Our projection for 2025
Is AI a bubble? Will it burst? And what will emerge in its aftermath? Which approach to technological sovereignty will prevail in Europe? Will AI agents re-shape digital markets?
Welcome to the last newsletter of the year. This time we decided to do something different by sharing the key themes and questions we’ll be thinking about in the new year. In other news we’ll be covering: increased AI defense spending, Open AI exploring advertising and partnering with telco Orange, new investments in European AI, and discriminatory AI systems used for detecting welfare fraud in Sweden and the UK.
Key questions for 2025
Is AI a bubble? Will it burst? And what will emerge in its aftermath?
Throughout the year, the writing has been on the wall: Serious investors and money managers from Vanguard, to Goldman Sachs, Sequoia Capital to the European Central Bank, have expressed concerns about a potential asset bubble forming in the AI market. Sky-high valuations, sustained by excitement, expectations and exuberance could lead to a correction in the markets in 2025, if profits remain elusive. Few of the dimensions that we will be watching are the following:
Valuations have reached dizzying heights:
Stock market frenzy: AI-related companies have seen their valuations surge, to the extent that a correction would affect the stability of the overall stock market. These increases are based on infrastructure plays and speculative expectations, rather than the demonstrated profitability of AI use cases. With alternative investment targets emerging, the capital flight to other assets and sectors (such as crypto, or defense) might intensify.
Costs remain staggering:
Infrastructure cost: Running the largest AI models at scale (e.g., OpenAI’s GPT-4) requires substantial rolling computational and financial investments, making profitability difficult to achieve for most companies. This differs from the previous waves of tech unicorns in the zero-percent-interest-rate era, where the marginal cost per user was minimal, allowing platforms like Uber to delay profitability for years and focus on pure scaling. The cooling off of the AI market poses problems for companies like Nvidia, which is now trying to lock-in customers and contracts to soften the potential ebb in demand in the near future.
Legal risks: Unresolved copyright issues could threaten the foundation of the nascent industry, or at the very least, significantly increase cost of training and operation.
The path to profitability remains unclear:
No revenue model: so far, most investments in AI have benefitted (a handful of) chipmakers, data centers, energy providers and cloud service providers, in other words to infrastructure. Meanwhile AI model makers and companies that build AI applications, including Open AI, are yet to find a working revenue model amidst skyrocketing costs. As AI Now’s journalist in residence Brian Merchant has argued in his latest report on the rise of AGI and the rush to find a working revenue model, Open AI heavily relies on enterprise customers who bet that generative AI will cut costs and improve productivity gains.
Overestimating productivity gains: Yet despite fervent hype to the contrary, AI has had almost no impact on the U.S. economy, with only 5-6% of businesses using it to produce goods or services (and only 6% planning to do so in the coming year). Moving from demos to demonstrable, high-value, high-productivity applications that would justify the cost remains a significant challenge. That leaves the most tempting corporate use-case: to cut labour costs with the help of AI - an austerity logic.
Market saturation: Many startups offer overlapping AI solutions with little to no competitive edge, further increasing market saturation and decreasing defensible profit margins. Combined with the signals of potentially laxer attitude towards mergers due to the increased influence of VC could lead to shakeout in the industry, as the financial firepower by the existing tech hyperscalers overpowers the fledgling AI startups funded by impatient VC capital looking for an exit.
Hoping for bigger models is not the answer:
Scaling laws in a real world: Expectations of ever-larger AI models having ever more capabilities rely on assumptions about the mathematical relationship between size and capability of the large-scale AI models. While the robustness of this assumption is suspect in the first place, in reality, financial and material constraints may soon curtail this trend.
Data saturation: Diminishing returns are likely as models reach the limits of internet-scale datasets.
What this means for the market:
We do caution against making strong predictions. Just because AI might be a bubble, it doesn’t mean it’s going to burst anytime soon, especially if parts of the new US administration now have a vested stake in keeping the market going and the cash flowing. The future is also shaped by power and not only by impersonal logic. We already see how the market is looking for alternative ways to extend the promise of the current large-scale AI trajectory, for example via a focus on ‘AI agents’, exploring with longer response times to make better use of existing models, a bigger focus on image and video generation in addition to text, and the exploration of new large-scale AI architectures.
A continued discrepancy between expectation and reality, however, comes with its own risks, especially when immature AI looking for profit is pushed to replace not just critical workers, but also the social systems where knowledge and established practices are situated, potentially leading to breakdowns that will be hard to rebuild (see Sarah’s comments at the Web Summit).
Even more interesting than speculation about an AI bubble burst, is what will emerge in its aftermath? While past tech bubbles, such as the dot-com bubble between 1995–2000 led to substantial disruption, it also laid the foundation for an arguably more mature market and sustainable growth to emerge.
What this means for Europe:
An existential question for policy makers in Europe is how much of the energy around industrial policy should be focussed on investing in specialized chips and cloud infrastructure that is optimized for developing and deploying large-scale AI. Instead of being on the ascendancy, the current hype cycle might be approaching its peak. Mistiming policy interventions and public investments can lead to waste and unused capacity, especially when the economics of the current AI market push towards smaller, more efficient models.
