Question: SpaceX's S-1 (filed 2026-05-20, debut 2026-06-12) disclosed a Cloud Services Agreement with Anthropic at $1.25B/month (~$15B/yr, ~$45B through May 2029) — the largest single customer-concentration disclosure in the filing. What is Anthropic actually getting, why this counterparty instead of hyperscalers, what does it mean for Anthropic's compute strategy, and what are the ticker implications across hyperscalers + neoclouds + NVDA?

Verdict (TL;DR):

1. The asset is xAI's Colossus 1 in Memphis, not a SpaceX-native cloud product. ~220K NVIDIA GPUs (mixed ~150K H100 + ~50K H200 + ~20-30K GB200), 300 MW, mixed-architecture inference cluster. xAI couldn't use it for training past a certain scale (straggler effect, reported ~11% effective utilization for training); Anthropic took the entire stranded asset. Post-merger, SpaceX is the contracting counterparty.

2. This is inference capacity, not training. Anthropic explicitly stated the deal doubles Claude Code rate limits and removes peak-hour throttling on Pro/Max. Anthropic's ARR went $1B → $30B in 16 months; capacity (not demand) was the gating factor.

3. It's ~10-12% of Anthropic's compute commitment stack but the only NON-circular line. Total disclosed compute book: AWS $100B+/10yr + Google/Broadcom $200B/5yr + MSFT/NVDA $30B + SpaceX $45B = ~$375-400B. The hyperscaler lines are structurally circular (AMZN $13B equity / GOOG $40B equity / MSFT-NVDA $15B equity → cash recycled back as compute spend). SpaceX is arm's-length premium-priced cash — exactly why it shows up as a clean line in an S-1.

4. Premium pricing, not strategic. Implied ~$50M per MW-year vs AWS Rainier ~$10M per MW-year-equivalent on the strategic-investor-priced commitment. Anthropic paid sticker for time-to-capacity — xAI was the only operator with live, idle, 300MW-scale capacity in Q2 2026. Everything else (5GW AWS, 5GW GOOG, MSFT/NVDA Vera Rubin) comes online 2027+.

5. Anthropic is becoming a circular-financing vehicle for compute. Roughly half of Google's and Amazon's 2026 artificial-intelligence segment profits came from their Anthropic stakes' mark-to-market, NOT operating margin (Fortune Apr 2026). Anthropic raised at $1T pre-IPO trading (Apr 2026 capture) against an $80B cumulative cloud spend commitment through 2029. Anthropic is effectively a leveraged bet on Claude becoming the dominant enterprise model — hyperscaler equity → Anthropic → hyperscaler revenue → hyperscaler artificial-intelligence segment "profit."

6. The 90-day mutual termination clause is the safety valve and the systemic risk. Either side can walk in 90 days. Probability of full ~$45B running through May 2029: ~55-65% bull / 35-45% bear. Most likely outcome: 2-3 year run with renegotiation in 2027-2028 when other capacity comes online.

7. Tenant-isolation IP language is the single biggest unknown. xAI is Anthropic's direct competitor (Grok vs Claude). Anthropic publishes red-team work on Grok routinely. Whether Anthropic's weights, training data, and customer prompts are operationally isolated from xAI staff is unspecified in the S-1. This is the single piece of evidence that would most change the read.

→ Primary-source upgrade — filed/amended S-1/A (2026-06-18)

The original streams below were built off the 2026-05-20 S-1 (secondary-read). The amended S-1/A (CIK 1181412, the version SpaceX priced + debuted on 2026-06-12) materially expands the deal and surfaces a second tenant. Web-verified 2026-06-18 (Cryptobriefing / Tesery / ActuIA / TechCrunch). Anthropic/SpaceX child-track work — routed here + to spacex-ipo-liquidity-event, NOT the AI-mega-IPO lane.

