Thank you for reaching out. We’ll be in touch shortly.
5150 E Pacific Coast Hwy
2nd Floor
Long Beach, CA 90804
Blackstart delivers GPU as a Service, private cloud, public cloud and dedicated superclusters from a Silicon Valley campus we run end to end — facility, power and hardware under one operator. That means you buy capacity from the people running the building, not from a reseller standing between you and the rack.
Most GPU clouds rent rack space in someone else’s building. We run the building.
Blackstart controls the campus, the power interconnection and the GPU fleet that runs inside it. One operator across the whole stack means fewer parties to negotiate with, pricing set at the source, and a capacity roadmap you can plan a multi-year workload against.
From a single reserved GPU to a dedicated supercluster on its own power block. Start where your workload is today and move up without changing provider, region or network.
Reserve GPU capacity by the hour, the month or the year. Bare metal or virtualised, provisioned in our own halls, billed on a straightforward rate card with no cross-region egress games.
A physically isolated environment inside our facility: your own racks, your own network segment, your own storage tier. No shared tenancy, no noisy neighbours, and a documented chain of custody from the loading dock to the GPU.
Self-serve, API-first access to shared GPU capacity with preconfigured environments. Spin up a training run, an inference endpoint or a notebook in minutes and pay only for what runs.
Large, non-blocking GPU fabrics built to order on their own power block. Designed with you, energised on our schedule, and operated by the same team that runs the facility around it.
Almost every GPU cloud you can buy from today is a tenant somewhere else. They lease the hall, buy power at a markup, and pass both to you. Blackstart controls every layer between the utility feed and your API key.
Why this matters to a tenant. Compute pricing is mostly power and real estate wearing a software badge. When a provider leases both, every rate they quote you carries a landlord’s margin and a utility contract they do not control.
Price. We set rates against our own cost of power and capital, not against a lease we have to cover.
Certainty. Expansion does not depend on a third party approving a fit-out or a power upgrade. If you need another block next year, we build it.
Accountability. One contract, one operator, one team on site. When something needs fixing at 3am, nobody is waiting on a landlord ticket queue.
Our flagship facility is at 650 Harry Road in San Jose, California — the former IBM Almaden research campus. It was engineered for heavy technical load long before AI needed it, and it sits inside the market where your customers, your engineers and your network peers already are.
25 MW is available now. Not queued, not subject to an interconnection study measured in years. Live capacity you can contract against this year.
The campus scales to 220 MW by 2028 across phased power blocks. Tenants who take capacity early hold a right of first refusal on the blocks that follow, so a workload that triples in two years does not have to move providers to do it.
Halls are designed for the rack densities modern accelerators actually draw, with liquid-ready provision for the next generation rather than the last one. Cooling is closed-loop, which keeps site water consumption far below conventional datacenter operation — a question enterprise procurement now asks on every deal.
Silicon Valley matters for reasons beyond prestige. Latency to the Bay Area user base is measured in single-digit milliseconds, carrier density means real route diversity, and your team can drive to the building.
See the facility →
The service tier follows the workload. Most tenants start on one and end up using two.
Multi-week runs across large, non-blocking fabrics with the storage throughput to keep every accelerator fed. Dedicated supercluster or private cloud.
Steady-state endpoints close to a dense West Coast user base. Reserved GPUaaS with burst headroom into shared capacity.
Short, frequent, unpredictable jobs. Public cloud with preconfigured images, moving to reserved capacity once the cadence settles.
Grant-cycle allocations with predictable annual pricing, and physical proximity for teams that want to see the hardware.
US-sited, single-operator infrastructure with documented physical custody. Private cloud, isolated network, evidence pack for audit.
Tightly coupled parallel workloads that need real interconnect rather than a virtualised approximation of one. Bare metal, dedicated fabric.
We are an operator, so procurement runs directly with the people who control the capacity.
Accelerator count, run duration, storage throughput, network shape and when you need it live. If you are not sure yet, a rough profile is enough to size a first quote.
We come back with a configuration, a power block, an availability date and committed-use pricing across term lengths. No brokered quotes, no wholesale markup.
Capacity is provisioned in our own halls with a named operations contact. Expansion blocks are reserved against the published campus roadmap.
25 MW live in San Jose today, scaling to 220 MW. Tell us what you need to run.
Request capacity →