Colossus (xAI / SpaceXAI)
Colossus (xAI / SpaceXAI)
One-line summary: SpaceX/xAI Memphis + Southaven training campuses — issuer-clocked first clusters in 122 / 91 / 64 days on a reused factory shell; Epoch (Jun 15, 2026 sat) books 340 MW IT / ~230k chips on Colossus 1 and 946 MW IT / 440k chips on Colossus 2. Issuer combined ~1.0 GW "compute power" is a different unit from Epoch's 1,286 MW IT.
What it is
spacex's June 11, 2026 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/local names (Macrohard / Minihard / Macrohardrr) are not in the 424(b)(4) glossary. From 2026-08-27-xai-data-center-build-tempo-vs-peers-and-ai-campus-capacity.
Why it matters to stock-market
Colossus is the concrete campus behind spacex's AI/compute line and elon-web-services-to-spacex-hyperscaler-rerate. The load-bearing grain from this source: the 122 / 91 / 64-day clocks are cluster bring-ups on a reused or rapidly converted shell, not greenfield campus clocks. See colossus-cluster-clocks and energized-vs-announced-campus-mw.
Key facts
Colossus 1 (issuer + NVIDIA + Epoch):
- Issuer: first cluster ~100,000 H100 / ~130 MW in 122 days, "repurposing the shell of an existing factory." From 2026-08-27-xai-data-center-build-tempo-vs-peers-and-ai-campus-capacity (SpaceX 424(b)(4)).
- NVIDIA Oct 28, 2024 PR: 100,000-Hopper cluster built in 122 days, 19 days first-rack-to-training; "in the process of doubling" to 200,000 Hopper GPUs as-of that PR — planned, not a 2026 energized count. From 2026-08-27-xai-data-center-build-tempo-vs-peers-and-ai-campus-capacity.
- Epoch (updated Aug 26, 2026; sat through Jun 15, 2026): road/retrofit of the former Electrolux building Apr 24, 2024; first training-useful load Aug 11, 2024 — 25k H100 / 35 MW IT; 100k H100 / 139 MW on Sep 2, 2024. Later scale-out (Epoch, not issuer-itemized): Feb 17, 2025 150k H100 + 50k H200 / 278 MW; Jul 22, 2025 +30k B200 → 276k H100-eq / 340 MW; still 340 MW IT / ~230k chips (150k H100 + 50k H200 + 30k B200) on the Jun 15, 2026 pass. From 2026-08-27-xai-data-center-build-tempo-vs-peers-and-ai-campus-capacity.
- Site-to-first-useful ≈ 3.5 months (Apr 24 → Aug 11, 2024); site-to-100k cluster ≈ 4.3 months. The 122-day clock is the cluster clock. From 2026-08-27-xai-data-center-build-tempo-vs-peers-and-ai-campus-capacity.
- Tenant as of May 6, 2026: Anthropic leased "all of the compute capacity" at Colossus 1 = more than 300 MW (over 220,000 NVIDIA GPUs) within the month. Epoch calls Anthropic's 220k vs its 230k chip table "roughly consistent." Leave 220k vs 230k open. From 2026-08-27-xai-data-center-build-tempo-vs-peers-and-ai-campus-capacity.
Colossus 2 (issuer + Epoch):
- Epoch land clearing Feb 28, 2025. Issuer first cluster ~110,000 GB200 / ~210 MW in 91 days; second cluster 110,000 GB300 / 220 MW in 64 days. From 2026-08-27-xai-data-center-build-tempo-vs-peers-and-ai-campus-capacity (424(b)(4)).
- Epoch maps the first-cluster S-1 line to Oct 19, 2025 (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. Prefer Epoch + S-1; leave the date open. From 2026-08-27-xai-data-center-build-tempo-vs-peers-and-ai-campus-capacity.
- Epoch maps the second cluster to Apr 6, 2026 (556k H100-eq / 490 MW IT). Jun 15, 2026 sat: 946 MW IT / 440k chips (110k B200 + 330k B300); Building 3 roof halfway. From 2026-08-27-xai-data-center-build-tempo-vs-peers-and-ai-campus-capacity.
- Site-clearing to first cluster ≈ 7.7 months (Feb 28 → Oct 19, 2025). The 91-day clock is again a cluster clock. From 2026-08-27-xai-data-center-build-tempo-vs-peers-and-ai-campus-capacity.
Combined power (do not collapse units):
- Issuer: Colossus + Colossus II approximately 1.0 GW of "compute power" (as-of the prospectus; many operating stats dated March 31, 2026), "with additional power capacity available for data center operations." From 2026-08-27-xai-data-center-build-tempo-vs-peers-and-ai-campus-capacity.
- Epoch: 340 + 946 = 1,286 MW IT. Different definition from issuer "compute power." Introl 2 GW / 555k GPUs (Jan 2026) is a contested vendor claim (search-only; not fetched). Do not write 2 GW as fact. From 2026-08-27-xai-data-center-build-tempo-vs-peers-and-ai-campus-capacity.
Announced, not treated as energized:
- Issuer next phase at COLOSSUS II: ≥220,000 additional GB300 and >400 additional MW of compute power "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 — Epoch projection, not an issuer date. From 2026-08-27-xai-data-center-build-tempo-vs-peers-and-ai-campus-capacity.
- Wikipedia (secondary): Southaven ~1.2 GW simple-cycle permitted March 2026; temporary turbines off by mid-2027. Not independently fetched from MDEQ. From 2026-08-27-xai-data-center-build-tempo-vs-peers-and-ai-campus-capacity.
Not retrieved: x.ai/colossus was Cloudflare-blocked. Grokipedia XAI_(company) infobox "200,000 NVIDIA H100s" and Introl 555k / 2 GW are stale/contested — not used as 2026 energized fact. From 2026-08-27-xai-data-center-build-tempo-vs-peers-and-ai-campus-capacity.
Open questions
- Colossus 2 first-online: Epoch Oct 19, 2025 vs Wikipedia ~January 2026.
- Which building hosts Google's 110k GPU CSA (facility not named in the Jun 5, 2026 FWP)?
- Official TVA order text; MDEQ permit PDF; court PDFs on the Clean Air Act suit — not fetched this pass.