What DeepSeek means for energy (and climate)
The AI acceleration is officially on. How screwed are we?
Last week, a small Chinese startup called DeepSeek accomplished the seemingly impossible: they unveiled an AI model that rivals ChatGPT while using dramatically less computing power, trained and developed with a fraction of the time and capital. Wall Street responded by dumping AI and energy stocks, betting on a future of lower AI energy demand.
But they’re wrong.
Instead, we’ll look back on this moment as one of the biggest catalysts for increasing energy demand in the digital era. DeepSeek's breakthrough is about to unleash an unprecedented surge in power consumption that will dwarf anything we've seen from the first wave of AI.
A staggering feat of cost destruction
DeepSeek's achievement is tremendous: they've matched ChatGPT's capabilities with one to two orders of magnitude less funding, chips, and training data by leveraging clever architecture choices and training techniques that squeeze more intelligence out of every calculation. Imagine building a car that runs on the energy of a bicycle — and cost as much to make. What would that do for mobility?
On the energy and climate side, many are already commenting that this will mean we’ll use less compute and therefore less energy.
In fact, the belief that there will be a lot less demand for power already drove a major selloff in power utilities stocks on Monday, with Constellation Energy shares down 20% and Vistra down 28%.
I won’t comment on trading strategies, but as a projection on energy demand, this is a mistake. Cheaper, more efficient, more accessible AI will exponentially increase the use cases for AI, thereby also increasing demand for its inputs.
Having a much cheaper model won’t change how retail users like you and I use ChatGPT, but for developers who are looking to embed AI into other use cases where businesses must eek out additional margin on top of what they’re already paying for inputs, this is a game changer.
A company like Perplexity could go from paying OpenAI $4.40 per million tokens to paying DeepSeek just $0.10 per million tokens. This is a 130X reduction in cost, and has massive ramifications for the kinds of companies that will be built on top of models like these, as well as how many of them get built.
For a robotics company looking to integrate AI into its product, which is both costly to invent and to produce, being able to shave down cost on an critical input like intelligence could help it cross the viability threshold.
In other words, cheaper, more efficient AI means more AI overall, not less. And exponentially so. That means more energy demand — so much more that it will dwarf any energy savings from using even dramatically more efficient models.
Let’s talk about why.
There’s a reason why all the VCs are talking about Jevons Paradox today
In 1865, economist William Stanley Jevons noticed something counterintuitive about coal-powered steam engines. When engineers made them more efficient, coal consumption went up, not down. It turned out that increases in efficiency made steam power cheap enough for whole new industries to be born.
Today, we're about to witness Jevons Paradox on an unprecedented scale with AI.
Jevons Paradox has been widely cited as an argument for why things like hybrid cars and LED lights don’t decrease energy usage even though they increase efficiency.
But in reality, these are two areas where Jevons Paradox hasn’t held up as well. LEDs have decreased energy consumption from the lighting sector, even as they increase demand for lights everywhere, all around the world. This is despite the dramatic expansion of use cases into everything from Christmas lights to Las Vegas to screens to phototherapy. LEDs are an exception to Jevons’ Paradox, or perhaps a misapplication.
AI, on the other hand, will be its poster child.
If AI went down to 1/100th of its current cost to produce and deliver, energy consumption from AI would increase rather than decrease — even if energy efficiency improved dramatically. That’s because Jevons Paradox applies much more aptly to AI than it does to LEDs, hybrid cars, or any other physical product.
Here’s why.
1. AI use cases are still in their infancy
AI is in its early stages, with massive untapped demand across industries and applications that we haven’t even imagined yet today. This is unlike lighting when LEDs were introduced, or cars, when hybrids came around.
We could see:
Real-time AI personal assistants for every person
Ubiquitous AI-generated media (videos, music, digital twins)
AI-driven automation in industries where inference costs are currently too high (e.g., robotic surgery, AI-powered factories)
Autonomous vehicles at scale
AI-enhanced gaming and virtual reality with persistent, AI-driven NPCs
An explosion in new use cases would drive higher net energy demand, even if individual AI computations became vastly more efficient. And again, these wouldn’t be linear increases.
2. AI is unbounded by physical constraints
LED-driven energy demand declined because the total number of useful lighting applications was already somewhat constrained.
For example, even though LEDs are 1/10th the cost to run, and lighting has certainly increased worldwide, it’s highly unlikely that the sum total of global lights has increased 10X, because each LED light still serves the same performance goals as an equivalent, 10X more energy intensive incandescent bulb. In other words, we have a natural, physical ceiling on how many lights we need and want.
Even with some significant new use cases and new users — from LiDAR to LED phototherapy to ultra thin screens, LEDs’ scalability is ultimately limited to physical devices that drive improvements within existing industries.
AI on the other hand has no such constraints.
Although it has physical inputs, AI itself is not a physical resource. It can scale indefinitely with demand.
AI also creates demand for itself in important ways. Cheaper, more accessible and ubiquitous AI would lead to AI-generated content, AI assistants, AI-generated code — in other words, AI would create AI-powered means of production that together generate more AI demand in a feedback loop.
And speaking of feedback loops, AI also benefits from a positive user-generated content loop where more users and more usage => more data => more precise, more relevant AI => more users and usage.
Finally, AI fundamentally unlocks new capabilities that are difficult to measure within the current market. How valuable will AI-powered drug or materials discovery be? What about self-improving manufacturing robots that learn as they build? How do we project demand for a brand new industry that doesn’t exist yet?
