In order to comprehensively observe what is happening in data and AI, and foresee what is next, we have to consider the context around the technology, and the impact of key factors such as regulation, national strategies, the availability of talent, and access to capital. It is also important to acknowledge the impact that increased AI activity has on the world around us, and how our understanding of those effects should shape the way we build and invest in the technology going forward — here we look at the climate impact of AI as it is already somewhat measurable; in future analyses we’ll also aim to dive deeper into understanding further aspects such as the emerging social impact of AI, effects on the labour market, unequal access to technology, and potential for overall increased inequality.

During this section, we will take a closer look at these areas in order to understand how they are contributing to shape, inhibit, or enable developments. Given that the subject scope has the potential to be so broad, we at times shine a light on the global context, and at other times provide a more focused view on the European context.

Around the world, national governments and international organisations are accelerating their efforts to create AI regulation. Some policymakers see it that they are fighting to stay ahead of the pace of technological development. And others, that they are rushing to keep up with it.

In theory, AI regulation aims to safeguard against potential risks and misuse, enable an ecosystem of rapid development and progress, and ensure that the overall impact on society is well-managed and net-positive. It is a difficult task to strike the balance of these aims, with restrictive policies finding their fit between being responsible and necessary versus constraining and anti-innovation, and enabling policies finding their fit between being progressive and supportive versus irresponsible and dangerous. All stances and opinions across these spectra can be heard loudly and passionately.

One of the biggest issues has been to establish who should shoulder the burden of responsibility, and thus where regulations should be targeted. In other words, should regulation be directed towards the builders of powerful AI models, or towards those that develop the applications to use those models in different settings? Leading AI researcher Andrew Ng and Yann LeCunn (Chief AI Scientist at Meta) have articulated this point well. Essentially, we are distinguishing between regulating a technology (e.g. powerful foundation models) and regulating applications (e.g. a recruiting automation tool connected to an LLM, or a personalised healthcare solution connected to an LLM). At its worst, regulating a technology would mean regulating research efforts, and creating obstacles for fast and open AI research. While a technology can be applied in many different ways to solve various problems, an application is a specific implementation of that technology designed to meet particular needs. Andrew Ng’s analogy is a perfectly good and clear one, so we will repeat it here:

“For example, an electric motor is a technology. When we put it in a blender, an electric vehicle, dialysis machine, or guided bomb, it becomes an application. Imagine if we passed laws saying, if anyone uses a motor in a harmful way, the motor manufacturer is liable. Motor makers would either shut down or make motors so tiny as to be useless for most applications. If we pass such a law, sure, we might stop people from building guided bombs, but we’d also lose blenders, electric vehicles, and dialysis machines. In contrast, if we look at specific applications, like blenders, we can more rationally assess risks and figure out how to make sure they’re safe, and even ban classes of applications, like certain types of munitions.”

Thus, safety is a property of applications, more than it is a property of technologies (or models). There are, of course, AI safety questions that need to be addressed at the model level, and the builders of advanced models must take responsibility with regards to topics related to training data, inherent biases, levels of transparency, and efforts towards alignment. But much of the safety depends on the context and way in which an AI model is applied. Thus, it is in recognising this distinction between technology and applications, and directing appropriate policy towards each layer of the AI stack, that AI regulation can be most clear, fair, and effective.

The last couple of years have really seen AI policy and regulation take centre-stage as a topic of heated discussion, and a prioritised agenda item for lawmakers. If 2023 was the year of creating plans, roadmaps, and aligning on vision, then 2024 can be seen as the starting point of action and implementation.

In the same way that the EU was a first-mover in introducing data protection regulation in the form of GDPR, becoming the de facto global standard subsequently mimicked by regions all over the world from India to California, the EU has acted quickly to introduce its EU AI Act in the hope that it can again go some way towards shaping the way the world does business and develops technology.

In March, the European Parliament voted to pass the AI Act, with a final green light coming in May, before it came into force on 1st August. Regulators are getting themselves set up in order to enforce the law (a new AI Office has been established), while companies will have up to three years to comply with it.

The AI Act applies broadly to “AI systems” - covering traditional machine learning as well as generative AI - and will affect all firms seeking to do business within the EU, regardless of where they are based. At its core, the regulation defines a three-tiered risk classification of AI systems, as either unacceptable risk, high risk, or minimal risk, based on their intended use and the potential impact on individuals and society.

