Tendrils, a Cambridge, U.K. startup designing a new kind of CPU, raised $30M in a round led by Benchmark, and Benchmark general partner Jack Altman joined its board. The company says today's CPUs are held back by compatibility with single-core designs, which forces complexity such as cache coherency protocols and reorder buffers. Tendrils is instead building a processor with thousands of cores around interaction nets, a model of computation in which a program is a graph that rewrites itself in many places at once, so work spreads across cores without shared mutable memory or a single instruction stream. It also develops Vine, an open-source programming language built on interaction nets, and has not disclosed a tape-out date (Tendrils)
Private Companies
Fractilefundraise
Fractile raised a $720M Series C led by Thrive Capital, Founders Fund, Lightspeed and Redpoint, at a $7B valuation according to a person involved in the deal. CapitalG, BlackRock, DST Global, the UK Sovereign AI fund and the British Business Bank also invested. The company stacks DRAM directly on top of logic, aiming for more memory bandwidth than HBM and more capacity than SRAM. It plans to launch its first commercial chip in 2027 (Reuters)
Togethercustomer
Together AI will run inference on Nvidia GB300 NVL72 systems operated by Rackspace, under a multiyear agreement with an anticipated total contract value of $380M. Rackspace supplies the GPUs, power, liquid cooling and networking, and Together AI runs its inference platform on top. Deployment is planned for 2027, subject to Rackspace financing the GPU purchase. In August, Together AI signed a $240M deal for an Nvidia HGX B300 cluster on IBM Cloud (Rackspace press release)
Sparxityfundraise
Sparxity, the Dresden chipmaker formerly known as SpiNNcloud, closed a Series A led by ICM HPQC Fund that brings its total funding to about $90M. The round size was not disclosed. Lockheed Martin Ventures, GlobalFoundries and the EIC Fund also invested. Sparxity builds chips for sparse AI inference, where most weights or activations are zero, and targets up to 44x faster inference on ultra-sparse models. It reports more than 30 customers, including the Jülich Supercomputing Centre (DCD)
Staera Semiconductorfundraise
Staera Semiconductor, formerly New Silicon Corporation, emerged from stealth with $4.6M in pre-seed funding led by Openspace Capital, with SGInnovate among the other investors. Its Lightwire product builds gallium nitride LEDs and waveguides into the bridge or interposer chip that links processors and memory, using standard foundry steps. Staera targets under 0.2 pJ/bit, against the roughly 5 pJ/bit it cites for co-packaged optics, and reach beyond 50 mm. Commercial release is targeted for 2030 (Staera press release)
Public Markets
CoreWeave reported after the close that third-quarter revenue rose 168% to $3.66B, and backlog reached $132.5B, up from $104B in June. It raised 2026 revenue guidance to $13.0B–$13.3B and expects $37B–$39B of capex (CNBC)
TSMC's board approved $40.9B of capital appropriations and up to $500M to set up a company for strategic semiconductor investments. It also cleared up to $500M each for OpenAI preferred stock and Anthropic shares at the IPO price (TSMC filing)
Emerging
Real-time quantum decoding: A paper from Daniel Barter and 14 co-authors, Machine-learning decoders for fusion-based quantum computing in real time, introduces PRISM, a transformer-based decoder for photonic quantum computers, where photon loss and failed fusion measurements erase qubits at known locations. Across surface, color and quantum LDPC codes up to distance 24, it matches the logical error rates of slower integer-programming decoders. Implemented on an AMD VP1902 FPGA, it decodes a distance-8 logical block with 6.59 microseconds of latency and 1.56 microseconds amortized per block (arXiv)