Nvidia, AMD, and CoreWeave rarely appear together on an investor list. Tensormesh's latest funding round is noteworthy not for its size but for targeting one of the biggest inefficiencies in AI inference: redundant computation.
On May 27, Tensormesh announced a $20 million funding round from investors including AMD Ventures, Nvidia's NVentures, CoreWeave, Valley Capital Partners, and Laude Ventures. Including previous funding, the company has raised approximately $24.5 million to date.
The same day, Tensormesh launched its inference optimization platform, Tensormesh Inference, into general availability. The platform addresses a straightforward problem: when models handle similar requests or long-context tasks, they repeatedly compute intermediate results that have already been calculated, wasting GPU cycles.
Tensormesh's approach is to cache these intermediate results—commonly referred to as KV cache—so that previously computed context is reused, and only new parts are computed. The company claims that in suitable scenarios, this can reduce latency and GPU costs by up to 90%, with cache hit rates exceeding 70% in some customer deployments.
This is especially critical in agent scenarios, where a complex task may require dozens of model calls, and earlier context is repeatedly carried into subsequent requests. Without caching, inference costs multiply quickly. KV cache turns that repetitive context into manageable, reusable infrastructure.
Tensormesh is not starting from scratch. It is built on the open-source project LMCache, which has over 8,000 GitHub stars and integrates with vLLM, SGLang, TensorRT, NVIDIA Dynamo, AWS SageMaker, and Oracle OCI. The commercial version packages this proven open-source foundation into a service that enterprises can deploy and purchase directly.
For Nvidia, AMD, and CoreWeave, the investment also makes clear strategic sense. Chipmakers and cloud providers want customers to run inference more cheaply and reliably, because the more controllable inference costs become, the more willing enterprises are to push AI applications into production. Tensormesh does not replace GPUs; it improves utilization on top of them.
Over the past year, most AI infrastructure spending has gone toward larger models and more GPUs. Tensormesh represents a different approach: on the same models and hardware, first recover wasted compute. As agents continue to drive up inference volume, this cost-saving infrastructure segment will become increasingly crowded.
The next question is whether KV cache can evolve from a point optimization into a standard layer in the enterprise inference stack. If that standard is established, Tensormesh will have discussed not just a funding milestone, but a long-term toll gate in the AI application era.
Sources: Tensormesh Raises $20M from Investors Including AMD Ventures, CocoLoop, CoreWeave, NVentures (Business Wire); Tensormesh taps Nvidia, AMD and CoreWeave for funding to fix AI model memory problems (SiliconANGLE)