Skip to main content

Learn

Why dynamic GPU pricing?

Most cloud GPUs cost the same per hour no matter when you run them, whether it's 3am or 6pm, a windy Sunday or a still weekday evening. Dynamic pricing is the idea that they shouldn't, that the price of compute should move with the real cost of the electricity running it, and fall when that electricity is cheap and clean.

What We're Offering

Sovereign, secure, and dynamically priced GPU compute

The future of AI infrastructure isn't just faster, it's smarter about when and how it runs. Leafcloud is pioneering GPU compute that works with the energy system, not against it. We keep your data sovereign and your workloads secure.

Dynamically Priced

Grid-Coupled Pricing

GPU pricing that moves with the grid. Prices fall when renewable power is abundant, and rise when it's scarce. This isn't just cost optimization. It's compute that helps build the energy system AI will need to scale sustainably.

Sovereign

European Infrastructure

European infrastructure designed for a world where data jurisdiction matters. Your compute stays under EU law. Your data never crosses the Atlantic. Your compliance is built in from the ground up.

Secure

Enterprise-Grade Security

Enterprise-grade security that doesn't compromise. ISO 27001, SOC 2 Type II, full GDPR alignment. The frontier of AI requires infrastructure you can trust without question.

The Real Cost

First, the thing nobody puts on the price tag: energy

When you rent a GPU, you're paying for more than the chip. You're paying for the power to run it, the cooling to keep it alive, the network around it, and the people maintaining it. Electricity is a real and rising share of that. At the scale AI is now reaching, it's becoming one of the defining costs of the whole industry.

Global data-centre electricity to double by 2030

From about 485 TWh in 2025 to around 945 TWh by 2030, slightly more than Japan's entire electricity consumption, with AI as the single biggest driver.

AI data centres demand tripling

Electricity demand from AI-optimised data centres alone is projected to more than triple over that period.

20%+ of all demand growth

In advanced economies, data centres account for more than 20% of all electricity-demand growth to 2030.

And most of that cost is hidden

The headline GPU spec tells you almost nothing about what you'll actually pay to run it. Energy isn't a line item; it's the majority of the bill, and it's growing with every generation.

The chip is only about a third of the bill

Over a multi-year life, the GPUs in a large cluster account for only around 35% of the true total cost of ownership. The rest is power, cooling, networking, and the people maintaining it, so the sticker price of the hardware hides most of what you actually pay.

And energy's share keeps growing

Each new generation of accelerator draws more power than the last; today's flagship cards pull roughly 700 to 1,000 watts each. The more capable the GPU, the larger the slice of its lifetime cost that is, quietly, electricity, which is why how that electricity is priced matters more every year.

Grid Dynamics

Why electricity prices move so much (and why that's an opportunity)

Wholesale electricity isn't one price. It changes every hour, and on a grid built around renewables it swings hard. In the Netherlands, wholesale power has increasingly gone negative, meaning the grid has so much clean energy it effectively pays users to consume it.

314 hours of negative prices in 2023

Rising to 458 in 2024, and 474 in just the first eight months of 2025.

800–1,200 hours forecast for 2026

As Dutch solar capacity rose roughly fivefold between 2018 and 2023 and wind roughly tripled.

Predictable patterns

Solar pushes prices down in the middle of sunny days; wind does it overnight and at weekends. There's a large, recurring window when Dutch electricity is both cleanest and nearly free.

The Mechanism

How grid-coupled pricing actually works

The mechanism is deliberately simple. Take the published base rate for a GPU and multiply it by how the grid is doing right now, measured against its recent norm: price = base rate × (current grid price ÷ 30-day average grid price)

Real-time grid tracking

When the grid sits at 18% below its 30-day average, the GPU costs about 18% less than its base rate. When the grid is at half its average, the GPU is roughly half price.

Floor and cap protection

The floor stops the price chasing the grid all the way down during negative-price hours. The cap protects you from grid spikes, so a cold still evening doesn't produce a shock bill.