Two corrections + one new fact that change the read:

1. Scale up: 220K/Colossus-1 → ~325K across Colossus I + II. The deal is bigger and spans both clusters — the price held at $1.25B/mo, so the implied per-GPU rate fell (more GPUs, same cash) vs the original single-cluster framing. The "$15B SpaceX/xAI" framing in circulation stays wrong: it is SpaceX = provider / Anthropic = customer (Colossus is the xAI-built asset SpaceX now monetizes), confirmed in the priced filing.
2. 90-day termination is after a 3-month initial period — not from day one. Marginally firms the near-term floor (a ~3-month minimum) while keeping the long-run optionality (and the systemic risk) the streams below describe intact. Bull/bear on the full ~$45B running to May 2029 is unchanged (~55-65% / 35-45%).
3. NEW — SpaceX is now a multi-tenant AI compute landlord. The same S-1/A discloses Google leasing ~110,000 GPUs at ~$920M/month from SpaceX (TechCrunch/MLQ, 2026-06-05). So SpaceX runs a two-named-tenant compute-lease book (Anthropic $1.25B/mo + Google $920M/mo ≈ $2.17B/mo / ~$26B/yr run-rate of disclosed AI-compute rent). This reframes SpaceX post-IPO as an AI-infrastructure-landlord, not only a launch/Starlink play — and it dilutes the Anthropic single-customer-concentration worry on SpaceX's side (two anchor tenants now).

The Anthropic↔SpaceX compute-landlord dependency (the angle this upgrade adds): Anthropic's compute book now carries a line where its landlord (SpaceX) owns xAI — Anthropic's direct competitor — the capacity is capped at ~11% of that landlord's park, and the lease is terminable on 90 days after a 3-month floor. That is a structurally weaker claim than an owned cluster or a hyperscaler-equity-backed credit: it's rented, competitor-adjacent, minority-share, and short-dated. It is the cleanest (non-circular, cash) line in Anthropic's stack and the most terminable — the two facts coexist. The tenant-isolation IP unknown (#7 above) bites hardest exactly because the landlord is the competitor.

What we're asking

The existing spacex-ipo-liquidity-event perspective and its 2026-05-26 deep-dive flagged the Anthropic contract as the largest customer-concentration disclosure but didn't dimension it. This investigation does the dimensioning: contract structure, asset class (compute / power / GPU type / network), Anthropic financials, hyperscaler counterparty dynamics, neocloud read, and bull/bear on the termination optionality.

What we found

Stream 1 — Deal structure + counterparty dynamics

Confirmed contract terms (S-1):

• $1.25B/month through May 2029
• ~$15B/year run-rate; ~$45B total contract value
• Cloud Services Agreement for COLOSSUS / COLOSSUS II datacenter access
• "Terminable by either party on 90 days notice"
• Anthropic retains ownership of content, models, and data

The asset (industry reporting — DCD, Tom's Hardware, SemiAnalysis):

• xAI's Colossus 1 in Memphis, TN
• ~220K NVIDIA GPUs (reported split: ~150K H100 + ~50K H200 + ~20-30K GB200 NVL72)
• 300 MW
• Mixed-architecture — Anthropic took the entire output of Colossus 1
• xAI moved Grok training to Colossus 2 (homogeneous GB200 NVL72, separate facility); Colossus 1 was effectively a stranded asset for training (mixed-architecture straggler effect, ~11% effective training utilization per actuia/Tom's Hardware) until Anthropic took it
• This is inference capacity, not training — Anthropic explicitly stated the deal would double Claude Code rate limits across Pro/Max/Team/Enterprise, remove peak-hour throttling on Pro/Max, and substantially raise Claude Opus API rate limits

Why xAI / SpaceX vs hyperscalers:

The honest answer is speed and power, not price. Anthropic's full compute portfolio (per DCD / Anthropic disclosures):

The xAI deal is ~10-12% of the total disclosed compute book by dollar but the only piece available within months, not years. Everything else (Google 5GW, AWS 5GW, MSFT/NVDA Vera Rubin) comes online 2026-2027+. The xAI deal is time-to-capacity arbitrage — Anthropic paid premium $/GPU-hour to fix the ARR-throttling problem in Q2 2026 instead of waiting for AWS/GCP/MSFT to finish construction.

Implied unit economics: at $15B/yr for 300 MW = ~$50M per MW-year. AWS Rainier at $10B/yr-equivalent on 1+ GW (subsidized by equity stake) = ~$10M per MW-year. The SpaceX line is 5x AWS strategic-investor pricing on a like-for-like basis. The gap is the time-to-capacity premium.

