TechCrunch · Oct 23, 2025
Russell Brandom explains Tensormesh’s bet on reusing the KV cache
Markdown · JSON · Original source
Russell Brandom’s TechCrunch article, published October 23, 2025, reports that Tensormesh launched with seed funding to build a commercial version of LMCache, an open source utility maintained by co-founder Yihua Cheng. The article describes the key-value cache (KV cache) as information that Tensormesh’s systems retain rather than discard after each query. Reusing it for a similar process in a separate query, Brandom writes, can provide more inference power for the same server load. He points to chat interfaces and , where models refer back to growing records of conversations or actions, as uses for this approach. Brandom’s account centers on a practical tradeoff: AI companies can build such systems themselves, but the technical complexity can make that difficult. Tensormesh is betting that companies will prefer an out-of-the-box product to hiring engineers and spending months on the work. CEO Junchen Jiang says teams have taken that route; he also describes keeping and reusing the KV cache without slowing the system as challenging. Brandom reports that LMCache can reduce by as much as 10x, but gives no measurement basis for that . The article leaves open whether customers will choose Tensormesh’s product and whether it can reuse the cache without slowing the system.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
TechCrunch · Article
Source details
Original source date: Oct 23, 2025 (2025-10-23).
Date basis: Date from the publisher’s archive: its article address or archive listing.
Added to Escape Meridian: 2026-11-11T14:33:25.536Z. First observed: 2026-10-29T00:55:37.219Z. Retrieved: 2026-10-29T00:57:37.250Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
Text as retrieved on October 29, 2026; originally published October 23, 2025. The article gives no measurement basis for its statement that LMCache can reduce by as much as 10x.
Open question
Can Tensormesh reuse the KV cache without slowing the system, and will customers choose its product over building their own?
Source-stated
Russell Brandom reports that Tensormesh launched with seed funding to build a commercial version of LMCache.
Tensormesh, launching out of stealth this week with $4.5 million in seed funding.
Source-stated
Brandom describes LMCache as retaining the KV cache so it can be reused for a similar process in a separate query.
Instead of discarding that cache, Tensormesh’s systems hold on to it, allowing it to be redeployed when the model executes a similar process in a separate query.
Source-stated
Brandom says reusing the cache can provide more inference power for the same server load.
the reward is significantly more inference power for the same server load.
Source-stated
Brandom reports that chat interfaces and agentic systems are relevant uses because they refer back to growing records.
The change is particularly powerful for chat interfaces, since models need to continually refer back to the growing chat log as the conversation progresses. Agentic systems have a similar issue, with a growing log of actions and goals.
Source-stated
Brandom reports that Tensormesh is betting companies will want a ready-made product because building the system is technically complex.
the technical complexity makes it a daunting task. Given the Tensormesh team’s work researching the process and the intricacy of the detail itself, the company is betting there will be lots of demand for an out-of-the-box product.
Source-stated
Brandom reports that LMCache can reduce inference costs by as much as 10x; the article gives no measurement basis for this figure.
Used well, LMCache can reduce inference costs by as much as 10x
Source-stated
Jiang says some teams have hired engineers and spent months building such a system, while Tensormesh offers its product as an alternative.
We’ve seen people hire 20 engineers and spend three or four months to build such a system. Or they can use our product and do it very efficiently.
Source and contributor context
TechCrunch · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Agentic AI
AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.
Model inference
Using a trained model to produce a prediction or response to new input.
Figure
Figure develops humanoid robots and publishes its own tests of how they perform physical tasks.
Editorial provenance
Version 1. Publication ID: 828b55ff-50f9-ba12-ea83-eec2a94d9196.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-11T14:30:10.795Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model.
Observation: f8dbd205791f4103c6ed16104d54c598f083c203130f7d7fd9f02423061e6de7. Source text hash: e4a67258ccffb90cf7ce839aa45e0518b07e968bac030e79698af940aace7e30.
Revision and source-check history
Version 1 · 2026-11-11T14:33:25.536Z
Russell Brandom explains Tensormesh’s bet on reusing the KV cache
Russell Brandom’s TechCrunch article, published October 23, 2025, reports that Tensormesh launched with seed funding to build a commercial version of LMCache, an open source utility maintained by co-founder Yihua Cheng. The article describes the key-value cache (KV cache) as information that Tensormesh’s systems retain rather than discard after each query. Reusing it for a similar process in a separate query, Brandom writes, can provide more inference power for the same server load. He points to chat interfaces and , where models refer back to growing records of conversations or actions, as uses for this approach. Brandom’s account centers on a practical tradeoff: AI companies can build such systems themselves, but the technical complexity can make that difficult. Tensormesh is betting that companies will prefer an out-of-the-box product to hiring engineers and spending months on the work. CEO Junchen Jiang says teams have taken that route; he also describes keeping and reusing the KV cache without slowing the system as challenging. Brandom reports that LMCache can reduce by as much as 10x, but gives no measurement basis for that . The article leaves open whether customers will choose Tensormesh’s product and whether it can reuse the cache without slowing the system.
