Flapping Airplanes, an AI lab run by twentysomethings, is reportedly in talks to raise money at a $5 billion valuation, up from $1.5 billion in January. I.M.J. McInnis, in a post on X reposted on 22 October 2026, asks what that price says about how soon powerful AI arrives. His own answer is: less than it seems.
The company pitched itself as a group that would invent algorithms needing far fewer training examples, which researchers call sample efficiency. McInnis says its pitch has since drifted toward new building blocks for GPUs that squeeze more output from each megawatt of power. That would point to cheaper running of models rather than cheaper training. He treats the raise as unfinished: the talks, as he describes them, involve Index Ventures, which led the January round and which he calls a prominent generalist firm, and Lux Capital, which he calls a well-regarded deeptech firm.
McInnis discloses a link to one of them. In his words: “COI note: the latter invested in my company.” The latter, in his sentence, is Lux Capital. He adds that his account mixes public information with reliable anonymous sources, and that any technical progress at the company is, as far as he can learn, secret.
The naive reading, he writes, is that insiders now rate the company’s odds of success 3.3 times higher. He thinks the stated reasons would be a bigger team and claimed progress. Then he lists reasons to doubt it. Investors told him there is a floor under a valuation set by what an acquirer would pay for an elite team, and that floor rises as the team grows.
His sharper claim is about the market itself. He says venture prices are far from rational and that the best-performing firms are better understood as brand-rent collectors. As evidence he cites Instinct, which raised a Series A at $0.5 billion and then a Series B at $2.5 billion about three weeks later, while its product was in limited testing. In his telling, little changed in between except heavy promotion by the early investors.
He says Flapping Airplanes is a prominent case. By his account it raised four rounds before its first publicly documented one, starting at $40 million, and in several of them the deal was that the current backer would help sell the stock to the next. He also reports that one investor told him the CEO’s fundraising skill was the reason he invested. That story has a counterweight: the previous round’s investor is still involved, though McInnis notes nobody knows at what size.
What would change his mind, he says, is credible technical backing from people with no stake in the outcome, current shareholders passing on a chance to cash out, or the same pattern at other labs that have not yet had a hype round.
The headline question rests on one scenario. If sample efficiency improved by many orders of magnitude, McInnis argues, effective scaling would leap ahead and timelines would shorten. A lone actor could also build very capable AI without a large GPU footprint, which would make treaties that count chips harder to enforce. He admits he has not run the numbers on how a million-fold gain would interact with known scaling laws and says someone should.
His own estimate stays loose. He believes the real rise in the odds of success is above 1x and below 3.3x, and that the rise in odds of a breakthrough within ten years is smaller still. The author ends by asking whether this is a shiny distraction or a big deal, and leaves it open.
For anyone tracking AI policy, the lesson sits in the evidence: until someone outside the deal verifies the technology, a compute-based treaty plan rests on a price set by investors, not on a demonstrated result.
I.M.J. McInnis, in a post on X, 22 October 2026.
