“Régis Hamelin”

 Good afternoon, everyone.

I’m Regis Hamelin, CTO of Blumorpho, and welcome to the Chip Think Tank. The Chip Think Tank is a series of webinars brought to you by the aCCCess project, which provides insights and reflections on the semiconductor industry from European leaders.

I’m Regis Hamelin, CTO of Blumorpho. I will be your host today, and I have the pleasure of welcoming Mathias Laumann,
co‑founder and CTO of SpinCloud Systems. Good afternoon, Matthias. Thank you for being with us. If you missed previous sessions, they are now available on our YouTube channel. You can scan the QR code to access them.

This webinar is recorded and will be made available after the presentation on this channel. For reference, the aCCCess Consortium is

composed of seven partners. You can see the seven partners on the left.
Blumorpho, the company I’m from, is the coordinator. On the right, you have all the associated partners of this coordination and support action. This CSA aims at coordinating the network of European Chips Competence Centres. There are in total 30 Chips Competence Centres in Europe. They have been selected by the European Commission and the Chips Joint Undertaking.

We have two Competence Centres in Spain and two in Belgium. The Competence Centre is a hub of excellence where companies can access expertise, business support, training, and skills development. A recent example of what we did is the Chips Venture Forum, which took place in November in Munich. It was organised in partnership with the EIC and DG Connect. There were around 170 pre‑scheduled one‑to‑one matchmaking meetings between startups and investors, with a lot of interaction.

These pre‑meetings helped us select 18 startups out of 58 participants. We also hosted panels and high‑level
discussions. Our objective was to facilitate and strengthen relationships between corporates, investors, and startups.

We received very positive feedback from this event, and we are very proud of the results. Matthias, you participated in this event,

and we will talk about that in a moment. This webinar is an interactive

discussion. You can enter your questions in the chat. We already received some questions. We would like to know a little bit more

about who is in the room today. Are you a startup, an investor, a Chips Competence Centre, a large company, or something else?

This helps us better understand who is participating in our event.

Thank you very much for your feedback.

We see that about one third of the participants are from startups and SMEs, around 20% are investors, and we also
have Competence Centres, large companies, and other profiles represented.

Thank you again.

Matthias, you participated in the Chips Venture Forum and were identified as one of the most critical reviewers of the
technologies presented during the competition. So maybe to get to know you a bit better, could you explain your background and where you are coming from? 

“Matthias Lohrmann”

Thank you. I will switch to the presentation about SpinCloud and then talk more about the other topics.

This presentation is an excerpt of the one we had at the Chips Venture Forum last November.

At SpinCloud, we are bringing the next generation of AI inference to the market as a complete solution for our customers.

When we talk about a complete solution, it starts with hardware. Over the last years, hardware demand and throughput requirements have doubled, especially since the deep learning revolution. The speed and complexity of models are growing faster than what traditional semiconductor scaling can support.

Models are becoming more performant and capable of complex tasks, especially with large language models.

On one hand, we need large models for complex tasks. On the other hand, you often only need a fraction of the computation to get the best result.

Today, everything is driven by GPUs, both for training and inference. We see inference growing faster and faster.

For sparse models used in inference, there is a missing hardware link. This is where we come in with SpinnEx, our next‑generation silicon designed to revolutionise the execution of sparse models.

Our background as a company, and mine personally, comes from SpiNNaker 2. We already have a product on the market that has been shipped to European and American companies and research institutions for HPC workloads.

The core feature of SpiNNaker 2 is scalability. You can scale to very large systems and work on very large datasets. The system is sparse and inspired by brain‑like computation models.

With SpinnEx, we target large language models and foundational models, with a release planned for 2028. SpiNNaker 2 is already commercially available and helps customers grow their applications.

SpiNNaker 2 was designed with the University of Manchester and TU Dresden. Seventy thousand units were produced in the first batch. Last year, we generated around six million euros in revenue from this hardware product.

At TU Dresden, we have built the world’s largest brain‑inspired supercomputer, on which we deploy larger applications.

Our team includes experienced people from deep tech, scale‑ups, and corporates. The challenge is not only hardware or software, but also algorithms and execution.

We started in 2021 and are now around 65 experts, mainly in Dresden and Munich, but also in London, France, and New Zealand.

That is a brief overview of SpinCloud.

Thank you.

