Autoresearch: SpaceX/xAI compute economics vs IPO folklore (as of 19 Sep 2026)

Generated by /autoresearch on 2026-09-19. Synthesized across 3 rounds from 12 web pages, anchored by a stale Grokipedia SpaceX primer (not used for 2026 claims). See Provenance. Treat as raw material — review before promoting. Honesty: stay hypothesis. Steps 2–3 stay partial. Do not graduate. Do not mint a new AI-infra chain. Do not re-rate SPCX. Context: vault/projects/stock-market (elon-web-services-to-spacex-hyperscaler-rerate)

Summary

Issuer 424(b)(4) (June 2026) and the Q2 2026 8-K / 10-Q now disclose an AI reportable segment and two named Cloud Services Agreements (Anthropic $1.25 billion/month; Google $920 million/month). That is the first primary segment P&L the chain asked for. It does not confirm the Baker/Fox folklore: SpaceX/xAI does not publish operating profit per gigawatt, and the $14B / $22–23B / $50B per-GW figures remain podcast-only. GAAP AI operating profit is still a loss (Q2 2026 $(1,257) million) even after CSA revenue; Connectivity is the profit engine. No issuer, NVIDIA, Anthropic, xAI, or OpenAI document ranks SpaceX as a top-4 AI hyperscaler or says it passed Oracle in ~30 days. Patel’s lease-to-Anthropic claim is now issuer- and counterparty-confirmed; the OpenAI half is still unconfirmed. Land/power: nameplate compute rose 0.4 → 1.0 → 1.4 GW, but the 424 risk factor, the Google 30 Sep 2026 delivery gate, and the Southaven turbine suit mean Baker’s Step-1 falsifier is not cleared.

Findings

First disclosed AI-segment economics (Marshall’s named test)

The 424(b)(4) prospectus filed 12 June 2026 is the first issuer print of AI as a segment (SpaceX 424(b)(4)):

• Q1 2026: AI revenue $818 million, loss from operations $(2,469) million, Segment Adjusted EBITDA $(609) million.
• FY 2025: AI revenue $3,201 million, loss from operations $(6,355) million, Segment Adjusted EBITDA $(1,237) million.
• Same prospectus, same page: Connectivity Q1 income from operations $1,188 million / FY 2025 $4,423 million. Space remained an operating loss.

The post-IPO print updates the same three-segment table (SpaceX Q2 2026 earnings release, Ex. 99.1; Form 10-Q for the quarter ended 30 June 2026):

Q2 AI revenue mix: advertising $367 million, “AI solutions & infrastructure” $2,194 million. The 8-K attributes the sequential jump to “new Cloud Service Agreements” and says those CSAs totaled $14.1 billion in contracted sales and $1.6 billion of incremental Q2 AI-infrastructure revenue. Note 2 of the same release defines contracted sales as the non-cancellable, enforceable period only — not the headline through-2029 arithmetic.

Q2 Connectivity income from operations was $1,656 million on $4,291 million revenue. Consolidated Q2 GAAP operating loss was $(143) million. The first disclosed segment economics therefore show compute/hosting as a material, still-loss-making-on-GAAP, Adj-EBITDA-positive-in-one-quarter line sitting beside a profitable connectivity business — closer to Marshall’s “data center play / very different business” than to a hyperscaler-multiple AI software story. They do not by themselves graduate Step 3.

Will Marshall (Planet CEO), 26 June 2026 Moonshots: “But that's not an AI play, that's the, that's the data center play, which is interesting, but it's a very different business.” That remains the named quality objection. First print does not refute it on GAAP operating profit.

Named CSA economics vs $14B / $22–23B / $50B-per-GW folklore

Those three per-GW figures are still podcast-only. They appear in Andrew Fox on BG2 (11 June 2026): “The implied monetization rate on that number is something like $14 billion per gigawatt per year for the AI business. They just signed Anthropic at 22 to 23. They just signed Google at 50.” Gavin Baker, same show: “xai's deal with Google for cloud computing generates more operating profit per gigawatt than Anthropic, than Meta, than Google, than OpenAI.” Issuer documents never state operating profit per gigawatt, never use those three numbers as a unit metric, and never rank counterparties on OP/GW.

