Pick the model. Pick the silicon. See the price first.
Basalt Compute lists NVIDIA, AMD and Tenstorrent machines in one catalog, each with its hourly price. The aim is to run open-weights models on whichever one is cheapest for the job. Deploying is not open yet; it comes in a later release. Until then, talk to us.
Prices are listed per hour. Nothing can be deployed or charged yet.
What $25 of credit would run
89 h 15 min00 h 00 min
About 89 hours 15 minutes with $25 of credit
Tenstorrent Grayskull e150, $0.28 per hourTenstorrent Wormhole n150, $0.48 per hourNVIDIA L4, $0.59 per hourTenstorrent Wormhole n300, $0.88 per hourNVIDIA RTX 4090, $0.89 per hourNVIDIA L40S, $0.99 per hourNVIDIA A100 80GB, $2.29 per hourAMD MI300X, $2.99 per hourNVIDIA H100 80GB, $3.59 per hourNVIDIA B200, $6.49 per hourAMD 8x MI300X, $23.90 per hour
Estimate from the catalog price, not a live deployment
Day 1
Auto-stop, 24 hours
Day 2
Day 3
Day 4
Elapsed
89:1500:00
Credit left
89:1589:15
Used
$25.00$0.00
Nothing used yet.Running. Billed in 5-minute increments.Stopped by auto-stop, about 24 hours in.Running. The credit runs out before the 24-hour auto-stop.Limit removed. Running until the credit runs out.Running. Under 24 hours of credit left.Running. Under 1 hour of credit left.Credit used up. The deployment has stopped.Credit used up before the 24-hour auto-stop. The deployment has stopped.
01
02
03
04
Cool: more than 24 h left
Warm: under 24 h
Hot: under 1 h
What $25 of credit would run on each machine in the snapshot:
Tenstorrent Grayskull e150, $0.28 per hour: about 89 hours 15 minutes.
Tenstorrent Wormhole n150, $0.48 per hour: about 52 hours 5 minutes.
NVIDIA L4, $0.59 per hour: about 42 hours 20 minutes.
Tenstorrent Wormhole n300, $0.88 per hour: about 28 hours 20 minutes.
NVIDIA RTX 4090, $0.89 per hour: about 28 hours 5 minutes.
NVIDIA L40S, $0.99 per hour: about 25 hours 15 minutes.
NVIDIA A100 80GB, $2.29 per hour: about 10 hours 55 minutes.
AMD MI300X, $2.99 per hour: about 8 hours 20 minutes.
NVIDIA H100 80GB, $3.59 per hour: about 6 hours 55 minutes.
NVIDIA B200, $6.49 per hour: about 3 hours 50 minutes.
AMD 8x MI300X, $23.90 per hour: about 1 hour.
What credit would buy, and for how long. Deploying is not open yet.
01
What $25 of credit would run.
What $25 of credit would run on this machine at its hourly price: $25 divided by the price, rounded down to 5 minutes. Estimate from the catalog price, not a live deployment.
02
Billed in 5-minute steps.
When deploying opens, usage is metered every 5 minutes and prorated when it stops.
03
Auto-stop.
The plan is that auto-stop defaults to 24 hours, counted from when the deployment starts running, and can be changed.
04
Or run until the credit runs out.
With the limit removed, it would run until you terminate it or your credit runs out.
Better hardware exists. Using it is the hard part.
Most AI inference runs on one vendor’s chips. Other accelerators cost less per dollar of compute or memory on paper, but a new open model often needs porting and tuning before it runs well on them, and few teams will rewrite their stack to find out.
Basalt Compute puts those chips in one catalog next to the usual ones, so the price of trying them is visible first. The work of making models run on them is done in the open, in a public registry. We are building the part that lets you deploy on them; that part is not open yet.
What is here today, and what is coming
One catalog, three chip vendors.
Accelerator, memory, CPU, RAM, region, hourly price and stock for every machine. Out of stock is shown as out of stock.
Open today.
An open registry.
Bring-up packages, bounties and engine images are public on GitHub.
Open today.
A person to talk to.
The contact form is open for inference, model porting and technical briefings.
Open today.
Deploying a model.
Choose a model and a machine, see the cost first, get an endpoint.
Not open yet.
From model to endpoint: the plan
Deploying is not open yet. This is the plan:
01
Choose a model.
Search by name and filter by type, family, memory needed and quantization.
02
Choose hardware.
See which machines are likely to fit, what each costs per hour, and whether it is in stock.
03
Deploy.
Review the summary, then deploy. Disk is 50 GB and auto-stop is 24 hours unless you change them.
04
Call the endpoint.
When the instance is up, copy its address into cURL or Python. The endpoint will have no API key, so anyone with the address can call it. Treat it as private.
