AI FACTORY

AI FACTORY

AI Factory Sim

DSX shell, SemiAnalysis lens: rack density in, hidden bottlenecks out.

DSX shell · SemiAnalysis stress lensDSX waypoints ↗XRFab ↗InferenceX ↗
hot aisleR1 · cold aisle · edge
Blackwell Ultra MGXGB300 NVL72135 kW TDP / up to 155 kW peak

Play film: satellite → campus → wafer line (LOAD → SORT → METRO → PACK) → hall → cabinet → rack → open tray. Same plant kernel. Wafer aisle is a logistics sketch, not a fab scan.. Front/rear details follow Lenovo public guides and 3D tour; U positions are compressed for screen.

source-backed
inspectGB300 compute tray

1U hybrid tray: 2 Grace + 4 B300, liquid on silicon, air on front I/O.

source-backed
drivesE1.S bays 0-7source-backedmgmtBMC RJ-45, USB, mini DPsource-backedDPUBlueField-3 B3240source-backednetwork2x OSFP cards, 800Gsource-backedrear left/rightcoolant return / supplysource-backedrear midcartridge + busbar clipsource-backed

live plant

load factor93%derivedIT / rack125.1 kWderivedbusbar2501 A / 1400 Aderivedreturn28.8 Cderivedliquid / air112.5 / 17.5 kWderivedCDU load56% × 4assumption
scalesatellite

Procedural globe. Pin marks the assumption campus.

Stylized globe and aerial. Not live Maxar/Google tiles, not a real tenant site.

assumption
chip introB300 GPU

C01 · GB300 compute tray

Four B300 in one tray.

Lenovo LP2357 / NVIDIA NVL72 ERA: the compute tray is the 4-GPU unit that fills 18 slots to 72 GPUs.

live 3.99 kW · share 63% of tray 6.3 kW

source-backed
cooling splitCPU/GPU/HBM liquid, PDB/OSFP air

Eight dual-rotor fans track OSFP, BlueField, PDB, BMC and drive temperatures, not GPU die temp.

source-backed
LEDspower, ID, NVLink, fault

Operator face is a real status surface: locate, link, fault, drive activity.

source-backed
power pathAC busway -> rack power shelves -> 50V rack busbar

conversion 10.0 kW / rack · residual 0.9 kW

source-backed
Lenovo GB300 NVL72 Product Guide ↗Lenovo GB300 NVL72 3D Tour ↗NVIDIA NVL72 AI Factory Components ↗SemiAnalysis TPU InferenceX Official Preview ↗
power draw5.5 MW
thermal index100/100
flow margin49%
power headroom2%
gap tax125 kW
busbar amps2501A / 1400A
return water28.8C
SemiAnalysis · 7 Sep 2026Ironwood: up to 50.4% more tokens/$ vs B200

InferenceX Official Preview. At 20 tok/s/user, TPUv7 Ironwood 9,364 tok/s/chip vs B200 8,903 / B300 8,925. Qwen3.5 397B FP8, aggregated. External TPU TCO vs hyperscaler GPU TCO. Not a rack kW in this plant.

source-backedSemiAnalysis ↗
SemiAnalysis InferenceX$0.181 / M tokens at 100 tok/s/user

Ironwood $0.181 vs B200 $0.222 (~19% lower) vs B300 $0.276 (~34% lower) at the same interactivity. Preview datapoint. Paid charts / Accelerator Model stay out.

source-backedSemiAnalysis ↗
SemiAnalysis InferenceXGB300 NVL72 disagg still wins the middle

External TPU stack is still aggregated. Semi: ~30% perf/$ edge for GB300 NVL72 disagg vs TPUv7 agg at mid e2e latency. TorchTPU private beta; OSS around mid-October. No TPU trays in simulatePlant.

source-backedSemiAnalysis ↗

scenarios

rack variant
power path

flaw toggles

Waypoint film copies DSX/XRFab camera grammar only. Wafer line is a FOUP logistics sketch — not lithography, not live tiles, not Omniverse.

Plant model. Sourced constants, derived P/V and Q=ṁcpΔT, marked assumptions. Not CFD, not a quote.

Read the rack thesis →
← tools