Sparxity closes Series A for chips that skip the parts of AI models that aren't working

Sparxity, the Dresden sparse-computing company formerly known as SpiNNcloud, said on Nov. 10, 2026, that it closed a Series A led by ICM HPQC Fund, bringing cumulative funding to $90 million, to build chips for fine-grained dynamic sparsity that it says could make inference up to 44 times faster on ultra-sparse models.

Most AI hardware assumes the model lights up every neuron every time; plenty of modern models don't. Sparxity is betting a chip built to skip the idle parts beats brute force on cost and power. "Up to 44x" is a target, not a benchmark — but 70,000 chips, Jülich and T-Systems say this is more than a slide deck. Note the one number the release skips: how big the Series A actually was.

Sparxity, a Dresden company that builds chips for sparse computing, said it closed a Series A led by ICM HPQC Fund. A Series A is an early private funding round, usually the first sizable one after a product is already in the world. The team has now secured $90 million in cumulative funding. The announcement does not say how large the Series A check itself was, and it does not give a valuation. The release is dated Nov. 10, 2026. Those lines are Sparxity's, carried on Business Wire.

Investors named with the lead are Lockheed Martin Ventures, GlobalFoundries, GS Futures, Cloudberry Ventures, TGFS, and the EIC Fund. Angel investors include Richard Socher of Recursive Superintelligence and Nancy Xu of Moonhub, which Salesforce acquired. Atlantic and Silicon Catalyst continue to back the company. Those names are Sparxity's.

The money is for specialized chips aimed at fine-grained dynamic sparsity. A neural network is a stack of many small calculations. Sparse, here, means a lot of those calculations are idle on a given step. Fine-grained and dynamic means the chip is supposed to notice which pieces are idle while the work is running, and skip them, instead of skipping a whole block a designer turned off in advance. Sparxity's phrase is "compute only what matters." The target is up to 44 times faster inference on ultra-sparse models, with better energy efficiency. Inference is running a trained model to get an answer. That 44-times figure is a target for the next chips. The announcement does not present it as a measurement on silicon that is already shipping. Those lines are the company's.

Sparxity says revenue quadrupled from 2024 to 2025. It counts three tapeouts, which means a chip design was sent to a factory to be made three times, more than 70,000 chips produced, 12 racks deployed, and more than 30 customers, including some of the largest supercomputing facilities. Those figures are the company's.

On selected workloads, Sparxity says its current systems are 10 times faster at sparse solvers, use 18 times less energy for event-based machine learning, and finish optimization problems in a twentieth of the time. A solver is a program that works a math problem through to an answer. Event-based learning updates a model when something happens, instead of on a fixed clock. Those multiples are what the company reports for systems it already has, on the workloads it chose. They are not the 44-times target.

A deployment at the Jülich Supercomputing Centre is described as multi-million-dollar. Jülich is home to Europe's first exascale supercomputer, a machine in the class that can do a billion billion calculations a second. Sparxity also has a partnership with Deutsche Telekom to bring the technology into T-Systems data centers. Ferri Abolhassan, chief executive of T-Systems, is quoted: "We are looking forward to having Sparxity on board in our AI factory soon." Those lines are in the announcement.

The company was founded in 2021 as SpiNNcloud, a spin-off of TU Dresden built on SpiNNaker2, a chip inspired by how the brain works. It renamed itself Sparxity last month. Those lines are the company's.

Hector Gonzalez, co-founder and chief executive, said the next era of AI will be defined not just by how much computing power can be delivered, but by how intelligently and efficiently it is used. Matthew Gould of ICM HPQC said Sparxity has shown it can manufacture the technology and deploy systems with major customers. Xu Owen He, a former Google DeepMind researcher and a strategic adviser, is quoted on PEER. He said that work showed dynamic sparsity turning on only a fraction of the compute across millions of fine-grained experts, and that today's hardware limits how far that idea can go. Those quotations are theirs, as the release prints them.

The picture is a Sparxity-branded chip package on a light gray surface. The credit is Sparxity.

In plain terms, a Dresden chip company formerly called SpiNNcloud closed a Series A and says the team has now raised $90 million in total. The size of this round is not in the release. The bet is silicon that skips the idle parts of a sparse model. Up to 44 times faster is a target for chips still being built. What the company says is already in the world is 70,000 chips, a deployment at Jülich, and a path into T-Systems data centers.