Issuer-clocked Colossus 122/91/64-day cluster bring-ups vs ~2-year greenfield peer clocks; energized vs announced MW/GPU for xAI, Microsoft, Meta, CoreWeave, Anthropic offtake.

xAI data-center build tempo vs peers, and AI campus capacity (2026)

Generated by Grok Bot research on 2026-08-27. WebSearch + fetch ladder + native X. Treat as raw material — review before promoting.

Vault priors (not re-asserted unless re-fetched): vault/projects/stock-market already lives the AI-power / Colossus / CEG stack (nuclear-baseload-for-ai-data-centers, constellation-energy, spacex, ai-capex-to-power-and-materials-cascade). Older All-In color of 220K GPUs / 300 MW is now a re-fetched Anthropic description of Colossus 1 (May 2026), not a current independent count. Overlaps threads/artificial-intelligence; this clip does not write wiki.

Summary

SpaceX's June 11, 2026 424(b)(4) is the best issuer primary: Colossus 1 first cluster ~100,000 H100 / ~130 MW in 122 days (repurposed Electrolux shell); Colossus II first cluster ~110,000 GB200 / ~210 MW in 91 days; second cluster 110,000 GB300 / 220 MW in 64 days; combined Colossus + Colossus II ~1.0 GW of "compute power" (issuer term; as-of the prospectus, with many operating stats dated March 31, 2026). SpaceX's own industry comparison is ~2 years for a 100 MW greenfield. NVIDIA's Oct 28, 2024 PR independently clocks the first 100,000-Hopper cluster at 122 days, and 19 days from first rack on the floor to training.

Epoch AI (updated Aug 26, 2026; satellite through Jun 15, 2026) is the best independent energized tracker: Colossus 1 340 MW IT / ~230k chips (150k H100 + 50k H200 + 30k B200); first training-useful load 25k H100 / 35 MW on Aug 11, 2024, 100k H100 / 139 MW on Sep 2, 2024. Colossus 2 946 MW IT / 440k chips (110k B200 + 330k B300) as of Jun 15, 2026, with first cluster dated Oct 19, 2025 against the S-1 110k-GB200 / 210 MW line — not Wikipedia's "around January 2026."

Peer clocks that were actually retrieved: Microsoft Fairwater Wisconsin was announced May 2024 and fully operational Jun 23, 2026 (~25 months), with equipment online in April; Microsoft did not publish a MW number. Meta Hyperion (Richland Parish) broke ground December 2024 and on July 13, 2026 announced an expansion to 5 GW IT — announced, not energized. CoreWeave reported 1.5 GW active power at Jun 30, 2026 (up ~500 MW in Q2) and ~3.7 GW contracted, with a ≥8 GW by 2030 target. EIA STEO (Aug 11, 2026) confirms data-center load is the demand driver and cut Texas 2027 load growth to 6% from 14% after the Aug 3 data-center pause; it does not publish a campus build-time series in the fetched page.

Contested: Introl's January 2026 "555,000 GPUs / 2 GW" (search-only; not fetched) and the folklore "100-day" clock (issuer + NVIDIA now say 122). Do not merge issuer "compute power," Epoch "IT power," and site-wide generation.

Findings

xAI build and scale-out clock

Naming / sites (issuer): SpaceX 424(b)(4) defines COLOSSUS as the flagship at Paul R. Lowry Road, Memphis, and COLOSSUS II as the Memphis + Southaven, Mississippi pair in one "coherent gigawatt-scale AI training cluster." xAI was acquired by SpaceX effective February 2, 2026 (same filing). Wikipedia and local reporting also use Macrohard / Minihard / Macrohardrr for nearby buildings; those names are not in the 424(b)(4) glossary.