What the hardware costs
List prices per hour for the machines in the catalog. A snapshot taken on ; the live hardware page has current prices.
$0.28Grayskull e150
$0.48Wormhole n150
$0.59L4
$0.88Wormhole n300
$0.89RTX 4090
$0.99L40S
$2.29A100 80GB
$2.99MI300X
$3.59H100 80GB
$6.49B200
Grayskull e150$0.28
Wormhole n150$0.48
L4$0.59
Wormhole n300$0.88
RTX 4090$0.89
L40S$0.99
A100 80GB$2.29
MI300X$2.99
H100 80GB$3.59
B200$6.49
8x MI300X$23.90
Machine
Vendor
Price per hour
Grayskull e150
Tenstorrent
$0.28
Wormhole n150
Tenstorrent
$0.48
L4
NVIDIA
$0.59
Wormhole n300
Tenstorrent
$0.88
RTX 4090
NVIDIA
$0.89
L40S
NVIDIA
$0.99
A100 80GB
NVIDIA
$2.29
MI300X
AMD
$2.99
H100 80GB
NVIDIA
$3.59
B200
NVIDIA
$6.49
8x MI300X
AMD
$23.90(not drawn in the chart)
Hourly list prices from the catalog on 18 November 2026. These are list prices, not benchmarks of your workload. The 8x MI300X machine, $23.90 per hour, is not drawn.
Three ways to work with Basalt Compute
Run inference.
Look at what a machine costs per hour before you commit to it, and compare the same class of job across vendors. Deploying on them comes in a later release.
Systems and ML engineers bring models up on alternative accelerators in the open, in the registry on GitHub. Bounties name the target hardware and the checks a port must pass. The intent is that engineers earn from the work as it is used; payments are not part of the site today.
If you make accelerators or run a datacenter with capacity, we want to talk.
What we can say today
There are no accounts and no payments. Nothing on this site asks you to sign in or pay.
Your messages go to our form service. If you send the contact form, your email address and message go to Formspree, a third-party form service, until we replace it with our own endpoint. The privacy page says what it receives.
No tracking cookies. We set none. Visit counts and page-speed measurements use cookieless tools, and only where they are turned on. The site keeps your light or dark choice on your device.
HTTPS with security headers, including a content security policy that limits what a page may load.
Error reports leave out secrets. If a page fails, a report goes through our own server to our error-reporting service, with tokens and similar values removed first.
Priced by the hour
Each machine has an hourly price, shown in the hardware catalog next to its specification. Nothing is charged today, because nothing can be deployed yet.
Hourly prices. See every machine and its price on the hardware page. See hardware and prices
How billing will work. When deploying opens: prepaid credit, metered every 5 minutes, prorated when a deployment stops.
Prices change. The table above is a snapshot of 18 November 2026; the live page is current.
Not here yet. Accounts, deploying, credit, endpoints, logs, API keys, one address for every model, per-token pricing, usage history and receipts. We will not describe them as available until they are.
A catalog of NVIDIA, AMD and Tenstorrent machines with their hourly prices, built so open-weights models can run on whichever is cheapest for the job. Behind it is an open registry where engineers bring models up on alternative hardware.
Can I deploy a model today?
No. Deploying, accounts and billing are not open yet and come in a later release. You can browse the hardware and its hourly prices, read the registry on GitHub, and talk to us.
Which hardware is listed?
The machines on the hardware page when you look. On 18 November 2026 that was NVIDIA, AMD Instinct and Tenstorrent accelerators, each with memory, CPU, RAM, region, hourly price and stock.
How much does it cost?
Each machine has an hourly price on the hardware page. Nothing is charged today because nothing can be deployed. When deploying opens, usage will be billed in 5-minute steps, prorated when a deployment stops, from prepaid credit.
When will deploying open?
There is no date. Sign-in and the console follow in later releases. If you want to know when, use the contact form and say so.
What happens to what I send through the contact form?
It goes to Formspree, a third-party form service, until we replace it with our own endpoint. We use it to reply to you. The privacy page has the details.
How do I contribute models, and can I earn from them?
Open the registry on GitHub. Bounties list the target hardware and the checks a port has to pass, and accepted work is published with its image digest. The intent is that contributors earn as their work is used. Bounty browsing, payments and earnings are not part of this site today.
Do you track visitors?
We set no tracking cookies. Page views and page speed are measured with cookieless tools, and only where switched on. Details are on the privacy page.
Can I switch from another provider by changing a base URL?
Not yet. Basalt Compute does not offer a single address for every model, API keys or per-token pricing. If that changes, this page will say so.
See what it costs. Then talk to us.
Browse the hardware and its hourly prices, or tell us what you want to run.