“Régis Hamelin”

My first question concerns the inception of the company. You were part of the scientific staff at TU Dresden. What was the moment when you decided to create a company?

“Matthias Lohrmann”

I would say it was longer than a moment.
I started 2015 and another startup and the founding team I supported them.
I switched to a corporate job working with Global Foundries than I was raised six years supporting international customers as a design consultant but I always was looking for what can I do different, where can I contribute, where can I push the boundaries and then in 2020 I had a very good conversation with Professor Meyer who is one of our founding fathers of the company and where I was then also later employed at the chair and we were for one year working on business case, but during really the holidays of 2020 to 2021, we came to a very clear view on where can we push something, where can we generate value for our customer.
And then two months later, basically, SpinnCloud was founded and the founding team was together. We started as four at the moment, operationally, and the company are two sector.

“Régis Hamelin”

And you showed us a range of products.
You showed us some chips, we saw also some balls and we saw also a supercomputer.
So which one is your product? Is it, are you currently testing the market or have you already chosen which one you want to sell?

“Matthias Lohrman”

That’s a good question.
The chip is basically the core component of our product.
The product we are selling is server nodes, It’s really something that is in a data center sitting next to your CPU server or GPU server to accelerate the computation, and the board is the unit in which you can scale up.
You can think of a single board as comparable to a state-of-the-art GPU, and with the number of these boards that you slide into our own notes, you scale the system capacity.
The maximum you can build with this technology is a 10 million compute core machine based those compute boards that you saw and one board has 7 ,000 cores the user can use.

“Régis Hamelin”

And you designed this whole supercomputer yourself ?

“Matthias Lohrmann”

So there were three of the main contributors but overall roughly 100 engineers were working on the design, but then there was also a lot of effort in how to make it available in the data center, how to integrate it with the network, onsite, cooling, and just need a lot of different expertise.
The benefit was that we had this network here, especially with the clusters we are in, with Manchester, with Dresden, but also colleagues in Munich and other places, and this made this overall possible at all.
But this is definitely one of the bigger challenges, especially because we started out university project we needed to have access to all of this know-how which is not a single university chair.

“Régis Hamelin”

 Sure and and you grow into the value chain making more and more value and maybe someone will purchase this demo this this can be this supercomputer is a kind of demonstration for you today but maybe it can become it could become a product and you could become a supercomputer or server supply.
Okay so when we were looking for the title of this webinar we were very quickly speaking about the evolution of the market and today we believe that in the discussion we had we believe that we are some kind of tipping point of the training and that the AI market is turning or in any case the AI market has to turn from training to inference because now we need to move to the production era.
So from your knowledge of the business what does support this kind of evolution at this stage?

“Matthias Lohrmann”

From a market side what we saw clearly was already some years ago a shift and the focus of the when was for training of AI models and only a fraction was for the real inference and this is now shifting more and more especially since we see these huge models with a few hundred billion of parameters that need this compute power and this is on the numeric side, a clear shift as inference market is growing, inference system market is growing tremendously but we see also a lot of traction with other companies and startups especially in the last week, So shortly before Christmas, an inference-only AI and cloud provider that had a very big deal was GROK, a license to NVIDIA that’s very close to a full sale.
And just recently Cerebras announced the partnership or the buying of hardware in larger quantities from OpenAI and Cerebras does training.
they also do inference and especially high speed inference for a chatbot application.

“Régis Hamelin”
In practice, since they have acquired the assets as well, this is the equivalent of a hybrid model, but the equivalent of an acquisition for $20 billion, which is a significant amount of money.
And indeed, Cerebras is also in the news.
And it happens at Cerebras, there are some connections between the founder of OpenAI and Cerebras, they are shareholders.
So we are still in this kind of circular, not economy, but circular financial model where people are placing orders to the others, we see the stock or the value increasing.
So that’s at the moment more financial than really business-driven.
Do you see really the evolution of inference more than these signs? Do you see really the applications taking off in AI?

