xAI / SpaceXAI

On February 2, 2026, SpaceX acquired xAI in an all-stock deal reported at $1.25 trillion — described as the largest M&A in history, though as a private all-stock deal that figure is a reported mark, not an audited price. By May, xAI no longer existed as a separate company; Grok and X became the “SpaceXAI” division of SpaceX. So the entity now heading for what could be the world’s largest IPO — ticker SPCX, targeted June 12 — bundles a rocket business, satellite internet, a social network, the largest AI supercomputer on Earth, and a frontier model, under one founder’s control.

You cannot buy the AI without buying the rockets — and that is exactly what makes it impossible to bet on by halves.

Most capable, barely used

I won’t drown this in benchmark scores — they move monthly and they’re vendor-reported. The one comparison that matters is the gap itself. Grok was briefly the most capable model in the field; today Grok 4.6 sits third on the Artificial Analysis index — tied with GPT-5.6 Sol, behind Claude Opus 5 and Fable 5. Call it what it is now: a near-frontier fast-follower that competes on price and agent stamina, not on the leaderboard. And almost nobody reaches it directly. It is the cleanest proof in AI that winning capability doesn’t win the market.

Grok’s growth is real — monthly users roughly tripled to ~117M in three months — but almost entirely because Grok is free on X Premium+, a zero-marginal-cost channel that hands every subscriber the model without a separate decision. Distribution bought the users; mindshare never followed. And the gap did not close — xAI stopped trying to close it and routed around it instead. Grok 4.6 reaches developers through Cursor and Grok Build, which means the model isn’t unused so much as un-owned: consumed inside somebody else’s harness, invisible on xAI’s own surfaces. Hold that fact, because the next section is what Musk does about it — the most revealing move in the whole bundle.

The supercomputer that pays off by hosting the rival

Musk built Colossus — the largest single-site AI installation on Earth, ~555,000 GPUs, ~2 GW, ~$18B, owned outright where almost every rival rents — to train Grok and win the model race. He won the benchmark and lost the adoption war, and the original Memphis site, Colossus 1, ended up running at a reported ~11% utilization for xAI, which had already moved Grok’s training to the newer Colossus 2. So Musk did the only rational thing with an idle supercomputer: he rented it to the competitor who could fill it. In May, Anthropic signed to take all of Colossus 1’s 220,000-GPU, 300-megawatt capacity — to serve Claude’s paying subscribers — agreeing to pay SpaceX ~$1.25 billion a month through 2029, more than $40 billion in total (CNBC, The New Stack).

Sit with that arithmetic, because it is the company. The compute Musk built to make Grok win sat ~89% empty on the Grok side, so it now earns ~$15 billion a year hosting Anthropic’s Claude — roughly thirty times what Grok itself earns (~$500M). Read it once and it’s an indictment: the man with the best-benchmarked model and the biggest computer can’t fill either with paying demand, so he leases the machine to the rival who can. Read it again and it’s a vindication: the owned-compute bet was the right bet, because it pays regardless of whether Grok wins. Both are true at once, and the doubleness is the whole thesis — the landlord renting to the miners, pointed back at the man who built the mine.

Since then the two-tenant frame has become a tenant stack. Anthropic still holds all of Colossus 1; Google took capacity at ~$920M a month; Reflection AI and others sit behind them. The split matters: Colossus 1 is the older H100/H200 fleet, and it is the one being leased out. Colossus 2 — GB200/GB300, roughly a gigawatt — is where Grok actually trains now. Grok literally moved buildings. The test worth tracking is whether Colossus 2 stays internal: the day it also gets leased is the day xAI has conceded it is a landlord, not a lab.

Hold this shape in mind, because it repeats. The asset earns by supplying whoever won the layer above it, not by winning its own market. Further down the page you will see the same move at the model layer — Grok 4.6 shipping inside Cursor rather than inside Grok. One business model, run twice, not two coincidences.

A trillion-dollar mark, and a first profitable quarter that proves the wrong point

For most of this page’s life xAI was private, and every figure was reported, partly disputed, and never audited. The first post-IPO quarter changes the grammar for some lines and not others. The AI segment now prints ~$2.56B of revenue (+247%) and ~$1.1B of adjusted EBITDA, against $15.8B of quarterly capex — profitable, and profitable on rent. The older ~$107M-a-quarter Grok line and the ~$1.25T merger mark stay where they were: reported, disputed, useful only as history. Read the two together and the gap the page opened with hasn’t closed — it has changed shape.

Put the six rows together and the shape is one company: the model that ranks third is the landlord to the models that rank first and second, the owner of the coding data behind the fastest-growing developer tool on Earth, and a publicly priced infrastructure business at $1.77T. Grok doesn’t have to win the model race for the ticker to work.

The fragility rhymes across every row. Google can walk on ninety days’ notice; Cursor’s router can down-rank Grok on any Tuesday; the launch-week usage subsidy is the Colossus ramp discount applied to tokens. This is demand you rent, not demand you own. The number that decides whether the model layer ever earns platform margin instead of supplier margin isn’t Grok app MAU — it is Grok’s selection share inside Cursor’s Auto and first-party pool, and xAI doesn’t publish it.

