Executive Summary
The AI hype is fueled by three actors: researchers advancing new technologies, investors deploying billions, and startups riding the trend. While the industry scales into language models like Chat-GPT, pioneers warn of technological limits and inefficient approaches. Financial flows are driven by FOMO (Fear of Missing Out) and unproven valuations, despite many companies lacking products.
People
- Llion Jones – AI researcher, co-developer of Transformer technology
- Margit Wennmachers – venture capitalist, Andreessen Horowitz
- Andreas Göldi – venture partner, B2 Venture
Topics
- Generative AI and language models
- Venture capital financing
- Technological limits of AI models
- Venture capital dynamics and market bubbles
Clarus Lead
Three groups of actors drive the current AI boom: researchers developing new technologies, venture capitalists investing billions, and startups needing an AI story to get funded. The article reveals a fundamental discrepancy: while investors, driven by fear of missing out, pump massive resources into resource-intensive language models, leading developers like Llion Jones warn of technological dead ends and the lack of genuine artificial intelligence. This asymmetry between capital flow and actual technological progress points to classic bubble formation.
Clarus Own Analysis
Clarus Research: The article documents concrete financial flows: globally, $45 billion flowed into AI investments in Q3 2025 – nearly double the previous year (Crunchbase data). Over 2,000 new AI companies were financed in 2024 (+8.4%). An early warning sign is the phenomenon of startups like Thinking Machines Lab achieving billion-dollar valuations before a product exists.
Classification: The AI hype follows known patterns from previous tech cycles (Dotcom, iPhone, Social Media), but is larger according to industry veterans. The central risk lies in the discrepancy between expectations and technological realities: language models don't understand the world like humans, but recognize statistically-probabilistic patterns. The hope for "the next breakthrough" could prove unfounded.
Consequence: For decision-makers this means: (1) caution toward inflated valuations, (2) distinguishing between genuine technological progress and marketing narratives, (3) attention to alternative AI approaches (swarm intelligence, energy-efficient hardware) as possible future technologies.
Detailed Summary
The Developers: Warning Against Saturation and Misplaced Priorities
Llion Jones, Welsh AI researcher and co-author of the groundbreaking paper "Attention is all you need" (2017), embodies growing skepticism among technology pioneers. Jones co-developed the Transformer architecture that Chat-GPT is based on. The paper describes how language models understand relationships between words and predict the most likely next word – an ability that increases with more training data.
Yet today Jones warns of saturation in AI research. In his view, the industry is wasting valuable time by focusing too heavily on language models. He has founded his own company called Sakana AI, which aims to make machine intelligence more similar to human intelligence. Jones emphasizes a fundamental problem: language models make surprisingly inhuman mistakes and don't understand the world like humans do. Example: a model struggles to depict a pelican riding a bicycle – because it has learned that normally the pelican is ridden. Children, by contrast, solve such combinatorial problems through imagination, with a fraction of the data and energy.
Jones criticizes investors and the tech industry for betting on the "safe bet": scaling more, larger models, more data centers. This leads to more innovative, efficient approaches being overlooked. The reasons for this focus are economic: those researching language models get more research funding, better publication opportunities, and the technology works and generates money.
Investors: FOMO and Rational Irrationality
Margit Wennmachers, long-time partner at venture capital firm Andreessen Horowitz, has witnessed several tech booms: Dotcom bubble, iPhone moment, social media. They are all history now. The AI wave, she says, is "probably bigger than all other technology waves combined."
For venture capitalists, the logic is understandable: yes, there will be many losers, but the few winners will generate billions. One or two superstars can offset the losses from all other investments. Therefore: "It would be irresponsible not to invest in AI now."
However, Wennmachers also identifies a psychological component: every venture capitalist lives in constant fear of missing a deal. This leads to FOMO-driven decisions. Add to this new investor types – "tourists," as Wennmachers calls them – who don't know tech but just jump on the hype. They come from banking or other industries and see AI as the next safe bet.
The numbers demonstrate the extent: in Q3 2025, globally deployed venture capital rose 38 percent to $97 billion, of which $45 billion went to AI – nearly double the previous year. Simultaneously, over 2,000 newly financed AI companies emerge per year.
