Jensen Huang Says $80–100B Per Gigawatt. Here's Why That Makes the Grid Connection the Investment Thesis.

NVIDIA CEO Jensen Huang stated at Computex 2026 that future AI data centers could cost $80 billion to $100 billion per gigawatt of compute capacity. One reply in the thread that followed did the arithmetic immediately: "$55M per MW, you do the math."

The math changes everything about how to think about grid infrastructure.

THE NUMBER — AND WHAT IT MEANS

Jensen Huang's $80–100B Per GW — Translated to MW $80–100 billion per gigawatt = $80–100M per megawatt
Conservative estimate: $55M per MW (infrastructure cost alone)
Nebius Béthune: 240 MW · €8 billion confirmed investment = €33M per MW
SoftBank Hauts-de-France: 3,100 MW · €45B first phase = €14.5M per MW

At $55–100M per MW, a 100 MW AI data center represents $5.5–10 billion in total asset value

At this cost per megawatt, every component of the deployment timeline has enormous economic weight. A 12-month delay on a 100 MW project at $80M/MW is not a scheduling inconvenience — it is $8 billion of capital sitting idle, generating no return, while the AI compute market moves ahead without you.

WHY THE GRID CONNECTION BECOMES THE INVESTMENT THESIS

When AI data centers cost $80–100B per GW, the value hierarchy of the project inverts. Traditional real estate and construction thinking puts the building at the center. Power-first thinking — which Jensen Huang's number confirms — puts the grid connection at the center.

Here is why:

The building can be constructed in 18–24 months. Steel, concrete, cooling infrastructure — all procurable with capital and project management.
The transformer can be sourced in 20–32 months from EU second-tier manufacturers at current lead times.
The grid connection cannot be bought. It must be inherited (brownfield) or earned through a queue (greenfield, 3–10 years depending on market). No amount of capital accelerates a grid operator's interconnection study.

At $80–100B per GW, the asset that cannot be bought with money becomes the scarce asset that determines everything else.

The Brownfield Value Equation — France Greenfield grid connection (Northern Virginia): 7–10 years · $0 saved but years lost
France brownfield (existing HV connection): 18–36 months · 5–8 years gained
Value of 5 years at $80M/MW on a 100 MW project: $8 billion in accelerated deployment
Cost of a Grid Deployment Risk Audit to verify this: on request

The audit cost is not a line item at this scale. It is rounding error.

WHAT THE AI BUBBLE CRITICS MISS

Several replies to the Jensen Huang tweet called it a bubble — "$1 trillion going in circles", "an AI Ponzi." This view misunderstands what $80–100B per GW actually represents.

This is not financial engineering. It is physical infrastructure: steel, copper, transformers, substations, fiber, cooling systems, buildings. The $80–100B per GW is the cost of constructing and operating the physical layer that runs AI models. It reflects real supply chain constraints — transformer lead times, grid connection queues, GOES steel monopolies — not financial speculation.

The infrastructure is real. The bottleneck is real. And France's brownfield sites with existing HV connections are a real answer to a real constraint.

THE IMPLICATION FOR SITE SELECTION

At $55–100M per MW, the due diligence on a data center site must answer one question before all others: can this site receive power, and when? Not approximately — precisely. Which RTE connection pathway applies. Which transformer manufacturer has slots for the target commissioning date. Whether the substation capacity is available or already absorbed by competing projects.

GridReadiness provides this answer in 72 hours. At the capital scale Jensen Huang is describing, that is not a cost. It is a rounding error on the rounding error.