SF Compute
San Francisco, US · 12 known investors
San Francisco Compute sells time on large-scale GPU clusters through contracts that customers can exit by reselling capacity on its liquid market. It operates the clusters directly with its own virtualization and bare-metal stack, serving AI and infrastructure customers who need scalable GPU compute without long-term commitments.
Also known as San Francisco Compute · SFCompute · The San Francisco Compute Company
Investors · 12
Also in the syndicate · 2
Valuation · disclosed
Disclosed eventsSource: SEC prospectus filings, and round valuations the company or its investors disclosed — follow each entry's link for the claim.
Company profile
researched Aug 2026SF Compute (the San Francisco Compute Company) builds data centers, the GPU clusters inside them, and the cloud layer on top, selling time on clusters of up to thousands of GPUs. Customers buy compute through contracts that can be exited by reselling the remaining capacity to other buyers on the company's order book, which the company describes as a liquid market for GPU offtake. Access is offered as virtual machines, bare metal, and managed Slurm and Kubernetes, via an API, a command-line tool ("sf"), or direct engagement with the company's team for larger deployments.
Contract sizes range from a single node for one hour to thousands of nodes for multiple years, and virtual-machine purchases can run from an hour up to six months. The CLI workflow is built around pools (which track a customer's compute allocation over time), buy and sell orders placed against hardware SKUs, bootable images, and instances; standing orders remain open on the order book until filled or cancelled. Compute ownership is separated from instances, so instances can be terminated and recreated without losing purchased compute time. Sell-back orders return credits to the seller minus a variable platform fee. Documentation lists SKUs in multiple regions including Amsterdam, Ashburn, Chicago, Seattle, and Tokyo, with H100 and A100 hardware, and public images ship with Ubuntu and CUDA preinstalled.
The company positions itself as an operator rather than a reseller, running the clusters with its own virtualization and bare-metal stack. It states it does not own the GPUs it provides access to but manages more than $100 million in hardware.
Founding story
The founders started the company after failing to buy compute on affordable terms while building a music generation product described as "Midjourney for music"; providers required minimum purchases and multi-year contracts the team could not take on, so they built SF Compute to sell short-duration cluster time to themselves and others.
Business model
SF Compute sells reserved GPU capacity under contracts of varying length and operates an order book on which tenants can resell idle capacity to other customers; the company operates the underlying clusters itself with its own virtualization and bare-metal stack rather than acting as a reseller. Sell-back proceeds are returned to the seller as credits net of a variable platform fee.
Customers purchase compute time at market rates quoted per GPU-hour or per node-hour, prepaid against an organization credit balance, with a variable platform fee applied when capacity is resold on the order book. Long-term reservations of custom H100, B300, or GB300 clusters are contracted directly with the company.
Traction
As of December 2025 the company employed around 30 people and said it manages more than $100 million in hardware while not owning the GPUs itself. Its site lists H100, B300, and GB300 clusters for long-term reservation, and documentation shows available H100 and A100 SKUs across five regions. Site pricing displays an average of $1.91 per GPU-hour over an August period, and press coverage notes listed pricing for Nvidia H100 and H200 GPUs.
Latest developments
In December 2025 SF Compute announced a $40 million Series A led by DCVC and Wing Venture Capital with participation from existing investors Electric Capital and Alt Capital, at a reported $300 million valuation, to expand its marketplace. Around the same period it appointed Eric Park, previously CEO of Voltage Park, as CTO and added executives from Sun Microsystems and Lambda. Its site indicates upcoming availability of B300 and GB300 clusters.
▸Full profile — market position, technology, go-to-market, geography, history, risks & controversies
Market position
SF Compute presents itself as a marketplace-based alternative to GPU clouds that require long fixed-price commitments, aiming to de-risk GPU offtake contracts and introduce liquidity and price discovery to compute. Press coverage describes it as a San Francisco-based startup in the AI compute marketplace segment; an investor quoted in that coverage frames the market as historically rigid and hard to access. Company commentary compares its model to providers such as CoreWeave, which it characterizes as focused on locked-in long-term contracts, and it named Modal, Together, and DigitalOcean in discussing GPU cloud economics.