Unacceptable risk refers to AI systems that pose a significant threat to people’s safety, rights, or fundamental freedoms, and are therefore banned. They include systems that:

• Use real-time facial recognition in public places

• (The police would need to get court approval for specific purposes such as anti-terrorism, finding a missing person, or preventing human trafficking)

• Use emotion-recognition technology at work or in schools

• Create facial recognition databases by scraping the internet

• Infer sensitive characteristics such as a person’s sexual orientation or political opinions

• Deploy “social scoring” and classify people based on behaviour, socio-economic status or personal characteristics

• Exploit vulnerable people or utilise cognitive behavioural manipulation to distort behaviour and impair informed decision-making

High risk refers to AI systems that pose a significant risk to people’s health, safety, or fundamental rights. To ensure their safe and ethical deployment, these systems are subject to stringent regulatory requirements. They include systems that relate to:

• Healthcare

• Education

• Law enforcement

• Human resources and recruitment

• Public services

• Administration of justice and democratic processes

• Transport systems

Some of the obligations that providers and deployers of “high risk” systems must adhere to are:

• Risk management across the whole AI lifecycle

• Data governance and ensuring high-quality datasets for training, validation, and testing

• Human oversight to minimise risks and ensure and ensure accountability

• Maintaining detailed technical documentation

• Ensuring that high standards are met for accuracy, robustness, and cybersecurity

• Reporting energy consumption, resource use, and other impacts throughout the systems’ lifecycle

Minimal risk refers to AI systems that pose little to no risk to people’s rights and safety. As a result, these systems are subject to the least regulatory scrutiny and can be freely used, though firms are encouraged to follow general principles related to fairness, non-discrimination, and the use of human oversight. Examples of minimal risk systems may include:

• Inventory management systems

• Predictive maintenance

• Personalised e-commerce recommendations

• AI-enabled video gaming

• Customer service chatbots

Regulating models

While the tiered risk classification addresses the application of technology towards different use cases, the AI Act also aims to regulate those firms developing the “general purpose AI models”, with different obligations depending on how powerful the model is considered to be.

All companies developing “conventional general purpose AI models” will need to maintain technical documentation showing how the model was built, publish a summary of the underlying training data, and prove that they respect copyright law.

Companies developing more powerful models, so-called “systemic risk general purpose AI models”, are subject to more stringent obligations, including needing to perform risk assessments and model evaluations, ensure high degrees of cybersecurity protection, undergo enhanced testing, and comply with reporting requirements such as any incidents of system failure.

The distinction between those models classified as “conventional” and those classified as “systemic risk” is made based on FLOPS - the “floating-point operations per second” measure of computational power that a model is capable of. The AI Act defines “systemic risk general purpose AI” models as those with more than 10^25 FLOPS, which includes models such as GPT 4, Gemini Ultra, and Mistral Large.

While there are currently not too many models meeting this classification (as of April earlier this year, Epoch AI estimated less than a handful), given the rapid pace of AI development and model scaling, it is likely that many more will fall under the regulation going forward. As a result, the AI Act retains the flexibility to adjust the threshold if the landscape sufficiently changes. Notably, in the US, the Biden Executive Order on AI sets the threshold at 10^26 FLOPs, which is a substantial increase in computational power, and likely equivalent to a 10x higher cost of model training.

So what a FLOPS-based threshold should be is clearly up for some debate. At the same time, whether such a measure makes sense or not is also a point of discussion, given that the quality of models that can be trained within this threshold is rapidly improving (e.g. with highly advanced smaller models), meaning that FLOPS may not be representative enough of model capabilities.

Big tech’s concerns and Europe losing out

In September, Mark Zuckerberg, Daniel Ek, Yann LeCun, and a number of technology leaders and researchers signed an open letter to the EU, asking for regulatory certainty in order for the region to not fall behind.

In accordance with claims that have long been raised, the signatories state that Europe has become less competitive and innovative, and risks falling further behind on AI, due to inconsistent regulatory decision-making - with, for example, interventions by the European Data Protection Authorities creating huge uncertainty around what kinds of data can be used to train models.

