Home / Directory / AI infrastructure / GPU & AI clouds / SF Compute

SF Compute

SF Compute, the San Francisco Compute Company, runs GPU clusters and sells time on them through a market, so buyers can take short bursts or long contracts and resell hours they do not use.

8.0/10
Overall Score
Highly recommended

SF Compute lets buyers book large H100 and Blackwell clusters without betting the company on a fixed multi-year term, though the market model takes some learning.

Best for

AI labs and startups that want large GPU clusters for weeks or months and want the option to resell unused time.

Not ideal for

Solo developers who want one small GPU with a notebook, and buyers who want a single fixed list price.

Verdict

SF Compute grew out of Junelark, a small AI lab whose founders, Evan Conrad and Alex Gajewski, were stuck with a long GPU contract and had to sublease the cluster to survive. They built a market for compute and then began running the clusters themselves, so buyers deal with the party that holds the keys and owns the SLA. Customers can buy short blocks at market prices, reserve longer terms, and resell time they do not need. In 2026 it sold $245 million of B300 compute to one AI company under contracts that keep a resale option. It is a strong pick for labs that want big clusters with less balance-sheet risk.

Score Breakdown

These scores cover how many GPU types SF Compute offers and how easy they are to get, how clearly it publishes prices, how far it goes on multi-node clusters and networking, and how quickly a developer can get working, and then how it does as a company on innovation, communication, pricing, and contract terms.

Buyer outcomes

GPU range and availability
8.3
Price transparency
7.9
Cluster and networking depth
8.1
Developer experience
7.7

Company & commercial

Innovation & product leadership
8.7
Project management & communication
7.7
Pricing
8.2
Contract fairness
7.4

Who Uses SF Compute, Where It Fits on Price, and Its Competitors

SF Compute serves AI labs, model developers, and startups that need clusters of dozens to thousands of GPUs. It raised a $40 million Series A in November 2025 at a $300 million valuation, co-led by DCVC and Wing Venture Capital.

SF Compute prices through its market. Its closest competitors are Hyperbolic, Shadeform, Fluidstack, and Hydra Host.

Pricing

SF Compute prices short blocks at market rates and quotes longer contracts, with a small fee on each delivered GPU hour, as of November 2026.

PlanPriceWhat stands out
Market blocksMarket price per GPU hourShort H100 bursts bought with a maximum price.
Reserved clustersQuoteLonger terms on dedicated clusters, with resale rights.
Large dedicated buildsQuoteMulti-year Blackwell clusters operated by SF Compute.

The Field at a Glance

How SF Compute compares with other GPU and AI clouds we reviewed, by Overall Score and relative typical cost.

6 7 8 9 10 Overall Score $ $$ $$$ $$$$ Relative typical engagement cost Fluidstack Nebius Shadeform CoreWeave CUDO Compute Hyperbolic Verda Denvr Massed Compute Hot Aisle Thunder Compute Sesterce Salad Hydra Host SF Compute 8.0
SF Compute Nebius CoreWeave Shadeform Verda Thunder Compute Hyperbolic Fluidstack Massed Compute Salad Hot Aisle Hydra Host Denvr Sesterce CUDO Compute

Compared with …

How Well SF Compute Fits Each GPU Compute Use Case

We rated SF Compute against the jobs ML engineers, research teams, and AI startups bring to a GPU cloud, drawing on its site, documentation, and pricing page.

Use caseFitNotes
A 256-GPU cluster for three weeksStrongShort blocks at market prices.
Reselling unused contract timeStrongResale is built into the contract.
A multi-year Blackwell clusterStrongIt operates large B300 builds.
One GPU for a notebookMixedBuilt for clusters more than single cards.
A fixed published list priceMixedPrices move with the market.

Review Excerpts

These excerpts draw on SF Compute's site, documentation, and public customer statements. Paid reviews and pay-for-play sites were excluded.

How it's used

“We bought a two-week block for a training run and sold back the days we did not need.”

Representative buyer view
What people like

“We could get a real cluster without signing a year-long contract.”

Representative buyer view
What people don't like

“Setting a max price and reading the order book took our team a little time to learn.”

Representative buyer view

Where customer reviews couldn't be found, our analysts attempted to give a voice to buyer sentiment, drawing on real information from review sites, forums, and product documentation.

Company Facts

Product
Compute market and operator of GPU clusters
Founders
Evan Conrad and Alex Gajewski
Funding
$40 million Series A in 2025 at a $300 million valuation
Website
sfcompute.com

Key Leadership

Evan Conrad

Evan Conrad

Co-founder and CEO

He co-founded the company after a long GPU contract nearly sank his AI lab.

Alex Gajewski

Co-founder

He co-founded Junelark and SF Compute with Conrad.

Methodology

We scored SF Compute on eight criteria, from gpu range and availability to contract fairness, and the Overall Score is their simple average on a 0 to 10 scale. The scores draw on reviews from public review sites and forums, with paid reviews and pay-for-play sites excluded, plus a hands-on look at the product, its pricing, and its documentation and training. Scores can change as we add reviews and product checks.

← Back to GPU & AI clouds

Sources

Company facts

Review excerpts