Mistral CEO says energy is not the current bottleneck for AI compute

The cost of AI compute

Mistral CEO Arthur Mensch recently addressed the ongoing debate regarding whether energy shortages will stall the progress of artificial intelligence. During a podcast, Mensch argued that energy is currently a minor component of the total expenses involved in running AI systems. He stated that energy is around 10% of the bill in the total cost of ownership of compute.

Because energy is such a small portion of the overall budget, Mensch believes it is not the primary obstacle for companies trying to scale their operations right now. He noted that for builders in regions like Europe, there is currently an excess of energy available. This means that the immediate pressure on the industry comes from other sources rather than a lack of electricity.

If Mensch's assessment is correct, the focus for the next few years should remain on the hardware required to run these models. He specifically pointed out that what matters most today is the cost of the chips and what matters is the cost of capital. This distinction highlights a shift in where the most significant financial hurdles lie for the tech industry.

Hardware vs. Power margins

The difference in profit margins between hardware and energy providers illustrates why the industry is currently hardware-centric. As TIKR.com reported, NVIDIA saw its revenue rise 106% to $96.2 billion in the quarter ended July 26, while maintaining a 75.0% gross margin. This high margin reflects the massive premium the market places on the specialized chips needed for AI training.

In contrast, a power producer selling into a market with plenty of spare energy cannot command the same level of profitability as a chipmaker. While power bulls have argued that scarcity will drive up prices, the current reality is that the most valuable assets are the ones that can actually process the data. The high margins at NVIDIA suggest that the market is currently rewarding the providers of the 'brains' of AI rather than the 'fuel.'

However, there is a potential long-term bull case for energy if the situation changes. A buyer who currently spends only 10% on power could see that cost rise significantly once energy becomes a scarce resource. Mensch suggests that this transition point for Europe may not occur until the 2029 or 2030 window, providing a buffer for current hardware-led growth.

The Texas power outlook

The conversation around energy bottlenecks also extends to the United States, specifically the Texas power market. Vistra CEO Jim Burke provided a perspective that aligns with Mensch's timeline regarding when energy might become a primary constraint. Burke noted that the company has seen some depression in ERCOT forwards, which are contracts for future electricity prices, due to slower load growth.

Vistra's outlook implies that the region is not facing an immediate shortage that would drive prices up instantly. Burke stated that the company is not making any assumptions about power price recoveries in its current financial planning. This suggests that the market is not yet reacting to a scarcity crisis, but rather to a steady growth pattern that may not peak for several years.

Analysts expect Vistra's normalized earnings per share to double between 2025 and 2027, but these estimates do not rely on a sudden spike in power prices. Because Vistra has hedged about 98% of its expected 2027 generation, its financial health is tied more to current contracts than to future scarcity. This reinforces the idea that the 'power trade' may be looking for a catalyst that hasn't fully materialized yet.

Regulatory and physical walls

While energy might not be the primary cost bottleneck, there are still physical and regulatory hurdles that data center builders face. Some companies have described the challenge of adding capacity as 'breaking the power wall,' which refers to the difficulty of getting enough electricity to new sites. Even if the energy exists, getting it to the specific location of a data center can be a slow and difficult process.

In Texas, for example, the grid approval process has faced significant delays. Governor Greg Abbott recently froze approvals for new data centers on August 3 while an audit is conducted. This creates a situation where there is load approved to energize, but it has not been allowed to energize yet due to administrative and regulatory hurdles.

These delays mean that even if energy is abundant, the speed of AI expansion is still being dictated by policy and infrastructure. With a new governor-elect taking office in January who wants to halt new data center projects, the regulatory environment could become a more significant bottleneck than the actual supply of electricity. This adds a layer of complexity to the timeline Mensch provided for when energy will finally become the limiting factor.

What to watch next

The next few years will likely be defined by the struggle to secure capital and high-end hardware. Investors should watch how NVIDIA maintains its dominance and whether other chip manufacturers can provide viable alternatives to meet the demand. As long as the cost of chips remains the primary expense, the hardware ecosystem will continue to see the most significant movement.

Additionally, the 2029-2030 timeline mentioned by both Mensch and Burke provides a specific window to watch for shifts in the energy market. If power prices begin to climb faster than current forward curves suggest, it could indicate that the bottleneck is arriving sooner than expected. This would shift the investment focus back toward the utilities and power producers that provide the infrastructure for the AI era.

Finally, the regulatory landscape in Texas and Europe will be a key indicator of how quickly data centers can be built. If government audits and new policies successfully slow down construction, the 'bottleneck' might be political rather than physical. Keeping an eye on these specific regional developments will help clarify whether the industry is facing a shortage of power, a shortage of chips, or a shortage of permission to build.