Colossus 1 — first useful training:

Construction start (independent): Epoch dates a road/retrofit of the former Electrolux building to Apr 24, 2024.
First training-useful MW: Epoch, Aug 11, 2024 — 25k H100 / 35 MW IT.
First 100k cluster: Epoch Sep 2, 2024 — 100k H100 / 139 MW; NVIDIA Oct 28, 2024 says the 100,000-Hopper cluster was built in 122 days, 19 days first-rack-to-training, and was "in the process of doubling" to 200,000 Hopper GPUs (as-of that PR — planned, not a 2026 energized count).
Issuer restatement: 424(b)(4) — first cluster ~100,000 H100 / ~130 MW / 122 days, "repurposing the shell of an existing factory."
Later scale-out (Epoch, not issuer-itemized): Feb 17, 2025 150k H100 + 50k H200 / 278 MW; Jul 22, 2025 +30k B200 (GB200 config) → 276k H100-eq / 340 MW; still 340 MW on the Jun 15, 2026 satellite pass.
Tenant as of May 2026: Anthropic, May 6, 2026 — "all of the compute capacity" at Colossus 1 = more than 300 MW (over 220,000 NVIDIA GPUs) within the month. Epoch notes Anthropic's 220k vs its 230k chip table as "roughly consistent."

Colossus 2 — first useful and scale-out:

Land/site: Epoch Feb 28, 2025 land clearing; Wikipedia (secondary) says an affiliate bought 5420 Tulane Road in February 2025.
First cluster (issuer): 424(b)(4) ~110,000 GB200 / ~210 MW / 91 days.
First-cluster date (independent vs encyclopedia): Epoch maps that S-1 line to the Oct 19, 2025 satellite state (batteries complete, turbines connected, cooling enough for ~200 MW). Wikipedia says the first GB200 cluster "came online around January 2026 after a reported build of about 91 days" — contradicts Epoch. Treat Epoch + S-1 as the load-bearing pair; Wikipedia as contested timing.
Second cluster (issuer): 110,000 GB300 / 220 MW in 64 days. Epoch maps this to Apr 6, 2026 (556k H100-eq / 490 MW IT).
Next phase (issuer, announced): "once fully operational, the next phase of expansion at COLOSSUS II will bring online at least 220,000 additional GB300 and over 400 additional megawatts of compute power." Epoch treats that phase as already in the Jun 15, 2026 state: 1,112k H100-eq / 946 MW IT / 440k chips (110k B200 + 330k B300), with Building 3 roof halfway.
Combined issuer claim: Colossus + Colossus II approximately 1.0 GW of compute power, "with additional power capacity available for data center operations."

Months from site/announce to first useful MW (this pass, not invented):

Colossus 1: Epoch Apr 24, 2024 start → Aug 11, 2024 first useful ≈ 3.5 months; → Sep 2, 2024 100k cluster ≈ 4.3 months. Issuer 122-day clock is a cluster clock on a reused shell, not a greenfield campus clock.
Colossus 2: Epoch Feb 28, 2025 clearing → Oct 19, 2025 first cluster ≈ 7.7 months. Issuer 91-day clock is again a cluster clock.

xAI official page: x.ai/colossus Cloudflare-blocked this pass. Grokipedia XAI_(company) preview succeeded on retry and is stale/contested on current counts (infobox still "200,000 NVIDIA H100s"; timeline cites Introl 555k / 2 GW). Use as encyclopedia color, not as 2026 energized fact.

Peer build clocks (site to first useful MW, then scale-out)

Issuer's own peer benchmark: 424(b)(4) — "an industry benchmark to bring online a 100 megawatt greenfield data center is approximately two years." That is SpaceX's comparison, not EIA.

EIA: Fetched STEO, released Aug 11, 2026 says U.S. generation is rising "to meet growing demand from data centers" and, after the Texas governor's Aug 3 pause, EIA cut Texas 2027 load growth to 6% vs 14% in the prior STEO. The fetched page does not give a site-to-first-MW construction table. Do not import Reuters' 4,195 / 4,268 / 4,391 BkWh figures — those were search-only.

Microsoft Fairwater (official, MW unpublished):

Jun 23, 2026 Source: first Mount Pleasant, WI facility announced May 2024; equipment online / startup April; fully operational Jun 23, 2026; "ahead of schedule"; ~10,000 construction workers "over the past two years"; second adjacent building scheduled 2028. No MW, no GPU count.
Fairwater Atlanta feature: Atlanta "began operation in October" (year not restated on the fetched page; industry recaps say 2025 — not treated as fetched). Architecture: GB200 NVL72, "scale to hundreds of thousands of NVIDIA Blackwell GPUs" across the Fairwater network, two-story, almost-zero-water liquid cooling, dedicated AI WAN. No campus MW.

