xAI Burned $6.4B Last Year: SpaceX Filing Analysis

This post anchors to the SpaceX S-1 filing released today (May 20, 2026) — specifically the xAI segment's $6.4B operating loss on $3.2B revenue and the capex curve underwriting Colossus 1 and 2. The headline number is staggering, but the trajectory is the real story: xAI's operating margin deteriorated from -60% in FY2024 to -200% in FY2025, and the Q1 2026 annualized run-rate implies -302%. Scaling is making the economics worse, not better.

For advertisers, operators, and competitive-intel teams, what matters is not the loss in isolation but the structure behind it. The Anthropic compute lease — $1.25B per month, 90-day termination clause — is the single most important revenue line in the filing and the one most likely to move before the June 12, 2026 IPO. The per-Grok-MAU unit economics derived from the S-1 ($27 blended revenue vs ~$191 fully-loaded cost) explain why that lease is so critical: xAI's own user monetization does not come close to covering its cost base.

What follows is the most granular public breakdown of xAI's burn available on publication day: the FY2024 → FY2025 → Q1 2026 waterfall, the operating-margin trajectory, derived per-MAU unit economics, a cross-lab burn comparison, the Anthropic lease risk analysis, the Colossus capex pull-forward story, and what the $60B Cursor option means for the vertical-integration thesis.

The SpaceX S-1 provides the first audited financial history of the combined X / xAI segment. The waterfall is stark: revenue grew 22% year-over-year from FY2024 to FY2025, while the operating loss quadrupled. Q1 2026 continued the pattern — $818M of revenue against a $2.47B operating loss in a single quarter.

To contextualize the FY2024 baseline: in FY2024, xAI posted $2.62B in revenue and a $1.56B operating loss — a -60% margin that, while deep, was within the range of early-stage SaaS buildouts. The following year, the operating loss more than quadrupled to $6.4B while revenue grew to only $3.2B. That divergence is the story. Per the S-1 as reported by TechCrunch (Rebecca Bellan, May 20, 2026), TechCrunch's itemised revenue breakdown shows AI solutions and infrastructure at $465M, X + Grok subscriptions at $365M, advertising at $116M, and data licensing at $88M — summing to roughly $1.034B in itemised lines against the $3.2B headline figure. The remainder is aggregated under broader S-1 reporting categories; the full revenue segmentation table requires the SEC EDGAR PDF to reconcile.

What does a -200% operating margin mean in practice? For every dollar of revenue xAI generated in FY2025, it spent three dollars. At the Q1 2026 annualized rate, it is spending four dollars for every one it earns. Unlike early-stage consumer software where unit economics improve with scale, xAI's trajectory is moving the wrong direction — because the cost structure is dominated by compute capital expenditure that grows with each new cluster deployment, not with user count. This is the central analytical challenge for IPO investors: revenue growth and loss growth are currently uncorrelated.

The critical forward question — which the S-1 does not answer — is what revenue growth rate would be needed to bend the margin curve. Based on the S-1 numbers, closing the gap to a -100% margin in three years would require revenue to roughly double while capex stays flat. Neither assumption is supported by the filing's own forward capex trajectory. Our analytics practice covers financial data interpretation and scenario modeling for teams tracking AI-sector investment.

The operating-margin trajectory is the most important single data series in the filing for competitive-intelligence purposes. Not because the loss number is large in isolation — every frontier AI lab is burning capital — but because the direction is wrong. Conventional SaaS and platform economics assume that scaling revenue reduces the loss-to-revenue ratio over time. xAI's S-1 shows the opposite pattern over the past two fiscal years.

The chart below compares xAI's disclosed margin trajectory against peer figures. OpenAI's FY2024 ratio is per The Information and NYT reporting summarized by independent analysts — not an audited S-1. Anthropic's figures are per outlet reporting (VentureBeat, SaaStr) and should be treated as estimates; Anthropic has no public S-1. Google DeepMind's operating margin is not cleanly isolable from Alphabet's segment disclosures and is excluded from this comparison.

The S-1 discloses that Grok had 117M monthly active users as of March 2026 — 21.3% of the X platform's 550M MAU. No other social platform has disclosed an in-app AI-feature MAU figure in a public SEC filing at this scale. Combined with the segment's $3.2B FY2025 revenue, this makes a per-MAU derivation possible — though the numbers require careful framing.

