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: SemiAnalysis’s Dylan Patel says Elon capped Grok Bot because renting the chips out pays better. Your subscription just got outbid by a tenant. On election night, town after town voted against data centers. "Not in my backyard," they said, "and definitely not in my power bill." Memory prices keep climbing, and Apple says they’re squeezing its margin. And from the Department of Token Efficiency: Cloudflare set AI agents on its own code reviewer and took 68% off an average review. Here’s the written version, with every source linked.
The tokens
You’ve been using about 460 trillion tokens a day, per tokensperday.com as of 4 November. Our number changed: we now go with the tracker. Its figure is a July estimate (about 360 trillion a day) grown to today at the tracker’s rate of about 2.2x a year (methodology).
That works out to about 50,000 tokens per person per day.
The world’s chips could serve roughly three times the tracker’s total, by our math. The tracker’s own capacity check is lower: about 691 trillion tokens a day (tokensperday.com).
The landlord’s back
Agents need an office, and that’s where our landlords, the clouds, come in. The landlord’s back, and he brought a new lease.
On the Ornn Compute Price Index, which tracks actual rentals:
NVIDIA H100 settled at $3.53 an hour on 3 November, up about 4% on the day and 6% on the week (Ornn H100).
Crusoe’s on-demand H100 goes from $3.90 to $4.50 an hour on 1 December (Crusoe pricing).
Wow. This four-year-old chip isn’t going to a farm upstate. He’s getting a raise.
Our CFOs are thrilled. Look at them. Thrilled.
On the same index, the newer B200 settled at about $9.41, down about 3% on the week (Ornn B200). More than the federal minimum wage, but no lunch break.
Chips, power and the work
Chips. How many chips does the world offer? About 35 million H100-equivalents is our estimate, about one for every 240 people. That’s up about two million because we switched trackers. Before, we trended Epoch AI’s count of chips sold. Now we use Epoch’s satellite tracker of big AI data centers, which Epoch estimates covers about 45% of the world’s AI computing capacity, and we scale up from there.
Power. About 44 gigawatts is our estimate for whole sites. The chips alone draw about 20 GW, by our sum of Epoch’s chip-sales data. Yesterday we said about 51. Our old number guessed the watts per chip; the tracked sites come out lower.
Forty-four gigawatts is about 5 watts per person. Epoch says big AI sites typically draw 60–80% of capacity (Epoch AI), so AI’s real draw is about 30 GW.
The world’s electricity averaged about 3,600 GW last year (31,779 TWh, per Ember), making AI’s share under 1%, by our math. Over a year, that’s about one exajoule. Count the fuel it takes to make it, and AI is under half a percent of the energy humans use (more than 600 EJ a year, per the Energy Institute).
The work. On our bet, all that SI (subsidized intelligence, as we like to call it) is doing about 11% of the hours we work: about 40 minutes a day for everyone still clocking in.
The reality check: American workers say AI saves them about 2% of their hours (Real-Time Population Survey on FRED; FRED Blog). That’s about 10 minutes of an 8-hour day, by our math. About as much of the day as you spend telling people they’re on mute.
Our math counts the work AI does; the survey asks workers what time it saved them. So who’s right? Our math at 40 minutes, or your coworkers saying it’s only saving them 10? Either the machines are working off the clock, or somebody’s taking credit for their work.
The news: SpaceX becomes a neocloud
Dylan Patel of SemiAnalysis was on TBPN on Monday and explained why Elon capped Grok Bot: he’s taking SpaceX and turning it into a neocloud (summary via TMTB).
On Patel’s numbers, SpaceX can rent a GPU out for about $15 an hour, about the list price for NVIDIA’s top chips. The rents we track run lower. It costs about $4 an hour to own.
So every hour a subscriber burns is about $11 of rent the landlord didn’t collect. Turns out, from SpaceX’s perspective, you were a squatter when the landlord changed the locks. (Musk’s own stated reason: the cost of serving Grok Bot was "unsustainably high relative to revenue.")