The company contrasts itself with providers that require long fixed-price contracts by letting customers exit commitments through resale on its order book, and with brokers by operating the clusters directly on its own virtualization and bare-metal stack. It sells slivers of time on large interconnected clusters (for example a month of a 1,024-GPU H100 cluster rather than a year-long commitment) and allows tenants to earn revenue on idle capacity.
Technology
The company operates GPU clusters and the cloud layer above them, offering virtual machines, bare metal, and managed Slurm and Kubernetes on its own virtualization and bare-metal stack. A scheduler and order book match buy and sell orders for compute time, including standing orders and auto-scaling that buys and sells time as fleets scale. Users interact through an API and the "sf" CLI, which manages pools, orders, SKUs, images, and instances, with SSH access to instances. Clusters offer InfiniBand networking (for example 1,024 H100s with 3.2Tb/s InfiniBand) and petabyte-scale, high-speed storage suited to large-model training.
Go-to-market
Self-serve signup with CLI and API access for virtual machines, plus a demo request and direct sales engagement for bare-metal and custom cluster deployments; the company also publishes pricing and cluster availability on its site and markets through its blog and podcast appearances. It is hiring supply-chain-focused account executive, customer support, HPC/GPU cluster architect, design engineering, legal, and people operations roles.
AI and infrastructure teams that need large-scale training compute without multi-year commitments, including small startups that cannot underwrite long take-or-pay contracts, as well as enterprises and data centers with excess capacity to sell. The company notes its InfiniBand- and storage-heavy clusters suit large-model training more than steady-state inference workloads.
Geography
Headquartered in San Francisco, with cluster SKUs listed in documentation across regions including Amsterdam, Ashburn, Chicago, Seattle, and Tokyo.
History
Founded in 2023. An early blog post announced the company selling short-duration H100 training cluster time, with its first cluster, Angel Island, online and a second, Bay Bridge, due within months. CEO Evan Conrad discussed the marketplace model publicly in an April 2025 podcast. In late 2025 the company hired Eric Park, former CEO of AI cloud provider Voltage Park, as CTO, along with executives from Sun Microsystems and Lambda, and raised a $40 million Series A announced in December 2025.
Risks & controversies
The company's own materials note that GPU cluster financing carries thin margins and high volumes, that risk is typically pushed onto customers through fixed-price long-term contracts, and that unmitigated customer risk implies a bubble. It also states its InfiniBand- and storage-heavy clusters are a poor fit for steady inference demand, and that resale prices depend on market conditions, with orders cancelled if bid rates are too low or availability is absent.
Compiled by commissioned research from 8 cited public sources — announcements, filings, and press listed under research sources below.
Key figures
latest reportedCompany-reported or press-reported figures, each dated to when it was claimed — not independently audited.
Timeline · 3
launches, deals, and filingsSeries A equity financing of $40 million to expand the company's AI compute marketplace, at a reported $300 million valuation, with participation from existing backers Electric Capital and Alt Capital.
$40M source ↗
SF Compute hired Eric Park, former CEO of AI cloud provider Voltage Park, as chief technology officer, and also hired executives from Sun Microsystems and Lambda.
The company announced availability of short-duration H100 training cluster time with its first cluster, Angel Island, online and a second cluster, Bay Bridge, expected to come online within months.
Dated company events from announcements, filings, and press; legal rows summarize public dockets and regulator releases.
In the news
▸Research sources · 8
primary sources listed
- SF Computesfcompute.com · web
8 public sources were cited for this profile; the first-party ones are listed here.
Frequently asked questions
- What does SF Compute do?
- SF Compute runs large GPU clusters and an order book that lets customers resell reserved compute they don't use.
- Who are SF Compute's investors?
- SF Compute's investors include AlphaLab Capital, Angel Collective Opportunity Fund, Caladan, Deep Acre, Electric Capital, Moth Fund, Predictive VC, Wing and 2 more.
- Where is SF Compute headquartered?
- SF Compute is headquartered in San Francisco, US.