Every figure below is a derived ratio from S-1 disclosures, not a SpaceX-disclosed per-user metric. The AI segment's revenue and cost base bundles X platform infrastructure, Grok inference, and xAI research overhead that are not split per-user in the filing. These ratios are useful as benchmarks and for understanding the monetization gap — not as ARPU figures for investor modeling.

The practical implication of this derivation is that xAI's user-side monetization — subscriptions, advertising, data licensing — generates roughly 14 cents of revenue for every dollar of fully-loaded cost per Grok MAU. The other 86 cents must come from somewhere. In FY2025, the answer was the Anthropic compute lease: $1.25B/month in external revenue flowing from Anthropic renting Colossus 1 while xAI moves its own training workloads to Colossus 2. That makes the lease not a bonus revenue line but the structural support for the entire cost-versus-revenue equation. See our Q2 2026 model-API pricing tracker for the broader per-token economics context that this burn is funding.

Comparing AI lab burn rates is methodologically fraught: xAI is the only frontier AI company with an audited S-1 as of May 20, 2026. OpenAI financials are from The Information and NYT reporting summarized by independent analysts. Anthropic figures are from VentureBeat, Sherwood News, and SaaStr summaries of financial leaks. Google DeepMind cannot be cleanly isolated from Alphabet's segment disclosures. The matrix below builds the best available picture — with disclosure quality flagged explicitly.

The most striking finding from this comparison is not that xAI loses more money than OpenAI in absolute terms — OpenAI's reported ~$5B FY2024 loss is substantial. It is that xAI achieves a worse loss-to-revenue ratio at a smaller revenue base. OpenAI's -135% FY2024 ratio was on $3.7B of revenue; xAI's -200% FY2025 ratio was on $3.2B of revenue. If OpenAI's own forward forecast of a ~$14B FY2026 loss holds — per Fortune's reporting on leaked financial documents — that loss is projected against a much larger revenue base, meaning the ratio may improve even as the absolute number grows. xAI does not publish forward loss forecasts. For broader context on Anthropic's own cost problem, the contrast with the compute-lease revenue model is instructive.

The SpaceX S-1 discloses that Anthropic will pay xAI $1.25B per month for compute access to the Colossus 1 cluster, running through May 2029 — a contract worth approximately $45B in total if maintained for its full term. Annualized, that is $15B per year, per reporting by Axios and TechCrunch on May 20, 2026. The structure makes sense as a market transition: Anthropic gains access to a 220,000+ GPU Colossus 1 cluster while building or securing its own capacity; xAI gains external revenue that helps offset the capex of building Colossus 2.

What most S-1 coverage buries: the lease contains a 90-day termination clause exercisable by either party. Tesery's teardown of the Anthropic deal terms, published May 21, 2026, surfaces this as the key risk factor. At $15B/year (~$1.25B per month), this contract represents approximately 46% of xAI's annualized Q1 2026 revenue run-rate ($3.3B annualized). If Anthropic exercises a 90-day termination in year one, the floor on the contract drops to roughly $3.75B — a $41B reduction from the full-term value. That scenario doesn't require bad faith; it requires only that Anthropic's own cloud compute agreements (reportedly ~$80B in commitments through 2029 per SaaStr) come online faster than expected, making Colossus 1 rental unnecessary.

The Colossus cluster buildout is the primary capex driver in the S-1. Colossus 1, located in Memphis, deployed 220,000+ NVIDIA H100, H200, and GB200 GPUs across more than 300MW of capacity — built in 122 days. Colossus 2 (Southaven, Mississippi) followed with a disclosed deployment of 91 days and combined nameplate capacity of approximately 1GW, per TechCrunch's S-1 reporting. A third Memphis-area building is under construction targeting ~2GW total capacity, per Tom's Hardware.

One critical distinction that S-1 coverage conflates: the S-1 does not appear to break out Colossus 2's GPU count or per-cluster build cost in line-item form. Industry estimates — from SemiAnalysis and Introl — put Colossus 2 at approximately 555,000 GB200-class GPUs and a headline build cost in the range of ~$18B. These are industry estimates, not S-1 disclosures. We flag them explicitly to avoid presenting market research as regulatory filings. The S-1's confirmed figures are the consolidated capex totals: $12.7B FY2025 and $7.7B Q1 2026.

