The xAI Compute Corridor

Colossus is xAI's AI training supercomputer in Memphis, Tennessee — now part of SpaceX after the February 2026 consolidation. Its scale-up is among the fastest in computing history: the first cluster stood up in 122 days and doubled to roughly 200,000 GPUs in another 92, a December 2025 mix of about 150,000 Nvidia H100s, 50,000 H200s, and 30,000 GB200s, with a further 110,000 GB200s at a second Memphis datacenter. By early January 2026 the current read was on the order of 555,000 GPUs drawing roughly 2 gigawatts of power across three buildings, with an announced target of one million GPUs. SemiAnalysis has described the Colossus 2 buildout as the first gigawatt-scale datacenter in the world. The million-GPU figure is a stated target rather than a deployed count, and the "first gigawatt" framing is analyst-stated — but even the verified figure is a power footprint comparable to a mid-size nuclear plant, which is why the question of where that electricity comes from is not a footnote but the whole story.

Gigawatt-scale AI clusters need large, stable power conditioning and storage, and the batteries doing that work for Colossus are Tesla Megapacks. xAI bought $269 million of Tesla Megapacks in April 2026 alone — a transaction documented in SpaceX's S-1 related-party section and part of roughly $1 billion of Megapack purchases across the Musk companies in under two and a half years. That is the literal mechanism by which the energy chain (Tesla) powers the AI chain (xAI): the carmaker's grid-storage hardware buffers the supercomputer's enormous, spiky power draw. The connectivity backbone leans on Starlink, and the capital comes from SpaceX. So xAI's compute does not run on standalone infrastructure; it runs on the other three chains. This is the clearest physical illustration of the dependency thesis — pull Tesla's batteries or Starlink's uplink or SpaceX's capital, and the training run stops. The AI node is not independent; it is downstream of the energy, connectivity, and launch nodes at once.

By drawing on the rest of SpaceX, and by separating two costs the headlines blur. xAI's operating loss — roughly $6.355 billion against $3.2 billion of revenue in FY2025 — is the cost of running the models, and it is the biggest single driver of SpaceX's consolidated $4.9 billion net loss. The capital expenditure to build the Colossus clusters is a different number: on the order of $12.7 billion in 2025, funded not from operations but from a financing stack of roughly $26 billion in debt and preferred equity, because the segment's capex outran its operating cash. The thing that makes this sustainable is the cash engine: Starlink generated about $4.4 billion of segment operating income in 2025, and that captive margin — plus the IPO's roughly $75 billion — underwrites the AI bet. A standalone AI lab must raise external capital for each gigawatt; xAI's buildout is financed by a satellite business it shares a balance sheet with. That is a structural funding edge, not an accounting trick.

Not without owning a cash engine the size of Starlink, which they do not. OpenAI and Anthropic finance their compute buildouts through external capital — successive mega-rounds, strategic investments, and debt — and each gigawatt of training infrastructure is gated by their ability to raise the next tranche. xAI's compute is gated by a different constraint: Starlink's margin. The connectivity segment's roughly $4.4 billion of 2025 operating income is captive internal cash that funds the AI buildout without a fundraising cycle, supplemented by SpaceX's own debt capacity and the IPO proceeds. That is why we describe xAI as a child of the connectivity node rather than a peer of other labs. A rival could match Grok's model quality on paper; it cannot match the financing structure without first building a profitable, globe-spanning satellite network to throw off the cash. The moat here is not the model — models converge — it is the balance sheet the model is attached to, and that balance sheet is the dependency graph.

It is the AI chain's physical proof. This site argues Elon Musk is unremovable because four chains — launch, connectivity, energy, AI — all route through entities he controls and are fused into one balance sheet. The compute corridor shows the AI chain is not a free-standing fourth bet but a dependent of the other three: Colossus runs on Tesla Megapack power (energy), Starlink uplink (connectivity), and SpaceX capital (launch margin), all inside one consolidated company since the February 2026 merger. You cannot remove xAI from the picture without removing the things it runs on, and you cannot remove those without breaking the businesses that depend on them in turn. That circularity is the unremovability claim in its most physical form — not a strategy slide but a power-and-capital diagram. The corridor is also where the bull and bear cases meet: it is the segment Morningstar discounts to near zero on our /bear-case/ page and the optionality /valuation/ says the multiple is buying. Both are pricing the same gigawatts.