“Matthias Lohrmann”

 We see this big use of large language models in the last year, and I think there’s lot of things ongoing right now in the industry. Where can I make use of it?
I give this new tool to the market and the end user has to find the application and establish it.
And there are complete flows where on the backside it’s running already in LLM.
Maybe it’s not advertised as AI enabled but it is not a value to the customer directly by using AI but the value that the tool based on an LLM provides.
So we do not see all of the applications but Everyone is aware, even as a developer, for example, to accelerate software development or to review documents, there are these applications, and they are quite big.
At the moment, we see also the number of data centers growing massively.
There’s also limitations in the speed, not only from the amount of hardware that can be installed, but also from other limitations like the electrical power being available.
This is a big sign that there is a high demand that enables also these new providers to gain a foot in the market.
Otherwise, the high-oscalers especially would dominate, but there is so much space for new players because of the steep growth we see today.

“Régis Hamelin”

 You were talking about AI infrastructure, so what is the bottom?
like you mentioned, energy, but when it comes to inference, is it the bottom exit performance, energy, cost, scalability as well?
Where do you need to put your value proposition?

“Matthias Lohrmann”

 For us at SpiNNcloud, it’s really the performance. We need more performant hardware.
This is something we can make use of from the models. The models already, they allow us to get more performance out of hardware, but I need to build a hardware with that in the back of my mind. If I build acceleration purely for dense models, I cannot make use of these especially large models that only 10 percent, 5 percent of the compute power, and this is where we come in.
This is where we can accelerate the computation, and we can also reduce the anti-energy bill.
But also, from a hardware perspective, the utilization goes up.
If I can live with a tenth of the compute power, I can get more out of it if you want.

“Régis Hamelin”

Of course and today, we are into what we call a GPU-centric model. It’s GPU everywhere for training and inference. So what do you think? Is it sustainable in long term for the workloads?
What’s your opinion on that?

“Matthias Lohrmann“

We will see more differentiation. Large training models increasingly benefit from approaches such as mixture of experts, which is one of the most promising ways to introduce sparsity today. This sparsity is not yet reflected in current hardware, which creates room for complementary infrastructure. This is exactly how we position ourselves in the market.

“Regis Hamelin”
You mentioned dynamic sparsity. Is this really your key enabling technology? And when it comes to product development, how do you balance chip design and software? How important is software for your business?

“Matthias Lohrmann“

It goes beyond a simple split. We have a very strong algorithmic team at SpinCloud. We take existing models and further sparsify them, which is a key part of our approach.

“Regis Hamelin”
Yes.

“Matthias Lohrmann“

What you see here is the sparsity and granularity of today’s models. Very dense models run extremely well on GPUs, as they are designed for that. But there are also models with a very high level of sparsity, sometimes up to 90%, especially mixture-of-experts models, which can be made even more sparse.

This granularity determines how many experts are activated and how efficiently the model can be broken down. GPUs do not perform well in these scenarios. This is where our algorithmic team enables customers to further exploit sparsity.

Beyond algorithms, we have software teams mapping models to hardware, hardware teams designing the chips, and infrastructure teams building large-scale systems. These four groups together make it possible to run models with hundreds of billions of parameters efficiently in data centers.

“Regis Hamelin”
Do you see your technology as complementary to GPUs, or as an alternative for specific use cases?

“Matthias Lohrmann“

We clearly see it as complementary. We do not address training and continue to rely on GPUs for retraining models. GPUs are also very efficient for dense inference.

Our value comes from large models with a high degree of sparsity, where we provide a different and complementary solution.

“Regis Hamelin”
Looking ahead, do you expect larger models or more specialized ones, for example for software development?

“Matthias Lohrmann“

It mainly comes down to the number of parameters. Larger models are more capable, but they are also very expensive to run in inference.

What we do is take large models and specialize and sparsify them for specific applications. This reduces parameters and compute while preserving performance. This requires a distillation stage, which you can think of as a model compiler. This approach allows us to keep the capabilities of large models at a much lower cost.

“Regis Hamelin”
Is there a specific type of inference where SpinCloud outperforms others?

“Matthias Lohrmann“

Large language models with very sparse mixture-of-experts architectures, where only a small fraction of experts is activated. Dynamic sparsity, which we already implemented in SpiNNaker 2, is a key strength in these use cases.

“Regis Hamelin”

So I make a pause here. I’ve seen that we have a question in the chatbot. So, it’s a question about the neuromorphic-based hardware. The question is, can it be used by private users as well? Is it possible to use neuromorphic computing for blockchain-based decentralized AI training? That’s very specific. The question is from Ilya Noshkov. Indeed, that is a specific question.

