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: the landlord worked through the weekend, and H100 rent jumped about 9% in a day. Your savings account just filed a complaint. Anthropic reportedly bought about $3.4 billion of gas generators. Bring your own power. It’s the new potluck. Nvidia’s board opened an independent investigation eight days before earnings. Somebody bring the board some snacks. Big Tech helped push US investment-grade bond sales to a record. The cloud just found out what an APR is. Claude still crowds the value tier list. And Anthropic put a cost auditor inside Claude Code. This time the call is coming from inside the house. Here’s the written version, with every source linked.
The tokens
Your weekend off did not apply to your agents.
You’ve been using about 463 trillion tokens a day, on tokensperday.com’s live estimate as of 9 November. That figure is a July estimate grown forward on the site’s own projection, so it’s up about a trillion since our last episode. The meter runs on a forecast. Your invoice runs on real money.
That works out to about 56,000 tokens per person per day, by our maths. You used yours before your first coffee.
The site’s own capacity check puts the ceiling at about 691 trillion tokens a day, as of July. Turns out the chips aren’t the bottleneck. Your budget is.
The landlord’s rate card
Uh-oh. The landlord has entered the chat.
“Sunday is a day of rest. For you. Not for my meter.”
On the Ornn Compute Price Index, which tracks actual rentals:
NVIDIA H100 settled at $3.37 an hour on the 8 November print, up about 9% from the $3.10 we read on Sunday’s show. That’s up about 12% on the week, and by more than a quarter since early August (Ornn H100). The landlord calls that a light week.
B200 settled at about $9.94 an hour. Priced like a mattress sale. Act now. That’s up less than 1% on the day, but about 33% on the month (Ornn B200).
“Up on the day, up on the week. Darling, I call it a winning streak.”
Our studio audience is CFOs. They’d like a moment.
Chips, power and the work
Chips. The world has about 34 million H100-equivalents, on Epoch AI’s chip-sales count, last updated 21 October. That’s one for about every 250 people, and somehow every one of them is always booked.
Power. AI’s data centres draw about 30 gigawatts, cooling and all. That’s our estimate, scaled up from Epoch’s data-centre tracker, and it’s unchanged since Friday. For scale, Our World in Data’s tally of the IEA’s figures puts AI at about 0.5% of the world’s electricity in 2025, and about 0.1% of all the energy humans use. Most of Ben’s goes to finding the mute button.
A small slice of the world’s power. Somehow a big slice of your invoice. And now some of it brings its own generator to the party (more on that below).
The work. On our bet, AI is doing about 11.5% of the hours we work, about 40 minutes a day for everyone on a payroll. Our trend has that share growing about 7% a month. So the bot works part-time and never asks for PTO.
The reality check: American workers say AI saves them about 2% of their hours, about seven minutes a day by our conversion, in the latest survey from July (Real-Time Population Survey via FRED). In late 2024, the same survey said about 1.4%, roughly five minutes. Two extra minutes in under two years. Progress at the speed of a loading bar. (FRED is a database, not a person.)
Our estimate counts all the work AI does; the survey asks workers how much time it saved them. Forty minutes for the machine, seven for you. The other thirty-three went to a meeting about AI.
The news: Anthropic buys its own power
Anthropic has bought about $3.4 billion of gas-fired generators, the Financial Times reported. That’s commitments on about 1.7 gigawatts of generators from OnePWR Solutions, a Texas equipment supplier. Reuters could not immediately verify the report, and neither Anthropic nor OnePWR responded to requests for comment (CNBC, via Reuters and the FT). The generators also had no comment.
To get that many, they must have gone to Costco.
That’s about $2 billion per gigawatt, by our maths. Suddenly your electric bill feels reasonable.
For scale, Caterpillar’s whole Power Generation business sold $3.42 billion last quarter (Caterpillar Q3 2026 results). One deal, about one Caterpillar quarter. Or, as DevilDogSaga put it on X:
Per-dollar verdict: the priciest kilowatt is the one you’re still waiting in line for.
The chip seller’s week
Nvidia’s board has opened an independent investigation. Every CFO’s least favorite pair of words.
The move follows New York Times reporting that Jensen Huang was the target of an alleged foreign intelligence effort (CNBC). These are reports, not findings: no one has been charged, and China’s foreign ministry called the original report “groundless speculation” (Euronews).
The board says independent directors will oversee the review with outside counsel, that it has no reason to believe national security or Nvidia’s secrets were compromised, and that it has confidence in Jensen and his team (Nvidia board statement). In a letter to staff on Sunday, Huang apologized, said he had been unfaithful to his wife, said he had fallen short of his responsibility to the company, and said he supports the investigation and will cooperate fully (Huang’s letter; CNBC).