Clock vs xAI: WI Fairwater is a ~25-month announced-to-operational official clock on a purpose-built hyperscale campus — closer to SpaceX's "two-year 100 MW greenfield" benchmark than to Colossus's 122-day reused-shell cluster clock. Not an apples-to-apples MW comparison because Microsoft published no MW.

Meta Hyperion / Richland Parish (announced, not energized):

Meta Data Centers, Jul 13, 2026: expanding the AI-optimized campus to 5 GW in compute capacity; "in advance of our Richland Parish Data Center coming online"; groundbreaking referenced as already done; >$50B regional investment. No first-MW date, no energized MW.
Louisiana Economic Development, Jul 13, 2026: 5 GW of IT capacity, nearly 10 million sq ft, originally announced December 2024. Still announced build-out.

Clock vs xAI: Hyperion is a multi-year, GW-scale, grid/utility-tied campus (Entergy gas plants, batteries, nuclear uprates). From Dec 2024 announce to Jul 2026 it is ~19 months in and not yet described as online. That is the conventional hyperscale clock Colossus skipped by reusing a factory + behind-the-meter turbines.

CoreWeave (fleet MW, not a single campus clock): Q2 2026 IR, curl — +~500 MW in Q2 to 1.5 GW active; ~3.7 GW contracted (as-of Jun 30, 2026). Fool transcript of the Aug 11, 2026 call (issuer remarks): >300 MW in June alone; first self-build "expected to come online later this year"; Indonesia 360 MW "coming online in approximately 18 months"; YE 2026 active-power guide >1.85 GW.

Not retrieved as campus clocks this pass: OpenAI/Oracle Stargate Abilene (SemiAnalysis "15 months" is search-only); Amazon Rainier / Trainium campuses; Google TPU campuses; Oracle official MW. If a campus has no citable official number below, it is left blank — not invented.

Current AI-campus capacity (energized vs claimed)

Who	What is energized (retrieved)	What is only claimed / offtake	As-of
SpaceXAI Colossus 1	Epoch 340 MW IT, 230k chips (150/50/30 H100/H200/B200). Anthropic: >300 MW / >220k GPUs leased.	Vault All-In 220k/300 MW is the Anthropic line, not a new count. Grokipedia 200k H100 is stale.	Epoch Aug 26 / sat Jun 15, 2026; Anthropic May 6, 2026
SpaceXAI Colossus 2	Epoch 946 MW IT, 440k chips (110k B200 + 330k B300). Issuer first two clusters = 110k GB200 + 110k GB300.	Combined issuer ~1.0 GW "compute power" vs Epoch 340+946 = 1,286 MW IT — different definitions. Introl 2 GW / 555k (Jan 2026) is a contested vendor claim, not fetched. Wikipedia "on the order of 1 GW" as of mid-2026.	Issuer prospectus Jun 11, 2026; Epoch Jun 15 sat
Microsoft Fairwater	WI-1 operational Jun 23, 2026; Atlanta "began operation in October." Network described as "hundreds of thousands" of Blackwell GPUs.	No official MW. Third-party 340 / 700 MW Atlanta figures were search-only — not used.	MSFT Jun 23, 2026 + Fairwater feature
Meta Hyperion	Not described as energized. "In advance of … coming online."	5 GW IT announced Jul 13, 2026. CNBC "2 GW by 2030 / 5 GW ~2032" is search-only — not cited as fact.	Meta + LED Jul 13, 2026
CoreWeave (fleet)	1.5 GW active power (Jun 30, 2026).	3.7 GW contracted (6/30); call-date 4.2 GW contracted (Aug 11).	IR + Fool transcript
Anthropic offtake (not campuses it owns)	Colossus 1 lease as above.	Amazon up to 5 GW, "nearly 1 GW … by end of 2026"; Google+Broadcom 5 GW "begin coming online in 2027"; Microsoft/NVIDIA $30B Azure; Fluidstack $50B U.S. infra.	Anthropic May 6, 2026
Google as tenant	—	SpaceX FWP, Jun 5, 2026 CSA: ~110,000 NVIDIA GPUs, $920M/month Oct 2026–Jun 2029; delivery gate Sep 30, 2026. Facility not named.	SEC FWP
Amazon / Oracle / OpenAI campuses	No official campus MW or GPU counts retrieved this pass.	Stargate / Ohio 8 GW (Bloomberg) left as search-only.	Provenance miss

