Live deals Have held stock, project excess or a cancelled order? Sell it
As little as 8 weeksOfficial partners

NVIDIASupermicroDell TechnologiesLenovoGIGABYTELogos identify the makers. Blade AI is an independent supplier.

NVIDIA HGX · 8-GPU servers

HGX B200 and B300 in as little as 8 weeks

8-GPU servers from Supermicro, Dell, Lenovo, GIGABYTE and others, through official partners: 8 to 10 weeks for a new build from compliance approval, against around 36 weeks quoted in the market for a new B300 build.

  • 8–10 wksBlade AI: B200 or B300 new build, from compliance approval
  • ~36 wksThe market, for a new B300 build
  • 2.1TBGPU memory in an HGX B300 (288GB per GPU)
  • 1.4TBGPU memory in an HGX B200 (180GB per GPU)
Supermicro SYS-822GS-NB3RT 8U NVIDIA HGX B300 8-GPU server (Intel)

The lead-time gap

How fast HGX ships through our partners

  1. NVIDIA HGX B200 / B300 8-GPU servers, new build
    Blade AI: as little as 8 wks
    The market: ~36 wks (new B300)
    ~27 weeks sooner

We sometimes have cancelled-order systems that can ship sooner: see live deals or ask. Why Blade AI is faster →

NVIDIA's figures, 8-GPU HGX

NVIDIA HGX B200 and HGX B300 compared
NVIDIA figuresHGX B200HGX B300
GPU memory1.4TB · 180GB per GPU2.1TB · 288GB per GPU
FP4 Tensor Core, dense72 PFLOPS108 PFLOPS
FP8 / FP6 Tensor Core72 PFLOPS72 PFLOPS
FP16 / BF16 Tensor Core36 PFLOPS36 PFLOPS
FP64 / FP64 Tensor Core296 TFLOPS10 TFLOPS
NVLink per GPU1.8TB/s1.8TB/s
Scale-out networkingConnectX-7 · 400GConnectX-8 · 800G

Figures from NVIDIA's HGX platform page, read 5 November 2026: Tensor Core figures other than FP4 are NVIDIA's sparse specification; FP4 is dense. Memory totals as NVIDIA states them; per-GPU figures are physical HBM capacity.

Choose B200 if

  • FP8 or BF16 training and fine-tuning
  • Your models fit in 180GB per GPU
  • You need FP64 for HPC or simulation

Choose B300 if

  • Very large models: 288GB per GPU
  • FP4 inference at scale
  • 800G ConnectX-8 from day one

Order the network with the servers. ConnectX cards, switches and optics have their own lead times. NVIDIA networking from UK stock

HGX platforms we source

Official partnersYou buy directly from the partner; the supplier pays our fee
Export checkedEnd-user screening on every order
Proof before paymentSerial numbers and photos before the balance is due

B200 or B300?

Tell us the model size and the date you need it

We come back with the platform, the partner with the shortest lead time and an indicative quote, within one working day. Right now we have projects running with customers in Europe, Asia and the USA.

Lead times

Lead times as quoted to us in November 2026; server lead times run from end-user compliance approval and are confirmed on each quotation. For GPU systems the authorised seller of record named on the quotation is the seller. Brand names are trademarks of their owners, used for identification only.

Email Get B200 or B300