Skip to content

Follow
the Power.

What the AI buildout does to power, water, and land — and who pays for it. Trace the number, or don't repeat it.

The Analysis Feed

Power, water, land, and grid politics — traced to primary sources.

Notes2026-09-09

The Second User Buys No Batching At All. The Fourth Buys a Third of What Theory Promises.

Ten rounds of four requests on a fanless 8GB M1, dispatched one at a time, two at a time, and four at a time. Four-way concurrency raises server throughput 39% and cuts every user's generation rate by 65%. The two-way case is stranger: its 9% gain is not batching, and generation actually gets 6% worse.

Analysis Brief →
Silicon & Chips2026-09-09

Apple's A20 Pro Is the First 2nm Phone Chip. Its On-Device AI Claim Is Missing the Only Number That Settles It.

Apple shipped the first high-volume 2nm smartphone SoC on September 9, with a Dual 16-core Neural Engine, 50% more memory bandwidth, and packaging that moves memory out of the chip's thermal path. Here is how the silicon, the flash-resident AFM 3 models, the developer frameworks, and Private Cloud Compute fit together — and why none of Apple's figures describes energy per token, the number that decides whether on-device inference is actually cheaper than a data center round trip.

Analysis Brief →
Notes2026-09-04

I Measured the Wrong Thing Three Times — What Animation Actually Costs a Reading Page

Swapping a JavaScript scroll reveal for the CSS one saved 16 ms of script per page scroll. Stopping eight animations that repeat forever saved nothing measurable. And the harness that found this told me a confident, wrong answer three times before it told me a true one.

Analysis Brief →
Nepal Desk2026-08-28

The Network Was Never the Problem — Measuring Local Inference Against the Cloud From Kathmandu

A cloud inference call costs 315 ms from Kathmandu before a single token is generated. A cold 7,200-token prompt costs my laptop 13.8 seconds before it does either. Three runs, two cache states, and a bug in my own harness that had the answer backwards.

Analysis Brief →
Notes2026-08-27

The Benchmark Number Is a Burst — I Ran a Fanless M1 Flat Out for 30 Minutes

746 back-to-back generations across two 30-minute runs on a passively cooled MacBook Air. Both settle on the same sustained rate — 51.7 and 50.9 tok/s — while approaching it from opposite directions, and macOS reports no thermal pressure in either.

Analysis Brief →
Notes2026-08-27

I Tried to Measure What One LLM Query Costs in Energy. The Instrument Lied Twice.

A failed measurement, published in full. The Mac battery gauge refreshes once a minute against generations lasting 2.8 seconds, and where a long enough window exists to beat that, its integral disagrees with the pack's own charge counter by 2.05x. Per-query energy is not measurable on this hardware without root.

Analysis Brief →
Notes2026-08-26

A 2.5× Bigger Model Costs Only 1.55× the Speed — I Measured Why on an 8GB M1

Fifteen benchmark runs across three local models on one MacBook Air, with the quantization variable held still. The result isn't the tokens-per-second table — it's what the table reveals about which physical resource local inference is actually spending.

Analysis Brief →
Silicon & Chips2026-08-16

TSMC's Earnings Call Reveals the Real AI Chip Bottleneck: Not the Fab, the Packaging

A primary-source read of TSMC's Q2 2026 earnings call: the CEO's own words say packaging capacity, not wafer fabrication, is now the constraint on AI chip supply — with a $60-64B capex raise to prove it.

Analysis Brief →
Notes2026-08-16

The Real Ceiling for Local LLMs on an 8GB Mac Isn't RAM — It's Disk

A follow-up benchmark on the same 8GB Apple M1: a genuine Llama 3.2 3B measurement, then a real, unplanned finding — macOS's own swap file, not GPU memory, is what actually stopped this machine from testing a 7-8B model.

