Whales, Black Swans and Snakes

Same as before, more than before

Remember the pandemic? The new normal, we'll come out of this changed, nothing will be the same again. For a while we really believed it.

Nathaniel Bullard has put together a set of data showing that through 2021 and part of 2022 we paid more attention to sustainability, we cut emissions, we consumed less, we picked up habits with a lighter footprint. Greentech companies received record funding, executives worked sustainability into their speeches and their reports.

Then it ended. From 2022, once the sharpest phase of the pandemic was behind us, energy-hungry habits and consumerist lifestyles picked up again.

The new normal looks a lot like the old normal, except with AI. Which is to say, a technology that needs enormous amounts of energy to work.

Fear and loathing in the markets

At the end of January the launch of DeepSeek, a new generative AI model developed in China, blew up the tech press, the stock indexes and the LinkedIn feeds.

The model cost less than 6 million dollars, a figure infinitely below the multi-million budgets of companies like OpenAI. It was trained using only 2,000 Nvidia H800 chips, much less powerful than the chips its Western competitors use.

Everyone talked about the collapse of Nvidia stock and of the other tech companies riding the AI boom. But the shock didn't spare the energy sector: the Chinese model, more efficient and therefore less energy-hungry, threatened to bring down the energy consumption tied to AI — and with it the profits of the utilities.

But is that really how it will go?

The day after DeepSeek launched, Nvidia stock closed down 17%. The Nasdaq, which is weighted towards tech, lost 3.1%; utilities lost 3.5%.

The consequences of efficiency

DeepSeek's AI model shocked the world because it showed performance comparable to the most advanced models, obtained at a far lower cost.

DeepSeek is a "lighter" model: it needs less computing power to run, less sophisticated hardware, and therefore less energy.

On top of that, it's an open source model: it runs on a decentralised network, and anyone can copy it, adopt it, and build new features and services on top of it.

Those very characteristics, though, can set off on the energy side a phenomenon known as the Jevons paradox:

greater efficiency drives AI to spread everywhere;
spreading everywhere increases the number of users and services;
total energy demand goes up, even though each single unit consumes less.

The wrong side of history

So far the spread of artificial intelligence has been an anomaly in the history of the digital revolution. All the startups that changed the rules of the game, the ones that created the myth of Silicon Valley, were born "from below", in some teenager's bedroom or in Californian garages, among cans of soda and pizza boxes.

What has driven the development of artificial intelligence, instead, are operations hyper-financed by large companies and investment funds — like OpenAI, backed by Microsoft from the start. Or the tech giants themselves, in-house, like Google or Meta.

So AI was born and grew up inside a proprietary, speculative logic. Very far from the open source logic that drove the digital innovation of the early days. So much so that Sam Altman himself came to say they had been on the wrong side of history, and that the future belongs to open source models.

DeepSeek did exactly that: it brought AI into the open source domain. A move that could change everything.

Swans and snakes

The launch of DeepSeek — whose logo is a whale — happened in the days of the Chinese new year, and the start of the year of the snake. But the effects it produced brought a different animal to mind: the black swan.

Nassim Taleb calls "black swans" the events considered highly improbable. The ones that, when they do happen, land with enormous force, capable of overturning every forecast and changing the expected course of history. Remember the pandemic?

The speed with which Large Language Models spread, and the disruption they promised, already looked a lot like a black swan. But the arrival of DeepSeek and the prospect of new open source models could be a black swan inside the black swan.

Because it opens an unprecedented pattern even in the forecasts about AI's development.

But above all because it would be able to accelerate innovation and open up applications that are unthinkable right now.

Are we ready?

We are fragile

Judging by what happened over the past few weeks: no, we're not ready. Our economic systems are fragile, to use another concept dear to Taleb: they don't have the flexibility and the resilience to absorb shocks and sudden changes of paradigm.

It happened with the pandemic, which found us unprepared. And it's happening with AI, which has caught pretty much everyone unprepared: the economy, work, schools, politics. And the energy system too, thrown into a fibrillation by the expectation of rising energy demand.

The solution, as Taleb says, is to build antifragile systems. Systems that aren't monolithic, that can react to volatility and change.

Open source and widespread access to technology, in this sense, can help us build our antifragility.

Because they increase optionality and the variety of solutions available to us.

Back to the garages

Concentrating the power of AI in a few hands makes the system extremely fragile. The opposite, "democratisation", could make it more resilient, because it would open it up to a great many entities experimenting with new solutions.

The founding principle of antifragile systems is: more alternative ways to reach the same result.

For AI technologies that means distributed models, more room to move for new startups and labs everywhere in the world. More garages, in short, and fewer concentrations of power.

But this movement towards openness and plurality is the same one we hope for in energy. Many nodes, small and large, connected in a network, which make the system more robust.

Going back to the garages doesn't mean taking a step backwards, it means exploring new paths without waiting for the giants to trace them. It's the key to increasing the alternatives and making the system more resistant. If one model collapses, a thousand others are left experimenting. If a large power plant stops, the local grid stays up. That's how we build antifragile systems: where open source and widespread creativity become our most solid defence against every next, unpredictable, black swan.