The hyperscaler-constraint thesis is real: GOOG/AMZN/MSFT are GPU-and-power-constrained on new commitments at this scale through ~2027. xAI was the only operator with live, idle, 300MW-scale capacity.

Counterparty risk both ways:

For Anthropic:

• xAI is Anthropic's direct competitor (Grok vs Claude). The 90-day termination clause is dual-edged.
• A reported (possibly paraphrased, not confirmed in S-1 language) "reclamation clause" if Anthropic "engages in actions that harm humanity" gives Musk subjective leverage over a political/cultural-positioning rival.
• Tenant isolation operationally: Anthropic retains content/model/data ownership per S-1, but the cluster is in xAI's physical facility with xAI's operations staff. Whether weights live encrypted at rest with Anthropic-managed keys is unspecified.
• Anthropic's safety-evals-publishing-on-Grok is a real and asymmetric risk.

For SpaceX/xAI:

• AMZN and GOOG indirectly fund $15B/yr to a direct competitor. AMZN's $8B equity stake in Anthropic generates revenue for an Elon entity.
• Customer concentration on SpaceX's side: Anthropic is ~80% of xAI segment revenue in the S-1 recast (xAI segment did $3.2B in 2025 vs $15B/yr from Anthropic going forward).
• If Anthropic exits, xAI segment economics collapse.

Stream 2 — Anthropic financial sizing

Anthropic 2026 financial snapshot:

The $1B → $30B trajectory in 16 months (80x growth) is the load-bearing fact. Pre-this investigation, the "$4-6B late 2025" framing was outdated by ~5x.

Revenue-vs-compute ratio:

If 2026 recognized revenue lands $20-30B (since ARR > recognized revenue when growth is 80x):

• $15B/yr SpaceX line = 50-75% of 2026 recognized revenue going to ONE compute vendor
• Total compute stack ($375B+) = ~1.5x Anthropic's implied 2026-2029 cumulative revenue at the current trajectory
• For comparison: OpenAI is reportedly spending >50% of revenue on Azure compute. Anthropic's compute-to-revenue is in OpenAI's neighborhood today, deflating fast.

Is it financeable?

The honest answer: not on operating cash flow today; yes on the capital stack as currently structured.

• Anthropic has raised $72-76B cumulative; pending round adds $30B at $900B+ valuation. Cash on hand post-Series G is plausibly $30-50B.
• The hyperscaler investments (AWS $13B, Google $40B committed, MSFT-NVDA $15B) are structurally circular — cash in from the cloud provider, cash out to the same cloud provider. Per Fortune Apr 2026, roughly half of Google's and Amazon's 2026 artificial-intelligence segment profits came from their Anthropic stakes' mark-to-market, not operating margin.
• The SpaceX line is the un-circular one. $15B/yr to a counterparty that is not also an investor. That's the new dynamic: as Anthropic outgrows hyperscaler capacity, it pays cash-out-the-door rates to neoclouds (CoreWeave, Nebius, IREN) and now to SpaceX/xAI infrastructure.
• The 90-day mutual termination clause is the safety valve.

The single most useful financial datapoint we don't have: the cash-vs-credit split on the $13B Amazon and $40B Google investments — specifically, how much is restricted compute credit vs unrestricted cash. This determines whether Anthropic's ~$50B+ cash position can actually fund $15B/yr of incremental SpaceX cash burn, or whether the SpaceX line forces another $30-50B equity raise within 12 months. If hyperscaler capital is 80%+ credit (likely), the SpaceX deal accelerates Anthropic toward IPO in 2026 rather than 2027.

Stream 3 — Hyperscaler ticker implications

• GOOG (Alphabet): Mildly negative directional, small magnitude. Anthropic-TPU still scaling to 1M chips; the xAI deal is incremental, not substitutive. But the narrative that "Anthropic is a captive TPU customer" is now demonstrably wrong — Anthropic will use whoever has live capacity. Magnitude: 1-3% headwind on GCP narrative, offset by $40B Anthropic equity stake appreciation.

