Data Center Pauses Driven by Hard Energy Constraints, Not Political Rhetoric

I noticed an interesting detail: Trump criticized New York Governor Hochul for pausing large data center construction, calling it a "crazy Green New Deal," but what really hit the brakes on data centers was the physical limit of the grid—not something any politician can solve with a snap decision.

Let's start with the conclusion: New York's pause is not an isolated event, but a turning point signal for the data center industry shifting from "computing priority" to "energy priority." Coming from a 5G baseband background, I'm all too familiar with this dilemma of "changing architecture when hitting the power wall." Data centers are crashing into their "Thermal Design Power Wall," and this time, the speed of process node and architectural innovation cannot keep up with the growth in computing demand.

Why New York? Why Now?

Last year, data center electricity consumption in New York State already accounted for over 8% of the state's total usage. According to current plans, this ratio will skyrocket to over 20% by 2030. This isn't a prediction by some environmental group; it's official data from the New York Independent System Operator (NYISO). Grid expansion isn't just about building substations; approval for transmission corridors, capacity gaps left by retiring coal plants, and intermittent fluctuations in renewable energy—each is a tough nut to crack.

Hochul's pause order targets "hyperscale data centers" (projects with rack power exceeding 20kW), essentially those AI training clusters that directly spike statewide load. I checked public info: total load for data centers under construction and planned in NY State exceeds 4GW, while the state's peak load is only around 32GW. Having one state feed over 12% of its electricity to data centers would make any governor think twice.

From Process Nodes, It's Actually a "Power Density" Problem

Chip designers know the core contradiction of advanced processes: transistor density doubles, but power density also doubles. From 7nm to 5nm, transistor density increased by ~1.8x, but power per unit area actually rose by ~20% (due to leakage currents and higher frequency requirements). TSMC N3's SRAM density improvement is only 1.2x, but power density continues to climb. This means stricter cooling requirements for the same chip area.

In data centers, single GPU power has surged from 300W to 700W or even 1000W. NVIDIA Blackwell B200's TDP is reportedly in the 700W range, and future Rubin architectures may break 1200W. Stuffing 8 GPUs into a rack, plus interconnects and storage, easily pushes total power over 60kW. Traditional air cooling can't handle it; liquid cooling becomes standard, but liquid cooling requires infrastructure retrofitting—a cost not every operator can bear.

The issue is that process nodes themselves cannot solve linear improvements in "performance per watt." There's an iron rule in chip design: more than half of the performance gain from each process generation is eaten up by increasingly high power densities. This is why, when working on 5G basebands, we spent huge efforts on Dynamic Voltage and Frequency Scaling (DVFS), clock gating, and power domain isolation—because physically, you can't let a chip run at max frequency forever.

Technical Reality Behind Political Maneuvering

Trump's criticism essentially says: Government shouldn't interfere with the market. But the underlying technical logic is: If the grid doesn't expand, even if chipmakers reduce power to the extreme, they can't fill New York's electricity gap. In reality, forward-thinking chip companies are already doing two things:

1. Heterogeneous Integration: Packaging chips from different process nodes together, connecting them via silicon interposers or Embedded Multi-die Interconnect Bridge (EMIB) to reduce overall power. For example, AMD's MI300 series integrates CPU, GPU, and HBM in one package, reducing power for inter-chip communication.

2. Specialized Accelerators: Abandoning general-purpose computing for ASICs. Google's TPUs and Amazon's Trainium are designed for specific computing patterns, offering 2-3x better power efficiency than general GPUs. This isn't black magic; it's engineering optimization—doing fewer things with fewer transistors.

But even these solutions can only slow power growth, not reverse the trend. Because AI training demand is exponential, while chip energy efficiency gains are linear. This scissor gap is the ultimate constraint on data center site selection.

What Impact Will New York's Pause Order Have?

Short-term, it will drive up construction costs in other states. If New York blocks construction, projects will flow to Virginia, Ohio, or even Texas. But grid capacity in those states isn't optimistic either; Virginia has already seen data centers queuing for approvals. Long-term, data center sites will concentrate near stable power sources like nuclear and hydro.

I've noticed a clear trend: Over the next five years, data centers will stop pursuing "more advanced process nodes" and instead pursue "lower price per watt." TSMC N2 (2nm) improves energy efficiency by 20%, but chip cost may double. For hyperscale data centers, making 100 chips with N5 process might be more cost-effective than 50 chips with N2—because data centers don't care about chip area; they care about total power and total throughput.

It's like how 5G basebands were optimized from 28nm to 7nm, but eventually, we realized antenna efficiency and baseband algorithms determined user experience, not just the process number. Same for data centers: process nodes are just one tool in the toolbox, not a universal key.

Trend Prediction: Data Centers Entering the "Liquid Cooling + Nuclear Energy" Era

I predict that within three years, data centers with rack power exceeding 100kW must have immersion liquid cooling as standard, otherwise cooling demands can't be met. Meanwhile, big tech companies will sign direct contracts with Small Modular Reactor (SMR) developers, like Microsoft's deal with TerraPower. Because only nuclear...

Original Link: https://www.theguardian.com/us-news/2026/jul/15/trump-new-york-datacenter-moratorium