Analysis Brief →
Power & Grid2026-08-16

What Constellation Energy's Own Earnings Call Says About AI and the Grid

A primary-source read of Constellation Energy's Q2 2026 earnings call and official results release: real quotes, real megawatt figures, and a CEO thesis — existing plants, not a nuclear building boom — that complicates the tidier version of the AI-power story.

Analysis Brief →
Nepal Desk2026-08-15

Nepal's New Government Wants AI Data Centers. Its Hydropower Has a Six-Month Asterisk.

Nepal's Rastriya Swatantra Party government wants to turn hydroelectric surplus into exported AI compute within five years. The math: a real 1,200 MW monsoon surplus, a dry season the country still covers with Indian imports, and a Sovereign AI Compute Center budget line with no allocation, timeline, or operator yet.

Analysis Brief →
Nepal Desk2026-08-15

How Far Is Kathmandu From the Cloud? I Measured Every Major Region Myself

Real TCP-connect latency, measured live from a Kathmandu residential connection to six AWS regions and three major AI APIs — Mumbai comes in at 44ms, Virginia at nearly 300ms, and the AI APIs are suspiciously fast for a reason worth explaining.

Analysis Brief →
Supply Chains2026-07-30

The Lithium Triangle in 2026: Direct Extraction Reaches Argentina, Codelco Takes Chile, Bolivia Still Waits

Eramet's Centenario plant proved direct lithium extraction works at industrial scale — in the same country where Chile's Supreme Court just settled the Codelco-SQM fight and Bolivia's DLE contracts remain stuck in the legislature. A primary-source status check on extraction technology and the three countries holding most of the world's lithium, as prices swing on named, traceable causes rather than sentiment.

Analysis Brief →
Silicon & Chips2026-07-30

The Memory Supercycle: How HBM Ate the DRAM Market and Made Your Laptop More Expensive

Samsung and SK Hynix both warned in their 2026 earnings that memory shortages will likely persist through at least 2027. Goldman Sachs calls it the worst DRAM supply-demand gap in 15 years — and the same shortage is why Nvidia is cutting RTX 50-series production by up to 40%.

Analysis Brief →
Silicon & Chips2026-07-30

China's Open-Weight AI Labs Are Winning the Model Race. Memory Is Still Winning the War.

DeepSeek, Kimi, GLM, Qwen, and MiniMax now fill eight of OpenRouter's top 10 slots. But Huawei's own chip supply tells a different story: its TSMC die stockpile is exhausted, and domestic HBM production caps Ascend output at roughly 250,000-300,000 accelerators a year.

Analysis Brief →
Cooling & Water2026-07-29

Twenty Olympic Pools a Day: How Johor's Data-Center Boom Ran Into Its Water Table

Johor holds roughly 80% of Malaysia's operational data-center capacity, largely thanks to Singapore's 2019-2022 moratorium. In February 2026, Malaysia froze new non-AI data-center approvals over water and power strain — the first time a Global South AI-infrastructure boom has hit this wall in public.

Analysis Brief →
Power & Grid2026-07-29

The Iberian Blackout, One Year On: What ENTSO-E's Final Report Actually Found

ENTSO-E's March 2026 expert-panel report on the April 28, 2025 Spain-Portugal blackout confirms it wasn't renewables or a lack of grid inertia — it was a voltage-control and reactive-power failure. Here's what the primary source says, against what got repeated in the first 48 hours.

Analysis Brief →
Supply Chains2026-07-29

The Gulf's Sovereign-AI Gamble Just Took Incoming Fire — Literally

Stargate UAE and HUMAIN are still building toward gigawatt-scale AI campuses on schedule. But after Iranian drone strikes hit data centers in Bahrain and the UAE earlier this year, the Gulf's AI bet now carries an openly acknowledged geopolitical risk premium — and investors are still pricing it.

Analysis Brief →
Data Centers2026-07-28

How Much Does It Cost to Run an AI Agent in Production? A Token-Economics Breakdown

Agent cost isn't tokens times price — it's the conversation history you resend on every turn. A cited pricing table across Anthropic, OpenAI, and Google, plus an interactive calculator that models what prompt caching actually saves.