Planned / announced future capacity

Leave these as announcement, not energized:

Colossus II next phase (issuer): ≥220,000 additional GB300 and >400 additional MW "once fully operational." Epoch already books most of that into the Jun 15, 2026 946 MW line and projects Q1 2027: 1,531 MW IT / 722k chips / $58.0B.
Colossus 2 Building 3: Epoch roof halfway Jun 15, 2026; Q1 2027 completion is Epoch's projection, not an issuer date.
Southaven permanent gas: Wikipedia (secondary, cites Wired/Gizmodo) — ~1.2 GW simple-cycle permitted March 2026; temporary turbines to come off by mid-2027. Not independently fetched from MDEQ this pass.
Microsoft WI-2: official 2028 completion. Fairwater network: more U.S. sites "under construction," no MW.
Meta Hyperion: 5 GW IT announced; first-energize date not on the Meta page.
CoreWeave: YE 2026 active >1.85 GW (guide); ≥8 GW by 2030 (CEO, call); Indonesia 360 MW ~18 months.
Anthropic: Amazon nearly 1 GW by end-2026 / up to 5 GW; Google 5 GW from 2027.
Google←SpaceX lease: 110k GPUs by Sep 30, 2026 or terminate/pro-rata.
Orbital: SpaceX 424(b)(4) "as early as 2028"; Anthropic "expressed interest" in multi-GW orbital with SpaceX. Not terrestrial MW.

Contradictions and open questions

122 days vs "100 days." Current issuer + NVIDIA say 122. The 100-day line is older marketing / recap lore. Do not write 100 days as the 2026 fact.
122/91/64-day clocks vs first-useful dates. Those clocks are cluster bring-up on an existing or rapidly converted shell, not "site control → first training-useful MW" on a greenfield. Epoch's first-useful Colossus 1 date is Aug 11, 2024 (25k / 35 MW), weeks before the 100k/Sep 2 line.
Colossus 2 first-online date: Epoch Oct 19, 2025 vs Wikipedia ~January 2026. Open. Prefer Epoch + S-1.
1.0 GW vs 1.3 GW vs 2 GW. Issuer ~1.0 GW "compute power" (GPU-server sense, per Epoch's reading of S-1 p.76). Epoch 340 + 946 = 1,286 MW IT. Introl 2 GW / 555k GPUs (Jan 2026) is a contested vendor claim (search-only; Grokipedia repeats it). Do not collapse these.
220k vs 230k on Colossus 1. Anthropic >220,000; Epoch 230k chips / 276k H100-eq. Same order; do not force a merge.
Elon vs local vs TVA. Native X this pass did not return dated Elon capacity posts (Dec 2025–Jan 2026 window empty; recent 20 posts are unrelated). SpaceXAI Memphis, Aug 21, 2026 (affiliate, primary-adjacent) claims "world's largest grid-connected battery pack" for Memphis/Southaven — no MW. S.E. Robinson Jr., Aug 21 (discourse) says TVA board on Aug 20 approved >100 MW firm to MZX Tech for "MAC datacenters." Not an official TVA page. Wikipedia/local recaps: temporary turbines, NAACP/SELC Clean Air Act suit, DOJ intervention — not fetched as court PDFs this pass (Clarion Ledger was search-only).
Energized vs leased vs announced. Colossus 1 is energized and leased to Anthropic. Colossus 2 is energized and still xAI/Cursor/Anthropic per Epoch. Hyperion 5 GW and Anthropic's Amazon/Google 5 GW blocks are announced. Microsoft Fairwater WI-1 is energized without a public MW.
Open: Google's 110k GPUs — which building? Oracle/OpenAI Stargate energized MW? Amazon campus MW? Microsoft Fairwater MW? Official TVA order text? MDEQ permit PDF? x.ai/colossus body (blocked)?

