Reserved by the month, power and bandwidth included. B300, H200, RTX 5090 and RTX 4090 in stock across North America, Southeast Asia and East Asia.
Most teams start on the API and pay per token. Reserve capacity when your load is steady enough to keep a card busy — the figure below is where that line usually falls.
| GPU | Memory | Notable | Regions | From /GPU·hr | Status |
|---|---|---|---|---|---|
| B300 | 288 GB HBM3e | ~8 TB/s · ~15 PFLOPS dense FP4 · up to 1,400 W | NA · SEA · EA | $6.50 | In stock |
| H200 | 141 GB HBM3e | 4.8 TB/s · proven for 70B–400B serving | NA · SEA · EA | $2.60 | In stock |
| RTX 5090 | 32 GB GDDR7 | Best cost per token on small and mid models | SEA · EA | $0.75 | In stock |
| RTX 4090 | 24 GB GDDR6X | Fleet workhorse, large pool available | SEA · EA | $0.45 | In stock |
Rates include power and bandwidth. Minimum term one month; 3, 6 and 12-month commitments carry a discount. Interconnect, local NVMe and private networking are quoted per configuration.
What fits without splitting a model across cards.
HBM3e on B300 and H200; GDDR7 and GDDR6X on the RTX cards. Bars are to scale.
Which way round is cheaper depends on how busy the card would be.
A rule of thumb, not a quote. Published crossover estimates range from 30% to about two-thirds; the right number for you depends on your traffic shape, model size and term. We will size it with you before you commit.
Whole machines or whole racks, yours for the term. Root access, your images, your orchestration. For training and long-running inference that needs predictable performance.
Guaranteed throughput and concurrency on our serving stack, with our optimization and scheduler doing the work. Monthly fee plus metered overage.
A named open-weight model deployed in your environment or in a region you pick, with inference confined to it. Project fee plus monthly operations.
| North America | US East · US Central |
| Southeast Asia | Singapore |
| East Asia | Seoul · Tokyo |
Inference can be contracted to stay in one region. Other locations evaluated on request.
Failure handling is three-part: detect → remediate → verify. A path with detection but no remediation is not considered fixed, and each one is held by an automated gate that injects the fault and asserts the remediation fires.
Releases roll back. One directory per version, so the previous directory is the rollback point. Read-only code is separated from runtime state; every node pulls the same build SHA.
Every model has a profile: weight provenance, version, quantization, context length and measured throughput, recorded before it serves traffic.
Requirements → quote → entity and export-control screening → contract and first payment → delivery with monitoring and an on-call contact.