Ep. 20: Deedy Das: Token Volume Proves AI Demand. Also, His Own Report Says Tokens Aren't Dollars.
On Monday morning, a venture partner posted a milestone. Ten quadrillion tokens.
0:00 / 8:48
Ep. 20: Deedy Das: Token Volume Proves AI Demand. Also, His Own Report Says Tokens Aren't Dollars.
On Monday morning, Deedy Das posted a number. Ten quadrillion tokens a year were running through OpenRouter, the model gateway he had helped his firm back. Token volume there has been doubling roughly every ten and a half weeks. By the next evening, three hundred thousand people had seen it.
That night, Chamath Palihapitiya quote-posted it with one line: counting tokens is like "counting the atoms in a barrel of oil rather than measuring the useful energy it produced." By Tuesday lunchtime, Das had answered him directly, arguing that atoms in a barrel of oil are in fact proportional to the energy it produces.
About three hours after that reply, Menlo Ventures published its annual enterprise AI report, with Das as one of four authors. Its headline number is a different unit entirely: a hundred and seventy-seven billion dollars of enterprise spending, growing five point four times in a year, with two companies taking eighty-six per cent of what enterprises spend on model APIs.
This episode is about which of those two rulers a venture firm reaches for, and when.
Three kinds of material sit in this guide. The statements are public and timestamped. The figures come from the report itself, from Menlo's own portfolio announcements, and from financial press on Alphabet's quarter and the selloff that followed. The incentive reading is analysis, and it is not a claim about anyone's intent.
Statement Timeline
All times show the channel's display timezone first, followed by UTC. Engagement figures are platform snapshots captured on 4 November 2026 at about 11:00, not counters from the moment of posting.
- 2 November 2026, 11:22 (09:22 UTC): Das posts the milestone. "Congratulations to OpenRouter for crossing 10 QUADRILLION annual token run rate today! Token volumes continue to 2x every 10.5 weeks. In the last 9 weeks, OpenRouter processed more tokens than all of time before that." Captured: 302,238 views, 843 likes. 1
- 2 November 2026, 22:11 (20:11 UTC): Chamath Palihapitiya quote-posts it. "This is the equivalent of counting the atoms in a barrel of oil rather than measuring the useful energy it produced." Captured: 346,070 views, 839 likes, 120 bookmarks — it outran the post it criticised. 2
- 3 November 2026, 12:26 (10:26 UTC): Das quote-posts the critique back, with a chart and one line. "Atoms in a barrel of oil is remarkably proportional to useful energy it produces." Captured: 80,051 views, 1,169 likes. 3
- 3 November 2026, 16:00 (14:00 UTC): Menlo Ventures publishes 2026: The State of Enterprise AI, by Tim Tully, Joff Redfern, Deedy Das and Derek Xiao. Enterprise AI deployment spending is put at $176.8bn for the year to September, up from $33bn — 5.4x, against 3.9x a year earlier. 4
- 4 November 2026, 04:17 (02:17 UTC): Das posts the report's three takeaways, leading with dollars: spend up 5.4x to $177bn, 69 per cent to the top three labs, and 86 per cent of LLM API spend to Anthropic and OpenAI. Captured: 18,835 views. 5
- Not a position that evolved: the token claim and the dollar finding were not separated by a change of mind over months. Both were Das's position within roughly thirty hours, in two different media — a feed post and a 15-part report — and neither was retracted.
What the episode decodes
The report is the most useful public measurement of this market that exists, and it is also a document by an interested party. Both facts are true at once, and the second does not cancel the first.
Its central finding is that money and tokens have come apart. Open-source models handle about one in five tokens and earn about one in twenty-five dollars. Open-source token prices fell roughly 3.3 times in a year to $0.13 per million; the blended price of closed-model tokens fell 24 per cent to $0.70. The report says the average price per token keeps falling as agents account for more usage, because most of what an agent sends is cached text at a fraction of the fresh-input rate.
So the token count and the dollar count are not rival measures of the same thing. They measure work attempted and money paid, and the distance between them is a price collapse. Das's Monday post and his Tuesday rebuttal were both about the first measure. His report is organised around the second.
Two footnotes carry most of the incentive structure.
