Source Report

Anthropic Compute: Committed Capacity Approaches Patel's Current Estimate, But Deployed Power Lags

Anthropic has secured over 2 gigawatts (GW) of committed compute capacity through multi-year deals with Google Cloud (1 million TPUv7 chips, over 1 GW online in 2026) and AWS Trainium, enabling rapid scaling via owned and leased infrastructure; this data moat—direct TPU purchases for $10 billion alongside rentals—allows deployment in custom Fluidstack data centers (Texas/New York, $50 billion total investment), bypassing cloud bottlenecks that slow rivals. Public reports confirm ~2+ GW committed as of early 2026, aligning with Patel's "2-2.5 GW currently," though actual deployed power is lower (e.g., Fluidstack sites ramping through 2026).[1][2][3]
- Anthropic's October 2025 Google deal: 400k TPUs bought outright (~$10B), 600k rented (~$42B RPO), totaling >1 GW in 2026.[1]
- AWS partnership adds multi-GW Trainium/Inferentia capacity; $50B U.S. infrastructure (Nov 2025) targets gigawatts-scale sites online 2026.[3]
- The Information (Feb 2026): Discussions for 10 GW total capacity over years, via rentals and owned space.[4]

For competitors entering AI infrastructure, Anthropic's hybrid model (buy + rent across clouds) creates a replication barrier: $50B+ CapEx commitments require hyperscaler equity (e.g., Amazon's $4B stake) for favorable terms, while pure cloud reliance exposes to capacity shortages.

OpenAI Compute: Public Figures Confirm ~2 GW Deployed, Slightly Exceeding Anthropic; Multi-GW Pipeline Matches Scaling Claims

OpenAI ended 2025 with 1.9 GW deployed capacity (9.5x growth from 200 MW in 2023), powering $20B+ ARR via Stargate (1.2 GW Abilene site operational) and Oracle partnerships; the mechanism—sharing economic risk on overruns via joint ventures—de-risks $500B/10 GW ambition, with 4.5 GW Oracle deal + Nvidia/AMD/Broadcom (each 6-10 GW) ensuring redundancy as single-site limits (e.g., power) force multi-campus training.[5][6][7]
- OpenAI CFO (Jan 2026): Capacity tripled to 1.9 GW in 2025 (200 MW '23 → 600 MW '24 → 1.9 GW '25).[6]
- Stargate: 1.2 GW Abilene (partial live), 7 GW across five new sites (Sep 2025), Oracle 4.5 GW expansion on track despite Abilene scaleback.[8]
- Bloomberg/Reuters: Partnerships target 10+ GW (Nvidia $100B, Broadcom 10 GW custom chips H2 2026).[9]

New entrants face OpenAI's first-mover lock-in: $400B+ committed across partners creates pricing power and site priority, forcing overbuilds or multi-cloud hedging at higher costs.

Revenue Claims: $20B ARR Confirmed Publicly, Monthly Additions Are SemiAnalysis Proprietary

Patel's $20B ARR for Anthropic matches Bloomberg/The Information (nearing $20B Mar 2026, from $9B end-2025), driven by Claude Code's enterprise ramp; however, "$4-6B ARR per month added" is unconfirmed outside SemiAnalysis/Dylan Patel statements (e.g., Jan $4B, Feb $6B adds), with public figures showing $4B pace (Jul 2025) scaling to $14-19B via projections, not explicit monthly deltas—indicating proprietary supply-chain reverse-engineering.[10][11]
- Bloomberg (Mar 2026): $19B+ run-rate, doubled from $9B end-2025.[10]
- The Information: $4B annual pace (Jul 2025); targets $20-26B 2026 (Reuters Oct 2025).[12]
- OpenAI: $20B+ ARR end-2025 (CFO), $25B recent—similar trajectory, no monthly claims.[6]

AI startups chasing ARR must prioritize compute-locked revenue (e.g., coding agents) over consumer subs, as Patel's unverified deltas highlight: public reporting lags proprietary tokenomics models tying usage to GW deployment.

