Big-4 hyperscaler capex, Q3 '26
$201.9B
▲ 108% vs Q3 2025
Company 10-Qs · Walnut
NEW · Issue #001 is out · Free every Saturday
The AI boom runs on watts & water.
A weekly newsletter for people breaking into AI infrastructure, written by an engineer who builds the equipment. Read it on Substack, then come here to play the weekly Sudoku and crossword and follow the startups worth watching.
RB
Revati Borkhade · Industrial Engineer building chillers for hyperscale AI data centers. 15+ peer-reviewed papers, 2 patents.
Latest Issue
#001 · The Gigawatt Week
The full stories, with the backstory, why it matters, and what to watch next, live on Substack.
🌡 Thermal & Cooling
Chiller makers are buying their way into liquid cooling
Alloy, PurgeRite, new AI-ready chillers, and why warmer water changes the chiller's job.
⚡ Power & GridThe "power island" becomes the default plan
Gigawatt campuses, a 3.2 GW project in Georgia, and why developers bring their own power.
💰 Capital & MarketRecord spending, thinner cash cushions
$201.9B in one quarter, bond sales, and what it means for suppliers.
This Week's Games
Play, score, brag
Answers to both puzzles come out in Issue #002. Share your score with #WattsAndWater.
Sudoku #001
Fill every row, column and 3×3 box with 1–9. Your score starts at 1,000 and drops with time and rule-breaking entries. Finish fast and clean.
Time
0:00
Errors
0
Best
—
Scoring: 1,000 points, minus 1 point every 3 seconds, minus 25 points for every entry that breaks a rule. Clashes turn red right away.
🔒 The full solution comes out in Issue #002. No peeking.
The Reader's Crossword #001
Every answer is in Issue #001. Read it on Substack cover to cover and you can solve all 12. Click a clue or a square, then type.
Solved
0/12
Checks
0
Across
Down
🔒 Full answers in Issue #002. Each check shows only which words are right.
The Pulse
Four numbers to know
The figures setting the pace of the buildout, each linked to its source.
Start Here
New? Twenty years in four eras
If you're new, read this once. Every story we cover fits somewhere on this line.
2006 – 2011
The cloud era
AWS launches in 2006 and computing becomes something you rent. Data centers are mostly air-cooled, and racks draw about 5–10 kW. Facebook founds the Open Compute Project in 2011 to share efficient designs.
Rack ≈ 5–10 kW2012 – 2019
Hyperscale & the GPU spark
In 2012, AlexNet shows deep learning works when it's trained on GPUs. Hyperscalers build ever-larger campuses and chase better PUE. NVIDIA ships its first DGX AI system in 2016.
Campus ≈ 100 MW2020 – 2024
The AI shock
The A100 (2020) and H100 (2022) push chip power up fast, and ChatGPT launches in late 2022. In 2024, NVIDIA's rack-scale Blackwell systems need around 120 kW per rack and liquid cooling.
Rack → 100 kW+2025 – now
Gigawatts & power islands
The first gigawatt campuses arrive. Power, not chips, becomes the bottleneck. Big-4 capex passes $200B a quarter, and liquid cooling becomes standard for AI.
Campus → 1 GW+💡
The one-line lessonEvery era multiplied the power per rack, and every jump turned a facilities problem (power, cooling, supply chain) into the real bottleneck. That's why AI infrastructure needs engineers, not just coders.
Startup Watchlist
Who's raising, and who to follow
Where venture money flows tells you which problems the industry thinks are worth solving next.
THIS WEEK
$500M
Torus · Utah
Launched Foundry, factory-built AI inference units of 1–20 MW. Torus says they go live in 12–16 weeks by using spare grid capacity instead of waiting in the queue. Torus also says the closed-loop cooling is waterless.
Source: Tech Times
LAST 3 WEEKS
$108M
Peak Energy · California
Makes passively cooled sodium-ion battery storage and is now moving into data centers. The company says it has more than $1B in customer orders to fill.
Source: ESG Today
The Watchlist ⭐
Tap Follow to track a startup. We'll flag its news in future issues on Substack.Startup
What they do
Category
Last round
Words of the Week
Say it like an insider
Tip: a couple of these show up in the crossword.
Power Usage Effectiveness PUE
Total facility power ÷ IT power. 1.0 is perfect. Everything above 1.0 is mostly cooling and power-conversion overhead.
Capital expenditure Capex
Money spent on long-lived assets such as buildings, chillers, turbines and GPUs. It's the number everyone watches on earnings day.
Chiller
A machine that makes cold water using a refrigeration cycle. It's the heart of most data center cooling plants, and it's sized in tons (1 ton = 3.517 kW of heat removed).
Direct-to-chip cooling DLC
Cold plates sit on top of the GPU and carry liquid through it. It's needed once racks pass roughly 50–100 kW.
Behind-the-meter BTM
Power made on-site and used directly, bypassing the utility grid and its interconnection queue.
Two-phase cooling
The coolant boils as it absorbs heat, and the vapor carries the heat away. ZutaCore and Accelsius both build this kind of system.
The weekly issue
One email every Saturday. Fifteen minutes. Smarter by Monday.
Built for engineers, students and career-switchers getting into AI infrastructure. Free on Substack.
- 3 stories with the backstory and why they matter
- Startups that raised, and a watchlist to follow
- Podcasts and a video for the week
- A new Sudoku and crossword, plus last week's answers
Subscribe on Substack →
Free. Unsubscribe anytime. Delivered by Substack.