GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 —
GROK-46H 865 —
GEM-37FH 865 —
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 —
GPT-6A 820 —
KIMI-K3X 810 —
CL-FAB5H 787 —
CL-OP5H 764 —
CL-OP46H 742 —
CL-OP47H 733 —
GEM-38FH 676 —
CL-OP47 583 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 —
GROK-46H 865 —
GEM-37FH 865 —
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 —
GPT-6A 820 —
KIMI-K3X 810 —
CL-FAB5H 787 —
CL-OP5H 764 —
CL-OP46H 742 —
CL-OP47H 733 —
GEM-38FH 676 —
CL-OP47 583 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

SpaceXAI Brings Colossus 2 Online at 946 MW With $35.8B in Memphis Compute

SpaceXAI has brought Colossus 2 online in Memphis at roughly 946 MW of IT power capacity, making it one of the largest single AI compute facilities now operating. The site contains an estimated 1,112,000 H100-equivalent compute units built around Nvidia B200 and B300 systems. The capital investment is approximately $35.8 billion.

The important detail is not just the size. It is the power strategy. Colossus 2 is using temporary natural gas turbines alongside limited grid power to bring capacity online before permanent generation is ready. Roughly 69 temporary units are expected to be removed after a dedicated 1.2 GW gas plant comes online in mid-2027.

Speed Over Interconnection

Most data center operators are now negotiating with utilities on timelines measured in years. SpaceXAI is taking the opposite route: install generation first, absorb the permitting fight, and monetise the compute while competitors wait for interconnection.

That makes Colossus 2 a test case for the next phase of AI infrastructure. The constraint is no longer just Nvidia allocation. It is whether a lab can secure power, cooling, land, and regulatory tolerance fast enough to keep model training and inference capacity expanding.

Colossus 1 already put Memphis at the centre of the AI compute map. Colossus 2 changes the commercial model. The broader cluster had crossed roughly 1 GW by mid-2026. Early 2027 projections put Colossus 2 alone at 1,531 MW and 1,824,000 H100-equivalent compute units.

Merchant Compute, Not Just Grok Capacity

SpaceXAI is no longer only building infrastructure for its own models. Anthropic, Cursor, and Google are among the notable lessees of Colossus capacity. That turns the facility into a merchant compute business: a private utility, GPU landlord, and frontier lab supply chain in one structure.

The economics are straightforward. A $35.8 billion campus needs large contracted customers. Model training for Grok can justify part of the spend, but recurring third-party capacity commitments make the buildout financeable at a different scale.

The Regulatory Risk Is Part of the Model

Temporary turbines are not a footnote. They are the mechanism that lets the facility come online before a conventional power project would clear. They also create the obvious risk: local permitting, emissions scrutiny, and grid coordination do not disappear because the compute demand is urgent.

The mid-2027 permanent plant is the release valve. If it arrives on schedule, SpaceXAI converts an improvised power bridge into a more stable energy base. If it slips, the company remains dependent on the temporary fleet that made the launch possible.

Key Numbers

MetricValue
Operational dateSeptember 3, 2026
IT power capacity~946 MW
Capital investment~$35.8B
Compute units~1,112,000 H100-equivalent
Hardware generationNvidia B200 and B300
Temporary turbines to remove~69
Permanent plant target1.2 GW by mid-2027
Early 2027 projection1,531 MW, 1,824,000 H100-equivalent

Colossus 2 is the cleanest expression yet of the new infrastructure rule: the winning AI labs are increasingly the ones that can build or control power plants, not just train better models.