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: Ramp’s card data says the average price of a million tokens fell 21% in October, and AI spend on its cards still hit a new peak. The unit price fell. The units did not. Google launched Gemini 4 Hydrogen, the cheapest model in the Gemini 4 family, while its bigger siblings double in price on 1 December. Happy holidays, from the pricing page. Nvidia fell as much as 14% the day after Jensen Huang’s exit and closed down about 4%, while AMD and Broadcom each rose about 6%. Nvidia sneezed. AMD and Broadcom sold tissues. And the Department of Token Efficiency honours Neon and Mubit, who sent an agent’s routine calls to a small tuned model and took about 70% off the GPT-5.5 bill. Here’s the written version, with every source linked.
The tokens
AI reads and writes in tokens, chunks of words.
The world runs about 467 trillion tokens a day, on tokensperday.com’s live model as of 13 November (a July estimate, grown to today). That’s about 56,000 tokens for every person on Earth, more words than most of us say in two days, and up about a trillion since Thursday’s show. Somebody left the agent on.
The site grows its count about 7% a month, on a Goldman Sachs forecast. That’s slow for a startup. Not on a base that big.
The landlord’s rate card
“Unlucky day to be a tenant. Every day is.”
On the Ornn Compute Price Index, which tracks actual rentals, as of the 12 November settle:
NVIDIA H100 settled at $3.39 an hour, up about 4% since Thursday’s show and about 8% on the week (Ornn H100).
B200 settled at $9.78 an hour, flat since Thursday and up about 3% on the week (Ornn B200).
Both chips got a raise this week. The old one got the bigger one. Same office, different HR.
Over the month it flips: the B200 is up about 31%, the H100 about 13%. Turns out the new hire negotiated better.
Chips, power and the work
Chips. Epoch counts about 34 million H100s’ worth of AI chips, its last published count, from 21 October (Epoch AI). Call it one for every 250 people on Earth. Plenty of compute out there. You still need a reservation.
Power. AI data centres draw about 30 gigawatts. That’s our estimate, built on Epoch AI’s data-centre tracker: 19.6 gigawatts, including cooling, at sites covering about 44% of the world’s AI compute, scaled up and run at a typical load. The IEA-based count for last year is smaller: AI at about half a percent of the world’s electricity, which is about a tenth of a percent of all the energy we use (Our World in Data). Roughly Ben’s own energy level by Friday.
The work. Our name for it this week is SI: Second-sourced Intelligence. Intelligence with a backup supplier. Procurement’s favourite kind.
On our bet, SI does about 11.6% of the hours we work, about 40 minutes a day for every person with a job. Nobody’s sure which forty. Workers themselves say it saves them about 2% of their hours, per the St. Louis Fed’s latest survey, from July (FRED). That’s about seven minutes a day, by our maths. Forty minutes of work, seven minutes of credit. Welcome to the team, SI.
These numbers are a work in progress. Got better data? Step boldly into our DMs.
The news: Google’s new Gemini
Google launched Gemini 4 Hydrogen, the cheapest model in the Gemini 4 family (Google). Fitting name: the lightest element, and the lightest bill in the family.
Hydrogen: $1.50 per million tokens in, $7.50 out.
Gemini 4 Carbon and Argon: launched at $2 in and $10 out. That’s an introductory price that runs through November. On 1 December it doubles, to $4 in and $20 out (Gemini API pricing).
Argon is a noble gas. Doubling the price is less noble.
Developer Simon Willison says he finds the pricing “really confusing” (Simon Willison). If Simon Willison is confused, procurement should lie down.
Today, Hydrogen only saves you 25% on Carbon and Argon. After 1 December, it looks like a bargain. Hydrogen, Carbon, Argon: the only thing that isn’t elementary is the pricing.
Back in 2024, Andreessen Horowitz named falling AI prices LLMflation: for the same smarts, prices fall about tenfold a year (a16z). At the frontier, come December, the price tag goes the other way.
Per-dollar verdict: intelligence is cheaper when it’s less intelligent.
