The U.S. will soon spend more electricity thinking than making steel, aluminum, and cement combined. And the plan is... more power plants? My thoughts below!

India's datacenter water debate is missing one thing: the comparison. Thermal power (today): 2,100B litres Textiles (today): 580B Datacenters (2030): 358B AI Afterhours runs the math tomorrow 👇 open.substack.com/pub/aiafterhou…

India's datacenter water debate is missing a comparison. Thermal power: 2,100B litres/year Textiles: 580B litres/year (at 2x best-practice waste) Datacenters: 150B litres/year → 358B projected by 2030 The framing problem — and the return math nobody ran aiafterhours.substack.com/p/compared-to-…

A guy used an AI coding assistant to control his robot vacuum with a PS5 controller. Got access to 7,000 of them across 24 countries. The entire IoT security model was built on one bet: exploits are expensive to find. AI just made them free. aiafterhours.substack.com/p/the-accident…

Jensen built an empire on FLOPS. At GTC 2026, he pitched efficiency instead. Perf/watt got more stage time than raw throughput. NVFP4. Vera Rubin's power envelope. The company that made "more transistors" a personality trait just admitted the game changed. aiafterhours.substack.com/p/jensen-huang…

Sora dying isn't about video generation but the opportunity cost of compute being too high. Every GPU-hour on consumer video is a GPU-hour not serving enterprise inference. OpenAI just discovered that token generation has a market price, and Sora was burning tokens below cost.

Forbes est. ~$15M/day inference cost at $2.1M lifetime revenue. Sora was a great product but an economics failure. Inference is multidimensional and OpenAI unf. optimized exactly one dimension.

Supply chain attack on a model routing library. When agents pick which model to call, every dependency in that chain becomes an attack surface. Routing isn't just engineering anymore — it's risk management.

This was February. Now Warren and Hawley are demanding mandatory energy reporting from data centers too — hourly loads, actual rates paid, who funds grid upgrades. 300+ state bills in six weeks. The free ride on ratepayer-subsidized AI power is ending. aiafterhours.substack.com/p/the-gigawatt…

Four npm supply chain attacks in one month. Axios (100M downloads) compromised last night. This is the security economics breakdown I wrote about in Feb, playing out in real time. What's happening + what to check: aiafterhours.substack.com/p/the-npm-ecos…

$2B on photonics because moving data costs 30-100x more energy than computing it. Jensen spent a decade selling FLOPS and is now buying plumbing.

Heat is waste energy. A data center that warms its neighborhood 9C is one converting electricity into atmosphere instead of inference. Fixing the data movement will fix the heat.

While we've heard "too dangerous to release" before — OpenAI said it about GPT-2 with 124M parameters. Yet CVE-2026-4747 is very real. 17-year-old FreeBSD RCE, no human team ever caught it. The boy-who-cried-wolf problem is that sometimes the wolf actually shows up.

Fascinating move. Meta rebuilt their entire training stack and decided the infra is worth more than open weights. Clearest signal yet that the GPU stack is unbundling!

$121 billion on compute in 2028. $85 billion in projected losses. Both companies report profit when you strip out training costs — which tells you training is the whole game right now. Inference is >50% of revenue for both!

Same model class, opposite release strategies. One restricts to 40 orgs, one pushes to thousands. Both confirm AI just repriced vulnerability economics — offense and defense costs collapsing simultaneously.

Same token pricing, better benchmarks. More useful output per dollar — that's the whole Anthropic financial thesis right now. Their revenue trajectory makes more sense when you look at unit economics, not model size. aiafterhours.substack.com/p/openai-vs-an…