Valentijn van Nieuwenhuijzen: Are sweet dreams made of AI?
By Valentijn van Nieuwenhuijzen, Investment Strategist and former CIO of NN Investment Partners/Goldman Sachs Asset Management
If you only listen to the news headlines, you might think AI has already rewritten the rules of our economy. Every earnings call talks about an AI strategy. Every board presentation has a slide on AI. Every country wants an AI plan.
At the same time, if you look at the hard data, like productivity, profitability or wage growth, the revolution is much harder to spot. Task-level experiments look impressive, but GDP and productivity still look stubbornly normal.
So, how to separate the AI signal from AI noise?
A good starting point is to realise that three very different conversations are constantly being mixed together: what the technology can do, what the business economics look like, and what the wider system effects might be. The signal lies in how those three layers interact. The noise is what happens when we collapse them into one big story.
On the technology layer, there is quite a strong signal already. Generative AI has gone from a quirky demo to an everyday tool at a remarkable speed. Coding copilots autocomplete and refactor code. Chat-style assistants summarise long documents, draft emails and help with research. Translation and transcription have quietly become good enough that many people stopped noticing them as ‘AI’ at all. There is no serious argument left that AI is just smoke and mirrors at the task level. That is signal.
The noise starts when we leap straight from the idea that AI helps with tasks to AI will also produce a productivity boom. History suggests a lag. Electricity and computers both took decades to show up in the productivity statistics because organisations had to be rewired around them. We are only at the beginning of that organisational process for AI.
The second layer is economics and business models. Here the signal is almost the opposite: the technology is dazzling but the business model is still murky.
The last big generation of internet winners benefited from a magical combination of near-zero marginal costs and strong network effects. Once the infrastructure was built, one more user or one more search query costs almost nothing, and every new user made the platform more valuable to everyone else. That is how we ended up with a handful of firms earning monopoly-like margins.
Large language models do not obviously share those economics. They have real variable costs as more usage means more computation, more energy, more hardware. Network effects are hardly there as an extra user doesn’t automatically make the model more valuable to others. Switching costs are also not that high, which caps how much any one provider can charge.
The signal is not that nobody will make money with AI. The signal is that the path to hyperscale economics – the kind investors became used to in search, social media and digital marketplaces – is far less clear this time. So, don’t focus on AI capex headlines, but focus on eventual return on invested capital. Until we see a few years of hard numbers on that, most of the valuation narrative is noise.
The third layer is system effects: what AI does to markets, information, and work. Here it helps to be neither utopian nor apocalyptic. AI will change work. Routine tasks in journalism, customer service, and basic admin are under pressure. Parts of consulting, software development and design are being redefined as ‘co-pilot’ roles. New AI related jobs are emerging. That is all signal.
But the story doesn’t end at jobs lost or gained. The deeper questions are how fast, where, and for whom? If new jobs appear in different regions and sectors than the ones where old jobs disappear, there will be frictions, retraining costs, and political backlash. If most of the gains accrue to a small group of firms and shareholders and are largely saved rather than spent or invested, AI can be ‘productive’ on paper while still doing very little for broad-based prosperity.
On the risk side, AI is clearly a complexity amplifier. Financial markets already run on layers of opaque algorithms and adding new black boxes on top does not necessarily make them safer. Information ecosystems already struggle with misinformation and automating the production of convincing fakes will not help.
The demos, the valuations and the big promises are the loudest part of the story. The signal sits quietly in in productivity data, in profit statements and in labour-market outcomes. The computer-story of the 1980’s reminds us how long it might take before the signals become clear. The music from that age teaches us how ‘Sweet dreams are made of this’ and how hard it is to disagree with them. But until those signals become clearer, the smart position on AI is to be excited about what it can do, sceptical about what it has already done, and selective on the messages to take away from the dream