As a result of this inconsistency and uncertainty, the signatories claim that the EU is set to miss out on two cornerstones of AI innovation - namely developments in open models, and developments in multimodal models. This has recently been seen to come to reality, with Meta not releasing its latest Llama 3.2 multimodal models in the EU, and OpenAI not making its Advanced Voice Mode available in Europe. If European builders lose out on access to the latest open models, then they can only fall behind others based elsewhere who are not prohibited and can continue to develop faster.

Looking forward

Now that the basis of regulation is set, how, and how quickly, the EU continues to refine and shape it will be key to enabling European tech companies to thrive and compete as AI winners are built. It’s worth noting that the five-year term of the EU Commission which has led the development of the EU AI Act is now coming to an end, meaning that a new college of commissioners will be formed and tasked with the implementation and refinement of the regulation. It remains to be seen if this will represent an opportunity for fresh perspectives to continue to evolve the regulation, or if it will contribute to a bumpy ride in its enforcement.

The policies of nation states will clearly impact the overall speed of technological development, where it takes place, and what kind of impact it will have. As we continue to proceed into this era of AI, governments around the world are tasked with setting out their strategies for seizing the opportunity, remaining internationally competitive, and ensuring that their societies benefit from new technology in both the near and longer term.

On a European level

The September publishing of Mario Draghi’s report into “The future of European competitiveness”, re-initiated calls for policymakers to ensure that the continent does not miss out on this AI platform shift. The report highlights the innovation gap between Europe and the US & China, and calls for

• Annual funding of €750-800bn for EU projects

• Strategic focus on innovation, and enabling European companies to become €100bn winners that remain European

• Improved EU coordination, addressing policy and regulatory fragmentation

• Seizing decarbonisation and the energy transition as an opportunity

Alongside the recent discontent from tech leaders regarding Europe’s approach to regulation, there is a growing feeling from some early-stage builders that they are much better off basing themselves in the US - given frontier innovation, access to more capital, more favourable regulation, better access to the latest AI models, and of course access to a large single market.

Evidently, there is work to be done to create the conditions that will enable Europe to thrive and be a relevant technological force in the coming years.

Comparing countries

Recently, a number of rankings have emerged, trying to compare countries in their approaches to AI and/or their potential to embrace it. Such rankings are of course highly complex and rather subjective, and thus very limited conclusions can be drawn from them, but they do make for an interesting point of departure in discussions around what can or needs to be done.

The Global AI Index looks at 83 countries and ranks them based on AI implementation (talent, infrastructure, operating environment), innovation (research, development) and investment (government strategy, commercial ecosystem). The 2024 ranking can be seen below.

While it’s unsurprising that the US and China sit at the top of the rankings, what is more interesting, and alarming for other nations and Europe as a bloc, is the size of the gap between those two at the top and the rest.

During the summer of 2024, the IMF published its AI Preparedness Index assessing the level of AI readiness across 174 countries, and considering factors including digital infrastructure, human capital and labour market policies, innovation and economic integration, and regulation and ethics.

This ranking does not paint the same extreme picture with regards to the frontrunners, likely because it focuses on ranking the ability to embrace and integrate the technology, rather than the ability to develop it, drive the innovation, and essentially own the value chain, as the Global AI Index addresses. According to the IMF ranking, much of western Europe is well-placed to embrace AI (Denmark scores 0.78, UK scores 0.77), alongside North America (US scores 0.77). While China lies a little behind with a score of 0.64. The big warning though is how AI could widen existing inequalities between wealthy and developing countries, and the IMF encourages efforts to open up access to state-of-the-art technology.

Recent national policies and actions

• Countries all over the world are pouring billions of dollars into new domestic computing facilities for AI

• The UK then cancelled $1.7bn worth of planned computing projects

• Countries and regions setting regulation and large investment commitments to boost their computer chip competitiveness

• The CHIPS Act in the US aiming to boost the US semiconductor industry, as well as Europe’s own EU Chips Act

• South Korea unveiling a $19bn package of incentives to boost its chip sector

• China’s $47bn semiconductor fund to ensure chip sovereignty

• Governments seeking to attract big tech firms such as Microsoft, Google, and Amazon, to make multi-billion dollar investments in cloud and AI infrastructure in their countries, as seen e.g. in France and Sweden

• Saudi Arabia allocating $100bn to invest in technology, including a massive $40bn fund to invest in AI, as it aims to become a global AI hub

With recent elections in France and the UK, and an upcoming election in the US, major economies are providing mandates to newly established leadership. The policies and actions of those governments will go some way to shaping who stands to win or lose the most in the years ahead.

