Anthropic pays to build the people who deploy its AI
Edition of 2026-10-03
# AI For All ### Saturday, October 3, 2026
As of today, a talent or operations lead at a large company can ask whether their own engineers can be trained to put AI to work inside the business.
The scarce part is not the model
Anthropic opened a training programme on Friday that is aimed at the people who deploy its models rather than at the models themselves. The Claude Frontier Academy is backed by a *$100 million commitment and aims to train 10,000 Frontier Deployed Engineers* by the end of 2027.
The design is borrowed from medicine. Organisations nominate their own engineers. Each candidate goes through a short in-person course and a graded practical, then enters a *12-week residency*, where they lead a real deployment inside their own company with Anthropic staff on call. Only then do they earn the badge. The first cohorts come from Accenture, Bain, Capgemini, Deloitte, McKinsey, Morgan Stanley and Novo Nordisk.
What stands out is the diagnosis. Anthropic is not selling a better model here, it is admitting that the model is not the bottleneck. A small group of people who understand both the tool and the business is what makes AI pay off, and that group is scarce enough that a model maker will spend nine figures to grow it. If you run a large team, the practical question is who inside it could pass that residency.
The desk AI computer, cut in half
Nvidia added a second version of its desktop AI computer on Friday. The new DGX Spark ships with *64GB* of memory, half of the original, and is built to run capable open models on a desk with no cloud connection at all.
The interesting part is what happens when one is not enough. Two of the smaller units link with a single cable to pool their memory into *128GB*, so the cheaper box is a starting point rather than a ceiling. Nvidia lists Acer, ASUS, Dell, Gigabyte, HP and MSI as the sellers, and says the machine runs from a normal wall outlet, the same stack as the larger model.
Hardware like this is the quiet counterpoint to the subscription. A business that pays for every query is renting a capability it could own outright, and the trade is only worth it if the box is genuinely useful on its own. Nvidia is betting that a model good enough to matter now fits in a case on the floor.
Weak signal
Apple said on Friday that it will add controls around Full Disk Access, the macOS switch that lets a single app read everything on a machine: files, mail, messages, browsing history. Its stated reason is new. As AI agents get more capable and more autonomous, Apple says, the risk of that blanket access grows, and granting it should take an explicit act rather than a default.
There is no date yet and nothing to act on today. It is worth holding because of the reach: this is a change that will touch every Mac a business owns, and Apple has now named AI agents as the reason it is moving.
The model gets the headlines. Friday, both Anthropic and Nvidia spent their news on the parts that decide whether it gets used: the person who can deploy it, and the box it runs on.