 “Matthias Lohrmann“

At the moment, not. At the moment, we have only selected larger projects, but later this year, we will have opening spinnaker technology, the neuromorphic technology that we’re selling already to a wider audience. So Q3, Q4, we could talk here also about applications, yes. 

“Regis Hamelin”

 Okay. So that’s opening a new market, so more developer-oriented markets. So, while we were talking, there’s another question that came up. So, it’s from Mattia Samiello. Sparsity can be translated in algebraic sparsity. How does spin Hacker perform in terms of mathematical precisions, finite element simulation? That’s maybe talks more to you than me. 

“Matthias Lohrmann” 

Yes, this is a good question. There is a lot of activity actually around the FEM implementation and we will see more at the end of this quarter around this. But there are also a lot of interesting publications around a neuromorphic way to implement finite element methods. And with Spinnaker we have hardware today that is scalable to really go into ranges where an FEM simulation is required. So it’s not only about the small models, but we can provide the Spinnaker tools, really large models and run an FEM application on this hardware. 

“Regis Hamelin”

 One thing to add here is also the performance that comes, of course, out of it. It was in one of the slides that we had here, the immense speed up also compared to CPU solutions because of the sparse nature of neuromorphic hardware. Another question from Stevie Cordette, are you using only back propagation for the inference or an alternative approach? 

“Matthias Lohrmann“

For the training of the models, we also use back propagation. And so the training algorithm for our Spindex inference systems is not touched, but the distillation pipeline touches parts of your model and adds a different cost model on top of it. But the back propagation is still the key. This is also why we can rely on GPU solutions that are available today.

 “Regis Hamelin”

And we have lots of questions. Thank you to the audience for sharing all these questions. Another question from Inessa Seifert. Inessa is an AI expert as well. How do you test your hardware? Is there a testing center with generic benchmarks? I guess this relates now to Spinnaker 2. 

“Matthias Lohrmann“

 For Spinnaker 2, we work a lot with external partners to implement actually the applications and let them do the benchmarking and evaluation of the system. We help them, of course, in the implementation of the algorithms. For For example, especially neuromorphic applications, optimization solvers, the FEM, but also others. But here, there’s no single benchmark available yet. There was activities in the last two years around benchmarking such systems, but we haven’t seen yet this taking off as for MLPerf, for example, for AI applications, which is quite established.

 “Regis Hamelin”

Benchmarking, we talked about that. And while we were preparing this session, I received a question also from LinkedIn, someone who sent me a question for you, from Shekhar Hameen Singh Pardeshi, and he’s an automotive embedded software developer, and he was talking about the energy efficiency of AI. Do we have a community-established parameter to measure energy efficiency for an AI solution? So, and the next question, should these parameters be monitored or even regulated?

“Matthias Lohrmann“

Yes, this is a very important question. We think that this must be monitored. We really need to think about how do we make use of the energy for these applications. But especially regulation, for example, at this moment would be entering the further development of this field. But what we need is really to have a discussion around this. We need those benchmarks, and we need to compare, and this is something that the community has to also come up with. And it’s been one of the key features that also customers look at is the energy efficiency of the system. So there’s also a lot of pooling and enablement for our customers to do this on their own. 

 “Regis Hamelin”

Sorry. Well, what struck me when I saw your presentation is a 78X. It’s very precise. You don’t say 70 or 80 or roughly 90. You say 78. So, how do you came up with that number? So, which benchmark did you use? For this BINX benchmark, we use a llama model. and 70 billion parameter Lama2 admittedly here. And we compared it to a GPU implementation. This is also where the exact number is coming from.  Another question this time from Laurent Dugojon where the question is coming more from the hardware. Which silicon node and which foundry are you working on? 

“Matthias Lohrmann”

So for Spinnaker 2, it is public knowledge that it’s for a 22 FTX Global Foundry’s technology where we implement it and where we also produce the chips. The next generation, it will be a more advanced node, especially moving to higher speeds and Also, the memory wall that we are also facing, even with sparsity, this is also not a secret to go into detail, to go into that detail. But the memory wall is also affecting us.

 “Regis Hamelin”

So coming back to benchmarks, a question from Inessa Seifert: “On the community for benchmarking. How is the testing organized? Is there any specific organization which is in charge of these benchmarks?”

 “Matthias Lohrmann“

About these benchmarks, I assume now it’s the MLPerth benchmarks, and there’s a community behind this. It’s driven by the large providers of hardware in the field, and they define the benchmark and everyone can contribute. So a lot of players also publish in this MLPerth leaderboard, and that is driven by the community. 

