Summary
The AI hype is fueled by three actors: ambitious researchers, risk-taking investors, and startups that can barely secure funding without an AI narrative. While billions flow into scaling language models, pioneers like Llion Jones warn of technological limits and inefficient resource use. The market is in a classic speculative bubble, driven by fear of missing out – yet real innovation could emerge elsewhere.
People
- Llion Jones (AI Researcher, Sakana AI)
- Margit Wennmachers (Andreessen Horowitz)
- Andreas Göldi (B2 Venture)
- Ba-Linh Le (Frontline)
- Mira Murati (Thinking Machines Lab)
Topics
- Generative AI and language models
- Venture capital financing
- Tech startup ecosystem
- Technological limits of AI
- Speculative bubbles in technology
Clarus Lead
The current AI boom is primarily a finance-driven phenomenon, not a technological one. Consider Thinking Machines Lab: a company with no product that raises two billion dollars five months after founding. This is symptomatic of a market where expectations exceed reality. Three actors drive this dynamic – and risk substantial losses while simultaneously hoping for the big breakthrough.
Clarus Original Research
Clarus Research: Venture capital investments in AI rose 38 percent in Q3 2025 to $97 billion globally; $45 billion flowed directly into the AI industry – nearly double the previous year. Simultaneously, leading AI researchers warn that current language models have fundamental understanding limits and learn significantly less efficiently than human brains.
Classification: The discrepancy between capital flow and technological progress reproduces classic bubble patterns (Dotcom, iPhone cycle, social media). Investors push scaling forward while researchers see alternative, more efficient approaches (swarm intelligence, photonic chips) underfunded.
Consequence: For companies and public institutions: selective investment decisions are required. For startups: AI as a buzzword is necessary but not sufficient – substance matters long-term.
Detailed Summary
The Researchers: Innovation Under Pressure
Llion Jones is co-author of the legendary paper "Attention is all you need" (2017), which describes the transformer architecture – the technological foundation of ChatGPT. Today he warns of saturation in AI research. Language models make systematic errors that are surprisingly "inhuman." They don't understand the world like humans do. A child can imagine a pelican riding a bicycle – an AI model cannot, because it needs trillions of data points to invert contexts. The human brain computes exponentially more efficiently with fractions of the data and energy.
Nevertheless, the entire industry focuses on language models. Why? Jones answers directly: Because they work and make money. Whoever researches transformer technology receives research funding, publication opportunities, and investor attention. Alternative approaches (artificial general learning, neuron-like architectures) are systematically underfunded. His company Sakana AI relies on "swarm intelligence" – many specialized AI agents working together. Whether that comes closer to biological intelligence remains to be seen. Jones is convinced: the next major breakthrough won't come from larger language models, but from fundamentally different architectures.
The Investors: Rational Irrationality
Margit Wennmachers has lived through the Dotcom bubble, the iPhone moment, and social media hype. She says clearly: The AI wave is bigger than all the others combined. For venture capitalists, the logical conclusion follows: many startups will fail and lose money. But the few winners will generate billion-dollar profits – and multiply the initial investment. That's the game. That's the hope.
However: current valuations are breathtaking. Startups without a product are suddenly worth billions. Wennmachers relativizes: successful companies "grow into their valuations" – investments justify themselves later. You lose money only once, but can multiply it tenfold.
The problem lies in psychology: Every venture capitalist lives in constant fear of missing a deal. FOMO (Fear of Missing Out) is the driving force. Add to this new investors – Wennmachers calls them "tourists" – who previously worked at Goldman Sachs and are now jumping on tech because it's hyped. This drives prices even higher.
Numbers show the acceleration: The number of funded AI startups rose in 2024 by 8.4 percent to over 2,000 (AI Index Report, Stanford 2025). Global venture capital invested in Q3 2025 climbed 38 percent to $97 billion – of which $45 billion went to AI startups.
Andreas Göldi (B2 Venture) experiences reality daily: a mediocre idea on a PowerPoint presentation is often enough to raise a few million francs. Many startups then fail to meet implicit expectations. AI evolves so quickly that investor theses become obsolete. Yesterday, legal tech with AI was the safe bet; today OpenAI and Google do it. Göldi searches for niches – like the Zurich startup Nautica Technologies, which uses AI robots to clean barnacles off ship hulls.