• AMZN: Negative, larger magnitude. AWS has loudly positioned itself as Anthropic's "primary cloud." The S-1 reveals that's no longer true on a marginal-spend basis. AMZN's $13B equity stake helps, but the AWS-Bedrock anchor-tenant narrative weakens. Magnitude: 2-5% AWS-narrative discount. The Fortune disclosure (50% of AMZN artificial-intelligence profits = Anthropic-stake-mark-to-market) is the load-bearing fact.

• MSFT: Neutral-to-positive. The xAI-Anthropic structure mirrors Azure-OpenAI (anchor compute + equity in an artificial-intelligence lab). It validates the model MSFT pioneered. MSFT's own $30B Anthropic deal is now one of four — Anthropic is multi-sourcing exactly as OpenAI claimed it would, derisking the MSFT-OpenAI dependency in reverse. Bonus: 2026-05-21 Sherwood report that Anthropic is in talks to use MSFT custom Maia chips would, if signed, validate Azure-Anthropic as another full-stack pair.

• ORCL: Neutral. Anthropic skipped OCI entirely. The OCI-OpenAI ($300B Stargate) relationship remains the comp. Oracle was never a serious Anthropic counterparty and isn't now.

• NVDA: Positive on demand-persistence narrative, neutral-to-mildly-negative on customer concentration. NVDA gets revenue twice in a sense (sold silicon to xAI, now selling Anthropic-funded compute consumption that drives more silicon demand for xAI Colossus 2 and beyond). But the optics of Anthropic backstopping a third frontier lab's hardware buying further concentrates NVDA's customer base. Net: still bullish on volume, watch the "five customers = 50% of revenue" narrative tighten.

• AVGO: Indirect read. BRCM filing confirms 3.5 GW Google-Anthropic TPU online 2027. Anthropic putting $15B/yr outside the TPU ecosystem partly validates AVGO's role as the second-source ASIC enabler (Anthropic + Google's $200B/5yr is BRCM's anchor). AVGO benefits from artificial-intelligence multi-sourcing structurally.

Stream 4 — Neocloud (NBIS / CRWV / IREN / APLD / CORZ) read

Mixed-to-negative. The bull thesis was: "frontier labs will go multi-source, neoclouds capture the marginal $." The xAI deal proves point 1 (multi-source is real) but disproves point 2 (the marginal $ goes to neoclouds) — Anthropic put $45B with xAI, not CRWV/NBIS/IREN/APLD. The largest single available chunk of frontier-lab compute demand in the market went to a competitor-turned-landlord, not to the neocloud cohort that's spent two years building exactly this pitch.

The narrative repricing:

• xAI is essentially valued at $230B partly because it's now positioned as a neocloud-plus-frontier-lab hybrid.
• CRWV, comparable compute footprint, trades at ~$70-80B.
• xAI's optionality is now repricing the comp set higher — but only if you believe additional frontier-lab anchor-tenant deals come to neoclouds.

The honest read by name:

• CRWV: most exposed (Musk-orbit anchor-tenant question). Q1 2026 revenue $1.9-2B; MSFT still ~67% concentration. Signed multi-year Anthropic deal in Q1 (small/incremental, not anchor).
• NBIS: most insulated (sovereign-EU positioning, Q1 2026 artificial-intelligence cloud revenue $390M +841% YoY). Customer-base diversification narrative intact.
• IREN / CORZ / APLD: pure plays without anchor tenants disclosed. artificial-intelligence segments still mid-single-digit % of total revenue. Capacity classes too small (~hundreds of MW each) to host a frontier lab solo.

The SpaceX/Colossus deal sits in a class of its own: the only frontier-class GW-scale anchor-tenant deal going to a non-hyperscaler counterparty. CoreWeave's MSFT relationship is the closest analog, but MSFT is also a CRWV equity backer — circular again. Anthropic-SpaceX is the first true arm's-length GW-scale frontier-lab compute deal disclosed publicly.

Verdict + reasoning

Bull / bear on the contract surviving 90 days

Bull (probability the full ~$45B runs to May 2029): ~55-65%.

