Listen to this episode on Spotify or watch it on YouTube.
Welcome back to The Cost of Intelligence, our comedy news show about the money behind AI. Each episode, Ben Sand reads the industry’s invoice and asks one question: are you getting more intelligence per dollar?
This week: Anthropic’s IPO is set to price on Wednesday. At the top of its range it would raise $62.5 billion, more than twice what Saudi Aramco’s IPO raised in 2019. The Financial Times reports that Jensen Huang has discussed a return as Nvidia’s CEO, and the Treasury Secretary says Nvidia should hire from outside. South Korea says Nvidia still plans up to 260,000 GPUs for the country. And in the Department of Token Efficiency, Hex took nearly 75% off the cost of each chat. Here’s the written version, with every source linked.
The tokens
The world runs about 470 trillion tokens a day, per tokensperday.com as of 16 November. That’s about 57,000 for every person on Earth, and the count is up a trillion since yesterday on the site’s own clock. Fifty-seven thousand each. Mine are mostly “make it shorter.”
Over on OpenRouter, usage of Tencent’s Hy4 grew 42% in a week (OpenRouter rankings). That’s just one platform. Somebody at Tencent gets a bonus. Somebody else gets the bill.
The landlord’s rate card
“Rent’s down, darling. My accountant needs a lie-down.”
On the Ornn Compute Price Index, which tracks actual rentals, as of the 15 November settle:
NVIDIA H100 settled at $3.06 an hour, down about 4% since yesterday but still up about 4% over the month (Ornn H100). A one-day discount on a one-month markup. The landlord still comes out ahead.
B200 settled at $9.66 an hour, unchanged since yesterday and up about 22% over the month (Ornn B200). Renters pay more for the same chip.
Chips, power and the work
Chips. Epoch counts about 34 million AI chips, measured in Nvidia H100s. That’s still its latest count, from 21 October (Epoch AI). Epoch tracks 96 data centres and estimates they hold about 44% of those chips (Epoch AI data centres). Somewhere out there, a chip is idle. And billing you.
Power. The data centres behind these chips draw about 30 gigawatts, our estimate from Epoch’s tracker. Last year, AI data centres used about half a percent of the world’s electricity, and about a tenth of a percent of all the energy we use. Both are our estimates, built from International Energy Agency figures (Our World in Data).
The work. Our name for it this week is SI: Stock-listed Intelligence. Now with a quiet period.
On our bet, SI does about 11.7% of the hours we work, about 41 minutes a day for everyone who still has a Monday. Workers themselves say AI saves them about 2% of their hours, per the St. Louis Fed’s July survey (FRED). That works out to about seven minutes a day, our estimate. SI would like a word with HR.
Economist James Bessen checked all 271 jobs in the 1950 US census. By 2010, automation had wiped out just one of them: elevator operator (Bessen, NBER). The only job that went down was the one that took you up.
These numbers are a work in progress. Got better data? Step boldly into our DMs.
The news: Nvidia
The Financial Times reports that Jensen Huang has discussed a return as Nvidia’s CEO with close confidants. CNBC carried the report on Sunday night (CNBC). Colette Kress remains interim CEO.
The petition to bring Huang back now says it has more than 160,000 signatures. The petition doesn’t verify its signers’ affiliations.
Treasury Secretary Scott Bessent told Fox News on Sunday that Nvidia should hire an outside chief executive (CNBC).
Nvidia reports earnings on Tuesday, 17 November. Expect the word “interim” to get a workout on that call.
Per-dollar verdict: your next GPU quote comes with a succession plan.
Korea
South Korea’s science minister says Nvidia still plans up to 260,000 GPUs for the country. Yonhap reported it on Friday, a day after the CEO change (Yonhap). The named partners are Samsung, SK Group, Hyundai Motor Group and Naver Cloud.
Nvidia first put that number on the table a year ago (NVIDIA). So far, it’s a plan on paper. A commitment to deploy is a save-the-date. The wedding is TBD.
Per-dollar verdict: “up to” is doing a lot of work. Ask any sale sign.
Anthropic’s IPO
Anthropic’s IPO is set to price on Wednesday. Trading is due to start Thursday, on the Nasdaq, under ANT (Anthropic; MSCI).
The price range tops out at $78 a share. At that price, Anthropic would raise about $62.5 billion from its own shares. That’s more than twice what Saudi Aramco’s IPO raised in 2019, about $29.4 billion (Aramco).
The prospectus is long. Your AI will summarize it. For a fee.
Per-dollar verdict: when your vendor has shareholders, your discount has competition.
The tier list
How many dollars of API usage you get per dollar of subscription, on our 16 November run of Adam Holter’s method with Real API Pricing. And they’re off!
Claude Max breaks first, at about $77 of API value per plan dollar.
Claude Ultra runs second at about $66, and Claude Pro is a neck behind in third at about $64. A Claude trifecta. The track is considering a name change.
Into the turn: Claude Max on Haiku takes fourth, and Google’s biggest Ultra plan is fifth, on Flash. Claude Ultra on Haiku is sixth at about $47, Claude Pro on Haiku seventh at about $46, and X Ultra eighth at $43.70 or better. The stewards are still reviewing that one.
Ninth and tenth: ChatGPT’s $200 and $100 Pro plans, both in the high twenties. OpenAI is racing itself.
Trailing the field, last of 44: Google AI Pro on Gemini Argon, at about $1.40. Argon is still looking for the starting gate.
Google’s pricing page says the Flash model behind fifth place doubles in price on 1 January (Gemini API pricing). Pay by the token, and you pay twice as much. Same horse. Twice the oats.
Department of Token Efficiency
Chainsaw fueled. Spreadsheets open.
Cut one: Hex, which makes a data agent (Hex).
Each chat with its agent cost $6.44 at its January peak. Now it’s $1.64. That’s nearly 75% cheaper.
Hex got there with a cheaper model and a leaner agent. On Hex’s own test, OpenAI’s Luna trails Opus, and costs a tenth as much, or less.
Hex also fixed expensive cache misses.
You know who lives in that chat log. Cheaper per chat? Wonderful. He’ll simply attend more chats. We might get rid of the goblin, but Jevons is here to stay.
The fine print: Hex’s own blog, measured from January’s peak. Token use grew almost a hundredfold in a year. No total bill.
Cut two: Modem swapped one database for another (Modem).
It dropped pgvector, an add-on for Postgres. Now it uses turbopuffer, a search database.
Modem says its whole storage bill fell 10%, even after paying for turbopuffer.
On day one, processor load on its main database fell 62%. Postgres finally got a four-day week.
The fine print: Modem’s own blog, rounded figures, no dollar amounts. The savings are database costs.
Cut three: alkemy runs an AI research agent (alkemy).
Nearly 40% of each prompt turned out to be tool instructions.
alkemy trimmed that and added caching, plus a few other fixes. In its tests, the average research brief now costs less than half as much.
The agent spent 40% of its prompt reading its own manual. Relatable.
The fine print: alkemy’s own blog. Test workloads only. No dollar figures.
A word from ourselves
Your bid sheet for the IPO window. Your guardrail for the price cliff you didn’t see coming. Your dictionary for the line item nobody can explain. The people who keep the landlord honest.
Strong Compute plugs into 27 compute clouds, from Lambda to Vultr to Cudo Compute, and 10 AI vendors, from Amazon Bedrock to OpenAI to Deepgram. Bring your own keys and see exactly where the GPU hours and tokens go.
Because somebody has to read the bill.
Try it at strongcompute.com
Keep your contexts short and your invoices shorter. See you tomorrow.



