AI inference efficiency and data center coordination in France, Digi Mart USA News: Since the February 2025 AI Action Summit and the announcement of 109 billion euros in commitments, France has ranked among the most ambitious European countries in artificial intelligence infrastructure. As these investments take shape, a central question is emerging: how can this installed capacity be turned into durable economic value?
For several years, the AI race was largely a matter of capacity: building the most powerful models and the infrastructure needed to train them. According to Deloitte, the market is now entering a new phase centered on usage.
Nearly two-thirds of AI-related computing is expected to go to inference, the phase that follows model training, by the end of 2026, compared with roughly half in 2025. Enterprise spending to access large models through AI services is expected to rise from $3.5 billion to $8.4 billion within one year.
Running these models at scale depends on millions of short requests, critical latency and fine-grained orchestration of resources. The challenge is therefore no longer just having computing power, but knowing how to use it efficiently.
One figure illustrates the issue. According to analysis from Creative Strategies, an AI computing rack represents an investment of more than $4 million.
In agentic AI environments, these specialized accelerators can remain idle for up to 22% of total inference time when the orchestration environment does not keep pace. That means nearly a quarter of the data center’s most expensive asset is underused, not for lack of power, but because of a failure of coordination.
This is where the complementarity between technological components becomes strategic. Specialized accelerators handle intensive computing workloads, while general-purpose processors orchestrate flows, manage data and distribute requests.
This is not a debate of opposing choices, but a question of balance.
This balance is also an energy issue. In the United States, electricity consumption by data centers dedicated to AI could rise from 176 TWh in 2024 to nearly 580 TWh in 2028.
In France, competitive and largely decarbonized electricity is a major advantage, provided that each watt produces the maximum useful value.
Three avenues for change emerge. First, rebalance capital expenditure (capex) decisions by integrating coordination capacity, not just raw computing capacity, from the design stage.
According to recent modeling, a more balanced allocation can improve a data center’s throughput by 2% and reduce the cost per token by 3.7%, for 1.7% additional capex.
Second, enrich performance indicators: actual accelerator utilization rate, cost per useful token, and value produced per watt consumed. Third, factor infrastructure efficiency into energy planning associated with AI development.
The best way to ease pressure on the grid is not only to add capacity, but to make better use of what has already been deployed.
The next battle in AI will not be won simply by accumulating capacity. It will be decided by the ability to turn that power into lasting economic value.
Marie Théry is President of Intel France and sales director for Southern Europe.
