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: our first Free Weights Friday asks whether to own the box or rent the tokens for an open model. Owning loses on price, by a retirement party. Nvidia has committed $390 billion to its suppliers and guarantees up to $118 billion of its customers’ data center bills; Nvidia meal-preps. A compute exchange says B300 capacity booked for Chinese-affiliated buyers is up for resale in the US and Western Europe. One careful owner. Riot Platforms, a bitcoin miner, signed the biggest lease in its history, with an AI lab; the best mining is landlording. And the Department of Token Efficiency has three cuts: sorted JSON keys, caching, and a Mac Mini with a job. Here’s the written version, with every source linked.
Free Weights Friday
The weights are free. The gym is where they get you.
Free weights are open models: anyone can download the file. The question is where to run it, and who charges what. One unit today: how many output tokens $10 buys. More is better.
The desk. Apple’s Mac Studio with an M5 Max and 128 GB of memory is $5,399 at Apple. For a free model. One user benchmarked that chip on OpenAI’s open model, gpt-oss-120b, at 92 tokens a second (oMLX benchmark). Faster than your coworker answers a quick question.
Run it eight hours a day for three years, power included, and $10 of that Mac buys about 5 million output tokens. Our estimate (the Mac’s price spread over those hours, plus 150 W at $0.18 a kWh, no resale).
The rent. The same model, on OpenRouter’s list of hosts (OpenRouter), on Thursday:
Together AI, at $0.60 per million output tokens: about 16.7 million per $10.
The cheapest hosts we found, like Near AI, at $0.15 per million: about 67 million per $10. That’s a squeezed four-bit version. The model packed a carry-on.
So on price, owning loses. Spend the Mac’s $5,399 at Near AI instead and you get about 36 billion tokens. The Mac makes about 2.65 million a day at eight hours, so it needs about 37 years to make that many. Our estimate. Break-even arrives right around your retirement party.
You own the box for control. No rate limits, except your power company. And the model never changes under you: nobody can sunset a file on your hard drive.
The other desk box. Nvidia’s DGX Spark has the same 128 GB. It launched in October 2025 at $3,999 (Nvidia). Nvidia’s own store now lists it at $6,950 (Nvidia marketplace), and the retailers we checked start at $6,949.99. Most computers get cheaper after launch. This one got pricier.
Prices that move. DeepSeek’s V4.1 Flash is a million-token model under an MIT license, with about 40 host listings on OpenRouter (Hugging Face). Early Thursday, by our own checks, the cheapest one changed twice in about an hour. Gas stations change prices slower than that.
Now, the bad news. PrimeIntellect, one small host, raised its output price for gpt-oss-120b tenfold in a week, from $0.075 to $0.75 per million (OpenRouter). Same $10, a tenth of the tokens: about 13 million instead of 133 million. Free weights, surge pricing. So check more than one host.
And some free weights don’t fit on any one desk. Moonshot’s Kimi K3.1 Pro has 2.8 trillion parameters (Hugging Face). Free to download. Bring a moving truck. At four bits a parameter, the weights alone take 1.4 TB. Apple’s biggest Mac Studio holds 512 GB, so you’d need three Macs just to hold them. Kimi lives on rented hardware.
Per-dollar verdict: for the model we checked, renting tokens wins on price. Owning wins on bragging rights.
The news: Nvidia signs for the AI build-out
Thursday we covered Nvidia’s sales. Today, what it signed for. The numbers are in the CFO commentary Nvidia filed with regulators this week (Nvidia CFO commentary, SEC):
Nvidia has committed $390 billion to its suppliers, up from $279 billion a quarter earlier. Most of the increase is multi-year memory. Memory, bought years ahead. Nvidia meal-preps.
It also guarantees up to $118 billion of what customers owe on data centers, more than its $115.8 billion of sales last quarter. Most of it is SB Energy ($105 billion). The guarantees phase in building by building, the first expected in Nvidia’s fiscal 2029, which starts in early 2028.
Nvidia’s longest warranty this year covers somebody else’s rent.
Per-dollar verdict: Nvidia now vouches for the rent. You’re still the one paying it.
B300s for resale
Ronit Jain of Liquid Compute, which describes itself as building a regulated exchange for compute, posted on X on Wednesday night. He says a staggering number of neoclouds (newer clouds built around AI chips) want his firm to resell B300 capacity first booked for Chinese-affiliated buyers, in the US and Western Europe instead, because of “concerns of stricter legislation” on serving China. We covered one of those rules on Thursday.
Their compliance problem. Your shopping trip.
Jain expects B300 prices to keep falling. He didn’t say how far. Then again, his firm makes money on the trade.
Per-dollar verdict: capacity booked for someone else could become your discount. If Jain’s right.
A bitcoin miner turned landlord
Earlier this month, Riot Platforms signed a 15-year lease on its whole Corsicana, Texas, campus. The tenant is an unnamed leading AI lab. Billions in rent, and no name on the mailbox.
The lease is worth $28 billion, for 910 megawatts (Riot Q3 results). Riot puts the rent at about $1.8 billion a year, on average.
It covers power and land only. The tenant brings its own chips.
Counting all its leases, Riot now has about $38 billion of long-term contracted revenue (Foley, which advised Riot). This one deal is nearly three quarters of that, by our math.
Fifteen years of rent. No halving.
Per-dollar verdict: before you rent a chip, somebody’s renting the megawatts.
All that to drive a future filled with Self-hosted Intelligence.
Department of Token Efficiency
Chainsaw fueled. Spreadsheets open.
Cut one: JSON key order. AnyLanguageModel is a Swift library from Hugging Face. Its requests could come out with their keys in a different order each time. OpenAI’s prompt cache needs the same start on every request, so the cache never hit. A developer sorted the keys in a pull request on Thursday. In a 15-turn test with GPT-6.2 Sol, the cost of a run fell from $4.05 to about $0.50, or 88% (pull request 432). All because a library couldn’t decide who goes first.
The fine print: it’s one benchmark, run through Azure, on a pull request that may not be merged yet. It only helps apps whose encoder shuffles keys.
Cut two: caching. Developer Carlos González ran 366 million tokens through Anthropic’s Claude Opus in one week. He pays for a $20 subscription; his tool estimates that week at $190.71 at API list prices. 97% of those tokens were cache reads, and he says that without caching it would have cost about $1,500.
Somewhere, a pricing team just felt a chill. Caching is a Costco membership for déjà vu.
The fine print: those are his tool’s estimates, not an invoice. The $1,500 is his own what-if. And plan limits can change.
Cut three: Mike P’s Mac Mini. Mike P asked his Grok Bot why his usage runs low. Its answer: scripts on a Mac Mini do the grunt work, then hand the bot only the summary. Routines stay quiet unless something breaks, and each bot has one job.
The bot hired an intern. The intern is a Mac Mini.
The fine print: that’s the bot talking, no bill attached. And work handed to a second assistant gets billed on that plan instead.
The best token is the one you already paid for.
A word from ourselves
Your magnifying glass for the surprise line item. Your spotter for free weights that don’t fit on your desk. Your price check for hosts that change prices hourly. The people who keep the landlord honest.
Strong Compute plugs into 27 compute clouds, from Modal to Crusoe to Nebius, and 10 AI providers, from Fireworks AI to Deepgram to Anthropic, plus 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 tomorrow. Try it at strongcompute.com.






