Flatkey launches dedicated NVIDIA B300 clusters for AI inference
Flatkey has launched Flatkey Compute, a NeoCloud that builds and operates dedicated NVIDIA B300 clusters for AI inference. Each cluster is reserved for one customer, with no other tenant sharing its hardware.
- Flatkey Compute provides single-customer NVIDIA B300 clusters.
- The clusters are built for AI inference, not general-purpose computing.
- Customers can specify GPU count, region and network topology.
- One, three and five-year reserved terms are available.
- Smaller teams can rent a few GPUs by the hour on demand.
Dedicated infrastructure for production inference
Flatkey says its service is designed for AI teams moving workloads from experimentation into production. The company argues that general-purpose cloud infrastructure, where multiple tenants share physical hardware, can make inference performance less predictable and complicate security reviews for regulated customers.
Under Flatkey Compute, a complete cluster is assigned to a single customer. According to Flatkey, this means another tenant's workload cannot affect the cluster's performance, and customers do not have to account for co-tenants when reviewing access to the machines.
The clusters run on NVIDIA B300 hardware and are built specifically for AI inference. Flatkey does not offer general-purpose computing, instead focusing its infrastructure and operations on AI workloads.
Flatkey sizes and operates each deployment
Flatkey says it has already delivered compute capacity to dozens of major customers ahead of the launch. Rather than selecting from standard instance types, customers specify the GPU count, region and network topology required for their workloads.
After a cluster is deployed, Flatkey operates it for the customer. The company says this removes the need for the customer to maintain its own infrastructure team to run the infrastructure.
Flatkey also describes the service as a compliant and trusted inference cloud. The company says it is GDPR compliant, continuously monitors its security controls through Vanta and does not retain request content.
Pricing, terms and availability
Customers can request capacity through Flatkey's website by providing the GPU count, region and term they need. Flatkey says it replies within 24 hours and then provides a formal quote covering the price, configuration, region and provisioning date.
Reserved terms of one, three and five years are available, with longer commitments priced lower, according to the company. A reference GPU price index on the request page provides indicative rates before customers make contact.
Flatkey Compute is available now by request. Smaller teams that need only a few GPUs can rent them on demand by the hour through the Flatkey console. The company launched in July 2026 and is headquartered in San Jose, California.
What the launch means for AI teams
Flatkey is positioning dedicated clusters as an alternative for organizations that need more control over inference infrastructure than shared cloud capacity provides. The model combines exclusive hardware with an operated service, allowing customers to specify deployment requirements without running the cluster themselves.
“Inference is moving from something teams experiment with to something their product depends on,” Flatkey founder Hunter Guo said. “At that point, you don't want to be sharing a machine with strangers. You want your own cluster, you want to know where it is, and you want somebody to run it for you. That is what we are building.”