 “Regis Hamelin”

Very good, the next question we had was about the apex of AI for inference, because actually the training is a one-time cost, but in the long term, inference will be the cost of AI. So, when it comes to your customers, do they have – are they aware of that first? Do they consider the OPEX as a criteria for purchase? And so, what is – when you are facing competitors, are you compared with regard to the raw performance or to all the efficiency? We’re talking about energy efficiency and benchmark. That’s why it’s important. the system architecture, so it’s more kind of scalability, I would say, or the total cost of an hardship. So what is really, I would say, the number one concern of your customers when they start talking to you? 

 “Matthias Lohrmann“

 The number one concern, I would say, is really the performance of a system that you get out. But they also knock out criteria. So if you cannot integrate into a data center, if you make this explicitly hard for a provider, It can also, let’s say, move back from this deal. The other thing is also with the power budget that is available because the data centers cannot scale as fast as they want, one limiting factor is energy. And then, of course, the question is how much electrical energy do you need? And on top of that is the question of cooling. Cooling gets more and more complex because the heat density grows and not every customer wants to go with liquid cooling, for example, and then this also becomes part of the discussion what can we do around it. And suddenly it’s more than just the hardware. It’s the overall solution to a customer that they look for. But of course the main metric that we are measured on is the tokens per second.

 “Regis Hamelin”

 Interesting. 

You said that the client has kind of an energy budget and cannot go beyond a certain level of energy. So, when we are reading the news, we have the impression that it’s working the other way around. We have the data center. We need to make it work, and we will take whatever amount of energy we need to make it work. And you seem to say that it’s working the other way around.

 “Matthias Lohrmann“

 Yes, there is a lot of activities around how do we make enough electrical energy available. Also, in the future, with renewable energies, there will be enough energy available. But what we have today in terms of grid, how do we distribute electrical power, how do we generate electrical power, this is one of the limiting factors for the growth of data centers today. Besides the fact that hardware is also not available in the numbers that are required. 

 “Regis Hamelin”

 So can we face a situation where data center cannot supply the computing power that they It is available in the hardware, but they don’t have the power to make it run, is it possible? 

 “Matthias Lohrmann“

Actually, yes. With the demand, we see a growth and the demand of AI hardware, and if we would extrapolate from where we are today, by 20, 40, 100 percent of the electrical power would be going into the data center. This is not a realistic future. That doesn’t make sense. It will not happen. For anyone listening who could be a little bit concerned, it will not happen. 

 “Regis Hamelin”

We have more questions from the audience. One is on software. Do you have plans to open source some models and where to collaborate and how to get hardware access? So it’s typically an that we needed to test for a use case, for example. So for Spinnaker 2, as the product is already available, you can give me a message or one of my colleagues, especially from product side or sales side to get access to Spinnaker 2 hardware. Or the future, sorry, what was the second part of the question? The question was more about, do you have open source models?

“Matthias Lohrmann“

All right, thank you. Yes, this is definitely part of our future releases. At the moment, we have not released the models open source, but we want to enable our customers to work with the hardware and part of using the hardware is providing the models and we don’t want to charge our customers for using the models we provide to them so they will go open source.

“Regis Hamelin”

And another question from Laurent Dubujon, what about your specific software plan offer?

“Matthias Lohrmann“

Software for us is a tool to enable our customer, and the software should be as open and transparent also as possible to our users. So there’s on the one side the distillation pipeline for Spindex that will be provided to our customers. They can compile the existing model and map it into our hardware. They will get a model in and a model out. And then the whole deck for deploying the model on the hardware business will also be released. For Spinnaker 2, we are also at the moment in the preparation for a major release of the software so everyone can plug in and have also the maximum of ownership of what they have. It’s not only the hardware. It’s also the software that they can change it, look inside, adjust it to needs if necessary as part of what we see for the craft community.

“Regis Hamelin”

 So let’s talk a little bit more about what it is today to be a European Fabless Startup. This is something that we have talked about during this CHIPS Venture Forum because one of the big ambition of this kind of forums is to support the startups to get their next round of funding to help them scale. And scaling large companies is something which is a little bit more difficult in Europe. So what is it like to be based in Europe? What does that enable? Where is it more difficult? Where is it easier than your competition elsewhere in the world?