SemiAnalysis Track Record: Accurate on Power Crises and Supply Chains, Revenue/Compute Often Directionally Right but Granularly Proprietary

SemiAnalysis (Patel) accurately forecasted U.S. AI power from 3 GW (2023) to 28 GW (2026) in Mar 2024 report—now validated by 1.9 GW OpenAI + 2 GW Anthropic ramps—via permitting/ERCOT analysis; predictions like xAI Colossus (1 GW), Anthropic Trainium (multi-GW), and TPU ramps preceded announcements, though revenue monthly adds ($4-6B) remain SemiAnalysis-exclusive without contradictions.[13]
- Power: ERCOT/Texas GW requests validated; onsite gas (xAI 500 MW turbines) as predicted.[13]
- Compute: Anthropic 1M TPUs, OpenAI multi-vendor 10 GW called pre-public; no major misses found.[1]
- No public retractions/inaccuracies; e.g., AWS-Anthropic resurgence (Sep 2025) spot-on.[14]

Rivals analyzing via public sources undervalue SemiAnalysis's edge—supply-chain modeling predicts ramps months ahead—but over-rely on Patel risks echo-chamber bias, as granular revenue lacks third-party verification.

Public Confirmation vs. Proprietary: Compute Deployed Figures Align, Revenue Scaling Matches but Lacks Monthly Granularity

The Information/Bloomberg/Reuters confirm ~2 GW deployed (OpenAI 1.9 GW, Anthropic 2+ GW committed) and $20B ARR trajectories, validating Patel's direction; unconfirmed: exact current GW (deployed vs. committed), $4-6B monthly adds—likely SemiAnalysis tokenomics-derived from cloud RPOs/deals, presented as fact without outlets citing independently.[15]
- Confirmed: Deals (e.g., Anthropic $50B infra, OpenAI Stargate 7 GW sites).[16]
- Proprietary: Monthly revenue deltas, precise scaling to 5-6 GW year-end (forward-looking).[11]

Entrants must blend public (deal announcements) with proprietary signals (e.g., SemiAnalysis permitting models) for edge; over-trusting unverified monthlies risks misallocated CapEx amid 3T USD data center debt boom.[17]

Confidence: High on compute (multiple outlets align); medium on revenue granularity (proprietary, but totals match). Additional cloud earnings transcripts would refine monthly claims.

---

Recent Findings Supplement (March 2026)

Anthropic Compute: Committed Capacity Exceeds 2GW, But Operational Scale Remains Under 2GW Amid Rapid Buildout

Anthropic's publicly committed compute via AWS Trainium and Google TPUs totals over 2 gigawatts (e.g., AWS Project Rainier scaling to 2.3GW full campus potential, Google Cloud's 1M+ TPUv7 chips equating to 1GW+ online in 2026), enabling multi-gigawatt training runs that leverage non-Nvidia chips for cost advantages—but current operational deployed capacity is not explicitly broken out as 2-2.5GW in outlets like The Information or Bloomberg; instead, reports emphasize future scaling to 5-10GW targets by hiring ex-Google execs and $50B+ infrastructure pledges, with no direct confirmation of Patel's exact "2-2.5GW current" figure as of early 2026.[1][2][3]
- AWS opened $11B Project Rainier (Oct 2025) with 500K-1M Trainium2 chips (half a million dedicated to Anthropic), full campus eyeing 2.2-2.3GW.[4][2]
- Google deal (Oct 2025): Up to 1M TPUv7 (~1GW+), $52B value, sites via Fluidstack in TX/NY online 2026; total committed >2GW per analyst estimates.[3]
- Recent: Internal plans for 10GW (Feb 2026, The Information), $50B U.S. infra (Nov 2025), Hut 8/Fluidstack up to 2.3GW (Dec 2025).[5][6]
Implication for competitors/new entrants: Patel's 2-2.5GW "current" aligns closely with committed/near-term operational (~2GW+ via partners), but scaling to 5-6GW YE2026 lacks public confirmation beyond ambitions—entering requires hyperscaler partnerships, as pure startup builds face grid/power delays (e.g., Anthropic pledging to cover consumer electricity hikes, Feb 2026).[7]

OpenAI Compute: Operational Capacity Hit ~2GW End-2025, Matching "Slightly Higher" Claim; Massive Future Pipeline (10-30GW+)

OpenAI's deployed compute tripled to 1.9GW in 2025 (per CFO Sarah Friar, Jan 2026), directly validating Patel's "slightly higher" than Anthropic benchmark, powered by diversified partners—but 2026 operational growth hinges on Stargate delays (e.g., Abilene expansion scrapped Mar 2026) amid 10GW+ deals (Nvidia, AMD, Broadcom); Reuters/Bloomberg confirm no single-source 2-2.5GW snapshot, but aggregate end-2025 deployment fits.[8][9]
- End-2025: 1.9GW operational (SiliconANGLE/Reuters), up from 0.6GW 2024; Microsoft Fairwater clusters ~GW-scale for OpenAI.[8][10]
- Pipeline: Nvidia 10GW ($100B, H2 2026 start), AMD 6GW (1GW MI450 H2 2026), Broadcom 10GW custom, Oracle 4.5GW, AWS 2GW Trainium; Stargate targets 10GW ($500B) but delays (e.g., no Abilene 2GW expansion).[11][12][13]
Implication for competitors/new entrants: OpenAI's ~2GW operational lead (end-2025) over Anthropic's ramping 2GW+ is real, but execution risks (Stargate stalls) create openings; new players need $100B+ funding/partners, as grid-locked builds (e.g., 250GW Altman vision infeasible short-term) favor incumbents.[14]

