SpiNNcloud: Shaping the Future of Brain-Inspired AI Supercomputing

As AI models scale, so does their demand for energy and compute, putting pressure on conventional hardware architectures that were never designed for efficient inference at this magnitude.

Mainstream systems like GPUs deliver raw performance but struggle with the energy and efficiency demands of emerging workloads. Brain-inspired approaches such as neuromorphic computing offer an alternative: leveraging principles like massive parallelism, event-driven communication, and sparse activation to improve performance per watt. These characteristics are becoming essential for applications like generative AI, optimization, and neurosymbolic reasoning.

SpiNNcloud was founded in 2021 by Hector A. Gonzalez (CEO), Christian Eichhorn (CCO), Matthias Lohrmann (CTO), Christian Mayr, building upon technology architected by Steve Furber, the creator of the original ARM architecture. The company builds large-scale, brain-inspired supercomputers for energy-efficient AI inference, powered by its proprietary SpiNNaker2 architecture. 

What sets SpiNNcloud apart is its ability to scale neuromorphic computing beyond the chip level to full supercomputing infrastructure, delivering up to 26 times the energy efficiency of GPUs in real-world AI workloads. Its systems are already in use across Europe and the US, including a major deployment at Leipzig University simulating over 650 million neurons.

In July 2025, SpiNNcloud secured €10 million in blended funding from the EIC Accelerator as part of the GenAI4EU initiative.

We’ve followed SpiNNcloud since 2022, and this is our third interview with the team. Learn more about the future of brain-inspired AI supercomputing from our conversation with SpiNNcloud co-founder and CEO, Hector A. Gonzalez:

What Inspired You To Start SpiNNcloud?

I started my career in engineering roles across industry and academia, working on chip design for biomedical and signal processing applications. During my PhD at TU Dresden, I focused on chip architectures for radar and machine learning.

It was during that time that I got involved in the SpiNNaker project, which aimed to build brain-inspired computing systems. Over time, it became clear that this architecture had commercial potential, especially for energy-efficient AI.

Editor’s note: The SpiNNaker project (“Spiking Neural Network Architecture”) is an initiative from the University of Manchester, designed to simulate large-scale spiking neural networks in real time.

In 2021, three co-founders and I spun out SpiNNcloud to commercialize the technology. Our goal was to build specialized, energy-efficient supercomputers for AI. Since then, we’ve grown to a team of 55 and attracted talent from companies like ARM, AMD, Intel, and IQM. We’ve deployed our systems across Europe and the US, and recently secured support through the EIC Accelerator as part of the GenAI4EU challenge.

How Is SpiNNcloud Advancing Brain-Inspired Computing for Large-Scale AI Inference?

SpiNNcloud develops specialized microchips and supercomputers inspired by biological brains. Our systems are designed for ultra energy-efficient AI inference and are particularly well-suited for dynamically sparse algorithms, which are becoming increasingly crucial as generative AI scales.

Editor’s note: Dynamically sparse algorithms activate only the relevant neurons or layers needed for a given input. Unlike dense algorithms that use the full model every time, this reduces computation and improves energy efficiency in large AI systems.

The architecture is neuromorphic and supports event-based communication, energy-efficient low-power hybrid processors, and a highly parallel topology that scales up to supercomputer level. 

Editor’s note: In this article, the terms neuromorphic are used interchangeably. Strictly speaking, neuromorphic refers to hardware mimicking biological neurons, while brain-inspired can include looser architectural analogies like parallelism or event-based communication.

Editor’s note: Event-based communication is a brain-inspired method where components exchange information only when needed, rather than on a fixed clock cycle. This asynchronous design helps reduce idle energy consumption.

Editor’s note: Low-power hybrid processors combine different types of processing cores or architectures in a single chip to optimize energy use and performance for various workloads.

Editor’s note: A highly parallel topology refers to a system architecture in which many processing elements operate simultaneously. This structure enables large-scale computation while maintaining low energy use, especially for AI tasks.