“Matthias Lohrmann“

One of the biggest strengths we see as a European fabless startup is the depth of the talent pool available in Europe. Most of our colleagues are distributed rather than all being based in Dresden. We have team members in France, in the UK, and we have also relocated several colleagues from other countries, for example Italy, to our main office here.

This is where Europe really has a strong advantage: access to highly skilled talent, a very solid academic system, and a strong mobility of people. Developers and managers are willing to move, and many experienced professionals return to Europe after working elsewhere in the world, bringing back valuable know-how from earlier stages of their careers. Having access to this talent pool is a major asset for us.

Another important aspect is the ecosystem of initiatives available at different levels. We benefit from regional initiatives, national programs, and European-level initiatives. For a semiconductor startup, European initiatives are particularly important because of the scale of investment, the expertise involved, and the teams that need to be brought together.

This landscape has evolved significantly over the past few years and has become much stronger. We started to see this momentum around 2021, and even before that we were already closely following these developments. Today, this is the ecosystem we operate in and actively contribute to. In particular, the support from the EIC, through both the Transition and Accelerator grants, has been instrumental in helping us scale.

In the first project, we developed a substantial amount of software together with experts from across Europe. In the second phase, we are now working on SpinnEx with continued support.

“Regis Hamelin”
Excellent. You mentioned something very interesting earlier. You said that you managed to attract people coming from outside Europe to work in Europe. Would you say that Europe has a real attractiveness when it comes to talent?

“Matthias Lohrmann“

Definitely, yes. One important factor for us was the network effect that emerged once the company was officially registered in Germany. We see ourselves as a European company, and Germans are actually a minority within the team.

What we noticed is that when we entered a local network, for example in Northern Italy, our visibility increased very quickly through these existing ecosystems. Tunisia is another good example for us, with a very strong pool of talented people.

As soon as there is awareness of what you are doing, it becomes much easier to attract talent. This was a key learning for us as a startup: communicating what you build is not only important for customers, but also for future team members. This clearly gives you an advantage when it comes to hiring.

“Regis Hamelin”
There is a question in the chat about how you organized such a distributed team, for example with people based in Italy. What were the main steps and milestones in setting up this organization?

“Matthias Lohrmann“

We actually started during COVID, which forced us to learn very quickly how to work fully remotely. As the team grew, some of our earliest colleagues were already based in Munich, so we always had to find ways to make this setup work.

One key element is over-communication. We probably communicate more than we would in a fully office-based environment, but this helps ensure that information is not lost. We also organize regular on-site events, where colleagues from different locations come together in Munich or Dresden.

During these workshop weeks, we focus on a specific topic, work very closely together, reach the next milestone, and then everyone leaves aligned on priorities and next steps. Communication is essential, and the second important aspect is actively building and sharing a common company culture.

“Regis Hamelin”
Do you sometimes feel that people are far away, or do you still feel close to your teams? Is it mainly a question of perception?

“Matthias Lohrmann“

I would say we feel close. One reason is that we have established hybrid meetings as a standard. People do not need to be physically in the same room to be fully involved. Everyone is always included, everyone can join discussions, and the setup is very open. This really helps reduce the feeling of distance.

“Regis Hamelin”
That also helps since the time difference is minimal, usually around one hour. Coming back to being a fabless company in Europe, sovereignty is a topic often raised by the European Commission. How important is sovereignty for you, both in terms of suppliers and foundries, and how important is it for your customers?

“Matthias Lohrmann“

On our side, we have put a strong focus on building a supply chain that is largely based in Europe. For example, our PCBs and compute boards are fabricated in Europe. We use the 22FDX technology, and we also have semiconductor fabrication in Dresden, where our offices are located.

The teams we work with for design enablement, mechanical engineering, and thermal engineering are also based in Europe. Having control over our supply chain and understanding how it operates was very important to us.

For our customers, sovereignty is an important factor as well, but it is not the only one. Performance still has to be there. European startups can clearly contribute on both fronts: technological performance and sovereignty.

“Regis Hamelin”
Okay, very good. Our friend Shekhar, who already asked a question via LinkedIn, has another very interesting one. Thank you for joining today. His question is about how you plan to organize workshops to collaborate with other startups with the goal of co-creating solutions.

