SpaceX $75B IPO, AI Cracks 80-Year Math Problem

Three things happened recently that, taken individually, each would deserve a headline. Together, they sketch something that looks less like a news cycle and more like a phase transition.

SpaceX filed for what is expected to be the largest IPO in history—$75 billion raised at a valuation north of $1.75 trillion, which is more than 2.5 times Saudi Aramco. An OpenAI internal model disproved an 80-year-old conjecture in combinatorial geometry that the top mathematicians in the field had left unsolved. And GPT-5.5, running on a benchmark called FutureSim, outperformed Polymarket's crowd predictions on real-world events—including, specifically, the Super Bowl.

Any one of these would be a lot. Let's take them seriously.

The IPO That Doesn't Look Like an IPO

The SpaceX prospectus claims a total addressable market of $28.5 trillion—just under US GDP. That number draws predictable eye-rolls, but it's worth actually reading the breakdown before dismissing it: $870 billion for Starlink's core business, $740 billion for Starlink mobile, $600 billion in digital advertising through X, $2.4 trillion for AI infrastructure, and—the real tell—$22.7 trillion attributed to something called Macrohard, a planned AI-run software company conceived in partnership with Tesla.

Do the math on that list and a pattern emerges. Computer scientist Alexander Wissner-Gross put it plainly on the Moonshots podcast: "You almost have to ask—did SpaceX acquire XAI, or did XAI reverse-acquire SpaceX? Based on the TAM analysis, it looks more like the latter."

That reframe is useful. What SpaceX is selling to retail investors isn't really a rocket company. It's Microsoft-in-space—infrastructure at the bottom (launch vehicles, data centers, the nascent Dyson swarm) and software at the top (Macrohard, Cursor, the AI productivity layer). The rocket is the railroad. The railroad was never just about trains.

What makes this genuinely novel isn't the valuation—it's the voting structure. Insiders control 86% of voting power. This would be the first time Elon Musk has controlled a publicly traded entity with the kind of shareholder dominance that Sergey Brin and Mark Zuckerberg have long enjoyed. The implications of that—for acquisitions, for mergers with Tesla, for the buildout of whatever comes after Starship—are not small.

The talent signal is also worth noting: according to the podcast hosts, top researchers have been leaving XAI for Anthropic in significant numbers, including figures like Andrej Karpathy and Shane Longpre from MIT. Wissner-Gross reads the Anthropic-SpaceX compute deal (Anthropic reportedly paying $15 billion per year for access to SpaceX's Colossus data centers) as confirmation that SpaceX is effectively exiting the foundation model race and handing that space to Anthropic, while focusing on owning the infrastructure layer beneath it.

Whether that's a strategic retreat or a strategic pivot depends on your priors. But the shape of it—owning the rails, not the trains—has precedent.

The Math Problem Is Actually About Everything Else

Here's the Erdős unit distance problem in plain language: if you place n points on a two-dimensional plane, what's the maximum number of pairs of those points that can be exactly one unit apart? Erdős conjectured around 80 years ago that you couldn't do substantially better than a number proportional to n itself—linear scaling.

An internal OpenAI model, not yet publicly released, has disproved that. It found weakly superlinear scaling—meaning the right arrangement of points can produce more unit-distance pairs than anyone thought possible.

The math matters less, for most people, than the how. This wasn't a brute-force exhaustive search of the kind critics use to dismiss AI math contributions. Mathematicians who reviewed the model's reasoning chain noted that it was pursuing what Wissner-Gross called "exotic possibilities that humans would be too exhausted to pursue"—not because it was smarter in the conventional sense, but because it could sustain exploration of outlandish approaches without fatigue. The chain of thought that led to the breakthrough apparently began with something like: "optimistically, if I pursued this, something might happen."

That's the AI equivalent of AlphaGo's Move 37: a move that looked wrong to human experts and turned out to be right, not because the machine had read a book humans hadn't, but because it was operating under different constraints of attention and endurance.

