valuation, and Box is betting that retrieval — not generation — is where the money is.
Spending Stalled, Usage Soared
New data from Ramp, the corporate spending platform that tracks what more than 85,000 American businesses actually pay for, tells a strange story about the AI economy. In July, spending on AI tokens jumped 42% in a single month. Then, across August and September combined, it grew just 1.5% — basically flat. But that is not because companies stopped using AI. Far from it. The volume kept climbing: between late September and late October, token use rose 37.1%, while spending rose only 8.9%. The gap is the price.
The Price War Behind the Numbers
The blended price of a million tokens across OpenAI and Anthropic fell from 76 cents to 56 cents in four weeks — a drop of 26.8%. OpenAI cut its prices hardest, with its blended price falling 42.4% after it halved the API price of two of its GPT-6 models in September. Anthropic cut its prices too, by 15.7%, for Fable 5.1 and Opus 5.5. Ramp’s lead economist, Ara Kharazian, told TechCrunch the slowdown is not a demand problem. His point is simple: the price war has just made it cheaper to use AI. The median company in Ramp’s sample spends $12 to $15 per employee per month on AI tokens, while the top 1% of companies — the most AI-obsessed — spend just under $10,000 per employee per month.
Anthropic Gains Ground
The market share moved with the prices. Anthropic’s share of the token money in Ramp’s sample rose from 51.3% to 60.5% in a month, while OpenAI’s fell from 44.2% to 35.1%. So here is the tension at the heart of the week: everyone is using more AI than ever, and the companies selling it are charging less and less for each use. That is great for buyers. It is a much harder business for the sellers, and it helps explain why Wall Street spent this week asking whether the AI boom can pay for itself.
fal Doubles to $16 Billion
If selling tokens is getting cheaper, somebody still has to run the models, and investors are pouring money into the companies that do the running. The big one this week is fal, a San Francisco startup that hosts and runs AI models for video, image, and audio. It raised $1.3 billion in a Series F round led by Bessemer Venture Partners, with Nvidia, Sound Ventures, and T. Rowe Price joining in. The deal values fal at $16 billion, double the $8 billion valuation it had in April, and the company announced its Series E and Series F rounds together, with a combined $1.65 billion raised across the two.
What makes fal interesting is not just the size of the check — it is what fal does. It sits in the inference layer, the unglamorous work of running a trained model every time someone asks it to make something. Demand for that work is surging, especially from customers building on open source models who would rather rent the capacity than buy the hardware. Bloomberg reported that fal passed $800 million in annualized revenue earlier this year, so this is not a bet on a promise. It is a bet on a business that is already big and getting bigger, because every AI video, image, and song that gets generated has to run on somebody’s machines. The founders are Burkay Gur, who used to lead machine learning at Coinbase, and Gorkem Yurtseven, a former Amazon developer. They started fal in 2021, which means they went from a new company to a $16 billion valuation in about five years, by selling the picks and shovels of the generative media boom.
Box Wants Your AI to Find the Right File
Cheaper tokens and faster inference are only useful if the AI agent can find the right information — and inside most companies, it cannot, because the information lives in millions of documents. That is the problem Box is trying to solve. At its developer conference on Thursday, Box showed off a new feature called Box Content Retrieval. It builds Nvidia’s Nemotron embedding system into the core of Box, so agents can pull the most relevant chunks of a document instead of reading the whole file.
The pitch is speed and accuracy. Nvidia’s vice president of generative AI software, Kari Briski, told TechCrunch the system delivers:
- Four times faster query answers
- 20% fewer errors
- 90% fewer tokens used
The trick is metadata. Box already knows who can see each file and what is inside it, so the agent gets the right slice of the right document, with the permissions intact. The example Box gives is lawyers, who could find the right clause or precedent without handing the agent the entire case file. The feature rolls out early next year to Box’s top-tier enterprise customers, and the Nemotron models themselves, starting with Nemotron 3 Ultra, are already available inside Box AI for those customers.
The Price War Has a Cost
The price war has a cost, and Anthropic just put a number on it. The company said in a recent filing that its price cuts shaved about $600 million off its revenue growth in the third quarter. Cheaper tokens are a strategy, but they are not free. Meanwhile, the Nemotron 3 Ultra model is already live inside Box AI for top-tier customers, even though the full retrieval engine does not arrive until early next year. If you are already paying for that tier, the new model is there now.
What It All Adds Up To
This is the enterprise version of the same story we have been telling all week. The models are cheap now. The hard part is pointing them at your own data without breaking the rules about who can see what. Box is betting that retrieval, not generation, is where the money is. Tokens are cheaper, inference is booming, and the race is now about who can aim all that cheap intelligence at the right information.
Sources
- Ramp AI Index: AI token spending stalls as prices fall: TechCrunch — https://techcrunch.com/2026/11/06/ramp-data-shows-ai-spending-stalling-just-as-anthropic-nears-its-ipo/
- fal raises $1.3B Series F at $16B valuation: TechCrunch — https://techcrunch.com/2026/11/03/ai-inference-startup-fal-doubles-its-valuation-to-16b-as-demand-surges/
- Box taps Nvidia Nemotron for agent-first content retrieval: TechCrunch — https://techcrunch.com/2026/11/05/box-taps-nvidias-nemotron-for-new-agent-first-content-retrieval-system/
- Anthropic price cuts cost about $600M in Q3 revenue growth: TechCrunch — https://techcrunch.com/2026/11/06/ramp-data-shows-ai-spending-stalling-just-as-anthropic-nears-its-ipo/
- Nvidia Nemotron 3 Ultra now available in Box AI: TechCrunch — https://techcrunch.com/2026/11/05/box-taps-nvidias-nemotron-for-new-agent-first-content-retrieval-system/
Adapted from the November 8, 2026 episode of Daily Tech and AI.
Daily Tech and AI is AI-voiced and produced by Angelo Ruiz.
Leave a Reply