10 national strategies to thrive in the AI era

In order to thrive in this era of AI - to enable technological development, economic growth, and widespread societal benefit amidst significant transformation - governments and regional blocs need to take assertive action rather than be asleep at the wheel. Here, we highlight a number of initiatives that we believe governments need to be acting upon in order to build positive momentum in this era of technological change and opportunity. The list is non-exhaustive, and some of these actions are more urgent than others, but to dive deeply into these topics is an exercise for another occasion. 

1. Secure sovereign access to computing resources

   Governments need to ensure that their countries are not left vulnerable by a lack of access, independence, or control when it comes to computing infrastructure. They need to work together with the private sector in order to ensure security across the entire value chain - from research and development, hardware production and supply, to data centres, AI models, and core tooling.

   While not all of these components need to be addressed and secured on the national level - some may be better tackled across regions, trade blocs, or bilateral agreements - governments must have clear strategies for the independent initiatives and cross-border collaborations that can secure the necessary compute resources and capabilities across the AI stack.

   This will enable a greater degree of influence and control towards national security (against foreign interference and cyberattacks), data sovereignty (better ensuring that data laws and regulations are adhered to), economic independence (not overlying relying on foreign technology providers), and resilience (to recover core infrastructure in the event of a crisis or disaster).

2. Enable the energy transition

   As we have seen throughout this report, there is extreme demand for increased compute resources in order to power AI development and applications, and this compute consumes huge amounts of energy resources (see more in the climate impact section). Significant investment is required to prepare the energy grid for AI power demand, and, as AI leadership may be strongly correlated with energy prices, it is essential that governments deliver on strategies to enable low cost, stable, and renewable power.At the same time, we have seen that computational efficiency can be improved and energy consumption reduced by developments in data centre infrastructure, chip architecture, models, and software tooling. Governments should be intentional in funding research efforts, and incentivising private sector actors, that have the potential to improve energy efficiency in relation to AI and compute.

   Beyond the specific challenges of AI power demand, we recognise that much of the broader energy transition is a financing problem more than it is a technology problem. Solutions such as batteries, solar, and heat pumps exist but often with high upfront costs, even if their lifetime costs are lower compared to fossil alternatives. Governments need to be consistent in their policy and long-term in their thinking, so as to best encourage companies and households to also think long-term and make sensible energy choices. At the same time, governments should encourage the private sector to develop and provide financing solutions that enable businesses and households to transition to renewables, so that upfront cost does not get in the way of delayed benefit.

3. Incentivise investment

   In order to fuel technological innovation, long-term economic competitiveness, and ensure homegrown category-leading AI companies, governments must ensure that their policies incentivise increasing amounts of investment in the necessary high-risk high-innovation technologies and companies.

   As we cover in the “Access to capital” part of this report, European asset managers still allocate only half as much towards venture capital as their US peers, while pension funds in particular are allocating very little to the asset class. This indicates the potential in Europe to unlock more capital that can fuel this wave of AI innovation. And given the scale of AUM, even miniscule percentage increases in amounts allocated towards venture capital have the potential to drive significant investment and innovation. Thus, government policies that aim to mobilise institutional capital can have an outsized impact. Examples of policies may be public-private partnerships (such as the Tibi Initiative in France), tax incentives, regulatory reforms, or government support for fund-of-funds structures.

4. Attain and retain talent

   It’s remarkable people that force breakthroughs at the forefront of research and technology, and that start and build category-leading tech companies. The best and most ambitious talent will go where they are empowered with the best chance of succeeding and having the large-scale impact they believe they can achieve. Thus, competitiveness starts from creating the conditions that best attract and retain world-class talent. These efforts must focus both on research and academia (increasing funding to universities and research organisations), as well as the private sector (increasing investment, incentivising entrepreneurship, creating the conditions for companies to compete).