Day-ahead visibility

The day-ahead market publishes tomorrow's hourly prices around 14:00 CET the day before, so the cheapest window isn't a guess. For flexible workloads, you can see exactly when to run.

The Status Quo

So why is flat pricing the norm?

If grid-coupled pricing can be cheaper for the customer, it's worth asking why nearly every GPU cloud charges a flat rate instead. The reasons are mostly structural, and understanding them is the clearest way to see what dynamic pricing is actually doing differently.

A flat rate is simpler to bill

One published number is easier to quote, budget against, and operate than a price that moves every hour. That simplicity is a genuine convenience, and for some buyers it's worth more than the savings, which is a fair trade to make with eyes open.

It lets the provider absorb volatility, for a fee

Energy prices swing hard, so a flat published rate is really a blended average with a margin built in to cover the provider's expensive hours. The customer pays that smoothed rate even during the cheap ones. It isn't sinister, it's just risk being managed by the seller rather than passed to the buyer, and the buyer pays for the service.

Most providers have little cheap energy to pass on

The large US clouds add a 20 to 40% premium on their European regions over their American ones, so their floor is already high. With expensive cooling on top (a conventional data centre spends roughly half again on cooling for every watt delivered to a chip), there's little headroom to drop the price even when the grid is cheap. Dynamic pricing only works if your cost base is low enough to follow the grid down.

Honest Trade-offs

When dynamic pricing helps you, and when it doesn't

Dynamic pricing is genuinely better for some workloads and genuinely worse for others. The deciding question is simple: can your work wait?

It helps most if you can time-shift

Batch jobs, fine-tuning runs, hyperparameter sweeps, overnight evals, and dataset processing are all examples. Anything that doesn't have to run right now can be scheduled into the cheap-clean window. These workloads get the lowest price and the greenest power at the same time.

It still helps if you can't shift, but less dramatically

A production inference endpoint serving live traffic has to run on demand, so it can't chase the cheap window. It still benefits from a competitive base rate and from the cheap hours it does happen to run through.

It's genuinely worse if your fixed schedule lands on expensive hours

If your workload must run at peak times and can never move, a flat rate might serve you better. You'd be exposed to the grid's expensive hours without being able to avoid them (which is exactly what the cap is there to limit).

Climate Impact

Why this is also about running AI greener

Saving money is the visible half. The other half is why Leafcloud built this in the first place, and it's inseparable from the cost story, because the cheap hours and the clean hours are the same hours.

The scale of AI's energy problem

The Dutch grid is so congested that the national operator's waiting list runs to 38 GW of requests. Large parts of Noord-Holland, including Amsterdam and the Zuidas, have essentially zero headroom for new large connections until around 2036.

Time-shifting eases the grid

When dynamic pricing nudges flexible workloads into the windows when renewable supply is high, it's doing exactly what grid operators are begging large consumers to do: use power when it's abundant, back off when it's scarce.

Heat reuse, not waste

Every watt a GPU draws becomes heat. Leafcloud's infrastructure is distributed into residential and care buildings, where that waste heat warms water for the people living there, displacing fossil-gas heating. The result is carbon-negative compute at −1,930 kg CO₂ per kW per year.

Common Questions

Everything you need to know about dynamic GPU pricing.

Very predictable for planning purposes. The day-ahead electricity market publishes tomorrow's hourly prices around 14:00 CET each day, so you can see exactly when the cheapest hours will be and schedule flexible workloads accordingly.

The floor prevents prices from dropping too low during negative-price hours (we still have hardware and infrastructure costs), while the cap protects you from price spikes during grid stress. This means dynamic pricing is bounded on both sides for predictability.

Dynamic pricing works best for flexible workloads like batch jobs, training runs, or dataset processing that can be scheduled during off-peak hours when renewable energy is abundant. If your workload must run at fixed peak times, a flat rate may be more predictable.

When the Dutch grid has excess wind or solar power, wholesale electricity prices drop dramatically (sometimes even going negative). Since electricity is a major cost in GPU compute, we pass those savings directly to you — making your AI workloads cheaper when they're also greenest.

Explore More

See it in action or talk to our team.