The Windstorm

A weekly dispatch from the eye of the storm

Week of October 19, 2026

Welcome to the Eye of the Windstorm.

This week the price of intelligence fell twice, in two different directions. On one side, Google made Gemini 4 Argon broadly available — a model built to grind through long, multi-step agent work — at a promotional $2 per million input tokens. On the other, NVIDIA released open models you can download and run on your own hardware, no cloud account required. Rent it cheaper than ever, or own it outright: both doors are open at once. Which quietly retires the last technical excuse and sets the whole game down on a single human question — what, exactly, are you going to point it at?

1. Renting intelligence got cheap — read the meter anyway

Argon landed this week at $2 per million input tokens and $10 per million output, with cached input 95% off. Enjoy it while it lasts: the promo expires November 30, and on December 1 the standard rate doubles to $4 and $20. Whatever the launch claims say about speed and cost per task, test it on your real work. And read the fine print: a Google Cloud billing change taking effect November 13 means that on long requests (completions over 32K tokens) you can be charged for tokens even when the job times out and fails. The sticker price was never the cost. How well you aim the thing is.

2. For the first time, you can own the engine

NVIDIA's Nemotron 3.5 is a 120-billion-parameter open model (about 12 billion active per token), shipped alongside smaller “Bolt” versions — one of which squeezes into 24GB on a desktop GPU, with a 4GB version due October 23. It is free to download on Hugging Face, and the Bolt versions run under everyday tools like Ollama and LM Studio. Why a business leader should care: you can now run capable agents inside your own walls, on your own data, without handing anything to someone else's cloud. Atlassian and Salesforce are already building on it, and Scale AI post-trained it for an EY due-diligence agent. Renting was never the hard part; deciding what is too sensitive to rent is a judgment call — and now you actually have the choice to make.

3. The capability raced ahead. Your org didn't.

GE Appliances said back in April that it is running more than 800 agents across manufacturing, logistics and supply chain. Impressive — until you sit with a second number. An IBM Institute for Business Value survey (a standing finding from June, not this week's) puts the share of CIOs and CTOs who feel fully ready for the agent scale they expect within a year at just 11%, with more than three-quarters of organizations admitting adoption is outrunning their ability to govern it. Notice which figure is the technology and which is the human. The agents are ready. The deciding, the governing, the winnowing of 800 agents down to the handful that actually earn their keep — that part is still entirely on us.

The Eye

Step back from the noise. Both of this week's headlines look like technology stories — a cheaper model, an open one. They are not. They are the same story told twice: the cost of raw capability is racing toward zero from two directions at once, and everything getting cheaper is the part that was never actually hard. Buying the horsepower was always the easy purchase. The expensive thing — the thing no price drop ever touches — is knowing what you want done, clearly enough to say it plainly to something that will do exactly what you ask and nothing you forgot to.

So the skill of this decade isn't prompting, and it isn't picking the right model. It's winnowing: standing in front of a thousand things an agent could do and choosing the few worth doing, then defining “done well” sharply enough that a tireless machine can't misread you. Nobody is a virtuoso at this yet — not the vendors, not the consultants, not me. We are all learning to conduct an orchestra that showed up already tuned and waiting on the downbeat. And here is the good news from the calm center: the most valuable move you can make this week costs nothing and runs on no GPU. It is simply getting clear about what you actually want.

Do this with your agent this week

Pick one real task from last week — something you'd normally hand a sharp new hire on their first morning. Before you let any agent near it, write three sentences: what a great result looks like, what it must never do, and how you'll know it worked. Then hand the agent the task and those three sentences. Watch how much of the quality came from your clarity rather than the model. That gap is the entire job.

From the eye of the storm,
Grant