5. Regulate with clarity and agility

   We cover the current state of regulation in the EU elsewhere in this report, where we also acknowledge the difficult balancing act that policymakers have in setting regulation that both safeguards against legitimate risks and enables a fast pace of innovation. The speed of technological development - with improving model capabilities and new applications - means that the goalposts are continuously moving, but the picture is also becoming increasingly clear. As a result, regulators need to be open to reassessing their frameworks - such as where to set FLOPS-based thresholds, how to distinguish between different risk levels, how data should be handled, or what mitigating measures are in fact reasonable and effective. This openness to refining regulation of course needs to be managed against the need to provide clarity and consistency so that companies and individuals can most easily plan and navigate. Again, another part of the difficult balancing act, but another key factor in laying the overall foundation for economies to thrive and seize the AI opportunity.

6. Prepare people, broadly educate, and facilitate dialogue

   Technological change can be daunting at the best of times, as uncertainty breeds concern. In the context of AI, this has been exacerbated by media scaremongering and discussions about risks that are at times perhaps extreme and distracting. This risks creating an environment of fear and ignorance, and leading individuals to widely reject new technology rather than interact with and embrace it → resulting in many being left unprepared when its implementation and impact starts to be more widely felt.

   Governments should work to build a positive (while of course objective) narrative and educate individuals regarding the potential of AI to improve people’s economic and life outcomes through increased earnings, and improved healthcare, education, and critical infrastructure. They should engage and consult citizens’ assemblies to identify fears and priorities, in order to build public trust and inform decision-making, and should encourage employers and employees to discuss the potential impact of AI in the workplace, in order to address fears related to job safety.

7. Integrate AI within government

   Governments should lead by example and integrate AI across government - working to successfully implement the technology, as well as to build a culture of progressive thinking towards it, such that government operations can increasingly become technology-first, and AI-enabled. Not only will this serve to improve the efficiency and effectiveness of public services and administration, but it will also provide greater insight and experience for those politicians and civil servants tasked with steering AI policy across the private sector and society.

8. Update the school curriculum and integrate AI within schools

   The school curriculum should be updated to emphasise, test, and evaluate the core skills that are required to make the most of AI tools and prepare for an AI-native job market. This includes focusing on

   (i) problem solving related to analysis and logical thinking, creative thinking, and quantitative and statistical skills.

   (ii) collaborating and judgement in the form of task delegation, machine understanding, and evaluation.

   AI should also be seen as a great opportunity to accelerate every child’s learning, as we move from a one-size-fits-all approach to teaching towards fully personalised and highly interactive education. This opportunity should be embraced, and programs should be established that guide and incentivise schools towards engaging with new tools, and support them in their safe implementation.

   The countries that more quickly and more deeply integrate AI into the delivery of education will see significant compounded effects in the knowledge and skills of their young people and labour market inflow.

9. Be predictive and proactive in forecasting job market trends

   Fund collaborative efforts across academia and industry that seek to understand (i) the unique characteristics, strengths and weaknesses of the economy, (ii) the changing nature of different industries, and (iii) the forefront of AI technology and its application. By putting significant resources towards this, governments can better predict how and where AI may disrupt the workforce, as well as where opportunities for job creation may emerge, so that actions taken can be proactive more than reactive, and lead to a net-positive effect on the labour market.

10. Become agile and generous in training and retraining

    To the extent that governments can be somewhat successful in point (9) above and identify changes within the labour market, as well as inherent strengths that the economy may have to further capitalise on, the next step is to take quick and meaningful action. The faster that training programs can be implemented to support workers in upskilling and seizing productivity gains of AI, the more advantageous to national competitiveness and economic growth. While the faster that retraining programs can be implemented to support those who suffer job losses as a result of technological change, the softer the landing for those individuals, and the cheaper it will be in the long run as those workers more quickly re-enter the workforce.

    These efforts will require various forms of public sector support as well as public and private sector collaboration, as governments enable and incentivise companies to invest in upskilling and retraining.

11. [Bonus] Find the “Home PC” reform of AI

    Over the past couple of decades, Sweden has established itself as a strong entrepreneurial tech hub, home to global category leaders such as Spotify, Klarna, King, Truecaller, and Kry. There are many reasons and contributing factors for why this small Nordic nation has been able to achieve such repeatable success, but one factor often referred to is the “Home PC Reform” policy of the late 1990s. This policy allowed employees to acquire a computer at less than half the retail price, through a tax-exempt arrangement provided by their employers. The reform had a significant impact on computer adoption in Sweden, as the number of households with at least one computer doubled in the first four years of the subsidies. The policy emphasised access to computers rather than just knowledge about them, and has been credited with contributing to Sweden becoming a leading nation in terms of IT literacy, and enabling a digital-native generation. This in turn helped to lay the foundation for the tech talent and founding of some of the aforementioned unicorn successes.