As AI gets cheaper and more efficient, energy demand will almost certainly expand at a pace that outstrips even the most impressive efficiency gains.
3. Open source means more players and more games
DeepSeek’s model is open source, meaning it’s publicly available, modifiable and auditable. Rather than being limited to company-chosen use cases, it can be adapted to any use case that anyone out there can think of.
The IEA has already said that data center electricity demand could double to 1,050 terawatts by 2026, driven by AI. But that was before an open source model blew the AI application layer wide open.
Could AI help drive the clean energy transition? Three possible futures
Is there a version of all of this where the energy efficiency benefits of a DeepSeek-esque approach buy time for renewables and nuclear to ramp up just in time to meet the explosion of use cases that such an approach to AI will inevitably usher in?
It depends on three key factors.
How fast new AI applications emerge (ie, the demand ramp-up speed).
How fast we deploy clean energy, upgrade grids and roll out today’s nascent innovations (ie, the supply response).
Whether AI can drive energy breakthroughs as AI leaders like Jensen Huang have suggested (self-reinforcing solutions).
Here are three scenarios:
Scenario 1: Clean energy catches up
In the first and best-case scenario, AI’s energy demands scale gradually enough to give the renewables and nuclear sectors a chance to adapt and grow into demand.
Maybe we get lots of amazing ideas, but enterprise adoption lags behind due to classic organizational change management challenges, along with the realities of existing infrastructure and the difficulties of adapting to business model changes.
Meanwhile, we’re deploying clean energy faster and faster while also making breakthroughs across every clear energy category, from perovskites to unlock game-changing efficiency gains in solar to next-gen nuclear fission with technology and form factors that unblock fuel and deployment bottlenecks.
In this scenario, AI still increases total power demand, but not faster than we can build clean energy supply. We get a net decarbonization alongside really cool, high-empathy AI companions rather than a runaway emissions increase accompanied by Terminator.
Scenario 2: Fossil fuels stage a comeback
In the second and worst-case scenario, AI demand grows exponentially on the back of an explosion of cheap supply, overwhelming clean energy deployment and empowering a fossil fuel resurgence and accompanying emissions spike.
Maybe we get all those ideas, people and businesses love them — a little too much — and the only way suppliers can adapt is to grow inputs by any means necessary. If AI becomes 100x cheaper overnight, demand for AI services could surge faster than infrastructure can keep up, pushing both utilities and behind-the-meter power consumers to fall back on the best known, fastest way to build new generation: combusting fossils.
In this scenario, AI not only accelerates emissions growth in the near term, but could also lock in new build fossil fuel infrastructure for decades to come. We get more emissions, and maybe better AI-driven climate modeling to help us understand precisely how screwed we are.
Scenario 3: AI drives the energy revolution
In the third and most techno-utopian scenario, AI helps to solve the energy challenge it creates — and then some.
AI is already being used to optimize energy systems — from grid balancing and demand-side management to optimizing renewable energy distribution and enabling better, faster research on potential deep energy breakthroughs in fusion and next-gen materials.
If AI can accelerate breakthrough energy technology faster than it increases demand, it could not only balance out its own energy impact, but actually create energy surplus with less environmental and climate cost.
In this scenario, AI might initially cause some power grid strain and friction, but over time, it’s able to drive previously unimaginable advances in energy that far outweigh the cost to get there, and those improvements might be so dramatic that they lead to the energy abundant, materials-independent future that techno-optimists everywhere have been manifesting for years.
To make this one happen though, we need to be making proactive investments in AI-for-clean energy, which reads as a good opportunity for climate investors who’ve been in the doldrums lately.
The AI acceleration has begun
Whether it’s DeepSeek or OpenAI or Meta with another open source model, it’s clear that AI innovation is happening at a pace that’s surprising even insiders.
If the future moves towards open source models, which is currently looking likely, then we’ll see an even bigger and faster proliferation of developers, applications, and demand.
Regardless of which model, which company or which country wins, AI at its core is an infinitely scalable machine that transforms megawatts into computational and geopolitical power.
DeepSeek’s breakthrough demonstration won’t change what AI is or does — it will accelerate it.
This acceleration brings us to a crossroads of (at least!) two revolutions: AI and clean energy. The decisions we make in the next 24 months will determine whether these forces amplify or undermine each other.
When I reflect on my own minuscule role in all of this, I’m both proud of the companies we’ve backed — companies like Swift Solar, Euclid Power and Bedrock Energy among others who are all run by kind, dedicated and brilliant clean energy entrepreneurs — and soothed by their very existence. That they are excellent and succeeding despite the gravitation pull that tugs on every startup gives me hope that smart humans with good intentions still have a shot at shaping our trajectory. We may be skeptical of the techno-optimists, but right now we need them to be right.
Somewhat weirdly, I do think that massively cheaper AI will give us a reduction in energy. Use. The reason is that although energy use for AI will undoubtedly go up, one of the applications of AI will be to reduce energy use everywhere else. And my guess is that the improvement in efficiency from applying AI to the rest of the economy will outweigh the increase in energy use on AI itself.
ngl I feel like Deepseek catalyst will in the short/medium run further incentivize tech firms to try and build out AI infrastructure (energy mainly) as fast as possible, which in the US means natural gas generators. There’s a reason why XOM/CVX have seemingly decided to reallocate some attention to nat gas generator build out. The worst case scenario is happening as we speak. Yes, tech firms will sign agreements to purchase power from renewable energy projects (which will take years to get running) and in the meantime compete for available 24/7 sources (fossil fuels). Great read nonetheless!!