Analysis Brief →
Notes2026-07-27

How to Run a Local LLM: Hardware, Cost, and Tokens-Per-Second Benchmarks

Real tokens-per-second numbers measured on an actual 8GB Apple M1 via Ollama, plus cited community benchmarks showing what changes once you move to a bigger model or a dedicated GPU.

Analysis Brief →
Silicon & Chips2026-07-24

India's $13 Billion Chip Bet: Inside the Semicon 2.0 Package

India's Cabinet approved a ₹1.27 lakh crore chip package on July 15, 2026, extending its semiconductor mission to twelve years — with Micron's Sanand plant already running and the Tata-PSMC Dholera fab targeting first silicon by late 2026.

Analysis Brief →
Power & Grid2026-07-22

Everything-to-Grid: How AI Data Centers, EVs, and Batteries Are Becoming the Grid

The DOE wants virtual power plants to grow from ~33GW to 80-160GW by 2030 while the IEA expects AI data-centre power demand to more than double — a sourced look at the grid dissolving into millions of two-way nodes.

Analysis Brief →
Data Centers2026-07-22

Cloud Repatriation: Why Enterprises Are Pulling AI Workloads Back On-Prem

37signals cut its cloud bill by roughly $2M a year, IDC finds 80% of enterprises expect some repatriation, and Microsoft finished its EU Data Boundary in February 2025, not 2024 — a sourced look at recalibration, not exodus.

Analysis Brief →
Silicon & Chips2026-07-20

Sovereign Compute, Splintered Silicon: How the Global AI Hardware Stack Fractured in 2026

The Biden AI Diffusion Rule was rescinded in two days. TSMC's foundry share rose anyway, to ~70%. A sourced look at how AI chip policy actually moved in 2025-2026.

Analysis Brief →
Silicon & Chips2026-07-20

The Inference Reckoning: Custom Silicon, Sovereign Compute, and the End of Public-Cloud AI Economics

Google's Ironwood, AWS's Trainium3, and up to 1M TPUs for Anthropic — custom AI silicon is real and shipping. Nvidia still owns the merchant layer, sourced and ranged.

Analysis Brief →
Data Centers2026-07-16

Cloud 3.0: The AI Inference Economics Reckoning

Per-token AI prices have collapsed up to 200x a year, yet enterprise compute bills keep climbing. Why standard public-cloud architecture is buckling under inference at scale, and why hybrid, sovereign, and edge deployments are rising to replace it.

Analysis Brief →
Power & Grid2026-07-12

When the Lights Came On: Global Electricity and the Electronic Era in 2026

The year renewables overtook coal, told from Kathmandu — a region-by-region tour of global electricity in 2025-2026, and how power-hungry AI chips, EVs, and batteries are reshaping the grid. Sourced from IEA, Ember, BloombergNEF, and Nepal's NEA.

Analysis Brief →
Data Centers2026-07-08

The World Is Being Rewired: A Data Center Tour of Every Region on Earth

From Northern Virginia's Data Center Alley to Saudi deserts, Nordic snow, and Nepal's hydropower dreams — a grounded, region-by-region tour of the global data center boom in 2025-2026, with real numbers on power, water, and the AI buildout.

Analysis Brief →
Data Centers2026-06-18

Global Tech Landscape 2026: AI Infrastructure & The Global Frontier

A country-by-country analysis of the 2026 global technology sector: a $1.51T semiconductor market, Microsoft's AI infrastructure build-out, Apple's manufacturing shift to India, Europe's defense-tech boom and delayed AI Act deadline, and Latin America's real-time payment rails.

Analysis Brief →
Data Centers2026-06-17

The Web Is Becoming an Agentic Cloud

Cloudflare, Vercel, and GoDaddy are racing to own the agentic cloud. With $2.59T in AI spending forecast for 2026 (+47% YoY), the infrastructure and tooling layers are merging — and the money is moving fast.

Analysis Brief →