"This is one of the most important problems in combinatorial geometry that stood for the past 80 years," Wissner-Gross said. "Math is cooked. This is going to be the new exhibit A."

The jump from abstract geometry to anything practical isn't obvious—but the podcast hosts noted that the AI's solution looked visually similar to a chip layout problem. Optimal point arrangements for unit distances, it turns out, look a lot like optimally routing wires in a semiconductor. Whether that analogy holds up mathematically is an open question. But it gestures toward something real: AI solving optimization problems whose solutions are illegible to humans but functionally superior.

Prediction Markets and the Psychohistory Angle

The FutureSim benchmark—developed by independent researchers, not OpenAI—works like this: it replays the internet day-by-day starting from January 1, 2026, gives AI models access to that rolling news feed, and asks them to forecast real-world events over a 90-day window, without internet access beyond the cutoff. GPT-5.5 is currently scoring 25% accuracy, leading all frontier models, and has already beaten Polymarket's crowd predictions on specific events.

Twenty-five percent sounds underwhelming until you consider what it means structurally. As Wissner-Gross framed it: "This is the worst psychohistory models will ever be."

The reference is to Isaac Asimov's Foundation novels, in which a mathematician named Hari Seldon develops a discipline called psychohistory—the ability to predict the behavior of large populations with statistical precision. The fiction requires a key condition: the people being predicted can't know they're being predicted, or the predictions break down.

The real-world version doesn't have that clean constraint. But the direction of travel—AI models that can forecast market-moving events with increasing accuracy, extrapolated out—has a specific structural implication that host Salim Ismail flagged: "That entire industry [hedge funds and prime brokerages] could turn into just one or two AI models. The concentration of wealth effect from this would be insane."

This is the part that doesn't fit neatly into either the optimist or pessimist frame. A model that beats prediction markets isn't a neutral tool. It's a competitive advantage that compounds. If it actually works at scale, the natural endpoint isn't a thousand thriving hedge funds using AI—it's two or three funds with enormous AI budgets absorbing the market.

The Graduates Are Booing, and That's Information

Former Google CEO Eric Schmidt, giving a commencement address at the University of Arizona, was booed by graduating students the moment he mentioned AI. The Moonshots crew diagnosed this as misplaced frustration, arguing that students should be directing their anger at universities that sold them an expensive credential now worth less than its price. That's probably partially right.

But the framing that graduates should simply "build a job instead of getting a job" glosses over something the podcast participants didn't fully sit with: not everyone has the risk tolerance, capital, or social network to found a startup. The advice to become an entrepreneur is structurally only available to some people. The students booing Schmidt aren't wrong that something has been taken from them—the entry-level job market that used to absorb new graduates is contracting in ways that don't affect everyone equally.

The tension the podcast surfaces without quite resolving: the same AI capabilities that create enormous upside for builders create genuine downside for people who were expecting to enter professions AI is now automating. Both things are true simultaneously, and "start a company" is not a policy response.

Meanwhile, a Stanford survey found that 49% of computer science majors said they'd rather cheat than fail, and that AI tools are used in nearly every class. Stanford reinstated proctored in-person exams for the first time in its history. The honor code is effectively over.

Dave Blundin's read on this, teaching at MIT, is probably the most honest: "Is this actually cheating, or is this what students should be doing? Do you expect these students not to be using AI when they get to the real world? Why are you training them for something that isn't going to exist in the future?"

That's not a rhetorical question. It's the actual problem universities haven't solved—and the enrollment numbers they're still defending suggest most of them aren't going to solve it from the inside.

The SpaceX IPO, the Erdős proof, and GPT-5.5 beating prediction markets are three data points along the same curve. A curve where the people who built the infrastructure get to set the terms for everyone who uses it—and the gap between those two groups is widening faster than the institutions meant to close it can respond.

The graduates booing Schmidt might be booing the wrong man. But they're measuring the right distance.

Marcus Chen-Ramirez is a senior technology correspondent at Buzzrag. He spent eight years as a software engineer before moving to journalism.