Tendrils Raises $30 Million to Build a CPU Around Automatic Parallelism

Tendrils also announced a board seat for Jack Altman, a Benchmark general partner and founder of HR software company Lattice. Benchmark has prior chip exposure: an October 1 report on the financing notes its participation in AI chipmaker Cerebras Systems' Series A in 2016.

Participants named by Tendrils include the UK Sovereign AI Fund. That brings a state investor into the round: the government's Sovereign AI initiative backs domestic AI companies through investment and support including access to supercomputers. The announcement does not disclose individual investors' contributions.

The October report, citing people familiar with the matter, had already identified Benchmark as the lead and said Tendrils was discussing a rapid follow-up raise that could value it above $1 billion. That was a possible future financing, not a completed deal or a disclosed valuation for this $30 million round. Benchmark declined to comment for that report. The distinction matters in a sector the report describes as capital-intensive, cyclical and subject to long development timelines.

Cloudberry VC’s company-reported architecture comparison depicts Tendrils’ proposed local-memory cores. Schematics, not chip floorplans or measured performance. Source: Cloudberry VC.

What the new architecture is meant to change

Interaction nets represent computation and data as a graph. Local rules rewrite connected nodes, and independent rewrites can proceed concurrently without changing the answer when their execution order changes. The participating investor's explanation dates the model to Yves Lafont's work in 1990; the underlying theory predates this financing by decades.

The proposed hardware advantage goes beyond finding parallel work. In his January 18, 2025 technical essay, Cremer argued that keeping data near the cores using it, then distributing growing computations to neighboring cores, could reduce data movement, latency and energy use. Conventional CPUs, he wrote, impose overhead through emulation and caches poorly suited to interaction nets' graph-based memory accesses. GPUs can face divergent execution paths and irregular memory accesses.

The early backer's account expects the design to dispense with machinery for coordinating shared data and discovering which instructions can run simultaneously, allowing simpler cores to scale into the thousands. Tendrils' own announcement likewise targets thousands of cores and says this requires moving beyond shared mutable memory and sequential instruction streams. These are design goals and architectural claims, rather than disclosed measurements from a finished chip.

Vine is the software side of that change. The investors describe an open-source language resembling Rust that compiles programs into interaction nets. Familiar syntax is meant to ease adoption; it does not establish that existing applications will run unchanged.

The participating investor's illustration uses an example it attributes to Tendrils: a nested-loop algorithm operating on lists of 200 items could execute an average of 274 small steps simultaneously on a hypothetical machine with unlimited cores. With lists ten times longer, the figure rises to 2,739. Those numbers describe available parallel work under that assumption, rather than observed speedups, energy savings or comparisons against a CPU or GPU.

Company-reported example, described by Cloudberry VC: average simultaneous interactions rise from 274 to 2,739 with larger lists under an unlimited-core assumption. These are not measured hardware speedups. Source: Cloudberry VC.

The competition includes existing parallel software

Cremer's essay identifies HVM, an interaction-net implementation developed by Victor Taelin and his team, and Bend, a language compiled to it, as existing efforts to expose parallelism automatically. HVM had also run on GPUs. Tendrils' proposed distinction is therefore hardware designed around the model, rather than the invention of automatic parallelism through interaction nets.

Cremer also acknowledged improving conventional programming approaches, including futures, channels and coroutines, which he expected to become almost as convenient to use. He cautioned that interaction-net programmers would still need to consider parallelism, for example to avoid unnecessary copying of values. The pitch depends on execution efficiency as well as programmer convenience.

The competitive baseline is demanding. Cremer warned that conventional architectures have benefited from decades of optimization by tens of thousands of engineers with billions in funding. The investment case points to earlier parallel-computing projects—dataflow machines, the Connection Machine and Transputers—that lost to improving conventional chips. Its investment case assumes that changes in chip design and distribution can make a new attempt more practical.

The next test is a workload customers will pay for

Cremer explicitly distinguished faster interaction-net execution from solving actual problems faster. His proposed entry point was small, self-contained applications with irregular parallelism, including optimization, sparse graph processing and simulation. That is a narrower initial opportunity than superiority across general-purpose computing.

The participating investor's commercial argument is similarly narrow: selling compute through an API can reduce the need to persuade customers to install unfamiliar chips directly. It identifies Tendrils as a silicon partner of Callosum, a company developing software to route parts of AI workloads to suitable processors. The UK government also named Callosum as the first equity investment of its Sovereign AI initiative in April. A processor could gain use through such an orchestrator by being best at one part of a workload; the investment account does not establish revenue or a production deployment for Tendrils.

Software migration and manufacturing remain conditions of that route. That account expects Vine's resemblance to Rust and AI-assisted porting to help, while acknowledging that a new chip ecosystem gives AI tools less existing code to learn from. It also argues that AI can reduce design effort, while noting that manufacturing masks still cost money. Neither argument supplies Tendrils-specific production economics.

The funding announcements do not disclose a fabrication partner, manufacturing process, tape-out schedule, customer pricing or independently tested hardware benchmarks. To substantiate the commercial case, Tendrils needs to show a useful application finishing faster or at lower cost than on available processors, with migration and system overhead included. That evidence would distinguish an architecture that exposes parallel work from a product customers have reason to buy.

Railway says stolen credentials fueled about $740,000 in fraudulent AI usage

Railway estimates stolen customer keys generated about $740,000 in unauthorized AI spending after an October intrusion. Secrets from up to about 100,000 workspaces could have been decrypted, and the company says it will cover losses providers decline to refund.