The Startups: AI is Survivability
Without a credible AI narrative, it is now almost impossible for a startup to be funded. Ba-Linh Le founded Frontline – a tool for analyzing violence risk for authorities in the violence protection sector. She didn't plan to use AI. She tested various algorithms and found: machine learning identified high-risk cases 30 percent more accurately than conventional methods.
The AI connection gives her visibility. Le says: "In the violence protection sector, we are known as the organization with the AI tool." That's an advantage – people understand the term, investors like to hear about it. But Le warns: AI must not be degraded to mere buzzword. In violence protection, responsible use matters. Nevertheless: currently there is hype, and startups must use it.
Key Statements
Finance-Driven Hype: Billions flow into scaling language models while alternative, more efficient technologies remain underfunded.
Psychological Bubble: FOMO and fear of missed deals drive irrational valuations – startups without products are worth billions.
Technological Limits: Leading AI researchers warn that language models have fundamental understanding limits and don't achieve the efficiency of human learning.
Necessary AI Connection: Startups without an AI story don't get funded – the narrative is vital for survival, not the technology.
Next Breakthrough Elsewhere: Swarm intelligence, photonic chips, and more efficient architectures could end the resource-hungry transformer era.
Stakeholders & Affected Parties
| Stakeholder | Involvement | Risk | Opportunity |
|---|---|---|---|
| AI Researchers | Benefit from funding and attention | Alignment with mainstream technology; innovation delays | Breakthrough with new approaches |
| Venture Capitalists | Massive capital allocation | Portfolio losses if bubble bursts | Trillion-dollar gains with winners |
| Tech Startups | Must have AI narrative | Financing pressure; artificial requirements | Rapid scaling if successful |
| Big Tech (OpenAI, Google, Nvidia) | Dominate with scale and chips | Cannibalization by more efficient competition | Entrench monopoly position |
| Public Sector | Regulatory risk | Too-late innovation; dependence on private players | Build technology sovereignty |
| General Public | Users of AI tools | Data privacy, bias, job losses | Productivity gains, better services |
Opportunities & Risks
| Opportunities | Risks |
|---|---|
| Breakthrough Technologies: Swarm intelligence and photonic chips could trigger efficiency revolutions | Bubble Collapse: Massive capital losses from unjustified valuations |
| Application Innovation: Startups like Frontline show AI solves real problems (violence prevention, ship maintenance) | Resource Waste: Trillions in inefficient data centers instead of fundamental research |
| Broader Accessibility: AI tools become more productive and cheaper | Existential Dependency: Startups must use AI whether sensible or not |
| Competition Reassessment: Alternative architectures could break Nvidia dominance | Investor Panic: FOMO can lead to catastrophic mistakes |
| Regulatory Clarity: Pressure could create frameworks (AI Act, EU) | Innovation Delays: Capital concentrates on safe mainstream bets |
Action Relevance
For Investors and Funds
- Action: Diversify technology thesis. Not just transformer scaling, but 20–30% of AI funds into fundamental research (swarm intelligence, quantum-inspired AI, photonic chips).
- Indicators: Monitor publications at top conferences (NeurIPS, ICML), patent activity in alternative architectures, energy efficiency metrics.
For Startups
- Action: Use AI narrative, but prioritize substance over buzzwords. Solve real problems (efficiency, accuracy, cost-benefit).
- Indicators: Measure productivity gains, error rate reduction, return-on-investment vs. baseline methods.
For Enterprises and Public Institutions
- Action: Selective AI adoption. Don't jump on the hype, evaluate: Does AI solve a critical problem? Is ROI demonstrable? Are there compliance risks?
- Indicators: Pilot phase with clear success criteria, cost-benefit analyses, risk assessments for privacy and bias.
For Regulators
- Action: Maintain technology neutrality. Regulate not just transformer models but also encourage alternative approaches.
- Indicators: Monitor patent landscapes, research funding distribution, AI infrastructure energy consumption.
Quality Assurance & Fact-Checking
- [x] Central claims and figures verified
- [x] Unconfirmed data marked with ⚠️
- [x] Web research for current data conducted
- [x] Bias or political one-sidedness flagged
Verified Data:
- SoftBank plans $30 billion investment in OpenAI (Bloomberg source, 2026)
- AI Index Report 2025 (Stanford): 8.4% increase in funded AI startups to over 2,000
- Crunchbase data Q3 2025: $97 billion global VC, of which $45 billion in AI