Tensormesh, a California-based startup focused on reducing the cost of running AI at scale, has raised $20 million in new funding. The company says it is solving a problem that quietly inflates AI bills for enterprises worldwide: every time an AI system answers a question, it reprocesses all its context from scratch, even when it has done that exact work before. Junchen Jiang, a University of Chicago professor and co-creator of LMCache, co-founded the company alongside CTO Yihua Cheng and Chief Scientist Kuntai Du, both PhD researchers from the same university.
AI inference, the process of running a trained model to generate responses, has become the dominant cost in enterprise AI. The market for inference services stood at $106 billion in 2025 and is expected to reach $255 billion by 2030. By early 2026, inference already accounted for roughly 55% of cloud AI spending, up from 33% in 2023. For large organizations, the numbers are stark: inference accounts for 80 to 90% of total AI compute spend, and that bill climbs with every new user, feature, or model added. The AI inference market reached $106 billion in 2025, with no sign of slowing.
This $20 million round extends Tensormesh's seed funding, bringing total capital raised to $24.5 million. The investor group includes AMD Ventures, CoreWeave, NVentures (NVIDIA's venture capital arm), Valley Capital Partners, and Laude Ventures. The company also announced that Tensormesh Inference, its main SaaS product, is now generally available.
To see the problem Tensormesh is addressing, it helps to understand how AI models handle requests. Each time an application calls a model, the model reprocesses everything: the system instructions, the conversation history, any documents or tool definitions it was given. All of it, from the beginning, every single time. For a one-off query, that is fine. For agentic workflows, where an AI system runs hundreds or thousands of chained requests per day, the cost of that repeated processing accumulates fast.
KV caching addresses this directly. The term stands for key-value caching, a technique for storing the intermediate results of a model's computation so that repeated inputs do not need to be recalculated. If the same system prompt appears in every request, the model processes it once and stores the result. Future requests pull from that stored result instead of starting over. Responses come back faster, and the GPU cycles that would have been spent on redundant work are freed up.
"KV caching represents one of the most consequential and underexplored opportunities in AI infrastructure today. Tensormesh has built the only platform that makes this technology production-ready for the enterprise, and we believe it will become a critical part of how every serious AI deployment is run." — Steve O'Hara, founder and managing partner at Valley Capital Partners and a Tensormesh board member
The capital will go toward product development, deeper hardware integrations with AMD, CoreWeave, and NVIDIA, and continued work on LMCache, the open-source KV caching project the company maintains. LMCache has over 8,000 GitHub stars and connects with widely used inference frameworks including vLLM, SGLang, TensorRT, NVIDIA Dynamo, AWS SageMaker, and Oracle OCI Data Science.
On the commercial side, Tensormesh Inference ships in two forms. Serverless inference gives teams immediate API access to a range of models, with an interface compatible with OpenAI tooling so no changes to existing code are needed. Reserved deployments offer dedicated capacity for organizations running AI at higher volumes, with predictable performance and custom support terms.
"Inference economics will define what is possible for the next generation of AI products. Tensormesh is tackling one of the most important challenges in AI infrastructure: helping companies reduce GPU spend without requiring changes to application code. The combination of meaningful cost savings and simple deployment is rare, it positions Tensormesh to become a critical layer in the AI infrastructure stack." — Hui Zhang, CTO and co-founder of Conviva and advisor to Tensormesh
The company draws its founding team from the University of Chicago, UC Berkeley, and Carnegie Mellon University. The founders' research background covers distributed systems and AI infrastructure, and their work on LMCache and CacheBlend earned the ACM EuroSys 2025 Best Paper Award. Tensormesh is headquartered in Foster City, California.
The platform stores computed results in a KV cache and reuses them across future requests that share the same inputs. When an application repeatedly sends the same system prompt, document context, or tool definitions, those elements are cached and served at no additional cost on subsequent calls. The company says well-optimized deployments can achieve cache hit rates above 70%, meaning more than seven in ten requests are served from stored results rather than full recomputation. Tensormesh claims this can reduce latency and GPU spend by up to 10x under those conditions.
A built-in Cost Savings Dashboard tracks performance in real time, showing cache hit rate, the proportion of cached to total input tokens, and the dollar value of savings as they accumulate. Most inference providers handle caching on their backend without telling customers it is happening, which means companies have no way to know what is being cached, how often, or whether they are benefiting financially. Tensormesh prices cached input tokens at $0, and the company says this is a permanent part of its pricing model, not a limited promotion.
"Tensormesh offers a new vision on the significance of the intermediate data that LLMs generate when processing prompts. Behind the term KV cache is a whole concept of AI interpretation of the question it is asked. This makes it a whole new class of data and a category Tensormesh is uniquely positioned to define." — Junchen Jiang, CEO and Co-Founder of Tensormesh
Tensormesh has raised $24.5 million in total, with this $20 million round completing the seed stage. The investor list reads like a cross-section of the AI hardware and cloud infrastructure market. NVentures, NVIDIA's venture arm, and AMD Ventures are the two largest GPU manufacturers. CoreWeave is among the largest AI-focused cloud providers. Valley Capital Partners and Laude Ventures round out the group.
Having GPU manufacturers and a major cloud provider invest in the same caching startup suggests they see this technology as infrastructure-level, not a peripheral optimization.
"As enterprises scale AI workloads, maximizing every GPU cycle is critical. Software innovations like KV caching are a powerful complement to raw accelerator performance. Paired with AMD Instinct™ GPUs, Tensormesh's platform can help customers drive value from their infrastructure investments." — Ramine Roane, Corporate Vice President of AI at AMD