    The “Home PC” reform was introduced at a time when the internet and computers were quickly changing the nature of work and life at home. Now, as AI is set to do the same, we ask “what is the Home PC reform of AI?” Can similar public-private schemes subsidise and provide households with greater access to AI tools? Or, more broadly, what policies can contribute to increased AI access and literacy such as to enable a generation of technically capable workers, innovators, and founders of AI category leaders.

It is clear that the US remains vastly superior to Europe and the rest of the world when it comes to the quantity of top AI talent. This is evidenced by the dominance of US universities at the forefront of AI research, and the dominance of US Big Tech firms in hiring engineers.

However, Europe’s AI talent foundation is strong, with positive momentum that will serve the ecosystem well going forward. We’ll explore each of these points in a little more detail below, but in short, with strong AI research institutions, a growing presence of leading tech and specifically AI firms, and an ability to retain much of the best AI talent and be a net importer of tech personnel, Europe has an increasingly strong flywheel for creating category-leading companies across the data and AI stack.

Europe’s AI talent flywheel is just starting to turn
European tech has been growing from strength to strength over the past couple of decades, and it is well established that each unicorn success contributes to the next wave of founders creating new companies. Founder mafias have emerged from companies such as Spotify, Revolut, Skype, Wise, and Klarna and given birth to dozens of new startups. As Atomico highlighted in their 2023 State of European Tech Report, this talent flywheel of unicorn alumni is accelerating, with nearly 9,000 companies having been created by alumni of European exited unicorns that were founded during the 2000s, almost 50% more than the number created by alumni of unicorns founded in the 1990s. Looking at the trend of the most recent cohort of unicorns founded in the 2010s, the pace only appears to be picking up.

When it comes to the specifically AI talent flywheel, we’re at an earlier stage but the momentum is certainly picking up. Large US tech firms such as Google and Meta have had established AI research presence in Europe for some time, while DeepMind, having been founded in London in 2010 has long been the jewel in the crown of the European AI ecosystem. It’s clear that these firms have served to provide the foundation for talent fueling the current wave of AI startup activity. A recent report by Accel and Dealroom, analysing 221 GenAI startups from across Europe and Israel, found that 25% of those companies have one or more founders who have worked at Amazon, Apple, DeepMind, Facebook, Google or Microsoft, and this increases to 38% when looking at the top 40 companies in terms of funding, and even further to 60% when looking at the top 10.

We expect that this flywheel is only just getting started as:

• US presence continues to grow with leading AI players such as Open AI and Anthropic (building models), and Scale AI and CoreWeave (providing enabling tooling and infrastructure), only recently establishing significant European offices

• Europe continues to establish its own cluster of companies operating at the frontier of developing AI models or serving AI to enterprise, with the likes of Mistral, Aleph Alpha, and H

• European leaders continue to emerge across the application and enabling layers of the stack - companies such as ElevenLabs, Synthesia, Wayve, Hugging Face, Poolside, and Parloa

These positive trends are creating the presence of more AI talent factories, and more factories will in turn lead to a higher volume of increasingly capable AI founders.

Disclaimer: While we fully believe in these positive dynamics over time, in the short-term the presence of relatively few, extremely well-funded companies is likely to pose a significant challenge to the rest of the startup landscape fighting for top AI talent. Anecdotally, this is already playing out in Paris, where the likes of Mistral, H, and Poolside represent formidable hiring competition for companies that have a fraction of their budgets.

The role of universities

When looking at which universities are contributing the most to AI research and the development of frontier models, the US clearly leads the way with Stanford, Carnegie Mellon, UC Berkeley, and MIT the biggest contributors. Europe does, though, have three universities out of the top 12 global contributors to notable AI models over the past 10 years, with University of Oxford, Technical University of Munich, and ETH Zurich each featuring.