What is primary:

1. Anthropic — May 2026. SpaceX 424(b)(4): Cloud Services Agreements with Anthropic PBC for capacity across COLOSSUS and COLOSSUS II; “approximately 325,000 NVIDIA GPUs”; “the customer has agreed to pay us $1.25 billion per month through May 2029, with capacity ramping in May and June 2026 at a reduced fee. After the initial three-month period, the agreements may be terminated by either party upon 90 days’ notice” (424(b)(4)). SEC staff correspondence asked SpaceX to quantify capacity and consider filing the contract under Item 601(b)(10); SpaceX said the CSAs are not contracts upon which the business is substantially dependent, citing the 90-day walk-away and redeployability (S-1/A correspondence).
2. Anthropic’s own page (6 May 2026) says it will use all Colossus 1 capacity: “more than 300 megawatts of new capacity (over 220,000 NVIDIA GPUs) within the month,” for Claude Pro/Max, plus interest in orbital GW (anthropic.com/news/higher-limits-spacex). That 220k / 300 MW figure is narrower than the issuer’s 325k GPUs across Colossus and Colossus II. Do not collapse them.
3. xAI / SpaceXAI (6 May 2026) matches Anthropic on Colossus 1, “over 220,000 NVIDIA GPUs,” and orbital interest (x.ai/news/anthropic-compute-partnership). It does not publish a monthly fee.
4. Google — 5 June 2026 free-writing prospectus. “On June 5, 2026, we entered into a Cloud Service Agreement with Google LLC… approximately 110,000 NVIDIA GPUs… $920 million per month from October 2026 through June 2029, with capacity ramping up through September at a reduced fee.” Failure to deliver the committed GPUs by 30 September 2026 lets Google terminate after a one-month grace or take a pro-rata cut. After 31 December 2026, either party may terminate on 90 days’ notice (spacexagreementfwp).
5. Cursor. April 2026 compute + $60 billion option; June 2026 call exercised; 10-Q says the Q2 net amount on the compute collaboration “was not material” (10-Q).

Annualizing the headline monthly fees ($15.0 billion Anthropic; $11.04 billion Google) and dividing by an assumed GW is not an issuer OP/GW. The 8-K’s $14.1 billion contracted sales is a defined, cancellable-period total — not Fox’s $14 billion per GW per year. Anyone mapping Fox’s 22–23 / 50 onto $1.25B and $920M per month is doing podcast arithmetic, not reading a line item.

Q2 AI operating profit is still negative. Q2 AI Adj. EBITDA of $1,146 million on 1.4 GW nameplate is a non-GAAP, one-quarter print that includes advertising, Grok/X subscriptions, and CSA hosting; it is not “highest OP/GW versus Anthropic/Meta/Google/OpenAI.” Customer B (AI-segment only) was 19.5% of consolidated Q2 revenue (~$1.52 billion of $7,814 million) — concentration, not a peer ranking (10-Q Note 3).

Hyperscaler rank / “passed Oracle in ~30 days” / compute as the under-priced IPO leg

Gavin Baker on BG2: “In 30 days we went from not being an AI hyperscaler to being number four. And we passed a lot of companies, including Oracle.” Brad Gerstner, same show: “we merged in X AI and we merged in Cursor, and then we announced these deals where it was very clear he was kind of building AWS right under our nose.” Both are shareholders on their own podcast. No corroborating primary:

• 424(b)(4) competitive language: SpaceX “intend[s] to sell our excess capacity… potentially positioning us to emerge as a competitor to AI cloud providers such as Coreweave and Nebius as well as hyperscalers.” That is forward-looking positioning, not a #4 rank or an Oracle pass (424(b)(4)).
• NVIDIA’s 28 October 2024 Colossus PR describes a 100,000-Hopper cluster and (in the newsroom PDF) a 122-day build; it does not rank SpaceX among hyperscalers (NVIDIA newsroom PDF; GlobeNewswire reprint).
• Anthropic and xAI announce a Colossus 1 lease, not a hyperscaler league table.
• OpenAI’s 2026 compute pages (Stargate / Oracle / Amazon / NVIDIA) do not name SpaceX (openai.com/index/building-the-compute-infrastructure-for-the-intelligence-age; openai.com/index/stargate-advances-with-partnership-with-oracle).

On mix: Q2 AI revenue ($2.6 billion) is now larger than Space ($1.0 billion) and smaller than Connectivity ($4.3 billion). AI capex ($15.8 billion in Q2; $23.6 billion H1) dwarfs the other two segments combined. IPO use-of-proceeds language lists “expansion of the Company’s AI compute infrastructure” first among growth uses (10-Q). That is not a primary statement that compute is the under-priced equity leg. GAAP profits still sit in Connectivity. Step 3 stays partial.

