Anthropic and the caravel problem · Valentin Radu
Open to project work, technical leadership, and engineer-in-residence gigs. Let’s talk
Get in touch
Name
Message
Send
Thanks, I'll be in touch soon.
← Back
Anthropic and the caravel problem
June 02, 2026
·<br>aiopinion
Lisbon, around the year 1500. A kingdom of perhaps a million people, no bigger than a single city today, had become the hinge of the known world. Portugal held no great army, no broad farmland, no mines worth the name. It held a boat. The caravel carried triangular lateen sails that let it claw against the wind where heavier vessels could only run before it, and in the hands of a generation schooled in new methods of navigation and cartography, it could do the one thing no other ship in Europe could: sail the length of Africa, round the southern cape, and come home again. When Vasco da Gama dropped anchor at Calicut in 1498, that single advantage had handed a small Atlantic kingdom the keys to the richest trade on earth.
For a while the lead looked eternal. Portuguese captains strung fortified trading posts along the coasts of Africa, Arabia, and India, and funneled the spice trade through a single clearinghouse at home. Pepper that cost a few coins in Malabar sold for a fortune on the docks of the Tagus. Lisbon swelled into one of the wealthiest cities in Europe, and the crown that ran it punched centuries above the country's weight. A million people were quietly taxing the commerce of half the planet.
When the latecomers came
The trouble with a technological lead is that it lasts exactly as long as it takes someone else to reproduce it. The caravel was a marvel, and it was also just a design; the routes were secret, and secrets travel. Within a few generations the Dutch and the English were sailing the same seas, and they had arrived carrying something Portugal had never thought to build. The joint-stock company let them pool the savings of thousands of merchants into a single venture, send more ships than any treasury could fund alone, and shrug off the loss of a fleet that would have ruined a king. They financed bigger voyages, garrisoned more ports, and undercut the pioneer in the markets the pioneer had opened. By the early 1600s Portugal had become a junior partner in the world it discovered.
Here is the part that usually gets misremembered as bad luck. Portugal had the best boat in the world until the end, and the best boat did not save it. The Dutch never built a finer caravel; they built the company, the credit, the warehouses, the whole apparatus that turned a voyage into compounding wealth. Portugal had mastered the act of sailing. Holland mastered the system that made sailing pay. One country stayed a guild of navigators. The other became a machine, and the machine ate the guild.
The same shape
Six centuries is a long time for a lesson to keep its edge, and yet it keeps surfacing whenever a small group invents its way to the front of a new frontier. It is surfacing again now.
Anthropic is sailing a caravel. For a stretch it has had the best models on the market: sharper, steadier, better at the work that matters this decade. It reached the frontier early and grew rich the honest way, charging for access by the token to everyone who had not trained a model of their own. The trade is real and the wealth is real. So is the precedent.
A model is, in the end, a file. The gap between the best system and a merely good one is measured in months now, and the floor keeps rising as open-weight models spread for free. A lead like that is borrowed time. And the ground underneath it is shifting: this week NVIDIA put a chip into ordinary Windows machines with enough unified memory to hold a frontier-class model and enough compute to run it, on a desk, off the network, inside the operating system. A capable model is about to live on the device the way search and spellcheck already do. The moment it does, the reason to send your data to a datacenter disappears, and for anyone weighing privacy, compliance, and a per-token bill, a model that never leaves the laptop wins the argument before it begins.
The deeper problem sits underneath the competition. Once a model is small enough to run locally, it is small enough to copy, and a company cannot meter what it has already handed over. Inference revenue, the thing that pays for everything, thins out the moment inference happens on hardware the customer already owns. A lab can spend a fortune training the most capable model in the world and then watch it run, for nothing, on machines it will never charge.
Who collects
The companies positioned to collect are the ones that already own the device. Apple, Microsoft, and Google arrived at good-enough AI late, but they hold the desktop, the phone, the enterprise seat, the silicon, and the capital to fund all of it at once. When the model becomes a feature of the operating system, the value flows to whoever owns the...