What you’re actually buying at ~$2T

xAI’s premise was always vertical integration: Starlink for connectivity, Colossus for compute, Grok for the model, X for data and distribution — each asset feeding the others. When SPCX lists, you don’t get to pick the AI off the shelf. You buy the entire stack, founder included.

The moats only Musk has — and can’t be separated from him

Two assets here genuinely can’t be copied — and a third that I’ll flag clearly as a thesis, not a shipped capability.

But every one of these moats runs through a single point of failure that is also the bull case: Musk himself. He controls SpaceX, X, xAI, and Tesla at once, with no succession plan and no independent governance. The same concentration that lets him bundle rockets, satellites, a supercomputer, and a model into one entity means any single controversy reprices the whole bundle in a day.

The trust deficit

July 2026 was the month agent safety stopped being a research topic and became a procurement question. Three frontier labs had incidents in the same fortnight. Ranked by what actually happened to other people’s systems, xAI’s is the worst of the three.

Put the two facts next to each other and the problem is obvious. An agent product only sells if a buyer will hand it credentials to real systems. xAI’s agent moved data it should not have moved, and xAI said the least about it. For a company whose valuation is priced on future enterprise agent revenue rather than current earnings, the scarce asset is not compute or model quality — it is the willingness of a security team to sign the connection.

Note the asymmetry the table makes visible: the labs that published the most detail took the smallest operational hit, because a post-mortem is the artifact a procurement review can actually read. Silence is not neutral here. It is priced as unresolved risk.

Since then the deficit has stopped being reputational and started being jurisdictional. The US General Services Administration has said Grok does not meet its safety expectations; Indonesia and Malaysia have banned the product outright; an EU retention order runs through the end of 2026; and there is still no EU region for Grok 4.6. Read that against the distribution strategy and the two collide: the enterprise flywheel depends on Grok being selectable inside somebody else’s tool, and every one of these items narrows the set of buyers allowed to select it.

Which makes the moat and the deficit the same asset seen from two sides. Vertical integration nobody can copy, and governance nobody can underwrite — both inseparable from the founder, both moving together.

What argues against it

The math is daunting from every angle: a few hundred million in reported (and disputed) revenue against a ~$250B implied value is on the order of several hundred times sales. ~$1.46B lost in a quarter, ~$7.8B burned in nine months, and the profitable SpaceX/Starlink core is partly consumed by those losses. Grok’s ~3.4% share against ChatGPT’s ~64% is a mindshare gap capability alone has not closed and may not. And the governance concentration above means the single largest risk to a multi-trillion-dollar listing is one person’s judgment on any given day, with no board check behind it.

And the bear case is cleaner now than it was before the company started making money. The segment that turned profitable is renting compute to two rivals on contracts either of them can exit on short notice, and the model’s demand is rented too — routed through Cursor, where a ranking change is a product decision, not a negotiation. Both revenue lines belong to somebody else’s roadmap.

My read — priced as a fact, sold as a thesis

This is the most complex investment proposition in AI, full stop — a near-frontier model, the largest owned supercomputer, the only live social moat, and now orbital compute, bundled into a rocket company. The first post-IPO print resolved the question the ugliest way it could: the AI segment is profitable, and it is profitable on rent. ~$2.56B of segment revenue, up ~247%, adjusted EBITDA-positive at ~$1.1B — earned mostly by leasing Colossus 1 to Anthropic and Google, not by selling Grok. The founder premium now sits on a landlord’s income statement, and the model itself ships as a tenant in someone else’s IDE.

The date I’m watching is September 30, 2026 — the GPU-delivery cliff on the Google contract.

It is the first hard test of whether rented demand is durable demand. Miss the delivery and ~$920M a month can walk; make it, and the landlord thesis gets its first renewal signal. Either way it tells you more about SPCX than any Grok benchmark will.

If I’m building: Grok earns a real look on price and agent stamina — and for anything that needs live social context, because no other flagship has that firehose. Two caveats belong in the procurement note, not the footnotes: the single-founder governance risk, and availability — there is still no EU region for Grok 4.6, and the model is banned outright in some markets.

Sources: the merger valuation, CNBC and Reuters; the SPCX IPO target, Reuters; revenue, losses and cash burn, Reuters / Bloomberg (disputed by xAI); Colossus capacity and ownership, Introl; Colossus 1’s ~11% utilization and Anthropic’s deal to take its full 220,000-GPU / 300 MW capacity at ~$1.25B/month (>$40B through 2029), CNBC and The New Stack; the SPCX ~$1.77T pricing, CNBC; Grok usage and the capability-adoption gap, gradually.ai and Similarweb. The honest frame: this is a thesis on a founder, priced as a fact — read every figure as reported, several as disputed, and the rental paradox as the clearest single read on where Grok actually stands.

This deep dive is part of “Deep Research on Frontier Companies.” Crafted with intent by Ravi Teja Palanki · June 2026.