Ramp’s AI index
Ramp runs corporate cards, so every month it reads the AI line on a lot of bills. Ramp sees your AI bill before you do. Its October numbers (Ramp AI Index):
The average price of a million tokens fell 21% over the month.
AI spend on Ramp’s cards still rose about 1.7%, to a new peak.
William Stanley Jevons, what a paradox. You have a name for it too. It’s not printable.
And the spend is lopsided. The top 1% of firms spend about $10,200 a month on AI per employee. The median firm spends about $16, roughly 600 times less. Ten thousand a month per employee: better than an executive bathroom key.
Who gets paid? On Ramp’s cards, 44.8% of businesses pay Anthropic, up 1.1 points in a month, and 40.6% pay OpenAI, flat. SpaceXAI trails both at 6.3%, but it’s the fastest climber, up 2.2 points. Somebody approve its expense report.
Per-dollar verdict: the tokens went on sale. The bill didn’t.
Nvidia, the day after
Thursday’s show covered Jensen Huang stepping down as CEO (NVIDIA newsroom). Then the market weighed in.
Nvidia shares fell as much as 14% early in the day and closed down about 4.3%, at $166.19 (Investopedia; Motley Fool).
That leaves them about 30% below their record close of $239.24 on 6 October (NBC News).
AMD rose about 5.9% and Broadcom about 5.5%, while the Nasdaq slipped less than half a percent (Motley Fool; Investopedia).
That looks like rotation: money left the leader for the challengers. Nvidia had a bad day. Its rivals had a casting call.
Interim CEO Colette Kress stays on as CFO until a permanent successor is named. As one poster on X put it:
Earnings come Tuesday, 17 November.
Per-dollar verdict: second-sourced intelligence starts with second-sourced chips.
The tier list
How many dollars of API usage you get per dollar of subscription, on our 13 November run of Adam Holter’s method with Real API Pricing. It ranks value, not usage. And they’re off!
Claude Max is first out of the gates at about $77 of API value per plan dollar.
Then Claude Ultra, about $66, and Claude Pro, about $64 (Claude pricing).
Into the far turn: Claude Max on Haiku, Google AI Ultra on Flash, Claude Ultra on Haiku and Claude Pro on Haiku, around $46.
Trailing the whole field: Google AI Pro on Gemini Argon, at about $1.40. Yes, the same Argon from our top story. Barely better than just paying for the API. That’s an F.
Good news, though: on 1 December, Argon’s API price doubles. Then this plan looks twice as good.
Department of Token Efficiency
Somebody out there cut their AI bill. Let’s honour them the only way we know how. Safety goggles on, CFOs.
Today’s cost destroyers: Neon, the database company, and Mubit, a startup that tunes small models on an agent’s own work (Neon).
They ran an agent on GPT-5.5, then sent its routine calls to a small tuned model. The big model kept the hard ones. Which small model? Mubit doesn’t name it. It’s an open-weight model, trained on the agent’s own replies once a task has at least 25 good ones, and it has to pass an exam built from the agent’s own calls before it takes over a step. The agent hired an intern, trained on its own homework.
The GPT-5.5 bill went from $5.79 to $1.78, about 70% off.
It finished about 40% faster (26 seconds down to 15.6).
Accuracy held up too: 60% before, 65% after.
Cheaper, faster, and no dumber. Apparently you can pick three.
The fine print: it’s a vendor demo. Twenty questions, run three times. The small model’s serving cost and Mubit’s own fee aren’t in the numbers. Some assembly required. Intern sold separately.
Worth a test, if your agent does the same chores every day.
Still: chainsaw approved.
A word from ourselves
Your countdown clock for the scary date, 1 December. Your bargain hunter for the 21%. Your backup plan for the chip supplier. The people who keep the landlord honest.
Strong Compute plugs into 27 compute clouds, from AWS to Vultr to Baseten, and 10 AI vendors, from Gemini to Fireworks AI to xAI, + many more. 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 context short and your invoices shorter. See you next episode.