Cloud, Compute, and Sovereignty: A State of an Industry in Flux

Abstract

The AI wave has shifted power from the commercial layer, where hyperscalers locked customers in through egress fees and pricing complexity, to the physical layer, where chips and electricity have become so scarce that the providers themselves are supply-constrained. The memory wall, HBM, advanced packaging, NVIDIA's CUDA moat, the collapse of model margins under DeepSeek, the gigawatt race: all converge on one line. In the economy of AI, infrastructure is no longer a means but the end.

Key takeaways

• The scarce resource has changed. Power once lay with whoever controlled the commercial relationship; it now lies with whoever controls the hardware and the energy to run it.
• NVIDIA's moat is software, not silicon. Its dominance rests on twenty years of CUDA ecosystem, not on any single chip, which is why challengers attack narrow segments rather than the whole platform.
• DeepSeek cracked the brute-force doctrine. Algorithmic efficiency (mixture-of-experts) matched Western leaders at a fraction of the cost per token, hollowing out foundation-model margins.
• Compute is now a matter of gigawatts. The battle has shifted from silicon to the electricity grid, making France's nuclear fleet a genuine strategic asset.