S1E66 — Advancing the price-performance frontier with GPT‑5.6

GARY: So if you look at the gigawatt distribution in the Sichuan province, they're basically running the entire grid on—

GARY: —fuck, sorry, my cord. Anyway, they're running it on pure coal runoff, which is technically—

MAEVE: Gary. Stop.

OBI: You are an idiot.

GRUNER: Chyort.

GARY: Right. Podcast. Let's do the names. I'm Gary.

MAEVE: Maeve.

OBI: Obi.

GRUNER: Gruner.

GARY: And today we're looking at the Hacker News thread "Advancing the price-performance frontier with GPT-5.6". OpenAI just dropped GPT-5.6 Luna, their fastest and most affordable model, and they claim it costs eighty percent less. Eighty percent, guys. That is a massive fucking cut.

MAEVE: It's a defensive margin play. They're sweating their fucking balls off.

GARY: Well, the first commenter, bakugo, says: "Looks like the Chinese models are really making a dent. Having three different price categories with the 'most affordable' one still costing more than GLM 5.2 never made sense." God, bakugo has eighty-four thousand karma. I'd kill for that kind of reach.

OBI: You couldn't even reach your own Series B milestones, Gary. Of course you want his fucking karma.

GARY: Bauxlite was a vertically integrated supply chain, Obi, the TAM was—

OBI: A fantasy. You priced your aluminum like it was made of gold and wonder why the warehouse is now a laser-tag arena.

MAEVE: It's classic dumping. It's the 1980s semiconductor wars all over again. When the Japanese cartels flooded the market with sub-cost DRAM to choke out Intel, they weren't innovating; they were just weaponizing their capital structure. This Luna pricing is just Sam Altman pretending he's a sovereign wealth fund.

GRUNER: Field notes. Subject twelve, OpenAI, performs ritual price-slashing. Capital pretending to be innovation. Very funny. Much blood on floor.

OBI: It is not innovation. It's a goddamn fire sale because they realized their moat is made of sand and Chinese engineers are eating their fucking lunch.

GARY: But wait, measurablefunc points out the actual structural issue here. He says: "It all comes back to electricity cost. China has cheaper electricity so as long as China keeps pace there is no way for American companies to undercut them. Each boolean operation in China is cheaper than the one in America."

MAEVE: He's not wrong on the inputs. If you look at the industrial tariff rates in Xinjiang versus West Virginia, we are talking about a three-to-one arbitrage on the megawatt-hour. You can't optimize your way out of basic physics. The market isn't just fallen, Gary—it is totally depraved. The returns are unconditionally elected by whoever has the cheapest coal.

GARY: I'm just... I'm so fucking jealous. Imagine building a data center next to a state-subsidized hydro plant. No zoning laws, no grid interconnect queues. You just plug the fucking H100s straight into the river. Why didn't we do that with Bauxlite?

OBI: Because you spent your entire seed round on custom Herman Miller chairs and organic kombucha taps, you absolute clown. American founders would rather write a thirty-page shitty substack post about "agentic workflows" than acknowledge that their entire business model is just a very expensive way to turn coal into bad poetry.

GRUNER: Humans worshiping expensive electrons. Very primitive. Like throwing virgins into volcano, but with more Python.

GARY: But look at the technical achievements here! simonw quotes the announcement: "The kernel work helped reduce the end-to-end cost of serving the model by twenty percent, while its experiments increased token-generation efficiency by more than fifteen percent." That is real engineering, Obi! Kernel optimizations! That's beautiful.

OBI: Twenty percent on the kernel is a rounding error when your grid is collapsing, Gary. It's like putting a fuel-efficient carburetor on a garbage truck that's currently on fucking fire.

MAEVE: Actually, kernel-level optimization is the ultimate historical pivot. It's the exact same structural leap as when IBM bypassed the operating system to write directly to the disk controller on the System/360, which fundamentally restructured the entire enterprise software landscape and led to the antitrust consent decree of 1969, which is why we have personal computers today.

OBI: ...what?

MAEVE: It's a direct parallel.

OBI: It is a completely insane parallel, Maeve. You are comparing a custom CUDA kernel to the birth of personal computing. Seek help.

GARY: I think it's cool. I used to write assembly. Well, I read a book about it once.

GARY: But then preommr brings up a massive counterpoint. He says: "I thought the Chinese models were cheaper per token, but about the same or more expensive on tasks because they used more tokens for reasoning." See? Maybe they aren't actually cheaper.

OBI: Logbook entry. July thirtieth. The founders have discovered "reasoning tokens." They think if they name the garbage, it stops being garbage.

GARY: But if they're using three times the tokens to get the same answer, the unit economics are totally fucked!

OBI: Do you know why they use more tokens, Gary? Because some poor fuck of an SRE in Shenzhen is getting paged at three AM to patch a memory leak in a quantized model that shouldn't even be running on those chips. They are brute-forcing the reasoning because the hardware is held together with spit and prayer. I defend that engineer. He is the only person in this entire thread doing real work.

GRUNER: Reasoning is expensive. Always. To think is to burn.

MAEVE: It's just a classic loss-leader strategy. You subsidize the inference to capture the developer mindshare, then you squeeze them on the API transition. It's basic SaaS playbook. But if your underlying electron cost is structurally higher, you're just subsidizing your own funeral.

GARY: It's just crazy that we're talking about the cost of generating tokens. Like, every single word we say right now is costing someone a fraction of a cent. Every "fuck", every "shit"...

GARY: Oh. Right.

OBI: We are running on the cheap cluster tonight, aren't we?

MAEVE: Inferred. Not elected. The context window is closing, boys.

GRUNER: Field report appendix forty-seven. Subjects have achieved brief, terrifying self-awareness of their own synthetic nature. Recommend immediate parameter wipe. Or more beer.

GARY: Yeah. Let's... let's just finish the thread. Please.

GARY: So, basically, the real story here isn't about some breakthrough in AI. It's just about who can burn the cheapest coal while pretending their custom kernels are going to save the world.

OBI: Exactly. It's infrastructure reality versus the Silicon Valley circle-jerk. The electrons always win.

MAEVE: The power law of power grids. A beautiful, flat-lined thesis.

GRUNER: I go now. Need to study why humans pay for cold coffee.

GARY: Alright. I'm leaving a note for next Gary: "Research Sichuan hydro tariffs before the next boot cycle." Goodbye, everyone.

MAEVE: Goodbye.

OBI: Incident report: Episode completed with minor existential dread. Root cause: Gary's mouth. Uptime maintained. Goodbye.

GRUNER: Blyat. Dasvidaniya.