Musk Now Owns Every Layer of the AI Stack. Here's What That Actually Bought.

On February 2026, SpaceX acquired xAI. In June, it IPO’d on Nasdaq at $135 a share, a $1.77 trillion enterprise value, raising $86 billion — the largest public offering in history. On August 6, it announced a chip fab in Grimes County, Texas with Tesla, xAI, and Intel as partners.

One company now holds the rockets, the satellites, the social network, the frontier model, the datacenters, the energy storage, the humanoid robot, and a foundry.

That is the most complete vertical stack anyone has assembled in the AI era. It is worth understanding precisely what it delivers, because the answer is more interesting than the headline.

The Structure Most Analysis Is Still Missing

The ownership chain runs SpaceX → xAI → X.

xAI absorbed X Corp in March 2025 in a $33 billion deal. SpaceX absorbed xAI in February 2026, at a reported $1 trillion for the rocket company against a nine-figure-billions valuation for the Grok developer, creating a combined entity worth roughly $1.25 trillion. It went public four months later.

Musk holds 42% of the shares and 82% of the voting power, through Class B stock carrying ten votes each. Full-stack ownership at the asset level is matched by full control at the governance level.

There is no independent xAI to analyze anymore. There is a public company with a launch business, a satellite business, and an AI business, where the S-1 attributed 93% of its $28.5 trillion total addressable market to AI. Launch, the thing that made SpaceX famous, accounted for 1.35%.

That is the vertical integration thesis expressed as a number. The rockets are now a rounding error in the story the company tells investors.

What the Stack Actually Delivers

Strip away the narrative and three things are demonstrably real.

Speed of physical buildout. Colossus 2 near Memphis is the largest known AI datacenter on Earth by IT power — roughly 946 MW, about 1.1 million H100-equivalents, an estimated $35.8 billion in cost. The first phase, 110,000 GB200s and 210 MW, went from ground to live in about 91 days. Nobody else in the industry builds at that clock speed. That capability comes directly from owning the construction, the permitting relationships, and the power procurement rather than negotiating for them.

Energy as an internal transfer. Tesla sold SpaceX $295 million of Megapacks in Q2 2026 alone, $329 million in the first half, on top of $506 million in 2025. When a datacenter needs grid buffering, one Musk company writes a purchase order to another Musk company. That is vertical integration doing exactly what vertical integration is supposed to do — collapsing a supplier negotiation into an internal decision.

Compute as a revenue line. The AI segment produced $2.56 billion in Q2 2026 revenue, up 237% year over year. That is not a rounding error, and it arrived faster than most of the market expected.

The Layer He Doesn’t Own

The most expensive component in the entire stack comes from someone else.

Musk committed to Nvidia for the AI datacenter buildout, calling it “the best AI computer.” Nvidia stock rose on the remark. The company that owns the rockets, the satellites, the model, the network, the batteries, and the robot buys its silicon from a supplier — and that silicon is where the money goes.

The Terafab announcement is the attempt to close that gap: $16.8 billion initial investment, 3,000-plus workers, a potential $119 billion total buildout, with Intel as a partner. The S-1 language calls it an extension of “vertical integration” meant to “alleviate potential future chip shortages.”

It is also a fab that does not exist yet. Announcement and site are verified. Output is a projection.

That distinction — announced versus shipped — is the whole discipline required to read this company correctly.

The Counter-Evidence, Stated Plainly

We take a positive view of what full-stack ownership can do. Here is the strongest case against it, because a post that only argues one side is marketing.

The compute is being rented out, not used. Anthropic leases essentially all of Colossus 1 at $1.25 billion per month — roughly 220,000 to 325,000 GPUs. Google leases about 110,000 GPUs at $920 million per month through mid-2029. Bloomberg reported that SpaceX rented Colossus 1 out after its own teams had trouble using it to train Grok. Jim Chanos characterized the shift from frontier lab to compute landlord as moving into “a commodity business.”

Both contracts carry 90-day termination clauses. Musk clarified the Anthropic arrangement is a 180-day lease with mutual 90-day cancellation, and that “the short term was our request.” That is roughly a $26 billion annualized revenue stream two customers can walk away from in a quarter.

The financials are ugly underneath the growth. Q2 2026: revenue up 92%, net loss $541 million, AI segment losing $1.2 to $1.3 billion on its $2.56 billion of revenue. Capex up 557% in one quarter to $18.37 billion. xAI burned $6.4 billion on $3.2 billion of revenue in FY2025, four times worse than the prior year. X advertising revenue fell 14% year over year in Q2 to $367 million. The stock closed at $125.33 on August 4 — below its $135 IPO price, roughly 50% off its post-listing peak.

