The best SF Compute alternatives for an NVIDIA H100 cluster are Nodus Reserved Compute (Beta) for a quoted reservation from one GPU to a connected cluster, Together AI GPU Clusters for the lowest published short H100 reservation we found ($3.69 per GPU-hour), and Lambda 1-Click Clusters for 16 to 2,000+ GPUs on InfiniBand. Nodus also lists the H100 SXM at $2.60 per GPU-hour for on-demand jobs, which beats every cluster rate here when your work fits on one 8-GPU machine. SF Compute stays the better pick if you want to set your own price on an open orderbook and sell back hours you end up not using.
We build Nodus, so weigh this accordingly. Every competitor fact links to its own page, and every cost is written out so you can redo it with today's rates.
How do SF Compute and its alternatives compare at a glance?
SF Compute sells compute time, not machines. Its docs describe pools that hold a schedule of node-hours, and orders that add or remove time from that schedule. One node is 8 GPUs, and you set a maximum rate in dollars per node-hour when you buy, per its orders docs. The alternatives below sell fixed terms at published or quoted rates.
| SF Compute | Nodus | Together AI GPU Clusters | Lambda 1-Click Clusters | CoreWeave | |
|---|---|---|---|---|---|
| How H100 price is set | Orderbook: your max rate per node-hour | On-demand list price; reservations quoted | Published per GPU-hour by term | Published per GPU-hour by cluster size | Published on-demand per instance-hour |
| H100 price | Shown after sign-in | $2.60/GPU-hr on demand; reserved by quote | $3.99 on demand, $3.69 for 3-30 days, $3.45 for 31-90 days | $5.85/GPU-hr at 64 GPUs | $49.24/hr for an 8 x H100 instance |
| Smallest unit | 1 node (8 GPUs) | 1 GPU | Not stated on pricing page | 16 GPUs | 8 GPUs |
| Term | Any window you order, e.g. 1 node for 1 hour | Jobs per run; reservations up to 365 days in the console | Pay as you go or reserved | 2 weeks to 1 year | On demand; commitments through sales |
| Sell back unused time | Yes, on the orderbook | No | Not on the pricing page | Not on the pricing page | Not on the pricing page |
| Failed hardware | Refunds failed nodes and can repack you onto healthy ones | Restarts the gang from the latest checkpoint | See provider docs | See provider docs | See provider docs |
Prices checked November 19, 2026: Nodus pricing, Nodus Reserved Compute docs, Together AI pricing, Lambda pricing, CoreWeave pricing, SF Compute specs and SF Compute orders docs. CoreWeave's figure is its North America on-demand price for the 8-GPU HGX H100 instance.
Why is there no SF Compute price in this comparison?
SF Compute does not publish a fixed H100 rate. Its pricing page asks you to sign in, and its orderbook API requires a Bearer token. The price you pay is whatever matches your order: buy orders fill if the book has supply at your max rate, and by default are cancelled if it does not, per the orders docs.
So the useful question is not "what does SF Compute cost" but "what rate does SF Compute have to beat". Multiply any per-GPU-hour price by 8 to get the node-hour rate SF Compute quotes in.
| Alternative | Per GPU-hour | Same price per 8-GPU node-hour |
|---|---|---|
| Nodus on-demand (one machine) | $2.60 | $20.80 |
| Together AI reserved, 3-30 days | $3.69 | $29.52 |
| Lambda 1-Click Cluster, 64 GPUs | $5.85 | $46.80 |
| CoreWeave on-demand 8 x H100 | $6.155 | $49.24 |

Before you place a bid on SF Compute, compare it with these four node-hour rates.
The decision rule we use: if SF Compute's best ask for your window is below the node-hour rate of the alternative you would otherwise sign, take it. If it is above, the published rate wins and you skip the bidding.
What does 64 x H100 for two weeks cost on each alternative?
We priced one common cluster purchase: 64 NVIDIA H100 GPUs (8 nodes) for 14 days, which is 8 x 336 = 2,688 node-hours, or 21,504 GPU-hours. We picked two weeks because it is the shortest term Lambda's 1-Click Clusters accept. Storage, networking extras and taxes are left out everywhere.
| Option | Rate | 64 x H100, 14 days |
|---|---|---|
| Together AI GPU Clusters, reserved 3-30 days | $3.69/GPU-hr | $79,349.76 |
| Lambda 1-Click Cluster, 64 GPUs | $5.85/GPU-hr | $125,798.40 |
| CoreWeave on-demand, 8 instances | $49.24/instance-hr | $132,357.12 |
| SF Compute | Your filled orderbook rate | 2,688 x your node-hour rate |
| Nodus Reserved Compute (Beta) | Quoted | Confirmed in the Nodus proposal |
Same sources and check date as above. The arithmetic: 21,504 x $3.69, 21,504 x $5.85 and 2,688 x $49.24.

The same two-week H100 cluster spans $79,350 to $132,357 at published rates, so the term you buy matters as much as the provider.
Nodus Reserved Compute does not publish a fixed price either. Its docs say each request is reviewed and answered with a proposal covering availability, price, access and delivery. Before you ask, the public rate index averages providers' published reservation rates for the exact term you pick, and shows nothing until at least two providers match. That gives you a market reference to hold any quote against, including one from SF Compute.
Does your job actually need a multi-node cluster?
Often it does not, and that is where the biggest saving sits. A fine-tune or eval sweep that fits in 640 GB of H100 memory runs on one 8-GPU machine, so a cluster term pays for interconnect and hours you will not use.
Here is one 8 x H100 machine for 24 hours, 192 GPU-hours, at on-demand list prices:
| Platform | Rate | 24 hours |
|---|---|---|
| Nodus | $2.60/GPU-hr | $499.20 |
| Together AI on-demand | $3.99/GPU-hr | $766.08 |
| Lambda 8 x H100 instance | $4.59/GPU-hr | $881.28 |
| CoreWeave on-demand | $49.24/instance-hr | $1,181.76 |
| SF Compute | Your filled rate | 24 x your node-hour rate |
Prices checked November 19, 2026, from the pages linked under the first table. The Nodus figure is the H100 SXM rate listed on the Nodus pricing page.