Original source
TechCrunch · Article
Source details
Original source date: Oct 23, 2025 (2025-10-23).
Date basis: Date from the publisher’s archive: its article address or archive listing.
Added to Escape Meridian: 2026-11-11T14:33:25.536Z. First observed: 2026-10-29T00:55:37.219Z. Retrieved: 2026-10-29T00:57:37.250Z.
Limits
Text as retrieved on October 29, 2026; originally published October 23, 2025. The article gives no measurement basis for its statement that LMCache can reduce by as much as 10x.
Open question
Can Tensormesh reuse the KV cache without slowing the system, and will customers choose its product over building their own?
Source-stated
Russell Brandom reports that Tensormesh launched with seed funding to build a commercial version of LMCache.
Tensormesh, launching out of stealth this week with $4.5 million in seed funding.
Source-stated
Brandom describes LMCache as retaining the KV cache so it can be reused for a similar process in a separate query.
Instead of discarding that cache, Tensormesh’s systems hold on to it, allowing it to be redeployed when the model executes a similar process in a separate query.
Source-stated
Brandom says reusing the cache can provide more inference power for the same server load.
the reward is significantly more inference power for the same server load.
Source-stated
Brandom reports that chat interfaces and agentic systems are relevant uses because they refer back to growing records.
The change is particularly powerful for chat interfaces, since models need to continually refer back to the growing chat log as the conversation progresses. Agentic systems have a similar issue, with a growing log of actions and goals.
Source-stated
Brandom reports that Tensormesh is betting companies will want a ready-made product because building the system is technically complex.
the technical complexity makes it a daunting task. Given the Tensormesh team’s work researching the process and the intricacy of the detail itself, the company is betting there will be lots of demand for an out-of-the-box product.
Source-stated
Brandom reports that LMCache can reduce inference costs by as much as 10x; the article gives no measurement basis for this figure.
Used well, LMCache can reduce inference costs by as much as 10x
Source-stated
Jiang says some teams have hired engineers and spent months building such a system, while Tensormesh offers its product as an alternative.
We’ve seen people hire 20 engineers and spend three or four months to build such a system. Or they can use our product and do it very efficiently.
Source and contributor context
TechCrunch · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Agentic AI
AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.
Model inference
Using a trained model to produce a prediction or response to new input.
Figure
Figure develops humanoid robots and publishes its own tests of how they perform physical tasks.
Editorial provenance
Version 1. Publication ID: 828b55ff-50f9-ba12-ea83-eec2a94d9196.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-11T14:30:10.795Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model.
Observation: f8dbd205791f4103c6ed16104d54c598f083c203130f7d7fd9f02423061e6de7. Source text hash: e4a67258ccffb90cf7ce839aa45e0518b07e968bac030e79698af940aace7e30.
passed · 2026-11-11T14:30:10.795Z · anthropic/claude-opus-5.5
Publication: 828b55ff-50f9-ba12-ea83-eec2a94d9196
attribution — pass: This is a single TechCrunch news article. The reading attributes the reporting to Russell Brandom and the quotations to CEO Junchen Jiang. Tensormesh's bet and its claims stay attributed rather than presented as established fact.
Tensormesh co-founder and CEO Junchen Jiang argues that this is an enormous source of inefficiency.
summary — pass: A general reader can follow the summary. It names Tensormesh, its seed funding and LMCache, and describes the KV cache as information retained across queries instead of discarded. It gives the chat and agent uses and the build-versus-buy tradeoff. Inference and agentic systems are defined on the page. Omitting the $4.5M amount and lead investor does not distort the meaning. The question follows from the challenge Jiang describes.
Keeping the KV cache in a secondary storage system and reused efficiently without slowing the whole system down is a very challenging problem
date — pass: The reading's date of October 23, 2025 matches originalPublished at day precision. The page's dateModified falls on the same day. The undated Most Popular sidebar is not used. This is a publisher page, so no archive disclosure is required, and there is no hindsight framing.
launching out of stealth this week with $4.5 million in seed funding
support — pass: The claims match their excerpts, and the 10x figure is correctly flagged as having no stated basis. One claim attributes cache retention to LMCache while the source says Tensormesh's systems; since the product is built on LMCache, this is minor. Smallest repair: say Tensormesh's systems. No unsupported benefits or costs are implied.
Instead of discarding that cache, Tensormesh’s systems hold on to it
media — pass: No media is selected, and no verified media candidates were supplied.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
AI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.
Definition referenceUsing a trained model to produce a prediction or response to new input.
Definition referenceFigure develops humanoid robots and publishes its own tests of how they perform physical tasks.
Explore FigureUsing a trained model to produce a prediction or response to new input.
Definition referenceAI systems that plan and carry out a series of steps toward a goal, often by using other software, rather than giving one answer.
Definition referenceUsing a trained model to produce a prediction or response to new input.
Definition referenceFigure develops humanoid robots and publishes its own tests of how they perform physical tasks.
Explore FigureUsing a trained model to produce a prediction or response to new input.
Definition reference