The Ecosystem: start-ups give cautious welcome to artificial intelligence innovation package

The start-up community is broadly positive about the European Commission’s artificial intelligence (AI) innovation package, announced at the end of January. But there are reservations about the speed at which it can be delivered, and the details of some of its initiatives.

The package includes a range of measures to support AI start-ups and innovation, the most dramatic being a plan to establish ‘AI factories’ around high-performance computers that are reconfigured for rapid machine learning and the training of large general purpose AI models. As well as providing computing power to AI start-ups and SMEs, these centres will offer support in algorithmic development, testing evaluation and validation of large-scale AI models.

Meanwhile, further financial support for generative AI is promised through Horizon Europe and the Digital Europe programme, leading to an additional public and private investment of around €4 billion up to 2027. This will be accompanied by initiatives to strengthen EU talent pool in generative AI, and policies on data to train AI models and to support the development of novel and emerging AI applications.

Stef van Grieken, co-founder and chief executive of AI for protein design company Cradle, sees positive elements in the AI package, such as the initiatives on talent and access to high-quality, multilingual datasets. But in other ways, the package leaves a lot to be desired. “I think that we’re going to see a lot of really cool research coming out of these announcements, but I don’t think we are going to see billion-dollar companies,” he said.

Founded in 2021, Cradle uses generative AI to help in protein design and accelerate research and development. With bases in Switzerland and the Netherlands, the company raised $24 million in a series A round in 2023.

One shortcoming van Grieken sees is that European funding instruments, such as Horizon Europe and the European Innovation Council, are not attractive to top-tier AI start-ups. “They are incredibly bureaucratic, they are very slow, and if you compare them to the amount of money being spent by governments in places like the US, the UK or China, they are a drop in the ocean,” he said.

Good intentions

The concern that the Commission’s good intentions will be marred by slow and burdensome delivery, is shared by other AI start-ups.

“We appreciate that more support is coming. However, the money and the computing support has to be easily accessible, with a minimum of bureaucracy,” said Tobias Kehl, who leads AI Startup Rising, which channels the start-up activities of the Hessian Centre for Artificial Intelligence.

“This has the potential to help start-ups a lot, but the processes have to be well-defined and easy to use, otherwise it becomes a barrier,” said Martina Silov, president of the Croatian AI Association, which brings together start-ups and established AI companies

Given the reputation of past EU start-up support initiatives, this is an area where the Commission has ground to make up. “The assumption of many start-ups is that they will never get what they are applying for,” Silov said. “If the Commission can get rid of that bias, then start-ups will approach these opportunities much more willingly.”

Silov also thinks it is important that the funding opportunities allow start-ups freedom to pursue their goals. “The calls need to be open, so that start-ups can be really innovative rather than being forced into some kind of box,” she said.

The most damaging box of all would be the vicious circle created by the new regulations of the AI Act. “If start-ups have to apply for the funding simply in order to catch up with the rules, then we are creating a huge gap between the real world and how AI should be developed,” she said.

Meeting the demands of the AI Act is uppermost in many minds, particularly given its emphasis on trustworthy AI. “The development of trustworthy applications, as well as building large-scale trustworthy models, will be a challenge for everyone, including start-ups,” said Martin Dostál, a member of the investment committee at AI specialist Look AI Ventures in Prague.

Validation of such models, for example, will require both relevant knowledge and a lot of computing power. “If the EU already owns supercomputing facilities, it’s a natural next step to provide their capabilities to start-ups to make the implementation of EU AI Act requirements more affordable,” he said.

He also insists that the way this resource is provided will make a difference. “Besides the availability, it is highly important to have flexible and efficient access to really enable start-ups to focus on their business and technology development while minimising administration overhead,” he said. “It is important not to forget that time and efficiency are often more important than the realisation cost in the innovation business.”

And there is a real need for further computing power among Europe’s AI start-ups. “For companies developing large-scale machine learning models, such as various large-language models, the supercomputer facilities are definitely useful,” said Dostál. “I estimate that around 40% of the projects under our consideration are working with large language models, which is a significant number that I see growing rapidly.”

