Sparxity raises $90m for ultra-sparse AI chip development

By Grace Kim

Sparxity has raised $90 million in a Series A funding round to develop custom chips for ultra-sparse artificial-intelligence workloads, the German compute company said.

The round was led by ICM HPQC Fund. Lockheed Martin Ventures, GlobalFoundries, GS Futures, Cloudberry Ventures, TGFS and the EIC (European Innovation Council) Fund also participated.

The capital will support chips built around “fine-grained dynamic sparsity,” meaning that a system activates only the computation required for a given model. Sparxity says this approach, which it calls “compute only what matters,” reduces unnecessary computation and data movement, easing system bottlenecks.

The company targets up to 44× faster AI inference in ultra-sparse models. It also expects customers to use the technology to address rising AI infrastructure costs and energy consumption. The 44× figure is a development target; the announcement did not provide a test methodology, baseline model, power measurements or shipping schedule.

What does the funding change for AI operators?

The immediate change disclosed is engineering capacity: Sparxity will finance custom silicon and full-stack systems for ultra-sparse workloads. For operators, procurement terms, pricing and availability remain undisclosed. Whether the promised speed and energy gains change total cost per inference will depend on those commercial details and published benchmarks.

The release did not allocate proceeds among chip design, manufacturing, software or rack integration.

Sparxity was founded in 2021 under its former name, SpiNNcloud. It builds full-stack systems around the sparse-native architecture. The company says the design minimizes computation and data movement, reducing bottlenecks while improving energy efficiency.

Between 2024 and 2025, Sparxity said its revenue quadrupled. It also disclosed the following operating metrics:

• Chip tapeouts, the stage at which a chip design is released for fabrication: three.
• Chips produced under the company’s oversight: more than 70,000.
• Racks deployed: 12.
• Customers globally: more than 30, including some of the largest supercomputing centers.

The figures cover design releases, chip production and system installations. Sparxity did not break out recurring revenue, shipment volume or customer concentration, so the revenue increase cannot be tied to a disclosed installed base.

In October 2026, Sparxity announced a deployment of a sparse-native computing rack system at the Jülich Supercomputing Centre (JSC) in Germany. Last week, the company unveiled a partnership with Deutsche Telekom to bring its sparse-native technology into T-Systems’ data center infrastructure.

Neither announcement specified a rack count, contract value or deployment timetable. The JSC project and the T-Systems agreement therefore add named sites to the company’s disclosed footprint without quantifying the associated revenue.

“The next era of AI will be defined not just by how much computing power we can deliver, but by how intelligently and efficiently we use it,” said Sparxity co-founder and CEO Hector Gonzalez.

He called the round “a powerful validation” of Sparxity’s sparse-native vision and the complementary positioning of its systems. Gonzalez said investor backing would accelerate the transition from dense, energy-intensive architectures to heterogeneous data centers, meaning sites that combine different processor types. Dedicated silicon for ultra-sparse workloads is intended to provide greater AI performance at lower cost.

The round gives Sparxity capital to scale custom-silicon development after three tapeouts, more than 70,000 chips produced and 12 racks deployed. It did not disclose valuation, the proceeds breakdown, benchmark methodology, production pricing or shipment dates. Until those figures arrive, the measurable baseline is $90 million in new funding, more than 30 customers worldwide and a target of up to 44× faster inference—not a verified production result.