The power, cooling, and water behind AI data centers and what's coming next.

Welcome back to the OptimusEdge. Every major tech company that once talked about efficiency is now signing nuclear power deals. That's not a sustainability initiative it's a sign the regular electricity grid genuinely cannot keep up with what AI data centers need.
___________________________________________________________
The Edge Upload: Today’s Insights

Nuclear went from niche to normal: hyperscalers have committed 9.8 GW of nuclear capacity across 13 separate deals as of May 2026 Microsoft, Google, Amazon, and Meta have each signed at least one, together worth tens of billions of dollars.

Europe just made heat reuse mandatory: under Article 26(6) of the EU's recast Energy Efficiency Directive, any new European data center above 1 MW is now legally required to assess whether its waste heat can be reused turning what used to be a sustainability nice-to-have into a compliance item.

AI is being used to cool AI: researchers are training reinforcement learning systems to jointly optimize data center cooling and waste heat recovery in real time using AI to solve a problem AI itself created.

We touched on power density in a previous issue; this is the supply side of that same problem. U.S. grid interconnection queues have grown to over 2,600 GW of projects.

Its waiting for connection, with average wait times around five years and roughly 80% of projects eventually withdrawing a backlog caused by insufficient transmission capacity and transformer shortages, not a lack of willingness to build.

Nuclear solves a specific problem regular grid power doesn't: it runs at 92%+ capacity factor, compared to roughly 34% for wind and 23% for solar meaning it's actually available around the clock, which matters enormously for a GPU cluster that can't simply pause training when the sun goes down.

That's why Microsoft is restarting Three Mile Island, Google has a 500 MW deal with Kairos Power, Amazon has invested over $700 million in X-energy's small modular reactors, and Meta is pursuing up to 6.6 GW across four separate nuclear developers.

None of these companies are doing this because it's cheap they're doing it because it's one of the only ways to guarantee power will actually be there.

The core mechanics: direct-to-chip liquid cooling attaches cold plates directly to the GPU die and pipes coolant through a Coolant Distribution Unit (CDU) that sits between the facility's chilled water loop and the rack itself.

This is now the default for anything above roughly 40 kW per rack. Immersion cooling goes further, submerging entire servers in dielectric fluid, and supports the highest densities of any current method.

Both approaches share a byproduct worth understanding: the coolant leaves a GPU rack meaningfully hot 55–65°C for direct liquid cooling which is exactly why heat reuse has become viable at all. Cold, diffuse heat from air cooling was never worth capturing.

Concentrated, hot liquid is.

This is where the public conversation gets muddiest, so it's worth being precise.

There are two genuinely different ways to count AI's water use, and most viral claims blur them together.

Direct, on-site water what a facility's cooling towers physically evaporate typically lands under a millilitre to a few millilitres per query.

Indirect water consumed at the power plant generating the electricity that runs the query adds tens of millilitres more, because thermal power generation itself uses water for cooling.

Both figures are real; neither alone is the whole picture, which is exactly why you'll see wildly different numbers cited in different places the disagreement is about accounting, not physics.

What is measurable and consistent: Google's own disclosed water use hit 10.9 billion gallons in 2025, up 34% in a single year, and roughly two-thirds of new U.S. data centers built since 2022 sit in areas already facing water stress.

The honest summary is that AI's water footprint is real, it's growing, and as we noted in a previous issue liquid cooling doesn't eliminate it so much as relocate it, from the rack to wherever the heat ultimately gets rejected.

A few genuine optimization shifts are already in motion, not just proposed.

Waste heat reuse is moving from pilot to standard: a single 1 MW IT load produces enough thermal energy annually to meaningfully heat nearby buildings, and facilities across Sweden, Denmark, and the Netherlands already pipe that heat directly into municipal district heating networks.

AI-optimized cooling control is the recursive piece reinforcement learning systems that continuously tune cooling and heat recovery together in real time are already outperforming static, human-set schedules in early deployments.

Power delivery itself is being redesigned, with NVIDIA's 800 VDC architecture aiming to support 1 MW racks by 2027 without the copper and conversion losses of today's systems.

None of these fully solve the power or water equation on their own but together, they represent the first real shift from "build bigger" to "build smarter" in how this industry is approaching its own resource footprint.

Takeaway: The power and water numbers behind AI data centers are genuinely large and genuinely growing that part isn't hype. What's changing is that the industry is now visibly investing in fixing the mechanism, not just scaling past it: nuclear for guaranteed power, heat reuse for the byproduct, and AI itself for the optimization layer on top.

That's today's briefing. If you've seen a "one ChatGPT query = one bottle of water" claim floating around, this one gives you the actual context behind it. Past issues are in the archive. See you tomorrow.

SMR Intel Deal Tracker: A continuously updated tracker of every nuclear power agreement signed by hyperscalers for AI infrastructure useful for seeing exactly who's committed to what, and when the power is actually expected online.

Nuclear Deal Data: Nuclear Energy for Data Centers 2026 the full 9.8 GW / 13-deal breakdown by company

Water Accounting Explained: AI Data Center Water Usage 2026 the direct-vs-indirect water framework and WUE benchmarks

Waste Heat Recovery: District Heating AI Data Centers: 2026 Waste Heat Recovery Guide the EU's Article 26(6) mandate and real district-heating case studies

Grid Constraint Data: AI Data Centers Hit Grid Wall: Big Tech Pivots to Nuclear in 2026 the interconnection queue numbers behind the nuclear pivot

That’s it for today !☀

Edge AI is levelling up are you? Until next time, stay curious, stay building, and don’t let your machines take over. 🤖😆

Enjoyed today’s issue? Share OptimusEdge AI with your engineering team. Subscribe to OptimusEdge AI

Your Edge AI Explorer,
Sharat Sami (Let’s connect on LinkedIn)