For the European AI market, a possible bubble means even more pressure to carve out a path for AI outside the narrow “bigger is better” paradigm that has dominated the sector in the past two years. In practice that could mean focussing on smaller models, or specialized use-cases.
Which approach to technological sovereignty will prevail in Europe?
2024 has been the year of Europe’s nascent industrial policy. Ideas of what this looks like in practice are still forming. Throughout the year, we’ve seen various proposals ranging from strengthening national AI champions, building a Eurostack, renewed calls for public digital infrastructure, digital public infrastructure, the Sovereign Tech Fund model of funding open-source fundamental technologies, and various proposals for public AI and public compute. Meanwhile, all leading hyperscalers have jumped on the bandwagon and are now selling “sovereign” cloud solutions.
The key distinctions between these initiatives are still murky, with different underlying political coalitions and beneficiaries. Some of the key points of contention are the centering of AI vs. broader digital infrastructure, the role of existing hyperscalers in these visions, the precise meaning of sovereignty (rules? capacities?), and how technology is embedded in broader visions for European societies. How these conflicting, and at times incompatible, questions will be solved, will be a defining feature for the European AI policy in 2025, as it will have tangible impact on how public resources get allocated.
Will AI agents re-shape digital markets?
“Agentic AI” is the fad of the moment and companies like Open AI would like to keep it this way throughout 2025. In Europe, this approach has been led by companies like the much-hyped H company in France. Simply put, AI assistants or agents are capable of autonomous decision making to achieve specified goals. What sounds like a massive leap in AI capacity, could be pushed as a tool to automated administrative tasks (think customer support, medical billing, or debt collection), or essentially as personal assistants that act as an intermediary between apps and websites.
The former raises a whole range of questions around bias, accuracy, and accountability, especially if used in sensitive social domains. The latter, however, also has the potential to further consolidate digital markets, as people would visit fewer apps and websites and instead concentrate their digital activity in fewer places. Combined with AI-powered search interfaces and chatbots that can further diminish website and app traffic flows, AI agents add an additional layer of uncertainty, especially for publishers.
Will the EU tackle concentrated power in the AI market?
From Draghi to von der Leyen’s political roadmap and mission letters, 2024 has been the year in which all signs point towards “a new approach to competition policy” in Europe. What this will look like in practice, is a key question for 2025.
In her first speech since taking office, European Commission vice-president for “a clean, just and competitive transition” Teresa Ribera signaled a revision of antitrust rules and state policy, with the goal of allowing EU companies to scale in global markets. We can expect a proposal on state aid rules by early March 2025.
What is less clear is to what extent this “new approach” will also lead to regulatory and enforcement actions that would increase contestability in the European AI market.
As Leevi has argued in his latest essay on Competition Policy and Large-Scale AI Markets in Europe, it is not enough to increase the number of players in the market. Curtailing concentrated power and fostering competition is necessary to “keep the space open for as-yet unforeseeable alternatives to emerge, to make the direction of technology amenable for democratic decision-making and shaping it towards applications that deliver in the public interest.”
In their October report on Preventing Monopoly Control Over AI, the Open Markets Institute and the Mozilla Foundation laid out a list of concrete remedies - from enforcing and strengthening merger control to forcing transparency through AI regulation and ensuring equal and fair access to critical inputs for AI development and commercialization - that could be a step in the right direction.
In other news
Increased defense spending
German AI company Helsing moves into the attack drone market.
Also, Germany’s biggest defense company Rheinmetall partners with US software specialist Auterion (funded by an influential German VC fund Lakestar) to develop necessary common operating standards for autonomous battlefield drones. These moves take place as the Ukraine war steers the tech sector in Eastern Europe from civilian products through to military uses ranging from weapons to civil defence systems.
Open AI exploring advertising
Open AI is exploring advertising in its drive to secure revenue, hiring advertising talent from big tech competitors such as Meta and Google.
Orange will become the first telecoms firm in Europe to have direct access to OpenAI's models.
Investments in the EU AI ecosystem
Silo AI cofounder Peter Sarlin, who sold his Finnish startup to AMD in July, is donating $10 million to back a new generation of AI researchers in Europe.
The French government has made a €500 million bid to acquire Atos’s advanced computing business, encompassing HPC, quantum, and AI activities, as part of a financial rescue plan for the French multinational IT services and consulting firm.
EuroLLM, a Horizon Europe and EuroHPC supercomputing supported 9B parameter AI model specialised in EU languages, is released and matches its comparative open source competitors created by Google and Meta.
London-based provider of AI-ready data centers and high performance AI cloud services, Nscale raises €146 million in Series A funding to accelerate expansion across Europe and North America.
In a striking example of how money is being made on the latest wave AI, we did also note that the Ireland-headquartered consulting giant Accenture currently generates more revenue of consulting on generative AI than OpenAI does building AI models.
Public sector use of AI
The Swedish and the UK government have deployed discriminatory AI systems for detecting welfare fraud, an investigation by the not-for-profit collaborative newsroom organization Lighthouse Reports, and the Guardian revealed.
Have a restful holiday season and see you in the new year!