The first is the asterisk. Menlo marks its own portfolio companies in the report's tables, and the mark sits beside the winner in nearly every category it surveys — including Anthropic, which at 55 per cent of enterprise LLM API spend is the report's single largest beneficiary. Menlo has held Anthropic since 2023, when the company was pre-product and pre-revenue, led its Series D, and has joined every round since; the firm has described it as the largest investment in its history. 6
The second is the token data. The agents-versus-humans exhibit — 80 per cent of traffic on OpenRouter, roughly seven times human usage — comes from one gateway, and the report's own footnote says that gateway's users "skew more toward individual developers and startups than the enterprise buyers in Menlo's survey." The exhibit and the survey describe different populations, and the report says so on the page.
OpenRouter is also the Menlo portfolio company at the centre of the first post. Menlo seeded it in March 2025 out of the fund it runs with Anthropic, led its Series A in May 2025 with Matt Murphy joining the board and Das joining as a board observer, and doubled down in its Series B. On 19 August 2026, OpenRouter announced an agreement to be acquired by Stripe; Menlo's announcement put its token run rate at "4.5+ quadrillion" that day and called it one of the most important companies of the AI era. 7
Set the two announcements against each other and the Monday post reads as a continuation: 4.5 quadrillion in August, 10 quadrillion in November, doubling on schedule. The metric that chart plots is usage. OpenRouter does not earn on tokens; it collects a fee on the money spent through it. Even the company being measured in tokens is paid in dollars.
Where the record stops
The strongest objection to the report is not that Menlo owns things it writes about. It is the $1.70 figure, which is the one most likely to be repeated.
The report states that every dollar going into building models is generating $1.70 of revenue from putting them to work. The arithmetic behind it is $176.8bn of deployment spending against $102.5bn of development spending. Those are two pools of money, spent by different sets of companies, and the ratio between them is not a return on an investment. It is a measurement of how much more is being spent on using models than on training them. The report does not claim otherwise; the sentence invites the stronger reading, and the stronger reading does not survive the numbers beside it.
What the report cannot answer is whether the spending holds. It notes that 55 per cent of AI dollars came from new budgets rather than existing software or labour budgets, which is the finding that makes this look like market expansion rather than substitution. Whether that remains true in a quarter where buyers are under pressure is a question a survey cannot settle.
The market supplied one answer the week before publication. Alphabet reported on 28 October: revenue of $127.58bn against $128.22bn expected, its first miss on both lines in nearly four years, free cash flow of negative $9.07bn against roughly negative $2.70bn expected, and long-term debt up about 479 per cent in a year to $125.1bn. The shares fell more than 10 per cent. 8
By 30 October, the Global X Artificial Intelligence & Technology ETF was 4 per cent lower on the week and 8 per cent lower on the month, 15 per cent below its summer record, while the iShares Semiconductor ETF was 29 per cent below its recent high. 9
That is a measurement of a different thing again: not what enterprises spent, but what investors think the spending is worth. The report and the tape are not in contradiction. They cover different populations at different moments, and the gap between them is the same gap this episode is about.
What founders and allocators can take from it
For a founder building on these models: read the concentration before the growth. Two vendors hold 86 per cent of enterprise API spend, and the single largest application market is coding, at $34.0bn of a $77.3bn application market, with the labs' own tools taking $24.0bn of that. If your product sits in a category the labs are entering, the growth number does not describe your risk.
For an allocator: when a firm publishes market sizing, look for the marks that say which companies in the tables it owns. A report written by an interested party is more informative than one written by a neutral, as long as you can see the positions. The asterisk is the disclosure, and in this report it is doing real work.
For anyone reading a growth metric from now on: ask what the unit costs. Tokens are a measure of activity, and activity is what falls out of bed first when prices move. Dollars are a measure of willingness to pay, and they lag. The report's own growth rates tell the story at three scales: enterprise deployment spending up 5.4 times, closed-model token volume up 12.1 times, and open-source token volume up 22.5 times. The distance between volume and money is the price of intelligence coming down.
Which is the honest version of the bull case, and it is stronger than the one that fits in a screenshot.
References
- 1
- 2
- 3
- 4
- 5
- 6
- 7Menlo Ventures: Stripe to Acquire OpenRouter
menlovc.com
- 8
- 9
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.