It is also clear that these academic institutions are just as essential to providing the foundation for Europe’s AI talent flywheel. Accel and Dealroom’s report illustrates the highly educated and rather academic nature of GenAI founding teams, with 38% of the companies analysed having at least one founder who holds or has held a position at an academic institution. This increases to 55% when looking at the top 50 companies in terms of funding, and even further to 80% when looking at the top 10.

When it comes to where the founders of GenAI companies are educated, Cambridge University, École Polytechnique, Imperial College London, University College London, and Oxford University lead the way, together educating one-third of GenAI founders in Europe.

Europe is a net importer of tech talent and is continuing to retain more of its top AI talent

It is often proclaimed that Europe is suffering a brain drain of tech talent, as people flock to join US companies or build in the US instead. Certainly, this is a sentiment and concern that most within the European venture scene are familiar with, however, there are a few data points that can potentially add nuance to the discussion and provide some optimism.
Atomico’s 2023 State of European Tech report shows how Europe continues to be a net beneficiary of talent movement, with 10k+ net new people joining the European tech scene during 2023.

While this inflow supports the tech ecosystem as a whole, Europe’s ability to build category-leading AI companies relies on the rate at which it can attract and retain world class data scientists and AI engineers. Over the past ten years, the US has been dominant in attracting AI talent from the rest of the world - powered by the growth of the likes of Amazon, Apple, Google, Microsoft, Meta, and Nvidia, the US has attracted and retained twice as much top AI talent as it has supplied to the rest of the world. However, as highlighted by the latest State of AI Talent 2024 report by research company Zeki, this trend appears to be slowing.

This shift in momentum is being driven by a combination of:

• More “national champions” in other countries managing to retain their talent

• The attractiveness for AI talent to leave Big Tech - seeking the opportunity to build something new and disruptive, while being compensated by significant equity and potential upside in highly valued VC-backed AI startups

• Factors related to politics and welfare steering some talent towards Europe over the US

A vast and well-functioning market of early and growth-stage capital is essential for funding and fueling the frontier companies emerging across the different layers of the AI stack. Overall, the picture for Europe is positive.

Data from Pitchbook shows how VC and growth funding in Europe has been growing over the past decade, and although down compared to the highs of 2021 the trend remains encouraging, particularly in comparison to US funding volumes. VC and growth funding in Europe is now at ~40% of the US, compared to only 16% as recently as 2015.

When it comes to AI funding, though, the US remains way ahead, with the San Francisco Bay Area clearly the undisputed leader in AI tech and funding dollars, as shown by Crunchbase data. During 2023, more than 50% of all global venture funding for AI-related startups went to companies headquartered in the Bay Area. This figure is heavily influenced by mega-rounds in companies such as OpenAI, Inflection AI, and Anthropic, and the SF share falls to 17% when looking at the number of rounds rather than total value. Dealroom also shows that since 2022, the US has invested 3x more than Europe in AI as a whole, and 7x more in generative AI. Despite this, huge funding rounds this year in European companies such as Wayve (UK, $1.05bn), Mistral (France, $645m), H (France, $220m), and Aleph Alpha (Germany, $500m) suggest that there exists the willingness and ability to heavily back European AI firms too.

The foundation is well set in Europe for even higher levels of investment, Atomico’s State of European Tech 2023 highlights the level of dry powder (the amount of committed but unallocated capital that funds have at hand to deploy) in Europe as having risen to $58bn for VCs, and $108bn when also including growth funds. Back in 2013, this number was just $14bn for VCs, and $23bn when including growth funds. Thus, European venture funds currently sit on by far the largest pool of deployable capital that the continent has ever seen, which should be encouraging for both the founders of startups, as well as the earliest backers of them taking the initial investment risk.

Atomico's report suggests that there exist significant opportunities for policymakers to work with asset managers to unlock even more funding for European tech. The report finds that European asset managers allocate 8% of assets under management (AUM) to venture capital, compared to 16% allocated by US asset managers. While pension funds continue to allocate very little to the asset class - in 2022, just 0.01% of pension fund AUM (which stands at $7.8 trillion) was invested into European VCs. With these numbers suggesting that there exists potential to unlock more capital to support technological innovation, a number of policies are aiming to facilitate and incentivise greater institutional investment in VC. For example, the Tibi 2 initiative in France, building on the success of the previous Tibi initiative, and aiming to attract up to €7bn of institutional capital into early-stage and deep tech investments. 