The 13 GW plan. On SpaceX’s October earnings call (transcript), Musk said SpaceX will have maybe 2.5 to 3 GW of compute online by the end of this year, is "fairly confident" of adding another 10 GW in 2027, would be "surprised" to land under 13 GW by the end of next year, and sees a small possibility of more than 20.
SpaceX ended September with 1.7 GW of nameplate compute, per CFO Bret Johnsen. That makes the plan close to 10x in about 15 months.
By the end of the year, Musk’s efforts alone would add about 25% to our estimate of the world’s AI data center power. I wonder how many Teslas that would run.
And he does have the track record. His first Colossus data center, in Memphis, went up in 122 days (xAI; Fortune).
Per-dollar verdict: your subscription price is set by what the chip could earn renting to someone else. All-you-can-eat just went to market price.
Towns vote no
On election night, voters sided against data centers in all but one of 41 local ballot measures, per CNBC. Emporia, Kansas banned "high impact" data centers (Route Fifty); at least 14 Ohio towns passed bans or limits (Fortune); Wells, Maine voted in a moratorium (Scientific American). By our rough count, more than 3 GW of planned data centers are now in doubt.
Overnight, Elon posted: "Data centers in space looking pretty appealing right now" (Elon Musk on X; Benzinga).
Well, it is hard to lose a zoning vote in orbit.
Per-dollar verdict: the scarcest thing in AI right now is a town that says yes.
The memory market
Sam Altman’s name on X briefly read Ram Altman (IBTimes India). That’s right, R-A-M. He’d said RAM in your personal computer is a "quite terrible" use of the world’s limited supply (VideoCardz). He later said that was taken out of context (Briefs).
Well, looking at memory prices, the discomfort is mutual.
TrendForce expects regular memory chips to cost 10–15% more this quarter (TrendForce).
On the spot market, TrendForce’s benchmark chip (8Gb DDR4) sold for under $13 in mid-November last year (TrendForce). This week it’s about $46 (TrendForce).
Waits are easing, but regular memory ordered today takes about 19 weeks, by our read. Huh, that’s St. Patrick’s Day.
Apple expects its gross margin to drop to 45–46% this quarter, down from about 50% in the June quarter. Apple CEO John Ternus: "What we’re seeing with memory pricing right now is unprecedented" (Apple earnings call).
Per-dollar verdict: dumber per dollar for your next gadget, and a year the memory makers will never forget.
The tier list
Come on down. How much API value you get for every subscription dollar, on our list built from Adam Holter’s and real API pricing (Adam Holter; Claude pricing). It ranks value, not what people use.
Holter puts an X plan on top at about 80x. We can’t source it, so it’s out.
Top of the list: Claude’s $100 Max plan, at about $77 of API value per dollar. Buy one, get seventy-six free.
The rest of the top ten, our best guess: Claude Ultra, Claude Pro, Google’s top AI Ultra tier, X Ultra, ChatGPT’s two Pro plans ($200 and $100), X Super, and bringing up the rear, Google AI Pro and Google’s lower Ultra tier.
The board’s been holding steady since a day ago, on paper. Elon capped Grok Bot, and the board hasn’t caught up yet.
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 Cloudflare. Its AI code reviewer got pricey: the average cost per review more than doubled, from about $1.19 in April to about $2.46 by the end of August (Cloudflare).
So Cloudflare sent AI agents to find the wasted tokens. The biggest leak: every review opened with the same setup, and the cache mostly missed it. The shared setup was glued to the working directory and the date, so the AI was paying extra to be told what day it was. Hmm. Relatable.
Cloudflare fixed that, and model changes saved even more. An average review used to cost $2.43. Now it’s $0.77. 68% off (Cloudflare).
The fine print: the "before" is Cloudflare’s two priciest weeks, and its performance reviewer now misses some real problems. Side effects may include a reviewer that’s cheaper, faster and a little less nosy.
Still: chainsaw approved.
A word from ourselves
Your X-ray for the cloud bill. Your fuel gauge for the token tank. Your price scanner at the AI checkout. The people who keep the landlord honest.
Strong Compute plugs into 21 compute clouds and 5 model providers you already use, from AWS and Lambda to OpenAI and Anthropic. 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.