We’ve built the world’s largest brain-inspired supercomputer to date, with five million cores across 34,000 chips and eight racks. Our systems are already in use at leading institutions in Europe and the US. One of our largest deployments, at Leipzig University, will be used for AI, high-performance computing, and drug discovery, simulating over 650 million neurons.  The hardware is powered by our SpiNNaker2 chips, which are based on a research project co-led by Christian Mayr at TU Dresden and Steve Furber, the creator of the original ARM architecture, at the University of Manchester.

Editor’s note: The ARM architecture is a widely used microprocessor design known for its energy efficiency. It powers over 95% of smartphones worldwide and is also common in tablets, IoT devices, and embedded systems.

How Do You Define Neuromorphic Computing in Your Approach?

Neuromorphic hardware exists on a spectrum. I like to emphasize this because when people hear the word ‘neuromorphic,’ they often assume we’re just doing spiking neural networks, as if that were the only computational primitive we support. That’s not the case.

Editor’s note: Spiking neural networks are a biologically inspired type of AI model where neurons communicate via discrete spikes. They aim to mimic the brain’s timing-based communication but are more complex to implement.

Neuromorphic systems vary in how closely they stick to biological inspiration. Our architecture has a looser commitment to biological realism. It supports extreme parallelism, event-based communication, and programmable dynamics, but that doesn’t necessarily mean spiking. For example, we can run conventional algorithms, but restructure them to leverage the properties of the hardware.

Editor’s note: Programmable dynamics refers to the ability to configure how components in a computing system behave over time, including how they process information or respond to inputs. This flexibility supports a wide range of algorithms.

So in SpiNNcloud’s case, neuromorphic means brain-inspired, but it’s not limited to spikes. We can implement a range of algorithm types.

What Are the Key Architectural Innovations Behind Your SpiNNaker2 and SpiNNext Products?

Our first product is SpiNNaker2, which is commercially available today. It’s a general-purpose neuromorphic microchip designed for a wide range of applications, including high-performance computing, optimization, drug discovery, and scaling foundational AI models.

The second product is SpiNNext, a spin-off of SpiNNaker2 that’s specialized for a specific class of AI workloads, in particular, dynamically sparse foundational models. These include sparse transformers and mixture-of-experts layers, which only activate parts of the model during inference to reduce computational cost. SpiNNext integrates strategic accelerators to support these workloads efficiently.

Editor’s note: Transformers are a type of AI model designed to process and understand sequences of data, like sentences. They are good at capturing context and meaning, which makes them especially powerful for tasks in language and generative AI.

Editor’s note: Mixture-of-Experts (MoE) models use many small specialized sub-networks, called “experts,” but only activate the most relevant ones for each input. This reduces the amount of computation needed and helps large models scale more efficiently.

Editor’s note: Strategic accelerators are specialized chips or hardware units designed to speed up specific computational tasks, such as matrix multiplication or sparse data processing. They help improve system performance and energy efficiency.

Both architectures are highly parallel and follow a brain-inspired, event-driven design that selectively activates only parts of the system as needed. This sparse, asynchronous approach enables better energy efficiency compared to traditional dense GPU or TPU architectures. These architectural principles are encoded not just at the chip level but across our large-scale supercomputing systems.

Editor’s note: TPUs (Tensor Processing Units) are specialized AI chips developed by Google to accelerate machine learning workloads. They are commonly used in large-scale AI systems to train and run neural networks efficiently.

What Performance Benchmarks Have You Achieved, and How Does Your System Compare to Conventional Hardware Like GPUs?

SpiNNaker2 is between 18 and 26 times more energy efficient than GPUs, depending on the application. In drug discovery workloads, it significantly reduces the energy required to run simulations and identify promising molecular matches. It also supports the analysis of patient profiles that are computationally designed to guide the development of new drugs. The architecture allows these models to converge quickly, meaning they reach high-quality results in less time and with a lower energy footprint. In optimization problems, the system has demonstrated up to 50 times faster execution times. 

