FundaAI’s 8 GW framing combined with the latest Minihard/Macrohard details indicate 8GW of SpaceXAI is doable by end of 2027-mid 2028. They indicate the 8 Gigawatts of chips and the 8GW of power are available to SpaceXAI

8 GW Demand (compute/chip-order side is locked in) as FundaAI calculates demand-side power need at ~8 GW total from ordered hardware already slotted into tier-1 OEM schedules.

~1.5+ GW of power already with Colossus 2 alone at ~946 MW via S-1s, Epoch AI satellite, and channel checks.
+13k GB300 racks delivering through 2026 (weighted to 2H): ≈2.5 GW.
+15k Rubin racks in 2027 with ≈4 GW+.
Sum ≈ 8 GW.

This maps roughly to a ~4 GW-scale step by around end-2026 (existing + GB300) and another ~4 GW+ step in 2027 (Rubin).

Chips, interconnects, and servers are not viewed as bottlenecks.

Current campus buildout (Colossus 2 / Macrohard + Macrohardrr/Macroharder + Minihard) aligns with a large slice of the near-term GB300 portion

Macrohard / Colossus 2 core is ~550k GPUs (GB200/GB300 mix).
Macrohardrr (also called Macroharder) is 220k GB300 GPUs.
Minihard identical 220k GB300 GPUs, ~25% of the floor area, 800 Gb networking, shorter interconnects, lower latency/heat/cost, fully liquid-cooled.

Combined ~990k latest-gen GPUs needing over 2 Gigawatts of power. Timeline for the campus set to be complete Q1–Q2 2027.

Power per 220k-GB300 modular unit ~400 MW IT load (≈3,056 NVL72 racks × ~130 kW). Efficiency gains (PUE improvement from ~1.3 toward 1.15) can free tens of MW. Campus total with these buildings is tracked in the ~2.0–2.3 GW range once online.

Supermicro referenced in modular/DCBBS liquid-cooled designs that support rapid GW-scale deployment (3–5 months cited for standardized approaches). Memory is the standard high-bandwidth HBM packages for these GPUs (288 GB-class per GPU in GB300 NVL72 configurations).

A Texas GW-scale site is also being explored.

$1.465B turbine order to the Jereh × FTAI Aviation JV for over 1 GW, delivery by November 2027.

Transformer supply is not a bottleneck.

Scenarios (YE2027 nameplate) are Bear 4 GW (~25%), Base 5 GW (~50%), Bull 8 GW installed / ~5.5–6 GW usable (~15–20%). Probability of full 8 GW installed ≈12–17% but would be completed in 2028.

The current Macrohard + Macrohardrr + Minihard trajectory is delivering a large GB300 tranche (approaching ~1M GPUs / ~2–2.3 GW campus) targeted for Q1–Q2 2027, with modular speed (Minihard-style builds in ~4 months).

Supermicro-style liquid-cooled approaches helping execution.

Getting to 1.5-1.6 million B300s would need another 3 Minihard and 1.1 GW. This might be the Texas location.

B300s for 4 Gigawatts

Using the practical 1.82 kW/GPU figure (IT load).

4,000 MW ÷ 1.82 kW ≈ 2.20 million B300 chips. Using FundaAI’s own scaling (13,000 GB300 racks ≈ 2.5 GW, and 13,000 × 72 = 936,000 chips).

4 GW is 1.6× larger → ≈ 1.50 million B300 chips (this uses higher facility-level power draw of ~192 kW per rack, which includes cooling overhead / PUE). Most people use the ~1.5–2.2 million range depending on whether they count pure IT load or full facility power.

Rubin portion (4 GW)Rubin (Vera Rubin)

Rubin has higher power draw. Public projections put TDP in the 1,800–2,300 W (1.8–2.3 kW) range per GPU.

Exact GPUs-per-rack (NVL72 vs denser NVL144-style) is not fully public yet, but FundaAI models 15,000 Rubin racks ≈ 4 GW+.

Assuming similar rack density to GB300 (72 GPUs/rack):
15,000 × 72 = 1.08 million Rubin chips for ~4 GW. If power per GPU is higher (~2.2 kW effective), the count would be lower (around 1.0–1.2 million). If racks become denser, the chip count rises.

This might be around four or 5 Rubin versions of the miniHard data centers.