The impact of AI on the climate started to gain more attention towards the end of last year, when estimates started to show that making an image with generative AI uses as much energy as charging your phone, a ChatGPT query needs nearly 10 times as much electricity to process as a Google search, and Google’s AI-powered search may be 30 times more energy intensive than a standard Google search. While there has since been development in models and computational efficiency, there is no doubt that the compute required to run AI is energy intensive. And, as we covered earlier in this report when looking at hardware, demand for AI computing power is skyrocketing, as firms train larger and larger models, and speed and volumes of inference also increase. In Google’s latest environmental report, the firm shared that its emissions had gone up 48% over the past five years, largely driven by increased efforts related to AI. While some claim that the data centre emissions from big tech firms may be up to 7.62 times higher than those firms officially report.

Data centre power demand is estimated to grow 160% by 2030, according to Goldman Sachs Research. This will see data centres worldwide go from consuming 1-2% of total power to 3-4% (in the US this has been estimated to be as high as 8%), with developed economies experiencing the kind of electricity growth that hasn’t been seen in a generation → and a doubling of data centre carbon dioxide emissions as a result.

Alongside the growing carbon footprint is the issue of AI’s enormous water footprint. In its 2022 environmental report, Microsoft disclosed that its water consumption increased by 34% YoY, to nearly 1.7 billion gallons (or more than 2,500 Olympic-sized swimming pools), largely driven by its AI research activities. When xAI’s “world’s largest supercomputer” gets up to full capacity, it is estimated it will need a million gallons of water per day. Researchers at UC Riverside forecast that in 2027, global AI demand may be accountable for 4.2-6.6 billion cubic metres of water withdrawal, which is equivalent to about half of the UK’s annual water withdrawal. This has the potential to pose very real and pressing challenges, as freshwater scarcity is already a growing issue as a result of depleting water resources and outdated water infrastructure.

Companies building solutions to AI’s climate impact

We expect to see a growing number of companies in Europe building solutions to address the different parts of the environmental impact of AI, in areas such as:

• Sustainable data centres powered by renewable energy, such as NexGen Cloud (UK), and evroc (Sweden)

• New innovative approaches to cooling and energy efficiency within data centres

• New more efficient chips, using photonics or other technologies, such as Black Semiconductor (Germany), and Lumai (UK)

• Chips optimised for inference and the running of AI models post-training, such as Fractile (UK)

• Software that enables developers to design their AI systems to train models faster, and get GPUs running more efficiently, such as FlexAI (France)

Transforming energy infrastructure

However, increased power demand will require significant investment in energy infrastructure. Goldman Sachs estimates that, as a result of the expansion of data centres and an acceleration of electrification, Europe’s power demand could grow by 40-50% between 2023-2033. Data centre power demand will predominantly rise in two archetypes of European countries:

1. Those countries with cheap and abundant power from wind, hydro, solar, and nuclear, such as the Nordics, Spain, and France

2. Those countries with large financial services and tech firms, and where tax breaks and incentives will be offered in order to attract data centres, such as the UK, Ireland, and Germany

Europe, though, has the oldest power grid in the world - fifty years old on average, compared to forty years old in North America, and twenty years old in China. As a result, Europe faces additional challenges compared to other regions when it comes to preparing the energy system for AI, and providing the AI data centre industry with what it needs, which in short is:

• Inexpensive electricity costs - given the immense amount of power to be consumed, particularly as inference needs compound over time

• Stability and robustness in the energy supply chain - safeguarded against geopolitical and weather disturbances

• Power generation with a low carbon intensity - providing huge amounts of renewable energy

Thus, it will require huge investment in order to keep new data centres electrified. Up to €850 billion may need to be invested in renewable energy such as solar and wind, as well as up to €800 billion in spending on energy transmission and distribution. As outlined by Sequoia, AI will catalyse an energy transformation, and, beyond the rise of renewables, a resurgence in nuclear energy will be one of the long-term effects of the AI wave. This can already be seen amidst recent activity from big US tech players, with Microsoft agreeing an $800m per year deal to reopen the Three Mile Island nuclear power plant, Oracle revealing a data centre design powered by three small nuclear reactors, and Amazon announcing efforts to recruit nuclear engineers.