A public apology, a vote of confidence, and outside counsel on speed dial.
Nvidia reports its quarter on 17 November (Nvidia). Its last guide told investors to expect about $108 billion in revenue, plus or minus 2% (Nvidia Q2 FY27 results). That’s before legal fees. Your GPU invoice is in there somewhere, waving.
The analysts have questions. The lawyers have a meter running. Both GPUs on today’s rate card are Nvidia chips, so your rate card holds its breath too.
Per-dollar verdict: maybe the clouds will start billing in six-minute increments too.
The money behind the buildings
Companies have sold $1.88 trillion of US investment-grade bonds this year through 4 November, a record, beating the $1.86 trillion set in 2020 (Reuters, via Investing.com). And it’s only November.
Amazon, Alphabet, Meta, Microsoft and Oracle sold more than $370 billion of it, roughly a fifth of the record. Somebody order more sign-here tabs.
In 2020, companies borrowed to survive. In 2026, they borrow to plug in.
All five are still investment grade, but Oracle sits one notch above junk at S&P, after a cut to BBB- in July (Computing). Oracle is sleeping on the top bunk.
Per-dollar verdict: the price of intelligence now includes interest.
The tier list
How much API value you get for every subscription dollar, on our list built from Adam Holter’s board and real API pricing. It ranks value, not what people use. The board was last checked on Friday.
Nine of the top ten seats belong to Claude. Claude almost booked the whole dinner table.
In order: Claude Max on Opus and Sonnet, about $77 of API value per subscription dollar; Claude Ultra on Opus and Sonnet, about $66; Claude Pro on Opus and Sonnet, about $64; Claude Max on Haiku, about $55; Google AI Ultra on Gemini Flash, about $49; Claude Ultra on Haiku, about $47; Claude Pro on Haiku, about $46 (Claude pricing).
Ranked by work per dollar instead, ChatGPT Plus on Luna buys the most: about 16 tasks per dollar, by our estimate. The dollar store just got competition.
Nine Claudes and one Gemini. The Claude family reunion has a plus-one.
And a reminder: Gemini 4 Argon’s API price doubles on 1 December, 22 days after the show, from $2 to $4 per million input tokens and $10 to $20 per million output (Google AI Studio pricing). Mark your calendar. Then mark it up.
Meanwhile, on X: on Friday night, Elon Musk posted that super intelligence leads to abundance. By Sunday morning, he said Grok Bot fixes for better token efficiency were underway, and he asked a user who keeps hitting his plan’s limit whether there’s an easy way to shift to per-token pricing. Friday: abundance. Sunday: per-token pricing. Monday: we read the invoice. Dogecoin Norway had the four-word review: fewer tokens, more brains.
Department of Token Efficiency
Somebody out there cut their AI bill. Let’s honor them the only way we know how. (Chainsaw. Watch the video.)
Today’s cost destroyer is Anthropic’s own cost auditor. A vendor helping you pay that vendor less. Frame this one.
It’s a command inside Claude Code called cost-optimize. Finally, a command CFOs can read. Anthropic’s Lance Martin ran it on four public benchmarks (Claude):
Retail customer service (tau2-bench): about 73% cheaper, with no change in the pass rate. Every dollar became about 27 cents. Same grade, smaller tuition.
Office documents (OfficeQA Pro): about 72% cheaper.
Legal (LegalBench): about 67% cheaper. The billable hour just felt a chill.
The trick is caching, batching, trimming and lower effort, not a new model. Lower effort: finally, an AI that gets Mondays. A sibling command, prompt-audit, goes after old “think step by step” scaffolding. Turns out the new model was already thinking. Somewhere, a goblin who lived off those pep talks has gone to haunt someone’s retry loop.
The fine print: these are the vendor’s own runs on public benchmarks, not a customer’s invoice. On a software-engineering test (SWE-bench Verified), it saved only about 24%. And yes, the vendor ships the auditor.
Side effects may include shorter prompts and a very quiet goblin. A bargain if your prompts were written for last year’s Claude.
Still: chainsaw approved.
A word from ourselves
Your thermostat for a rent freeze. Your crowd control for a cache stampede. Your stimulus check in this economy. The people who keep the landlord honest.
Strong Compute plugs into 21 compute clouds, from Oracle to Vultr to CUDO Compute, and 5 model providers, from Anthropic to Gemini to Fireworks. 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.