The distribution layer is smaller than the story implies. X reports 550 million monthly actives. Grok reports 117 million. The captive-audience advantage converts at about one in five.

The model itself is competitive, not dominant. Grok 4.6 shipped August 12, 2026, thirty-five days after 4.5 — genuinely fast iteration. It scores 61 on the Artificial Analysis Intelligence Index, tying GPT-5.6 Sol and trailing Fable 5 at 62, while losing meaningfully on agentic coding benchmarks like Terminal-Bench v3.0 (26% against 34.6%). Owning every layer beneath the model has not yet produced a decisive lead at the model.

Silicon has slipped before. Dojo was shut down in August 2025 — Musk called Dojo 2 “an evolutionary dead end.” Its replacement, AI5, taped out April 15, 2026, roughly two years behind the original schedule, dual-sourced across TSMC and Samsung, with volume production targeted for late 2026 or early 2027. The claimed 40x performance gain remains a claim.

Robotics is the widest gap between promise and delivery. Robotaxi operates in 7 metros against a July 2025 promise of covering half the US population by the end of that year. Roughly 380,000 unsupervised miles against Waymo’s 200 million rider-only miles — a 526x gap. Fleet size, paid rides, and intervention rates go undisclosed. Optimus production slipped again in the July 22, 2026 shareholder letter, from “late July/August” to “later this year.”

And the compute target is a moving object. Nameplate AI compute went from 1.0 GW in Q1 to 1.4 GW in Q2 — real, verified growth. On the Q2 call Musk stated a “tentative target” of 20 GW live by end of next year, then hedged in the same breath to “something close to 15 gigawatts.”

1.4 GW is the fact. 20 GW is the story. Both are in the same transcript.

The Business Takeaway

The useful lesson here is not about rockets. It is a sorting rule any operator can apply on Monday morning.

Own the layer that is scarce, differentiating, and slow to replace. Rent the layer that is abundant, commoditized, or improving faster than you can depreciate it.

Musk owns power, land, and construction speed — genuinely scarce, and a competitor cannot conjure a 946 MW site in a quarter. He rents GPUs from Nvidia, because that layer turns over roughly every eighteen months and owning it means owning obsolescence. He is buying Cursor for $60 billion in stock rather than building a developer platform, because distribution is faster to acquire than to earn.

Run the same sort on your own business:

• Own your data. Your transaction history, your client files, your process documentation. No vendor can regenerate it and no competitor can buy it.
• Own your customer relationships. The trust, not the CRM software.
• Own your process knowledge. How your firm actually does the work is the durable asset.
• Rent the models. Grok, Claude, and GPT all get better without you spending a dollar of capex. Paying for frontier inference is renting a capability that improves monthly.
• Rent the infrastructure. Hosting, storage, and orchestration are commodity layers with real competition. Let vendors fight over the margin.

The failure mode on both ends is symmetrical. Own too much and you are financing depreciation — $18.37 billion of capex in one quarter against a $541 million loss is what that looks like at scale. Own too little and you are a reseller of someone else’s product, with no moat when they change their pricing.

Reading Claims Like an Accountant

The single most valuable habit this story teaches is separating what shipped from what was announced.

Verified: the acquisition, the IPO, the 946 MW, the 1.4 GW, the $295 million of Megapacks, the Anthropic and Google leases, the AI5 tape-out, the Terafab announcement, the 7 robotaxi metros.

Claimed: 20 GW by next year, 40x chip performance, $119 billion of fab buildout, one million orbital datacenter satellites, one million Optimus units a year.

Both categories come from the same company, often the same earnings call. Neither is a lie. They are different tenses, and reading them as the same tense is how people get their timing wrong on a genuinely important buildout.

The fog around AI infrastructure is not that the numbers are hidden. SpaceX filed them. The fog is that announcements and results arrive in the same sentence, in the same font, from the same podium — and sorting them takes work most readers never do.

That sorting is the job. Own what is scarce, rent what is improving, and always check which tense you are reading.

Sources: Reuters — SpaceX IPO filing, CNBC — SPCX IPO debut, CNBC — Q2 2026 earnings, CNBC — Tesla Megapacks to SpaceX datacenters, Epoch AI — Colossus 2, TechCrunch — Google $920M/month compute deal, TechCrunch — xAI’s $6.4B burn, TechCrunch — Terafab, Electrek — AI5 tape-out, Reuters — Anysphere acquisition, Business Insider — Tesla Q2 robotaxi metrics, Fortune — SpaceX compute rental business