Land and power (Step-1 falsifier: “if they can get the land and the power”)

Issuer nameplate compute draw — GPUs × all-in power, not utilization, not facility overhead — was 0.4 GW at 30 June 2025, 1.0 GW at 31 March 2026 (424), 1.4 GW at 30 June 2026 (8-K / 10-Q). The 8-K: “Expanded compute capacity to 1.4 GW… with the continued build-out of Colossus II and with significant incremental capacity under construction.” The 424: Colossus + Colossus II “approximately 1.0 gigawatt of compute power, with additional power capacity available”; next Colossus II phase “an additional 400MW”; Colossus II “capable of operating entirely by our self-built behind-the-meter gigawatt-scale natural gas power.” Cluster clocks in the same filing: first Colossus cluster 122 days on a reused factory shell; first Colossus II cluster 91 days; greenfield 100 MW benchmark ~2 years. This pass’s 424(b)(4) text does not restate a 64-day third-cluster clock.

Constraints that keep the falsifier open:

• 424 risk factor: scale “increasingly constrained by the availability of power and water,” long-lead arrangements, gas-turbine supply, and “we currently rely significantly on natural gas and gas turbine technology” (424(b)(4)).
• 10-Q: NAACP suit (14 April 2026) alleging Clean Air Act violations for mobile gas turbines powering Colossus II in Southaven, Mississippi; preliminary-injunction motion 6 May 2026 (10-Q Note 16).
• Google FWP: committed-GPU delivery by 30 September 2026 or Google may walk.
• 10-Q AI risk: CSAs are monthly fees, cancellable on 90 days after ramp; “power constraints… permit delays could impair our ability to deploy capacity.”

Baker (BG2): “Excited about it if they can get the land and the power.” Nameplate +0.4 GW in Q2 is evidence of some pace. It is not evidence that land/power is secured at a multi-GW implied run-rate, and the issuer itself flags the constraint. Step 1’s cluster-clock confirmation (already tagged) is unchanged; the campus/power falsifier is neither triggered nor cleared.

Patel: SpaceX will lease compute to Anthropic and OpenAI

Anthropic — confirmed on primary. Counterparty blog, xAI blog, and issuer 424/S-1 staff letters all describe a May 2026 Colossus CSA. That is no longer a SemiAnalysis-only claim.

OpenAI — not confirmed on primary as of 19 Sep 2026. OpenAI’s official infrastructure posts name Oracle, SoftBank, CoreWeave, Amazon, Microsoft, and NVIDIA — not SpaceX. SpaceX filings name Anthropic, Google, and Cursor; they do not name OpenAI as a compute tenant. Customer B is unidentified. Trade press speculation about a later unnamed $1.11 billion/month hosting deal is not treated as fetched primary here.

Patel’s wording (Dwarkesh, 25 August 2026 — this is the 2026 claim, not an August 2025 SemiAnalysis lease paper): “SpaceX is building a ton of compute, and they're actively going to lease quite a bit of it to Anthropic and OpenAI, most likely because they're the ones who have the marginal capability to pay the highest price.” Same episode: “Elon showed, actually, no, I can sell my compute for $25 million a megawatt or $40 million a megawatt to Anthropic in Google.” The $25–40M/MW figures are Patel’s, not issuer OP/GW.

The August 2025 SemiAnalysis piece that search returns (xAI's Colossus 2) is a build-out note (Colossus 1 ~300 MW; Colossus 2 toward GW), not a documented Anthropic/OpenAI lease. Do not back-date the 2026 lease claim to August 2025.

Contradictions and open questions

• OP/GW folklore vs issuer P&L. Fox’s $14 / $22–23 / $50 per GW and Baker’s “highest operating profit per gigawatt” have no issuer, NVIDIA, Anthropic, Google, or OpenAI counterpart. Q2 AI GAAP OP is still a loss. Stay partial on Step 2.
• Anthropic GPU/MW mismatch. Issuer 325k GPUs across two campuses vs Anthropic/xAI “all of Colossus 1” / 220k GPUs / >300 MW. Both can be true if Colossus II capacity is in the CSA and not in the consumer blog; the documents do not reconcile the split.
• Cancellable “through 2029” headlines. Staff letter + 424 + Google FWP + 10-Q risk factor: 90-day termination after a short initial period. The 8-K’s $14.1 billion contracted sales is the honest duration metric.
• #4 / Oracle. Still single-source Baker. Issuer will only say it may compete with CoreWeave, Nebius, “as well as hyperscalers.”
• OpenAI tenant. Patel “most likely”; no primary CSA. Unnamed Customer B and later conference rumors are not a substitute.
• 64-day clock. Prior wiki note; not found in this pass’s 424(b)(4) body (122 and 91 only).
• Land/power. 1.4 GW nameplate and +400 MW next-phase language vs turbine litigation, gas/water risk factor, and Google’s 30 Sep 2026 delivery gate.
• Marshall vs Baker. First segment print is live. GAAP AI losses + Connectivity profits + short-notice hosting lean Marshall on business quality; CSA dollar scale and one quarter of positive AI Adj. EBITDA keep Baker’s revenue claim from dying. Neither side gets a graduation.

