SpiNNcloud and Its EIC Accelerator Brain Chip for Sparse AI

Most of a large language model sits idle for any given word it produces, yet the graphics processors that run these models are built to compute everything at once. SpiNNcloud Systems, a 2021 spin-off of TU Dresden, wants a chip that only works where the model needs it. Its EIC Accelerator project, SpiNNext, is to build that chip on top of SpiNNaker2, the brain-inspired processor its founders developed in the EU's Human Brain Project.

The EIC Accelerator selected SpiNNcloud at its March 2025 cut-off under the GenAI4EU challenge for generative AI. The EIC's own company page and the Dresden accelerator SpinLab both put the package at €10 million of blended finance, a grant plus an equity investment. CORDIS records the grant part: €2,499,997.50, signed on 9 November 2025, for December 2025 to November 2027. No amount or date for the equity part has been published.

How SpiNNaker2 differs from a GPU

A SpiNNaker2 chip is a grid of small, low-power processors rather than a few large ones. eeNews Europe reported in May 2024 that each chip carries 152 Arm M4F microcontroller cores plus neural network accelerators, is made on GlobalFoundries' 22nm FDX process and runs at 0.45 to 0.6 volts depending on load. Sandia's computing research centre gives 153 cores per chip; the other sources give 152. The design lets cores fire only when an event arrives, which is how spiking neural networks and the brain itself work, and it can also run conventional deep learning.

The efficiency case rests on company figures. Hector Gonzalez, the co-founder and chief executive, told BeBeez in July 2025 that "our systems are 18 times more energy efficient than current GPUs", and the German news agency dpa, reported by apotheke adhoc, repeated the factor of 18 against conventional high-performance computers. Igor's Lab wrote in August 2025 that SpiNNext was being advertised with a factor of 78, without saying against what. No independent benchmark of either figure was found.

Who has bought the systems

SpiNNcloud has hardware in the field, and the buyers have said so themselves. TU Dresden put a system of 35,000 SpiNNaker2 chips and more than five million cores into operation in April 2025, funded by the European Regional Development Fund and the Free State of Saxony, with SpiNNcloud's software stack developed under the EIC Transition project. Sandia National Laboratories unboxed its system, NERL Braunfels, on 3 April 2025, funded through the NNSA's Advanced Simulation and Computing programme, and puts it at 175 million neurons.

In July 2025 SpiNNcloud announced a system of about 4,320 chips and 656,640 cores for Leipzig University's drug discovery research, per BeBeez; dpa called the contract worth millions and said it would be implemented within two months. In November 2025 Data Center News reported a system able to simulate more than 393 million neurons delivered to the THOR initiative at the University of Texas at San Antonio, funded by the US National Science Foundation. Whether the Leipzig system went into service is not stated in any source found. The company's current website claims 3 tape-outs, more than 70,000 chips produced, more than 12 racks deployed and more than 30 customers, naming Sandia and the Jülich Supercomputing Centre; the other figures have no second source.

What SpiNNext promises

The CORDIS summary describes a mismatch: advanced AI algorithms compute sparsely and selectively, while conventional hardware processes everything densely. SpiNNext is to add native hardware support for dynamic sparsity to a chip derived from SpiNNaker2, aimed at generative AI and large language models, with lower energy use and European technological sovereignty as stated goals. The project started on 1 December 2025 and ends on 30 November 2027. CORDIS lists no total cost, no reports and no results, and the page was last updated on 18 November 2025; at this stage of a two-year project that is expected, not a gap.

1. Next-generation chip with hardware support for dynamic sparsityNot yet dueProject runs to November 2027; no tape-out of SpiNNext announced.
2. Generative AI and language model workloadsClaimWebsite charts speed gains of 30 and 44 times in tokens per second per user, without a stated baseline.
3. Lower AI energy useCompany figureFactor of 18 for SpiNNaker2 against GPUs, per the CEO; no independent benchmark found.
4. Platform the new chip builds onExistsSpiNNaker2 systems at TU Dresden, Sandia and UTSA, per the operators.

Grants, a small profit and a larger loss

SpiNNcloud's public money is easier to trace than its private money. CORDIS records two EIC grants of about €2.5 million each: the closed SPINNODE EIC Transition project, which ran from May 2023 to September 2025 to bring SpiNNaker2 from cloud systems to edge devices, and SpiNNext. North Data lists three smaller public grants besides: €516,313 from November 2022, €221,381 from January 2024, and €655,131 of ERDF/JTF technology funding from April 2026. These figures come from different records and are shown side by side, not added up.

The private side is thinner. The EIC page says the company secured its first venture investment in 2024, without an amount. The databases CB Insights and Caplight name Banyan Ventures and SpinLab among the investors, and the venture firm behind atlantic.vc lists SpiNNcloud in its portfolio, but no round with an amount and a named lead has been announced by the company. The register records increases in share capital to €32,864 in December 2024, €43,739 in September 2026 and €47,364 in October 2026; who subscribed the new shares, and whether the EIC Fund is among them, is not public.

The filed accounts, as reproduced by North Data, show a net loss of €249,333 for 2022, profits of €220,660 for 2023 and €138,621 for 2024, and a loss of €1,132,404 for 2025, published on 21 October 2026. Revenue and headcount from the same filings are not freely displayed, so how much of the earlier profits came from grants rather than sales cannot be read from the public record.

A new name on the website

The address spinncloud.com now redirects to sparxity.ai, which presents "hardware for sparse AI models" with processors, boards, software and multi-rack systems, and whose imprint names SpiNNcloud Systems GmbH, HRB 41467, with Hector Gonzalez and Matthias Lohrmann as managing directors. The legal entity is unchanged in the register. The company has published no announcement explaining the new name, and none could be found in a web search on 28 October 2026. The new site's pitch, hardware for sparse AI models, is the subject of the EIC project as CORDIS describes it.

Why the EU funds an AI chip from Dresden

The European Chips Act of 2023 sets out to strengthen Europe's semiconductor ecosystem and reduce its dependence on chips from elsewhere. SpiNNext was funded under GenAI4EU, a challenge for European generative AI, and its CORDIS summary names European technological sovereignty and lower AI energy demand as goals. SpiNNcloud designs its own AI processor in Dresden and sells it as complete systems; SpiNNaker2 is made on a GlobalFoundries process, per eeNews Europe, and where SpiNNext will be manufactured has not been announced.

The reporting is based on publicly available sources, listed below.