Thunder Compute secured a $13 million Series A, led by Matrix Partners, to commercialize its network based GPU virtualization technology, enabling cloud providers and enterprises to dramatically raise utilization of underused GPU fleets.
Thunder Compute’s $13 million Series A funds the scaling of its GPU virtualization software to unlock idle capacity in enterprise and cloud data centers. The round was led by Matrix Partners, with participation from Y Combinator and CEAS Investments. The company positions the raise as enabling a shift from operating its own virtualized GPU cloud (as a technology testbed serving more than 10,000 users) toward partnering with external cloud providers and enterprises that already own large GPU fleets. Funds will support hiring systems researchers, enterprise deployment engineers, and a sales team, plus continued advancement of the virtualization technology.
What is Thunder Compute?
Thunder Compute was founded in 2022 (some profiles list 2024) by CEO Carl Peterson (formerly a Bain & Company management consultant) and Brian Model (formerly a quantitative developer at Citadel Securities). Headquartered in San Francisco, the company describes itself as a systems lab commercializing GPU virtualization research. Its team draws from backgrounds at Citadel Securities, Aquatic, and AWS.
The core thesis is that GPUs lag other hardware resources: CPUs, storage, networking, and memory have long been virtualized, enabling pooling, dynamic allocation, and high utilization, while GPUs remain largely allocated as dedicated bare metal resources. Enterprise GPU utilization averages only 5–20% (per the CastAI 2026 State of Kubernetes Optimization Report), leaving an estimated ~$200 billion in idle capacity. Trillions continue to be spent on GPU CapEx even as large fractions of existing silicon sit idle.
Thunder’s proprietary software implements network based GPU virtualization (sometimes called GPU over TCP). Key technical characteristics include:
- GPUs are treated as network resources rather than tightly coupled PCIe devices. Workloads communicate with GPUs over the data center network fabric via TCP.
- Virtualization sits at the CUDA layer and is invisible to applications: existing inference, training, or other GPU code runs unchanged while CUDA calls are translated into network messages.
- Sole tenancy is preserved while a process is actively using a GPU (full VRAM and compute access); when the process exits or idles, the GPU is reassigned.
- Connection setup incurs ~10–20 ms latency (once, at startup). Ongoing impact on runtime is optimized at the systems level and is often negligible for common workloads; edge cases may see up to ~2x slowdown, which the company argues is outweighed by fleet level efficiency gains.
- Security includes wiping GPU memory and resetting the card between jobs.
This differs from existing approaches such as NVIDIA MIG (single node partitioning) or traditional GPU passthrough/vGPU (still essentially dedicated allocation). Network based pooling enables a cluster wide scheduler to fill utilization gaps across an entire fleet, conceptually similar to network storage systems like Ceph or SANs.
In its own cloud, Thunder reports serving ~1.8x more users on the same GPU fleet than without virtualization; some enterprise pilots have seen gains of 4x or more depending on workload and baseline utilization (especially I/O-bound or long running reserved instances). The company currently has two enterprise pilots underway.
Recommended: Antora Energy Raises $550 Million In Series C Funding Round
Peterson has framed the goal as building the “VMware for GPUs”, making the underlying hardware disappear so that capacity can be scheduled whenever it is available rather than reserved continuously for a single workload. Economic value accrues primarily to the GPU owners (cloud providers or enterprises), who can extract more work from existing CapEx and potentially lower prices; developers experience little change beyond simply requesting a GPU.
The Series A marks an intentional transition: the company spent roughly four years in stealth developing prototypes and then operated its own cloud as the primary proving ground (“selling to ourselves”). The new capital supports productizing the software for external fleets and building the commercial organization required for that go to market motion. Longer term, the virtualization layer is viewed as a foundation for a broader GPU infrastructure platform.
Public data indicate prior seed funding, including an early Y Combinator-backed round and subsequent seed activity (estimates of total capital raised before or including this round range around $17–18 million across multiple rounds). The $13 million Series A sits toward the smaller end of recent AI infrastructure financings but targets a clear, high value bottleneck: not merely acquiring more GPUs, but raising utilization of those already deployed.
The company ranks among a competitive field of GPU optimization and cloud infrastructure players but differentiates itself through systems level, network based virtualization rather than purely workload layer scheduling or single node techniques. Its website and communications emphasize abundance through software efficiency rather than solely through additional silicon manufacturing.
The raise provides runway to move from internal proof of concept and self serve cloud operations into enterprise and provider partnerships, with the potential to materially improve data center economics if the reported utilization multipliers hold at scale.
Please email us your feedback and news tips at hello(at)superbcrew.com