For work that fits on one machine, Nodus on-demand costs $499.20 a day, and SF Compute needs a fill below $20.80 per node-hour to match it.
On Nodus that run is a Job with a --max-cost spending limit, as the commands in the multi-node guide show. Nodus multi-node training is in Beta: gangs have an API ceiling of 8 nodes, and the actual size is limited by which provider pairs Nodus has qualified. For a 64-GPU InfiniBand run today, a reservation is the honest answer, and our reserved vs on-demand guide shows how to find the utilization where it pays off.
What happens to the hours you do not use?
This is the one factor where SF Compute is different in kind, not just in price. Its reselling guide lets you place a sell order for any part of the time you own, including the end of a reservation or a gap in the middle. You receive credits only when a buyer fills it, minus a platform fee with a fixed per-node-hour part and a percentage part. Buy orders carry no fees, per the orders docs.
Take the two-week cluster above and suppose training converges on day 10. The last 4 days are 8 nodes x 96 hours = 768 node-hours, or 6,144 GPU-hours. At Together AI's $3.69 that is $22,671.36 of reserved time, and at Lambda's $5.85 it is $35,942.40.
Unless your contract says otherwise, a fixed term bills those days. On SF Compute you can offer them on the orderbook, and what you recover depends on the bid and the fee.
Two caveats from SF Compute's own docs. Reselling is in public preview, and its roadmap says preview features are best-effort and not covered by the uptime SLA. And sell orders are not guaranteed to fill.
What happens when a node fails mid-run?
SF Compute says plainly that nodes will go down. Its specs FAQ says it refunds failed nodes and can repack you onto nodes with healthy hardware. For your own state, its docs point you to object storage for saving training checkpoints between runs.
On Nodus the restart is part of the job. The multi-node guide says that if a member's machine is lost or reclaimed, Nodus stops the gang's current epoch, keeps the surviving machines, gets new ones only for the lost ranks, and restarts every rank with the latest checkpoint restored. recovery.maxAttempts counts those restarts and defaults to 3. The Nodus vs CoreWeave comparison walks through the same question for Slurm and Kubernetes users.
Where is SF Compute the better choice?
SF Compute is the better choice when you want price discovery instead of a rate card. You can bid below the current ask with --allow-standing and wait for a fill, which no published price list offers.
It is also the better fit when your plans change often. Selling back the end of a reservation, or a holiday week in the middle, is a documented feature with CLI commands. Callable buy orders, in public preview, let you reserve past your compute limit in exchange for letting others buy that time from you at up to 1.2 times your max rate, per the orders docs.
SF Compute buys you very short terms too. An order can be one node for one hour, and the specs page lists no ingress or egress fees, bare metal and managed Slurm, and InfiniBand on those clusters. Its multi-node guide runs torchrun across nodes over an isolated InfiniBand partition.
If you need 64 H100s on InfiniBand next Tuesday for 36 hours, SF Compute's model fits that request, subject to what the orderbook holds. Lambda's 1-Click Clusters start at two weeks, and our Nodus vs Lambda comparison covers when Lambda's longer terms make sense.
How do you move an SF Compute training job to Nodus?
Keep your training script. SF Compute's multi-node guide launches torchrun on every node by hand over a subnet and InfiniBand partition. On Nodus, the gang is assembled for you and torchrun reads its rendezvous from environment variables Nodus sets, so it needs no flags.
Install the CLI and sign in, using the commands from the install docs:
pip install nodus-compute
nodus login
Then run the two-node example from the multi-node guide. Start with the guide's smoke-test train.py, then swap in your own script:
nodus run --name torchrun-2-node --gpu H100 --nodes 2 --launcher torchrun --max-cost 2.00 -- torchrun train.py
With the guide's script, rank 0 prints world_size=2 when both nodes joined. For a reserved cluster instead, submit your GPU model, quantity, duration, location and start timing at Request reserved capacity. Submitting does not reserve inventory or authorize payment; Nodus confirms dates, configuration and terms first.
FAQ
Can I connect my SF Compute account to Nodus?
No. SF Compute is not one of the providers listed in the Nodus cloud accounts docs. You can use both side by side, but Nodus does not manage SF Compute capacity for you.
Does SF Compute charge to buy capacity?
Buy orders have no fees; you pay the fill price when the order fills. Sell orders pay a platform fee with a fixed part per node-hour and a percentage of the sale value, per SF Compute's reselling guide.
Is there a limit on how much I can reserve on SF Compute?
Yes, accounts carry a compute limit in node-hours that counts both held time and unfilled buy orders. Its limits docs describe how standing sell orders and callable buy orders offset it, and how to request a higher cap.
Can I reserve just one H100 through Nodus?
Yes. The Reserved Compute docs say you can request a single GPU, and the final offer confirms how GPUs are arranged across physical servers. Console requests accept durations up to 365 days.
Originally published on the Nodus blog.
Top comments (0)