“Matthias Lohrmann“

This is usually handled on a case-by-case basis. Our typical approach is to start from a concrete use case, most often around SpiNNaker 2. We then evaluate whether we can genuinely help a customer address a specific problem.

If that is the case, we look for the right framework to move forward: it can be a joint project, a collaborative research call, or sometimes a more exploratory product development. As a startup, you are always looking for use cases that allow your solution to deliver the most value.

The business case or the underlying application is therefore critical. Having an idea alone is not sufficient; you need enough context and substance behind it to decide how to proceed.

In addition, there are already many established frameworks in the field of neuromorphic computing. When companies already work within these frameworks, migrating to SpiNNaker is relatively straightforward and requires limited support from our side. The more customized the use case becomes, the more closely we need to work together to assess whether it makes sense and can be implemented effectively.

“Regis Hamelin”
As you know, aCCCess is a network of Chips Competence Centres, but we also aim to facilitate access to other tools, such as pilot lines and design platforms. How important are these tools for you? Are you actively using or considering them?

“Matthias Lohrmann“

Yes, absolutely. Working with chip enablement teams is very important for us. When developing a chip, you need access to different design and development teams associated with the design platforms.

What really matters is having access to the right expertise at the right stage of development. As a startup, you need many different skills, but not all of them at the same time. Sometimes what you need is a fractional expert, and this is exactly what these platforms can provide.

Chip design also comes with a significant overhead. Questions around compute power, data access, and infrastructure need to be addressed, but as a startup you do not want to focus on these aspects too early. These platforms allow you to concentrate on your core design work while still having access to the necessary tools and support.

There are also additional funding and support opportunities associated with these platforms. The CHIPS Venture Forum is a good example, as it brings together startups, investors, and decision-makers to exchange, collaborate, and better understand each other’s challenges. From our perspective, this makes the platform extremely valuable.

“Regis Hamelin”
Very clear. Looking ahead, where do you see SpinCloud in three to five years?

“Matthias Lohrmann“

We see ourselves as a key driver of the sparse AI inference revolution. This is clearly the direction we are taking.

“Regis Hamelin”
And commercially speaking, how big are you today, and how do you expect the company to grow?

“Matthias Lohrmann“

Last year, we generated around six million euros in revenue. We aim to double that this year. Looking further ahead, with SpinnEx around 2028, we expect a significant increase. Even a market share of five percent would already represent several hundred million euros, potentially up to one billion, in the sparse AI inference market.

“Regis Hamelin”
If you had a magic wand and could make a wish within the framework of the Chips Act, what kind of support would you expect from the European ecosystem to help you reach your objectives over the next three to five years?

“Matthias Lohrmann“

I would say stronger connections between the different platforms that startups like us are part of. Better visibility and access to relevant groups across HPC, chip design, startup incubation, and related opportunities would make a real difference. Many opportunities exist, but they are not always easy to identify or access.

“Regis Hamelin”
That makes sense. The network of Chips Competence Centres was partly inspired by what EuroHPC has built. Once this network reaches a more stable phase, after the current period of rapid growth, it should be easier to create bridges with other initiatives and facilitate access for companies like yours.

One final question: if you could go back to 2020 and give yourself one piece of advice, what would it be?

“Matthias Lohrmann“

I would advise myself to spend more time engaging with the market, especially as a technical founder. Talking to potential customers early is essential. It is less about extracting the last percentage point of performance and more about addressing real customer needs, particularly in the early funding stages.

“Regis Hamelin”
So even before having a finished product, your advice would be to go and talk to customers, even with just a PowerPoint?

“Matthias Lohrmann“

Exactly. You may already have a clear vision of what your product can and cannot do, but once you truly understand what your customers want, development becomes much more efficient and you avoid many unnecessary iterations.

“Regis Hamelin”
Very good. We are now reaching the end of this interview. Thank you very much, Matthias, and thanks as well to the team behind the CHIPS Venture Forum 2025. Thank you to the audience for your strong participation and all your questions.

If people would like to contact you, what is the best way to do so?

“Matthias Lohrmann“

LinkedIn is the easiest option. My email address is also available on the last slide of the presentation.

“Regis Hamelin”
Thank you. This webinar was recorded and will be available as a replay in the coming days. If you would like to contact Matthias, his email address is shown on the slide.

Thank you again for joining us today. It was a pleasure having you with us. The next session will take place at the end of February, and the topic is currently being discussed.