AService
GPU compute leasing
DaoWorks finds GPU capacity that fits your workload — reserved clusters, dedicated nodes or on-demand — compares offers on like-for-like terms and negotiates the lease. You sign and pay the provider directly; the buyer pays us a commission only when a deal closes.
Updated
At a glance
- You sign with
- The provider
- You pay
- The provider, directly
- Our fee
- Commission from the buyer, on close
- Every deal
- Sanctions and export-control checks
01What we source
Three ways to lease compute.
Reserved clusters
Multi-node GPU clusters with a high-speed interconnect, reserved for a fixed term, typically from a few months to three years. The lowest price per GPU-hour, in exchange for commitment.
Good for
Pre-training, large fine-tunes, steady inference fleets.
Dedicated nodes
Single-tenant bare-metal servers, usually eight GPUs each, on monthly or longer terms. Full control of the software stack and no noisy neighbors.
Good for
Fine-tuning, research, production inference.
On-demand capacity
Hourly capacity with no long commitment, for bursts and tests. It costs more per GPU-hour, and the newest GPUs are often unavailable this way.
Good for
Evaluations, traffic spikes, trying hardware before you commit.
02Hardware
Accelerators we source.
Availability and pricing move weekly. We report what is actually on offer, not what is on a spec sheet.
| Accelerator | Memory and typical use |
|---|---|
| NVIDIA H100 | Hopper, 80 GB. The most widely available training and inference GPU; the default when price and availability matter most. |
| NVIDIA H200 | Hopper, 141 GB. H100-class compute with more and faster memory, for larger models and longer contexts per GPU. |
| NVIDIA B200 · B300 | Blackwell. Current-generation training and high-throughput inference. |
| NVIDIA GB200 · GB300 NVL72 | Blackwell, rack-scale: 72 GPUs in one NVLink domain, for the largest models. |
| AMD Instinct MI300X · MI325X · MI355X | 192–288 GB. High-memory accelerators, where your software runs on ROCm. |
Sources NVIDIA H100 · NVIDIA H200 · NVIDIA GB200 NVL72 · AMD Instinct accelerators
03Your brief
What to send us.
A short list is fine. What you don't know yet, we work out with you.
- GPU model and count, or the workload if you're not sure
- Interconnect: InfiniBand or RoCE, and whether you need a full NVLink domain
- Term and start date
- Location and data-residency requirements
- Storage: type, capacity and throughput
- Networking: egress volume, public IPs, private connectivity
- How you'll run it: bare metal, Kubernetes or Slurm
- Target price per GPU-hour and payment terms
- Who you are and where the cluster's users are — needed for export-control screening
04Comparison
How we compare offers.
Two quotes at the same headline price can differ by a lot once storage, egress and commitment are counted. We put every offer on the same terms first.
| Dimension | What we pin down |
|---|---|
| Price | All-in price per GPU-hour, with storage, networking and support either included or priced separately — never mixed. |
| Commitment | Term, prepayment, deposit, and what happens if you need to end early. |
| Fabric | InfiniBand or RoCE, bandwidth per GPU, rail-optimized topology, NVLink domain size. |
| Storage | What's included, what more costs, and sustained throughput. |
| Network | Egress fees, public IPs, private links to your cloud. |
| Location | Country, data center and facility tier — and whether that works for your data. |
| Reliability | SLA, node-replacement time, and the credits you actually get. |
| Flexibility | Rights to scale up, extend, or assign the contract. |
05Export controls
Where the GPUs are, and who uses them.
Advanced GPUs are export-controlled. For every compute deal we check where the hardware is, who will use it and for what, against US export controls (the Export Administration Regulations, including the Entity List) and sanctions lists. A deal that fails a check doesn't proceed through us.
Sources Export Administration Regulations, 15 CFR 730–774 · BIS Entity List
06Questions
GPU compute FAQ
Does DaoWorks own GPUs?
No. DaoWorks doesn't own, lease out or resell hardware. We find capacity from GPU clouds and data-center operators, and you lease it from them directly.
Who do I sign the lease with?
The provider. We help you negotiate, but the contract, the invoices for compute and the payments are between you and the provider.
Is your commission added to the provider's price?
No. The provider's price is the provider's price. Our commission is a separate agreement between you and DaoWorks, set in writing before we start and payable only when you sign.
Can the cluster be outside the US?
Yes, if that is where you need it and the deal passes export-control and sanctions screening for that location, end user and end use.
Can you find the newest GPUs?
Availability of the newest accelerators changes quickly and is often reserved capacity on longer terms rather than on-demand. We report what is actually on offer, at what term and price, not what has been announced.
→Next step
Send us a compute brief.
GPU type and count, or model and monthly tokens. Term, region, budget. We come back with what is actually available and on what terms.