Provenance

Rounds run: 3 of 3

Sub-questions by round:

Round 1 (broad survey):

1. Do SpaceX 424 / S-1 / IR filings disclose AI-segment economics, OP/GW, or the $14B / $22–23B / $50B-per-GW figures?
2. Have NVIDIA, Anthropic, OpenAI, or xAI issued primary statements that SpaceX is a top-4 hyperscaler, passed Oracle, or is leasing compute to the labs?
3. What do issuer filings say about land/power versus the implied campus pace?
4. Has a first segment print confirmed or refuted Marshall’s “very different business / not an AI play”?

Round 2 (drill-down):

1. Extract exact Anthropic / Google / Cursor CSA terms from 424, FWP, and staff letters — targeting folklore vs filed monthly fees.
2. Pull Q2 8-K / 10-Q AI P&L, nameplate GW, contracted sales, and customer concentration — targeting first post-IPO segment economics.
3. Locate Patel’s Aug 2025 vs Aug 2026 lease-to-OAI/Anthropic wording — targeting whether the dual-lab claim is primary.

Round 3 (resolve remaining uncertainty):

1. Confirm Google CSA exact EDGAR URL and OpenAI official silence — targeting any missing named tenant.
2. Confirm issuer never states #4 / Oracle / OP/GW — targeting Step 2–3 stay-partial.

Anchor source (Grokipedia, fetched before round 1):

• SpaceX — 8,017 chars extracted — stale (still treats SpaceX as private as of 24 March 2026). Not used as a 2026 source.
• xAI topic resolved to the CBC series X Company — no usable xAI encyclopedia anchor.

URLs fetched (12 successful, 2 failed):

Round 1:

• SpaceX 424(b)(4) — official / issuer — segment P&L, 122/91-day clocks, Anthropic CSA, 1.0 GW, power risk.
• SpaceX S-1/A#1 — official / issuer — same economics as 424 draft.
• SEC staff letter on Anthropic CSA — official — $1.25B/month, 90-day terminate, Item 601(b)(10) analysis.
• Anthropic: Higher usage limits and a SpaceX compute deal (6 May 2026) — official — Colossus 1, >300 MW / >220k GPUs.
• xAI: New Compute Partnership with Anthropic (6 May 2026) — official — same Colossus 1 lease; no fee.
• NVIDIA newsroom PDF, 28 Oct 2024 — official — 100k Hopper, 122 days, 19 days rack-to-training.

Round 2:

• SpaceX Q2 2026 earnings release — official / issuer — Q2 AI P&L, 1.4 GW, $14.1B contracted sales, $1.6B incremental infra.
• SpaceX Form 10-Q, quarter ended 30 June 2026 — official / issuer — CSA accounting, Customer B 19.5%, nameplate definition, NAACP suit, IPO proceeds.
• Dwarkesh / Patel transcript (Mar 2026 episode, for contrast) — interview — $10–13B/GW rental color; not the Aug lease claim.
• 404k reprint of later Patel remarks — secondary reprint of the Aug 2026 lease-to-both-labs line.

Round 3:

• SpaceX Google CSA FWP (5 June 2026) — official / issuer — $920M/month, 110k GPUs, 30 Sep 2026 delivery gate.
• OpenAI: Building the compute infrastructure for the Intelligence Age (29 Apr 2026) — official — Stargate / Oracle; no SpaceX.

Failed / not used as 2026 sources:

• [Failed: https://investor.nvidia.com/news/press-release-details/2024/NVIDIA-Ethernet-Networking-Accelerates-Worlds-Largest-AI-Supercomputer-Built-by-xAI/default.aspx] — Cloudflare challenge; same PR recovered via NVIDIA newsroom PDF + GlobeNewswire.
• [Failed: https://ir.spacex.com/] — timeout; Q2 numbers taken from the EDGAR 8-K (same release the IR page points at). IR event stub exists (4 Aug 2026).

Not fetched (secondary / off-whitelist / encyclopedia): CNBC/Yahoo/DCD recaps of the CSA; Scientific American / Daily Memphian turbine features; Epoch campus MW estimates; Wikipedia Colossus page; SemiAnalysis paywalled Colossus-2 post (search-snippet only). Grokipedia not cited for 2026 facts.

Tools used: WebSearch, WebFetch, grokipedia-fetch (_lib/grokipedia.py). No --include-x. Generated: 2026-09-19 21:30 UTC