800V DC, Liquid Cooling, SMR: The 3 Technologies Reshaping AI Datacenters in 2026

The year 2026 marks a technological turning point for artificial intelligence datacenters. Three innovations are converging to redefine physical infrastructure: direct 800V DC electrical architecture, advanced liquid cooling, and Small Modular Reactors (SMR). Together, they promise to halve energy consumption and multiply computing density tenfold. FINXIA Capital analyzes their impact on infrastructure investments.

The first revolution comes from power supply. Direct 800V DC architecture, popularized by Amazon and Google hyperscale centers, eliminates traditional AC/DC conversion steps. Result: energy efficiency rises from 85% to 97%, representing 12% savings on the electricity bill. For a 50 MW datacenter, this means 6 million euros in annual savings.

This architecture becomes essential with NVIDIA Blackwell Ultra GPUs. Each 200 kW rack requires stable, massive power that conventional AC infrastructure struggles to provide. Electrical equipment manufacturers — Vertiv, Schneider Electric, Eaton — see their 800V DC order books exploding, with lead times reaching 18 months.

Liquid Cooling: From Optional to Mandatory

The second revolution is thermal. Forced air cooling reaches its physical limits at 40-50 kW per rack. Yet current AI configurations easily exceed 150 kW per rack. Only liquid cooling, through immersion or dielectric fluid circulation, can maintain acceptable operating temperatures.

The gain is spectacular. A datacenter with liquid cooling achieves a PUE of 1.08 to 1.15, compared to 1.4 to 1.6 for a traditional air-cooled center. On a 50 MW power load, this difference translates to 12.5 MW of energy savings — the equivalent consumption of a city of 15,000 inhabitants. Hyperscalers now demand liquid cooling as standard for any new contract.

SMRs: The Energy Ace of Tomorrow's Datacenters

The third innovation is nuclear. Small Modular Reactors (SMR) are compact reactors of 50 to 300 MW that can be deployed in the immediate vicinity of datacenters. Microsoft signed a contract with Constellation Energy to power its Pennsylvania centers. Amazon and Google are exploring similar partnerships in Europe.

In Europe, France and Great Britain lead the race. EDF is working on the NUWARD reactor (340 MW) for large industrial installations. Rolls-Royce SMR targets European certification by 2029. For a 100 MW datacenter, a dedicated SMR eliminates all dependence on the national electrical grid and guarantees stable energy pricing for 40 years.

FINXIA Capital's Investment Thesis: The Triple Convergence

FINXIA Capital believes the convergence of these three technologies creates a new asset class: the fourth-generation AI datacenter. These sites combining 800V DC, liquid cooling, and SMR potential benefit from a structural competitive advantage translating into rents 50% to 80% higher than legacy datacenters.

Our financial model shows that a greenfield site integrating these three technologies, while requiring a 25% to 35% higher initial investment, achieves a PUE of 1.05 and generates revenues per MW 60% higher. The payback period remains contained at 4.5 years thanks to multi-year contracts signed with hyperscalers, who willingly pay the premium for energy resilience.

What Investors Must Understand

In 2026, investing in a datacenter without an 800V DC plan, without liquid cooling, and without thinking about energy autonomy is investing in obsolescence. Next-generation GPUs impose these constraints as standards, not options. Assets that don't follow this technology curve will see their value melt by 30% to 50% by 2028.

The investment window for fourth-generation datacenters is 2026-2027. By 2029, 800V DC equipment and liquid cooling systems will be in global shortage, and SMR sites will be reserved for large-scale contracts. Whoever builds today according to these standards will capture the first-mover premium and secure revenues for two decades. This is FINXIA Capital's technological conviction.