8× B3002.3 TB HBM3e · Blackwell Ultra
Highest memory and FP4/FP8 throughput per node.
$39,192 /moOnly 9 left
Configure
Solutions · Pre-training
Pre-train on dedicated 8-GPU nodes. Pay by the month.
B300, B200 and MI355X servers with up to 2.3 TB of HBM per node, full NVLink or Infinity Fabric inside the server, and a flat monthly price that does not punish long runs.
- Servers from $19,592/mo
- At least 25% below market
- Root access in < 60 s
The problem
What slows teams down today
Hourly bills on long runs
A four-week run on an hourly cloud is billed for every hour, including checkpoints, restarts and evaluation gaps.
Capacity you cannot plan
Flagship GPUs on hourly clouds come with quotas, waitlists or “contact sales”.
Shared or virtualised hardware
Noisy neighbours and virtualisation overhead slow collective communication.
How CoreGPU fits
Dedicated GPUs, a monthly price, no surprises
Every server is racked, tested and imaged before it goes on sale, so you get root access in under a minute and a bill you know in advance.
- Dedicated HGX and UBB servers with the full NVLink 5 or Infinity Fabric mesh
- A fixed monthly price per GPU, at least 25% below the market median
- Allocations you can see: live counts and the next restock date
- Add servers month by month as the run grows
Recommended servers
Configurations our engineers recommend
Live stock and the monthly price of the whole server. Open a GPU page to configure it.
8× B2001.44 TB HBM3e · Blackwell
Mature Blackwell stack with full FP64.
$35,592 /moBack Nov 20
Join waitlist
8× MI355X2.3 TB HBM3E · CDNA 4
2.3 TB of HBM3E and strong FP8 on ROCm.
$19,592 /mo21 available
Configure
FAQ
LLM pre-training: common questions
Another question? Our engineers answer within one business day.
Contact usCan I run multi-node training?
Yes. Rent several 8-GPU servers; for private networking between nodes or allocations above 8 GPUs, contact sales so we can place them together.
Which precision should I target?
FP8 is the default for modern pre-training on Hopper, Blackwell and CDNA 4. B300 adds the most FP4 throughput if you train or serve in FP4.
What happens if a GPU fails mid-run?
Every allocation is burned in before sale. If hardware fails, we replace the server and extend your month for the downtime.
Other solutions
Tell us about your workload
Model size, context length, deadline and budget: we reply with the right GPU and server count within one business day.