This page breaks the AI boom into four pillars you can measure in the real world: (1) compute & packaging (chips, memory, and advanced assembly), (2) data center infrastructure (buildings, power delivery, and cooling), (3) power supply + long-term offtake contracts (gas, nuclear, renewables), and (4) crypto miners pivoting into HPC/AI hosting (reusing sites, power, and operations). In practice, the biggest gating items are usually MW (megawatts of power), grid interconnect approvals, and cooling capacity. [1][2][7]
These are confirmed items from primary documents or reputable reporting that change timelines, cost, or execution risk. If you’re skimming, this section is the fastest way to update your mental model.
Big Tech is leaning harder on gas and nuclear alongside renewables as grid delays and timing risk rise. That shifts the advantage toward dispatchable / baseload power (energy you can count on when you need it) and the contract mechanics of who pays for upgrades, interconnect, and reliability. [1]
TrendForce projects liquid cooling penetration in AI data centers rising from 14% (2024) to 33% (2025e), driven by GB200-class rollouts. Practically: hotter racks push more spending into pumps, piping, controls, and ongoing installation/maintenance. [2]
Nidec highlighted Project Deschutes CDU prototypes based on Google’s OCP specification. A CDU (coolant distribution unit) is the plumbing “hub” for liquid-cooled racks—so standards matter because they reduce one-off engineering and speed deployment. [3]
APLD’s S‑3 (SEC filing) describes an initial closing of 112,500 preferred units ($112.5M) and a right to invest up to an additional $4.9B under an amended Unit Purchase Agreement (UPA). Staged funding can accelerate construction—but the details (triggers, controls, and dilution) matter. [4]
CoreWeave priced an upsized $2.25B convertible senior notes offering (due 2031), scheduled to settle Dec 11, 2025. Converts are a common growth tool—but they make outcomes more sensitive to rates and equity pricing. [5]
ERCOT’s Tier‑1 Eastern Backbone recommendation cites ~$9.384B estimated cost and ~1,108.8 miles of new right‑of‑way (ROW), with expected in-service 2030–2032. This is what “grid lead times” means in practice. [7]
Each pillar is a “real world” constraint: what gets tight, what gets expensive, and what tends to decide timelines.
Think of it like a power‑hungry factory for computation: many buildings of servers, connected by high‑speed networking, fed by dedicated power infrastructure, and cooled like industrial equipment. When companies announce a big AI campus, the limiting steps are rarely the walls—they’re usually power delivery, interconnect approvals, and cooling installation/maintenance capacity.
This is why AI strategy is increasingly infrastructure strategy: procurement, construction sequencing, and service logistics matter as much as the model roadmap.
Transmission, substations, and equipment lead times set the delivery clock. ERCOT’s Tier‑1 765‑kV plan provides a concrete example: multi‑year builds and multi‑billion-dollar cost. The nearer-term bottleneck often becomes substation capacity + transformer/switchgear availability, not just generation. [7]
This is why the AI story isn’t only chips: power delivery, EPC (engineering + construction), and the contractual “who pays” mechanics can be decisive.
These charts use only numbers explicitly stated in cited sources. Think of them as “orientation,” not forecasts.
M&A is a recurring signal that “cooling” is moving from a one-time hardware purchase to lifecycle services and systems integration. Examples include Vertiv’s announced plan to acquire PurgeRite (~$1B, per Reuters) and Eaton’s Boyd Thermal deal ($9.5B, per Reuters / Eaton). [10][11]
A fast-reading table: what each name represents, what changed recently, a 3–6 month view, and the single biggest way the thesis breaks.
The market often talks about AI in abstract terms. On the ground, it turns into site selection, substation studies, equipment orders, and construction sequencing. The table below is a quick lens on where different companies sit in that stack—and which failure mode matters most for each.
If you only track one thing: follow the power + interconnect timeline as closely as you follow the chip roadmap.
| Ticker | Role in thesis | Recent highlight | 3–6 month stance | Top risk |
|---|---|---|---|---|
| CEG | P3 baseload/clean + contracting | Calpine-related exchange offer timeline: Early tender/withdrawal deadline Dec 22, 2025; expiration Jan 8, 2026. [9] | Positive | Regulatory + integration + power price volatility |
| VST | P3 dispatchable/merchant + contracting | “All-of-the-above” power strategy raises value of dispatchable MW amid grid delays. [1] | Positive | Policy / cost allocation / rate class backlash |
| TLN | P3 reliability + PJM narrative (select assets) | DOE 202(c) order 202‑25‑6A for PJM/Wagner continues until Dec 31, 2025 (per DOE page). [8] | Positive | Regulatory/operational events; precedent risk |
| EQT | P3 gas supply optionality | “All-of-the-above” implies near-term gas demand relevance for AI load build. [1] | Positive | Gas price + regional price spreads (basis) |
| MOD | P2 thermal/cooling leverage | Liquid cooling adoption ramp supports attach + service ecosystem expansion. [2][3] | Positive | Competitive intensity; execution on programs |
| ONTO | P1 inspection/metrology leverage | December signal mainly macro: watch capex confirmation vs sentiment noise. [6] | Neutral | Wafer‑fab equipment cycles; customer capex timing |
| INTC | P1 compute/platform + packaging | December: keep focus on execution milestones rather than “AI tape” volatility. [6] | Cautious | Execution vs peers; margin/roadmap |
| AVGO | P1 networking/custom silicon | AI sentiment wobble is valuation risk; watch confirmations in order commentary/filings. [6] | Positive | Order volatility via hyperscaler cadence |
| APLD | P4 HPC hosting build (structured capital) | $112.5M initial preferred units; right to invest up to $4.9B; warrants mechanics described. [4] | Speculative | Funding timing, dilution/terms, execution |
| IREN | P4 mining→HPC optionality | December is chiefly macro/sentiment; treat cost of capital as gating variable. [6] | Speculative | BTC exposure + capex funding |
| CORZ | P4 hosting optionality | December is chiefly macro/sentiment; contracts + power access decide outcomes. | Speculative | Execution + counterparty + power availability |
| CRWV | P4 AI cloud build | $2.25B upsized convert (due 2031) underscores financing sensitivity. [5] | Cautious | Leverage/financing sensitivity; concentration |
These are explicit dated deadlines in cited primary sources. Use them as “checkpoints” for confirmation, not as predictions.
Where projects slip (or fail): financing, grid timelines, and operational capacity.
AI infrastructure is capital‑intensive. If rates rise or markets tighten, builders can’t fund equipment and construction even if demand stays strong. Treat “cost of capital” as an operational bottleneck, not a finance footnote. [6]
“Interconnect” is the permission + engineering work needed to connect a large new load to the grid. It often implies substation upgrades, new equipment, and (sometimes) new transmission. ERCOT’s $9.384B / 2030–2032 example shows why timelines can stretch. [7]
Scaling liquid cooling requires not just hardware but installation, commissioning (bringing systems online), and maintenance. Standards help reduce custom engineering—but they can also concentrate demand into certified vendors and their service networks. [3][2]
A simple checklist you can run weekly/monthly to stay grounded in what’s actually happening (links in Sources).
Bracketed numbers throughout the page link to the references below. Use these to verify claims or go deeper.
