SpaceX is making minihard, an improved modular 220K GB300 data center. It will be small building to the right of Macrohard and Macrohardrr. It is the most important thing Elon and SpaceXAI has said about its infrastructure this year. Minihard is about 25% of the floor area of Macrohardrr but the same 220K GB300s.

Minihard is designed from day one to be efficient for both internal training, inference and leasing.

When Macrohardrr and minihard are both done, then the first 550K GPU colossus 2 would be redundant and could be fully leased to others for an additional $30-35 billion per year. SpaceX could mass produce the 220K GB300 and then 220K Rubin chip data centers. They could rent out some and keep others for internal use and their own business.

220,000 GB300s ÷ 72 per NVL72 rack ≈ 3,056 racks
× ~130 kW per rack ≈ 400 MW

810,000 ft² → ~494 W/ft²
25% of that → ~1,975 W/ft²

Cooling efficiency could improve power usage by possibly 15%, saving up to 60 MW. There are also other energy savings.

Why squeeze at all? Floor space is cheap in Mississippi.

The answer is cable length, communication speed and the cooling efficiency.

At 800 gigabit, networking distance is the single most expensive variable in the building.

Passive copper at 800G reaches roughly 1–2 meters. Active copper maybe 3–5.

Past that you need an optical transceiver. Roughly $1,500–3,000 and 15–20 watts per port. Two ports per link. A 220k-GPU non-blocking three-tier fabric runs on the order of 600k links — over a million port ends.

Move even a fifth of those from optics to copper and you’ve saved several hundred million dollars and several megawatts of pure transceiver heat.

They have C mapped the 220K B300s to get 10X more training and inference speed. They are standardizing on that size for those benefits. Colossus 2 area has the original 550K GPU building and two 220K GPU buildings. Almost 1 Million GPUs. Should all be done Q1 or Q2 of 2027.

Tail latency in synchronous collectives compounds. Improvements make the AI faster and more efficient. Pipeline parallelism (which xAI has emphasized) is especially sensitive to the slowest hop. Shrinking the physical diameter tightens the latency distribution and reduces pipeline bubbles across millions of steps.

Cooling moves fully to direct-to-chip liquid, enabling higher coolant temperatures, more free-cooling hours, and PUE (power utilization efficiency) improvements that are material at 400+ MW. Moving PUE from ~1.3 to ~1.15 on a 400 MW IT load would save ~60 MW .

Replicability is the strategic point. An exact topological match means the specialized training stack that maps to 220k GB300s with 800G NICs runs unchanged on both buildings. The two sites become interchangeable capacity units rather than one-off snowflakes.

Colossus 1 failed as a training cluster because it mixed three chips H100/H200/GB200. Reportedly ~11% utilization. xAI gave up and moved training to Colossus 2.

Implications for the AI leasing business

SpaceX/xAI has already turned into one of the largest AI capacity lessors on the planet. Anthropic is paying ~$1.25 B/month for Colossus 1 capacity and Google ~$920 M/month for ~110k GPUs. Multi-billion-dollar recurring revenue with multi-year contracts.

Higher density + identical topology improves the leasing economics in several ways. Lower cost per delivered FLOP/token. Copper savings, lower PUE, and reduced networking complexity drop the fully-loaded cost of capacity. They can either price more aggressively to win more deals or keep prices high and expand margins.

Standardized 220k GPU data center packages at double or more revenue charged to Google. $22 billion per year. Customers or internal teams can be offered plug-and-play high-performance clusters instead of bespoke builds. This shortens sales cycles, simplifies operations, and supports rapid capacity expansion.

Higher utilization and better inventory management. When training moves to a new denser site (as happened when Colossus 1 utilization was low and capacity was leased out), the older or parallel site becomes clean, high-value lease inventory without software rewrites.

Premium positioning. A cluster that is both denser and topologically identical to the internal Grok training fabric is more attractive than generic GPU hours. Lessees get performance closer to what the model developer itself uses.

Impact for Grok and Cursor (LLM businesses)

Grok (SpaceXAI) and Cursor already operate in a tight partnership. Grok 4.5 was jointly trained using Cursor’s real-world coding interaction data on xAI compute.

Density and topological identity amplify this. Faster and cheaper model iteration for Grok. Identical high-performance clusters mean the optimized stack can be spun up on whichever building has free capacity. Training runs finish sooner, experiment throughput rises, and the cost of the next Grok generation drops. Pipeline-parallelism gains from lower tail latency are especially useful for the large models they are chasing.

Better inference economics. Higher density + efficient networking reduces the cost of serving tokens. That improves margins on Grok API usage and makes high-volume coding workloads (Cursor’s core use case) more sustainable.
Cursor-specific upside. Cursor supplies unique developer data. xAI supplies the compute.

More efficient, interchangeable clusters mean more joint training capacity can be allocated without starving either party’s production needs. It also strengthens the strategic logic of deeper integration (the market has already floated large acquisition valuations for Cursor precisely because of this compute-for-data synergy).

Competitive moat. Rivals that must retune software every time they move to a new building or density regime lose time and performance. A family of topologically identical 220k-unit clusters lets Grok/Cursor treat compute as a fungible, high-performance resource rather than a constant re-engineering project.

In short, the denser “Minihard” configuration is not just an engineering optimization—it turns the Memphis campus into a more modular, lower-cost, higher-performance factory for both leased capacity and internal frontier models