The deployment-speed acceleration (122 days to 91 days) demonstrates that xAI is operationally capable of compressing build cycles — but the cost structure of each cycle grows with the cluster size. This is what we describe as “capex pull-forward”: the $7.7B Q1 2026 single-quarter capex suggests Colossus 2 construction costs were concentrated in that quarter, not spread evenly. Once the transition is complete (Anthropic on Colossus 1, xAI training on Colossus 2), the annualized capex rate may moderate — but the S-1 provides no forward capex guidance to support that projection. For the broader context of this buildout within the 9-18 GW AI infrastructure shortage, see our analysis of AI infrastructure's gigawatt shortage.

xAI does not publish forward loss forecasts. The S-1 provides FY2024, FY2025, and Q1 2026 actuals — nothing beyond. The only forward loss projection available for any frontier AI lab as of this filing date is OpenAI's own ~$14B FY2026 loss forecast, sourced from Fortune's reporting on leaked financial documents (November 2025). OpenAI targets profitability by 2030 per those same documents.

If xAI's Q1 2026 operating loss of $2.47B annualizes — which we cannot assume but can use as a reference point — the implied FY2026 loss would be approximately $9.9B. OpenAI's own projected $14B loss would occur on a materially larger revenue base ($34B+ per various 2026 analyst estimates), implying a better loss-to-revenue ratio. xAI on a $9.9B annualized loss trajectory would be running at a worse ratio unless revenue accelerates significantly from the Q1 2026 $818M quarterly run-rate.

The revenue acceleration scenario depends primarily on three levers: Grok MAU growth and paid conversion (from the current 117M MAU / 1.9M paid-Grok-sub base), API and enterprise AI-solutions revenue (the $465M itemised line in FY2025), and third-party compute revenue (the Anthropic lease). Of these, the lease is the most near-term certainty; the Grok MAU monetization story requires a step-change in ARPU that the S-1 does not yet show. See our analysis of Grok 4.3's platform asymmetry for the product context behind the MAU numbers.

On April 21-22, 2026 — approximately a month before the S-1 filing — SpaceX disclosed a $60B acquisition option for Cursor, the AI-native coding tool, exercisable approximately 30 days after the IPO closes, with a $10B breakup fee. The strategic logic connects directly to the financial picture in this filing: Cursor's coding-model layer provides the user-facing product surface that could convert Colossus 2's raw training capacity into recurring enterprise revenue.

The thesis, as we read it from the filing sequence: xAI trains Grok on Colossus 2 (the infrastructure layer), Cursor provides enterprise developer distribution (the application layer), and Grok's 117M MAU on X provides the consumer social proof. The combination would position SpaceX / xAI as the only entity with training compute, frontier model, social distribution, and a developer-tool GTM under one cap table.

The financial reality check: a $60B Cursor acquisition at current loss trajectories would add meaningful debt to a company already burning at $2.47B per quarter. Whether IPO capital ($75B target raise at the $1.75T valuation) is sufficient to absorb both the acquisition and the ongoing capex curve is a question the S-1 does not model explicitly. For the broader Grok product picture that the Cursor thesis depends on, see our analysis of Grok 4.20's 2M-context release.

The S-1's language on forward intent is the clearest statement xAI has made publicly about where the capex is going. Per TechCrunch (May 20, 2026) surfacing the S-1 disclosure language: xAI plans to scale Grok toward “multiple trillions of parameters,” described as enabling “a step change in reasoning in depth and overall intelligence.” This is company-authored marketing language from the filing itself — not an independent technical assessment — and should be read as the company's stated thesis, not a verified technical capability projection.

The orbital compute angle is real but distant. The S-1 references xAI's plans for orbital AI compute satellites, with deployment “no earlier than 2028.” SpaceX's Starship and Starlink manufacturing capacity makes this a credible long-range roadmap item — but it has no near-term impact on the FY2025-2026 financial picture we are analyzing.

The analytical interpretation of “multiple trillions of parameters” as a stated goal is that the current Colossus 1 + Colossus 2 combined infrastructure is not the terminal state. The third Memphis building (targeting ~2GW total combined) and whatever follows are each additional capex events. This is the structural argument for why the spending will not self-limit: Musk has publicly stated a goal of having more AI compute than any other entity, and the S-1 formalizes that ambition as a disclosed strategy — which means IPO investors are being told explicitly that the capex curve is the intended trajectory, not a temporary phase.

For competitive-intelligence teams, the relevant question is not whether xAI will stop spending — the filing makes clear it will not — but whether the Anthropic lease, the Grok MAU monetization curve, and future enterprise AI-solutions revenue can grow fast enough to narrow the gap before the IPO capital runs out. The broader S-1 financial picture across Starlink, launch, and AI is covered in the companion Day 5 analysis.