Editor’s note: In machine learning, converge means that a model has learned an optimal or stable solution during training or inference, typically measured by reaching a low error rate or high accuracy.

In neurosymbolic models, which combine neural networks with symbolic reasoning, we have observed between 10 and 100 times greater cognitive capabilities. These models are seen as more robust than purely statistical approaches like GPT-style models. Our architecture enables them to scale more effectively, supporting more complex reasoning and larger model structures while maintaining energy efficiency.

Editor’s note: Symbolic reasoning is an approach where AI uses rules and symbols to solve problems, similar to how humans use logic. Unlike statistical methods that learn from patterns in data, symbolic reasoning follows clear, step-by-step rules that are easier to understand.

Editor’s note: Statistical approaches in AI, like deep neural networks, solve problems by learning from patterns in data. Instead of following fixed rules, they make predictions based on probabilities and examples, unlike symbolic reasoning, which uses predefined logic.

Looking ahead, with our next-generation product, SpiNNext, we estimate energy efficiency gains of over 80 times compared to GPUs. This product is not yet on the market, but architectural simulations indicate that it will be even more disruptive than our current offering.

What Are Your Main Use Cases?

As a hardware company, we’re not tied to a single use case. Our role is to provide infrastructure that can enable a wide range of applications. It’s hard to say that our architecture is only meant for drug discovery or optimization.

That said, these are two areas where we’ve demonstrated significant improvements over the state of the art. But rather than defining ourselves by those use cases, I’d describe our offering as a more energy-efficient way of deploying and generating workloads across different domains.

What Are the Main Technical Bottlenecks in Designing Neuromorphic Hardware?

One of the main technical challenges is ensuring interoperability. That means allowing models developed for other architectures to be ported easily to our hardware, while still leveraging their unique characteristics. To do that, you need robust software tools to convert and sparsify models, and to support a wide range of operators. Keeping that operator set up to date so it supports state-of-the-art models is an ongoing effort and a common challenge for any hardware company.

Editor’s note: In AI, operators are the basic functions that models use to process data, such as adding numbers, multiplying matrices, or applying activation functions. Hardware needs to support these functions to run modern AI models properly.

How Do You Differentiate Yourself From Other Neuromorphic Companies and Mainstream AI Hardware Platforms?

From a strict neuromorphic standpoint, we’re essentially the only company focused on designing supercomputers that take a strong inspiration from the brain. In our case, we’re building the world’s largest systems in this domain.  Many other companies in the neuromorphic domain are focused on building energy-efficient chips for edge devices or edge systems. 

Compared to mainstream AI platforms, which are typically based on conventional architectures like GPUs, our system is brain-inspired at every level. From the way a single microchip operates to the full-scale supercomputer, our architecture reflects principles such as parallelism, event-based processing, and energy efficiency.

In that sense, we’re positioned somewhere between the neuromorphic field and the mainstream AI domain.

What Strategic Milestones Have You Reached, and What Are You Working Toward Next?

We’re already selling infrastructure and have been profitable for the past two years, which isn’t easy for a startup. We’re selling on-premise infrastructure, and those sales are already in the double-digit millions.

These sales come from physical installations, but we’re also working toward enabling compute access. That is one of the strategic milestones we’re focusing on, to provide broader access through that model.

What Advice Would You Give to Fellow Deep Tech Founders?

The business component is just as important as the technical one. Deep tech companies often focus heavily on developing their technology, which is understandable, but it’s also crucial to surround yourself with the right business people early on; people who understand the market and can offer strategic advice.

As a founder with a scientific background, it’s tempting to keep refining the product endlessly. There’s always something to improve. But that can turn into an endless cycle. You need to channel that energy into clear, market-driven conversations and make sure you’re building a commercially viable company, not just an R&D project inside a lab. That’s the difference between a successful company and one that stays stuck in development.