The bottleneck has moved off the chip and onto the grid. This lesson follows a watt from the substation to the silicon, names the metrics that decide who can actually build, and opens a whole layer of investable names beyond your core seven.
For three lessons you've watched bottlenecks migrate: from litho (L17), to HBM and CoWoS packaging (L12), to networking (L13). In 2025 the binding constraint moved again — and this time it left the building. You can buy the GPUs; you still can't power them. The scarce resource is now electricity delivered to a place, on a schedule. This lesson makes that constraint quantitative, so you can read a "10 GW buildout" headline the way an engineer reads a node name.
Core thesis: The end of Dennard scaling (L19) pushed the industry toward ever-denser, ever-hotter accelerators. A GB200 die now dissipates 500–600 W/cm² and a single rack pulls ~120–130 kW. Stack thousands of those and a frontier cluster needs 300–500 MW — the load of a small city — on a grid where a new connection takes 7–10 years. So the new moat isn't FLOPS; it's secured power, fast. That reprices utilities, grid equipment, cooling, and nuclear — a value-capture layer financial media is only now waking up to.
Recall L19: around 2005 voltage stopped scaling, leakage took over, and power density became the ceiling on a single chip. The industry's escape was parallelism and specialization — more chips, packed tighter. That escape works at the chip level but simply relocates the heat problem upward. The same physics that made one transistor hard to cool now makes one building hard to power.
The numbers compound at every level. A modern accelerator is ~700–1200 W. NVIDIA's GB200 NVL72 packs 72 GPUs into one liquid-cooled rack drawing ~120 kW nominal (deployed racks report 130–132 kW) — roughly 10× the density of a traditional ~12 kW air-cooled rack. [NVL72 specs] Multiply by thousands of racks and you arrive at the cluster-scale figure that now governs the industry.
Sources: die heat flux, cluster power, connection queues.
To reason about cost and efficiency, trace a watt from the grid to the gate. At each layer some power is lost to conversion, distribution, and — above all — cooling. The single most important accounting tool is PUE.
Headlines quote gigawatts and hundreds of billions. Translate them with four numbers — the energy-economics analog of "density + PPA" from L19:
| Metric | What it measures | Reference value (2025) | Why an investor cares |
|---|---|---|---|
| PUE | Facility efficiency (overhead) | ~1.1–1.2 best · ~1.5 avg | Lower PUE = more sellable compute per MW procured = better unit economics |
| $ / MW (or $/kW) | Cost to build capacity | ~$10–12M / MW all-in | The capex denominator; sets the depreciation that must be earned back |
| $ / GW campus | Full gigawatt-scale build | ~$45–55B / GW | Why only a handful of balance sheets can play at the frontier |
| Energy / token | Inference efficiency | ~0.3 Wh / query (GPT-4o) | The marginal cost of serving; ties power to the P&L of every API call |
Sources: $/GW (Turner & Townsend via BloombergNEF), energy/query.
The energy/token figure is the bridge back to L8 (inference economics). At ~3–4 joules per output token for a mid-size model, energy is a real and growing line in the cost of every served response — and the lever (FP8/FP4 quantization, better MFU, MoE sparsity) that the most efficient operators pull to widen margin. Hold this thread; it reappears in the inference-economics capstone.
Here is the structural fact that reorders the whole investment landscape. A hyperscale campus can demand as much power as an aluminium smelter or a mid-size city — but a new large-scale grid connection in the US or Europe now takes 7–10 years, with some projects waiting 13. In the PJM region (the largest US grid market), the application-to-operation timeline rose from under 2 years in 2008 to over 8 years in 2025, with a queue exceeding 2,600 GW of pending requests. [grid-impact study]
Meanwhile demand is vertical: AI datacenter power is projected to rise from ~10 GW (2025) to ~68 GW (2030), +160%, and over 23 GW of capacity was under construction globally at end-September 2025 — three-quarters of it in the US. [BNEF] When demand grows that fast against a decade-long connection queue, the scarce asset isn't the GPU — it's a shovel-ready megawatt.
The diligence reframe: The right question about an AI buildout is no longer "did they secure the H100/Blackwell allocation?" — it's "where is the power, is it contracted, and when does it energize?" Power-purchase agreements, behind-the-meter generation, and existing grid interconnections have become the genuinely scarce, defensible asset. This is the 2025 successor to "do they have CoWoS allocation?" from L12.
That scarcity is driving three workarounds, each an investable theme:
Cooling is the largest non-IT slice of the PUE numerator, so it's the overhead operators fight hardest. The L19 physics forces the issue: at 500–600 W/cm², a GB200 die produces heat flux 40–100× beyond what moving air can remove. Air cooling is not merely inefficient at this density — it is physically impossible. [liquid-cooling rationale]
| Approach | Removes up to | Where it's used | Trade-off |
|---|---|---|---|
| Air (CRAC/CRAH) | ~15–20 kW/rack | Legacy & general compute | Cheap, simple; hits a hard density wall |
| Direct-to-chip liquid | ~120–150 kW/rack | GB200 NVL72 & all frontier AI | Cold plates + plumbing; new failure modes, leaks |
| Immersion | ~200+ kW/rack | Emerging / niche | Highest density; dielectric fluid, serviceability cost |
Liquid cooling does double duty: it both lifts the density ceiling and cuts PUE (water carries ~3,000× the heat per volume of air, slashing fan and chiller energy). That is why it shifted from optional to mandatory in one GPU generation — and why a discrete, fast-growing cooling supply chain (Vertiv, nVent, cold-plate and CDU makers) became a picks-and-shovels play on the buildout.
This is the lesson's payoff. The power constraint creates value-capture outside your core seven (TSMC, NVDA, AVGO, SK Hynix, ASML, AMD, Cadence/Synopsys) — a parallel set of beneficiaries most chip-focused investors underweight.
How this connects to your thesis: the power layer doesn't replace the core seven — it gates them. If utilities and grid equipment can't deliver megawatts, NVIDIA can't ship the GPUs into operation, which caps realized (vs. ordered) demand. So power-buildout data (interconnection approvals, PPA volume, turbine backlogs) is now a leading indicator for your existing positions — not just a separate trade. Add a power-availability falsifier to the demand side of the NVDA/AVGO thesis.
Your recurring "follow one dollar of hyperscaler capex" trace now has a power branch. Of the ~$750B the 14 largest operators are spending in 2026 (up from ~$450B), a large and growing share never touches a GPU: at ~$45–55B per GW, the campus shell, substation, transformers, switchgear, cooling, and the power contract itself absorb the dollar before silicon does. [BNEF capex] The bottleneck (power), the captured name (utilities + electrical equipment), and the leading indicator (interconnection + PPA volume) are now explicit hops in the trace.
Read first: SemiAnalysis — "AI Datacenter Energy Dilemma: Race for AI Datacenter Space" — the highest-signal cost-and-power model of the buildout, from the newsletter in RESOURCES.md. Then: the IEA's 2025 datacenter electricity update for the macro demand/grid picture, and BloombergNEF's buildout tracker for capex and $/GW figures.
Select the best answer for each.
1. A datacenter with PUE = 1.5 means that for every watt of compute, the facility draws:
2. Why has liquid cooling become mandatory for racks like the GB200 NVL72?
3. In 2025, the binding constraint on AI capacity growth has primarily become:
4. The most defensible scarce asset created by the power constraint is:
5. How does the power layer relate to your core seven holdings (e.g. NVDA)?