• Anthropic's ARR trajectory ($30B → likely $60B+ by EOY 2026) makes $15B/yr compute spend defensible.
• xAI's segment economics need Anthropic too badly to terminate (Anthropic = ~80% of xAI segment revenue post-deal).
• Both sides have strong commercial reasons to honor the term.

Bear (early termination / renegotiation): ~35-45%. Triggers:

1. Anthropic safety evals on Grok generate political fight → Musk invokes reclamation clause if it exists
2. GOOG or AMZN accelerates capacity delivery → Anthropic exits xAI premium pricing
3. Colossus 2 GB200 capacity comes online and xAI wants Colossus 1 back for next-gen Grok inference
4. SpaceX post-IPO board demands customer diversification away from Anthropic concentration

A clean 4-year run-out is NOT the base case. A 2-3 year run with renegotiation in 2027-2028 is more likely — when AWS Project Rainier 1GW + GOOG-BRCM 3.5GW + MSFT Vera Rubin all come online and the time-to-capacity premium evaporates.

What this changes in the SpaceX IPO perspective

The 2026-05-26 spacex-ipo-liquidity-event perspective flagged Anthropic as "the bull case for xAI segment economics + the 90-day termination clause is the risk." This investigation sharpens both legs:

• Bull leg sharper: $15B/yr from Anthropic = ~80% of xAI segment revenue going forward. xAI segment becomes Anthropic-tenant-cluster + frontier-model-lab. Profitable on Anthropic alone if direct costs <$10B/yr (plausible at premium pricing).
• Bear leg sharper: a 2027-2028 renegotiation is the base case, not a tail event. The xAI segment economics break if Anthropic walks before T+24mo. SpaceX IPO investors need to model partial renegotiation, not full-term lock.

What this changes in adjacent perspectives

• ai-power-bottleneck: validates the "frontier labs are power-bound, not capital-bound" thesis. Anthropic paid 5x AWS strategic pricing because xAI's gas turbines + grid contracts were the only available power in Q2 2026.
• ai-power-delivery: the Memphis Colossus power-density profile (~1 MW per ~750 GPUs at peak) is the operating data point for the cohort.
• optical-supercycle: Colossus 1's mixed-architecture inefficiency is the cleanest argument yet for why next-gen clusters need homogeneous fabric (Colossus 2 GB200 NVL72 = homogeneous, training-capable). Validates the optical inter-sat thesis indirectly via cluster fabric requirements.

Trade implications

This is mostly a perspective-shift, not a single-stock trade. But the implied directional reads:

1. Long NBIS / Short AWS-Anthropic-narrative — the multi-source thesis is now structural fact. NBIS's sovereign-EU positioning + customer diversification is the cleanest neocloud bet. AWS's Anthropic-anchor-tenant narrative weakens.
2. Long MSFT vs short ORCL — MSFT validated the anchor-tenant + equity model that ORCL has tried to replicate at OCI without same legitimacy. Anthropic-MSFT custom Maia talks (Sherwood 2026-05-21) hardens this if signed.
3. Watch GOOG TPU-share narrative — if Anthropic stays on TPU at 1M-chip scale, GOOG narrative survives. If Anthropic shifts to NVDA-on-xAI as primary inference platform, TPU narrative weakens.
4. NVDA customer-concentration watch — Anthropic now drives NVDA revenue via xAI clusters AND via direct NVDA purchases on Azure. Five-customer concentration tightens.
5. AVGO long via Anthropic-TPU stack — BRCM is the silicon enabler for Google's TPU at the Anthropic anchor-tenant scale.