Provenance

Method: Grok Bot / WebSearch / fetch ladder / native X Generated: 2026-08-27 Rounds: 3 (R1 clocks + capacity → R2 contradictions → R3 announced vs energized; no further questions after R3) URLs fetched: 16 successful unique URLs + 2 Grokipedia API + 1 blocked official (x.ai/colossus) + 1 CF-then-curl (CoreWeave IR) + 1 SEC-tool-block-then-curl (Google FWP) Fetch notes: x.ai/colossus Cloudflare 403 — Provenance miss, no invented body. CoreWeave IR WebFetch hit CF challenge; same URL recovered via box curl. SEC Google FWP WebFetch hit "undeclared automated tool"; recovered via curl + browser-ish UA. Grokipedia page-preview?slug=XAI_(company) found: false once, retry succeeded. @xai get_users_posts unauthorized. Elon get_users_posts Dec 28, 2025–Jan 15, 2026 returned 0. search_posts_all skipped. EIA Aug 2026 STEO PDF not pulled (HTML used). No *.gov court PDFs. Recap blogs (Introl, MeasureDAI, SiliconReport, Bloomberg Ohio) used only to discover URLs — not treated as fetched pages. Web sources:

https://www.sec.gov/Archives/edgar/data/1181412/000162828026042639/spaceexplorationtechnologi.htm (SpaceX 424(b)(4), Jun 11, 2026)
https://www.sec.gov/Archives/edgar/data/1181412/000162828026041150/spacexagreementfwp.htm (Google CSA FWP)
https://epoch.ai/data/ai-data-centers/directory/colossus-1
https://epoch.ai/data/ai-data-centers/directory/colossus-2
https://investor.nvidia.com/news/press-release-details/2024/NVIDIA-Ethernet-Networking-Accelerates-Worlds-Largest-AI-Supercomputer-Built-by-xAI/default.aspx
https://www.anthropic.com/news/higher-limits-spacex
https://news.microsoft.com/source/2026/06/23/microsoft-completes-construction-on-first-datacenter-facility-in-mount-pleasant-wisconsin/
https://news.microsoft.com/source/features/ai/from-wisconsin-to-atlanta-microsoft-connects-datacenters-to-build-its-first-ai-superfactory/
https://datacenters.atmeta.com/2026/07/deepening-our-investment-in-richland-parish-louisiana/
https://www.opportunitylouisiana.gov/news/meta-commits-more-than-50-billion-for-north-louisiana-project-becoming-one-of-the-largest-data-centers-in-history
https://investors.coreweave.com/news/news-details/2026/CoreWeave-Reports-Strong-Second-Quarter-2026-Results/default.aspx
https://www.fool.com/earnings/call-transcripts/2026/08/18/coreweave-crwv-q2-2026-earnings-call-transcript/
https://www.eia.gov/outlooks/STEO/
https://en.wikipedia.org/wiki/Colossus_(data_center)
https://grokipedia.com/api/page-preview?slug=XAI_(company) (stale on current MW/GPU) X sources:
https://x.com/SpaceXAIMemphis/status/2090883043368018298 — affiliate/official-adjacent; battery claim; no MW (primary-adjacent)
https://x.com/SERobinsonJr/status/2090934263017357511 — discourse; TVA >100 MW firm / MZX Tech / Aug 20 board
https://x.com/cb_doge/status/2090884359796687294 — discourse reprint of battery claim
Elon recent posts: no capacity numbers this window. Historical Colossus posts not returned by dated get_users_posts. Grokipedia: XAI_(company) preview after one 502-equivalent miss (found: false). Useful for 122-day / Electrolux / June–Sep 2024 narrative; do not use 200k-H100 infobox or Introl 555k/2 GW as current. No dedicated live Colossus slug confirmed.

Rounds: 3 / URLs fetched: 16 unique successful (+2 Grokipedia API) / X posts hydrated: 3 citable (1 affiliate-primary, 2 discourse)