The same need is apparent to Kehl. “Hessian.AI has roughly 700 graphics processing units in place that we can provide the teams we accelerate, and we can already see that this is in high demand,” he said. “So, this support is crucial. And for many start-ups, it’s also important that those GPUs are in Europe, and that the data does not leave the EU.”

However, van Grieken is concerned that the proposal to adapt existing high-performance computers as AI factories may be a poor substitute for bespoke infrastructure. This is because start-ups may have to reconfigure their systems to run on these machines, at significant cost in time and resources.

“My recommendation to start-up founders who consider getting this free compute is: be absolutely certain that you are not going to spend more than three hours getting your things to run on it,” he said. “Otherwise, you are going down this rabbit hole of changing your models to fit the shape of the infrastructure provided, and that is an absolute waste of time.”

A significant computing mismatch could undermine the intentions of the package. “The potential risk is that you dramatically delay European start-ups, because they are waiting for compute that was promised to them to show up in a way that is useable,” he said.

In addition to computing power, Kehl thinks it is important that start-ups have access to clean, open data in order to train their models, and that the results of this work are openly available. Otherwise, the fall-back position for start-ups that need to use a foundation model for a specific application will still be US giants like OpenAI.

“Access to computing is one thing, but application programming interface access to open and transparent models, based on European values, that are trained on infrastructure and hosted here would be a really nice addition,” he said.

Silov also thinks that these AI factories should be as open as possible. “Many start-ups may not need this computing power, but if the application process is easy, it will encourage them to try it and see where it leads them,” she said. “It’s not necessarily about the final product, but the experience of having access to cutting edge technology and being able to use it. And maybe that leads to new ideas, to new partnerships and collaborations.”

AI and economic strategy

A broader concern is that Europe may be spreading its funding for AI too thinly. “One start-up cannot do a lot with a couple of thousand euros, or even a million,” Silov said. “Maybe we should start betting on bigger ideas, and perhaps put tens of millions of euros into a handful of selected start-ups.”

Van Grieken agrees. “In Europe, most of the funding instruments are effectively addressing things the market will not fund, but that’s the long tail [of companies] right now,” he said. “I would like to see more weight behind the winners, because we are incredibly behind in terms of investment.”

When it comes to Cradle’s own sector, AI and biotech, the prospects are better. “This is an area we could still win, and that’s because we have an incumbent pharmaceutical and biotechnology industry here,” said van Grieken.

However, other opportunities are slipping through Europe’s fingers, and not because of a lack of support for AI. In sectors such as cultured meat or other novel foods, product regulation is the problem. “For some reason, we find it very hard to lift those types of restrictions, which means that all the European synthetic meat companies are launching products outside of the EU,” van Grieken said.

Dostál also thinks the EU should not lose sight of the broader business context. “Providing and sharing AI-related capabilities and bringing new opportunities for AI companies is great,” he said. “But it would help to continually streamline the environment for conducting lean, efficient, and fast-paced innovative businesses.”

Elsewhere in the Ecosystem...

• A programme to help start-ups supported by the European Innovation Council access international procurement markets will be launched on 22 February. Called SPIN4EIC, the programme will run innovation procurement academies and offer tailored assistance to take part as bidders in public and private sector tenders in and outside Europe. It also includes support to for public buyers looking for innovative suppliers.
• PulmoBioMed, a 2020 spin-out from Northumbria University in the UK, has raised £1.4 million in seed funding for the development of a breath sampling technology that can distinguish fluid droplets originating deep in the lungs from others in the airway. This will help in the early diagnosis of asthma. The investment round was led by the North East Venture Fund, and included Northumbria University, SFC Capital and private investors in the US, EU and UK.
• The Commission is putting together a study of the language technology market, in order to assess the influence of Europe actors and the development potential in the global market. Start-ups and scale-ups in this sector will want to contribute to the survey currently running that will inform the study. Contributions are due by 23 February. The preliminary findings of the study will be presented in May.
• A study carried out for the Biotechnology and Biological Sciences Research Council (BBSRC) in the UK has found that spin-outs created from its research have contributed £5.2 billion to the economy over a 25-year period. In 2021-22, the review found that there were 402 spin-outs supported by BBSRC funding, employing over 8,000 people across a range of UK markets.