Open follow-ups (filed for separate work)

1. Track the tenant-isolation IP language — through Anthropic enterprise sales briefings + competitive intelligence from Google DeepMind / Meta artificial-intelligence. The single most useful evidence we don't have.
2. Anthropic IPO timing watch — if cash-vs-credit on hyperscaler stakes is 80%+ credit (likely), Anthropic raises again 2026-Q3/Q4 OR IPOs in 2026 not 2027. Watch S-1 filings.
3. Microsoft Maia chip Anthropic deal confirmation — Sherwood 2026-05-21 report says talks; confirm signed/not.
4. Memphis Colossus 1 / Colossus 2 capacity verification — SemiAnalysis is the cleanest source for GW capacity numbers; cross-check before the SpaceX IPO T-7 fire-drill.
5. Anthropic ARR Q2 2026 update — Dario disclosed $30B in April; Q2 2026 print (if any) determines whether the $60B EOY 2026 trajectory holds.
6. xAI segment revenue recast verification — if S-1/A amendments split out Anthropic-specific revenue from rest-of-xAI, that's a material disclosure for SpaceX IPO modelers.
7. Reclamation clause primary-source verification — the "actions that harm humanity" language is paraphrased in industry commentary, not confirmed in S-1 language as seen. Need to find in primary filing.
8. Anthropic-MSFT custom Maia chip deal status — Sherwood 2026-05-21 reporting; watch for confirmation in Anthropic 2026-Q3 disclosures.

Sources

Primary disclosures:

• SEC EDGAR — SpaceX S-1 (CIK 1181412): https://www.sec.gov/Archives/edgar/data/1181412/000162828026036936/spaceexplorationtechnologi.htm
• Anthropic — Series G $30B at $380B post: https://www.anthropic.com/news/anthropic-raises-30-billion-series-g-funding-380-billion-post-money-valuation
• Anthropic — AWS 5GW Trainium expansion: https://www.anthropic.com/news/anthropic-amazon-compute
• Anthropic — Google/Broadcom $200B/5yr: https://www.anthropic.com/news/google-broadcom-partnership-compute
• CoreWeave Q1 2026 10-Q: https://www.sec.gov/Archives/edgar/data/0001769628/000176962826000222/crwv-20260331.htm
• Nebius Q1 2026 6-K: https://www.sec.gov/Archives/edgar/data/0001513845/000110465926064092/nbis-20260331xex99d1.htm

Industry / sell-side reporting:

• TechCrunch — Anthropic pays xAI $1.25B/mo: https://techcrunch.com/2026/05/20/anthropic-will-pay-xai-1-25-billion-per-month-for-compute/
• Axios — Anthropic-SpaceX $15B/yr: https://www.axios.com/2026/05/20/anthropic-spacex-compute
• DCD — SpaceX S-1 reveals $1.25B/mo Colossus: https://www.datacenterdynamics.com/en/news/spacex-ipo-filing-reveals-anthropic-set-to-pay-musks-firm-125bn-a-month-to-rent-xai-data-center-space/
• DCD — Anthropic cloud spend $80B through 2029: https://www.datacenterdynamics.com/en/news/anthropic-cloud-spend-expected-to-reach-80bn-through-2029/
• Tom's Hardware — 220K GPUs, 300 MW, mixed-architecture inference: https://www.tomshardware.com/tech-industry/artificial-intelligence/musks-spacex-has-rented-out-access-to-its-supercomputers-220-000-nvidia-gpus-and-300-megawatts-of-ai-compute-power-to-rival-anthropic-musk-says-no-one-set-off-my-evil-detector-antrhropic-also-interested-in-orbital-data-centers
• Tom's Hardware — Colossus 1 mixed-architecture training inefficient: https://www.tomshardware.com/tech-industry/artificial-intelligence/musks-colossus-1-ai-supercomputers-inefficient-mixed-architecture-design-couldnt-be-used-to-train-grok-so-anthropics-using-it-for-inference-instead-musk-readies-unified-blackwell-only-colossus-2-for-frontier-training-and-potential-ipo
• SemiAnalysis — xAI Colossus 2 First Gigawatt Datacenter: https://newsletter.semianalysis.com/p/xais-colossus-2-first-gigawatt-datacenter
• VentureBeat — Anthropic $30B ARR / 80x growth: https://venturebeat.com/technology/anthropic-says-it-hit-a-30-billion-revenue-run-rate-after-crazy-80x-growth
• Sacra — Anthropic revenue / valuation / funding: https://sacra.com/c/anthropic/
• SaaStr — Anthropic $14B ARR Feb 2026: https://www.saastr.com/anthropic-just-hit-14-billion-in-arr-up-from-1-billion-just-14-months-ago/
• Bloomberg — Anthropic in talks $900B valuation: https://www.bloomberg.com/news/articles/2026-05-12/anthropic-in-talks-to-raise-30-billion-at-900-billion-valuation
• TechCrunch — Anthropic-AWS $100B / $5B AMZN: https://techcrunch.com/2026/04/20/anthropic-takes-5b-from-amazon-and-pledges-100b-in-cloud-spending-in-return/
• TechCrunch — Google $40B Anthropic investment: https://techcrunch.com/2026/04/24/google-to-invest-up-to-40b-in-anthropic-in-cash-and-compute/
• TechCrunch — Anthropic 2028 revenue $70B projection: https://techcrunch.com/2025/11/04/anthropic-expects-b2b-demand-to-boost-revenue-to-70b-in-2028-report/
• TechCrunch — is xAI a neocloud now: https://techcrunch.com/2026/05/06/is-xai-a-neocloud-now/
• Fortune — Half of GOOG/AMZN artificial-intelligence profits from Anthropic stake mark-to-market: https://fortune.com/2026/04/30/google-amazon-ai-profits-anthropic-stake-bubble-earnings-2026/
• Constellation Research — Anthropic-MSFT-NVDA $30B compute pact: https://www.constellationr.com/insights/news/anthropic-microsoft-azure-nvidia-ink-30-billion-compute-pact
• nextomoro — Compute rent / $/GPU implied economics: https://nextomoro.com/compute-rent/
• actuia — Anthropic Colossus 1 / 11% effective utilization for training: https://www.actuia.com/en/news/anthropic-rents-colossus-1-for-125-billionmonth-on-an-xai-park-capped-at-11-capacity/
• MindStudio — Anthropic compute empire timeline: https://www.mindstudio.ai/blog/anthropic-compute-deals-timeline-amazon-google-spacex-microsoft
• VC Corner — SpaceX SPCX IPO S-1 full teardown: https://www.thevccorner.com/p/spacex-spcx-ipo-s1-teardown-valuation-2026
• Dealroom — xAI $6.4B loss on $3.2B revenue 2025: https://app.dealroom.co/news/feed/xai-lost-6-4b-in-2025-despite-3-2b-revenue-as-grok-ai-scales-to-trillions-of-parameters
• Sacra — OpenAI $25B ARR Feb 2026: https://sacra.com/c/openai/
• Forge — Anthropic IPO timing: https://forgeglobal.com/insights/anthropic-upcoming-ipo-news/
• Introl — Colossus 2 GW expansion / 555K GPUs Jan 2026: https://introl.com/blog/xai-colossus-2-gigawatt-expansion-555k-gpus-january-2026

Internal cross-references:

• 2026-05-26-spacex-ipo-deep-dive-separate-trades-and-oxygen — parent SpaceX investigation, included Anthropic deal as open follow-up #4
• log — 2026-05-26 deep-dive update entry
• published/perspective-updates/2026-05-26-spacex-ipo-fast-entry-and-separate-trades.json — published package on parent perspective
• inputs/articles/2026-05-06-sherwood-anthropic-s-scramble-for-compute-now-includes-rival-xai.md
• inputs/articles/2026-05-21-sherwood-report-anthropic-is-in-talks-to-use-microsoft-s-custom-ai-ch.md
• inputs/articles/2025-09-03-semianalysis-amazon-s-ai-resurgence-aws-038-anthropic-8217-s-multi-gigawa.md
• inputs/reddit/2026-05-20-wallstreetbets-anthropic-is-paying-spacex-15-billion-per-year.md
• inputs/reddit/2026-05-21-stocks-according-to-spacex-ipo-papers-anthropic-is-paying-spacex-15.md
• inputs/reddit/2026-05-21-wallstreetbets-anthropic-secures-45bn-spacex-deal-for-claude-ai-computing-p.md
• inputs/reddit/2026-05-20-investing-anthropic-about-to-turn-profitable-in-q2-2026-wsj.md
• inputs/reddit/2026-04-28-valueinvesting-anthropic-valuation-hits-1-trillion-on-pre-ipo-trading-up-73.md