Anthropic Revenue: Explosive Growth to ~$20B ARR Confirmed, Aligning with Patel's Directional Claims But No "$4-6B/Month Add"

Anthropic's ARR trajectory—$9B end-2025, $14B Feb 12 2026, $19B end-Feb, nearing $20B Mar 2026 (Bloomberg/Reuters/The Information)—validates Patel/SemiAnalysis' early calls on Claude Code driving $4-6B monthly adds (e.g., SemiAnalysis Feb 2026: $6B Feb alone), but public sources report cumulative GAAP revenue ~$5B (2023-Dec 2025) vs. run-rate extrapolations; no outlet confirms exact "$4-6B/month" as fact, treating as estimates amid $30B raise at $380B valuation.[15][16][17]
- Growth: $1B Dec2024 → $4B mid-2025 → $9B end-2025 → $14B Feb2026 → $19-20B Mar2026; Claude Code $2.5B+ ARR (doubled YTD).[18][19]
- Funding/Validation: $30B Series G (Feb 2026, $380B val), CFO court filing: $5B cumulative GAAP revenue; run-rates per insiders/Bloomberg.[15][16]
Implication for competitors/new entrants: Patel's revenue foresight (e.g., SemiAnalysis outgrowing OpenAI) proven accurate directionally, but exact phrasing proprietary—enterprise focus (80% revenue) creates moat; entrants must hit 10x YoY or partner (e.g., Anthropic's AWS/Google revenue shares $80B thru 2029).[20]

SemiAnalysis/Patel Claims: Proprietary Estimates Proven Accurate on Trends, But Specifics (e.g., Exact GW/ARR Adds) Unconfirmed Publicly

Patel's SemiAnalysis reports (e.g., Feb 2026: Anthropic $6B ARR Feb add, outgrowing OpenAI; datacenter models tracking ramps) have high directional accuracy (e.g., pre-announced Anthropic TPU 1GW, AWS Trainium multi-GW), but exact Patel claims like "2-2.5GW current" or "$4-6B/month" appear as proprietary estimates—not verbatim in Reuters/Bloomberg/The Information—framed as "estimates" with public data aligning post-hoc (e.g., OpenAI 1.9GW end-2025, Anthropic ~2GW committed).[21][22]
- Hits: Predicted Anthropic ARR surges (now $20B), OpenAI delays (Stargate stalls), compute ramps (TPU/Trainium pre-announce).[20]
- No misses noted post-Sep2025; self-reported "remarkably accurate" historicals (e.g., power demand).[10]
Implication for competitors/new entrants: SemiAnalysis excels at supply-chain foresight (e.g., TPU external sales), blending public/proprietary—reliable for trends, but specifics require subscription; new entrants use for benchmarking, but public outlets lag (e.g., no GW operational breakdowns), risking underestimation of ramps.

OpenAI Revenue: Confirmed $20B+ ARR End-2025, Now $25B (Mar 2026), Validates Scale But Trails Anthropic's Acceleration

OpenAI hit $20B ARR 2025 (CFO Jan 2026, Reuters), now $25B (The Information Mar 2026, up 17% MoM), tying to 1.9GW capacity growth—publicly confirmed via blogs/court filings, no Patel tie but aligns with SemiAnalysis tokenomics models; enterprise push narrows Anthropic gap.[23][24]
- $20B end-2025 (from $6B 2024), $25B Feb2026; $110B raise (Feb 2026, $840B val).[25]
Implication for competitors/new entrants: Confirmed hyperscale revenue funds GW builds, but Anthropic's faster ARR adds (per SemiAnalysis) pressure—new players target niches like inference to avoid training CapEx arms race (~$280B burn to 2030 est.).[26]

Confidence & Gaps: High on trends/public figures (e.g., 2GW-scale operational via partners); medium on Patel's exact phrasing (proprietary, directionally accurate per SemiAnalysis track record). No post-9/13/2025 regulatory changes; additional SemiAnalysis access strengthens proprietary validation. All $ in USD.[21]