Andreas Göldi, partner at Swiss fund B2 Venture, observes a new phenomenon: a somewhat interesting idea on a PowerPoint presentation is enough to raise a few million francs. The problem follows quickly: many startups fail to meet implicit expectations. Since AI develops so rapidly, it has also become harder for investors to formulate clear theses. Large corporations and established AI providers quickly dominate niche areas like legal AI applications or programming software. Göldi therefore looks for genuine niches – such as his investment in Zurich startup Nautica Technologies, which uses robots to clean ship hulls.
Startups: AI as Funding Necessity
For founders, the message is clear: "Startups that don't have a credible AI story have a hard time getting funding."
The startup Frontline by Ba-Linh Le shows how this works in practice. Le and her co-founders developed a risk assessment tool for authorities to prevent domestic violence. It estimates physical, emotional, and financial violence risks based on detailed inquiries. Originally, AI wasn't planned – they tested various algorithms and statistical methods. Then they realized machine learning improved their application: the AI tool identifies particularly at-risk individuals 30 percent more accurately than conventional tools.
The AI connection gives Frontline visibility. Le says: "In the domestic violence prevention sector, we're known as the organization with the 'AI tool.'" But she also warns against reducing AI's added value to a catchphrase. In violence prevention, responsible use is crucial. The current hype is an advantage for startups, but should be approached with caution.
Key Messages
Capital flow exceeds technological progress: $45 billion flowed into AI in 2025, yet many startups have no products or clear business models.
Fundamental technological limits: Language models don't understand the world like humans; they recognize statistical patterns. Acknowledging this is central to realistic valuations.
FOMO drives irrational decisions: Investors fear missing deals and make risk-happy decisions without solid foundations.
Alternative technologies are ignored: Researchers warn that swarm intelligence, more efficient AI chips, and other approaches are overlooked because scaling language models is more profitable in the short term.
Historical pattern: AI hype follows known bubble cycles (Dotcom, iPhone, Social Media), but could be bigger this time.
Stakeholders & Affected Parties
| Group | Profile | Opportunities | Risks |
|---|---|---|---|
| Researchers & Developers | Pioneers like Llion Jones | Research funding; visibility | Loss of direction from market pressure; resource flight to large corporations |
| Venture Capitalists | Investors like Wennmachers, Göldi | Potential excess returns on right bets | Portfolio total losses; bubble bursts; reputational damage |
| AI Startups | Founders like Ba-Linh Le | Easier access to funding; market attention | Overvaluation; pressure to fulfill promises; rapid decline when hype ends |
| Large Corporations (Tech) | Nvidia, Google, OpenAI | Hardware demand; data monopolies | Infrastructure over-investment; alternative technologies could replace |
| Public / Regulators | End users, authorities | Efficiency gains; security innovations | Blinded by hype; unsafe, misunderstood systems in critical areas |
Opportunities & Risks
| Opportunities | Risks |
|---|---|
| Breakthrough to genuine AGI: Massive investments could lead to genuine artificial intelligence. | Market bubble: Valuations are unsustainable; major crash if hype collapses. |
| Efficiency gains in industry/authorities: AI tools like Frontline's could bring real benefit. | Capital misallocation: Billions flow into resource-hungry, inefficient technologies instead of innovative alternatives. |
| Hardware innovation: New AI chips with photons instead of electrons could reduce energy consumption. | Big Tech dependency: Large corporations monopolize AI applications; startups have fewer chances. |
| Alternative approaches could prevail: Swarm intelligence or more efficient architectures could replace older approaches. | Technological failure: AI models don't really understand the world; fundamental limits could be unreachable. |
| Industry standards & regulation: Clarity could slow hype cycles and promote sustainable innovation. | Crisis of trust: If bubble bursts, AI sector could take years to regain trust. |
Action Relevance
For Investors
- Step 1: Distinguish between genuine product progress and marketing narrative. Check: Does the startup have a functioning product or just an idea?
- Step 2: Actively monitor alternative technologies (swarm intelligence, efficient hardware). Not all winners must be language models.
- Step 3: Warning sign: valuations exceeding 50x revenue for pre-revenue companies. Set milestones.
- Indicator: AI startups without customer metrics after 18 months are high risk.
For Founders
- Step 1: AI story is a door opener – but not business