SpaceX building improved modular AI data centers ('Minihard') with ~220K GB300 GPUs each at 25% footprint of larger facilities while maintaining equivalent GPU density
# Minihard: SpaceX's Modular 220K GPU Data Center Strategy
SpaceX is building Minihard, an improved modular data center housing 220,000 GB300 GPUs, positioned adjacent to its existing large-scale facilities. The announcement represents one of the most significant infrastructure disclosures from Elon Musk and SpaceXAI this year. Minihard is designed to occupy approximately 25% of the floor area of the larger adjacent facility while housing the same number of GB300s.
The facility is engineered from inception for efficiency across internal training, inference, and third-party leasing. Once both Macrohardrr and Minihard are operational, the original 550K GPU Colossus 2 building could transition entirely to commercial leasing, potentially generating $30–35 billion annually in additional revenue. The strategic vision extends beyond a single data center: SpaceX intends to mass-produce standardized 220K GB300 modules and later 220K Rubin-chip variants, retaining some for internal use while leasing others.
## Technical Specifications and Power Profile
The 220,000 GB300s, distributed at approximately 72 units per NVL72 rack, yield roughly 3,056 racks. At approximately 130 kilowatts per rack, this configuration consumes approximately 400 megawatts total. Housed in 810,000 square feet, the baseline power density reaches approximately 494 watts per square foot; the optimized dense configuration achieves approximately 1,975 watts per square foot. Cooling efficiency improvements could reduce consumption by up to 60 megawatts, with additional savings from complementary optimizations.
## The Architecture of Density: Why Minihard's Footprint Matters
Floor space in Mississippi is inexpensive; the driver for density lies elsewhere. Cable length, communication speed, and cooling efficiency are the bottlenecks. At 800 gigabits per second, networking distance becomes the single most expensive variable in facility design.
Passive copper connectivity at 800G spans approximately 1–2 meters; active copper extends to 3–5 meters. Beyond these distances, optical transceivers are required—costing roughly $1,500–3,000 per unit and consuming 15–20 watts per port, with two ports per link. A non-blocking three-tier fabric for 220,000 GPUs demands approximately 600,000 links, translating to over one million port endpoints. Shifting even one-fifth of these links from optical to copper delivers hundreds of millions of dollars in savings plus megawatts of eliminated transceiver heat.
## Standardized Topology and Performance Gains
SpaceX and xAI have engineered a specific mapping of the 220K GB300s that delivers approximately 10 times improved training and inference speed. This performance premium justifies standardization on this size. The Colossus 2 complex will ultimately comprise the original 550K GPU building plus two 220K GPU buildings—approaching one million GPUs total—with completion expected in Q1 or Q2 2027.
Tail latency in synchronous collective operations compounds across distributed systems. Shrinking the physical diameter of a cluster tightens the latency distribution and reduces pipeline stalls across millions of training steps. This benefit is especially pronounced for pipeline parallelism, the training strategy xAI emphasizes. Identical topologies across multiple facilities mean that the specialized training stack optimized for 220K GB300s with 800G interconnects runs unchanged on each building, rendering them interchangeable capacity units rather than unique snowflakes requiring constant re-optimization.
## Cooling and Power Efficiency
Direct-to-chip liquid cooling enables higher coolant temperatures and more hours of free-cooling operation. The resulting improvements in power usage effectiveness (PUE)—moving from approximately 1.3 to 1.15 on a 400-megawatt IT load—translate to roughly 60 megawatts of saved energy at scale.
## Learning from Colossus 1
SpaceX and xAI's first large-scale cluster, Colossus 1, failed as an internal training platform due to mixing three incompatible GPU types (H100, H200, and GB200), resulting in approximately 11% utilization. The company abandoned internal training on Colossus 1 and shifted to Colossus 2. This history underscores the strategic necessity of standardization.
## The Leasing Business: Scale and Recurring Revenue
SpaceX and xAI have become among the planet's largest AI capacity lessors. Anthropic pays approximately $1.25 billion per month for Colossus 1 capacity; Google pays approximately $920 million per month for approximately 110,000 GPUs. These represent multi-billion-dollar recurring revenue streams supported by multi-year contracts.
Higher density combined with identical topology strengthens leasing economics in multiple dimensions. The fully-loaded cost per delivered FLOP and token declines through copper savings, lower PUE, and simplified networking. This cost advantage can be deployed either aggressively to win additional deals or maintained at existing pricing to expand margins. Standardized 220K GPU modules positioned at premium pricing relative to generic GPU hours could command revenue nearly double Google's current spend—approximately $22 billion annually.
Operationally, standardized clusters enable customers and internal teams to deploy plug-and-play high-performance infrastructure rather than custom builds, shortening sales cycles and supporting rapid expansion. When training workloads migrate to denser facilities (as occurred when Colossus 1 utilization proved insufficient), older sites transform into clean, high-value lease inventory without requiring software rewrites or optimizations.
Topologically identical clusters also confer premium positioning. Lessees receive performance characteristics closer to those used by frontier model developers themselves, not generic GPU hours.
## Impact on Grok and Cursor
Grok and Cursor operate in close partnership. Grok 4.5 was jointly trained using real-world coding interaction data from Cursor on xAI's compute infrastructure.
Denser, topologically identical facilities amplify the value of this partnership. Faster model iteration becomes possible: the optimized training stack can be deployed on whichever Minihard facility has available capacity, accelerating training completion and experiment throughput while reducing the cost of subsequent Grok generations. Pipeline-parallelism improvements from lower tail latency particularly benefit large-scale models.
Better inference economics follow from higher density and optimized networking, reducing the per-token serving cost and improving margins on Grok API usage and Cursor's high-volume coding workloads. Cursor's unique value—supplying developer interaction data—combined with xAI's compute advantage enables more joint training capacity allocation without starving production services. This compute-for-data synergy strengthens the case for deeper integration, addressing market speculation about acquisition valuations.
Standardized, interchangeable clusters create a competitive moat. Rivals forced to retune software each time they operate on different facilities or densities incur performance and schedule penalties. SpaceX and xAI can treat compute as a fungible, high-performance resource rather than a constant engineering challenge.
## Strategic Synthesis
Minihard is not merely an engineering optimization. It transforms the Memphis campus into a modular, lower-cost, higher-performance factory capable of efficiently serving both third-party leasing and internal frontier model development—with standardized building blocks that can be replicated and scaled.