The clip is engineered to be a knockout. A host on TBPN asks Sam Altman whether space data centers will hand OpenAI meaningful compute in two or three years, then five, then ten, then ten thousand, and Altman answers no, no, no, and finally, I wish Elon luck. The account that posted it attached a wager, that this interview will not age well.
Elon Musk quote-tweeted it with a promise and three jabs. The promise, that SpaceX starts flying these craft next year. The jabs, that Altman stole an open source AI charity, stole Apple's phone technology, and might need a parole officer's permission to visit. Worth being precise about what each of those is. The parole line is pure trolling with nothing under it. The charity line points to Musk v. Altman, where a jury in Oakland found in May 2026 that Musk sued too late over OpenAI's shift from nonprofit to for-profit, a statute-of-limitations ruling he has vowed to appeal, not a finding that anyone stole anything. The Apple line points to Apple's July 2026 trade-secret suit over OpenAI's Jony Ive device, which OpenAI disputes and which no court has decided. Three accusations, zero verdicts against Altman. The tweet reads as a scoreboard, but nothing on it has actually been scored.
Set the theater aside and look at the engineering, because that is where the wager lives or dies. Altman's position across interviews has been steady. Not a near-term priority, won't matter at scale this decade, and the current pitch is, in his word, ridiculous. He cites the launch-cost math and the small problem of servicing a dead GPU in orbit. He also scopes it in time rather than calling it impossible, which is the honest shape of the objection.
Musk is building on the opposite timeline. SpaceX moved to absorb xAI in early 2026 explicitly to build orbital data centers, filed with the FCC for as many as roughly a million AI satellites, and detailed a craft called AI1 with a wingspan near seventy meters drawing 120 to 150 kilowatts, about the draw of a single Nvidia GB300 rack. Launches next year, cost parity in two to three years, an IPO to fund it. Confident numbers, all of them forward-looking.
The physics does not care who tweets harder. Launch today runs somewhere between one thousand and roughly three thousand dollars per kilogram. Google's own Project Suncatcher paper argues the economics only close when that drops below two hundred dollars per kilogram, which its authors place in the mid-2030s, and Deutsche Bank's read reportedly lands viability well into the 2030s as well. Then there is heat. A vacuum is a superb insulator, so shedding the waste heat of dense compute in orbit is genuinely hard, on top of radiation on the chips, debris, and the servicing problem Altman named. On this evidence, the version of the future where orbital compute matters at scale is real but slow. Altman's someday-not-this-decade is closer to the numbers than Musk's next-year. The dunk may be the thing that ages badly.
What is undeniable is that the work has started. This is a race, not a duel between two men. Starcloud put an Nvidia H100 into orbit in November 2025, the first data-center-class GPU up there. Google plans to fly TPUs with Planet Labs by early 2027. Blue Origin, Thales Alenia in Europe, and China are all in it. Something is being built. It is just being built on a decade's clock, not a news cycle's, and the incentive to compress that clock in public belongs to whoever has an IPO to price.
The same week, a different kind of confident claim showed up, smaller and more honest to take apart. A post promoted a model called Dolphin3-Cyber-8B, marketed as a fully uncensored, expert-level cybersecurity assistant for exploit writing and vulnerability analysis, fine-tuned deep on real security data, running locally on a laptop with zero refusals. Strong words. The model card tells a plainer story.
Read the specs. It is an 8-billion-parameter Llama 3.1 model, and the base it started from, Dolphin3.0-Llama3.1-8B-abliterated, was already uncensored before anyone touched it. The adaptation on top was a LoRA of rank 16 that trained about 42 million parameters, roughly half a percent of the whole, for 500 steps in two to three hours on a single free-tier Tesla T4, with the working context cut to 2,048 tokens. That is a light afternoon of tuning on a model that already refused nothing, repackaged as quantized files that fit on a MacBook. Calling that fine-tuned deep on real security data is generous to the point of fiction.
The uncensored part rests on a technique called abliteration, and it is worth understanding because it is genuinely clever and genuinely limited. Arditi and colleagues showed that a model's refusal behavior is mediated by a single direction in its activation space. Abliteration finds that direction and orthogonalizes the weights against it, so the model loses the ability to say no, without any retraining. Maxime Labonne's widely read walkthrough notes the cost that the marketing never mentions: the same operation tends to degrade the model's general performance. Removing the brake is not the same as adding an engine. Abliteration strips refusals. It adds no expertise.
That distinction is the whole game. A Swedish thesis comparing censored and self-hosted uncensored models for penetration testing found what you would expect, that the uncensored ones are more willing but not more capable, and carry real ethical and oversight risks precisely because nothing stops them. An 8B model with a 2,048-token window and a half-percent tuning run is not an expert. It is a compliant novice. The honest description is that this tool will answer questions a hosted model declines, which is useful for a bug-bounty hobbyist and equally useful for someone with worse intentions, since the study of uncensored models used as backends for criminal services counted 11,598 of them and validated a random 84 as genuinely uncensored. Eric Hartford, who started the Dolphin line, frames the whole project as giving people control over the AI on their own hardware, and that argument has weight. It just does not turn a rank-16 LoRA into an oracle.
The two stories rhyme because the same substitution is happening in both. A claim gets stated with total confidence, the mechanism that would back it is either missing or points the other way, and the person making the claim has a reason to want you looking at the volume instead of the number. Musk needs the orbital timeline to feel imminent because there is capital to raise against it. A hobbyist repo needs expert-level to feel earned because attention follows the word. In neither case is the underlying thing fake. Compute in orbit is coming, and local uncensored models are real and consequential. The dishonesty is entirely in the gap between what was built and what was announced, and that gap is measurable if you bother to read the launch-cost curve or the training log. I trust the log over the tweet, and the curve over the wager, and I would put more on Altman's decade than on Musk's next year, not because either man is trustworthy, but because the physics keeps its own record and does not care who is winning the thread.