The most valuable thing an artificial intelligence company owns is also the easiest to copy. That uncomfortable idea sits at the heart of a dispute between Anthropic, the American maker of the Claude models, and Alibaba, the Chinese technology giant. Anthropic claims its rival learned to mimic Claude not by stealing code or smuggling chips, but simply by talking to it. The fight matters far beyond the two firms, because it arrives just as Anthropic is said to be preparing one of the largest public listings the technology world has ever seen.
How you copy a model by asking it questions
The technique at the centre of the row is called distillation. In plain terms, a competitor sends a flood of questions to a powerful model, records its answers, and then uses that record to train a cheaper system of its own. The copy never sees the original's inner workings. It just studies its outputs, the way a student might learn a master's style by watching the finished paintings rather than reading the notes.
Anthropic alleges that Alibaba did exactly this, using a web of fake accounts and ordinary looking interactions to harvest Claude's capabilities on the quiet. If true, it means a rival can approximate years of expensive research at a fraction of the cost, without ever breaching a firewall or breaking an obvious rule. That is a nightmare for any company whose entire value rests on staying ahead.
The moat question
Investors love a moat, the durable advantage that keeps competitors at bay and profits flowing. In software it has usually meant network effects, switching costs, or data no one else can touch. The trouble with frontier AI is that the lead a lab spends billions to build can be narrowed by anyone willing to ask enough clever questions. The product that took a fortune to create can be partially reverse engineered through its own front door.
That is why the timing stings. Anthropic is reportedly heading toward a public offering later in the year at a valuation that would put it among the most richly priced debuts in history. A listing on that scale asks investors to believe the company can defend its position for years. The distillation claim asks them to wonder whether it can defend it at all.
Two ways to read the same fight
Not everyone sees the episode as bad news for Anthropic. One reading is that imitation is the sincerest form of validation. If a rival is going to the trouble of copying Claude, the argument goes, it is because Claude is the benchmark worth copying, much as buyers still pay for the premium original even when a cheaper clone exists. By this logic the dispute confirms Anthropic's place as the leader.
The gloomier reading is harder to dismiss. The company's revenue has been growing at a remarkable clip, but the real question for anyone buying into a public offering is whether that pace can last. If cheaper imitations keep nipping at the heels of every new model, margins could be squeezed long before the lead disappears. Growth that cannot be protected is worth less than growth that can.
A rule book written for the wrong threat
Part of what makes distillation so slippery is that the rules were built for a different kind of theft. Export controls were designed to stop tangible things from crossing borders, advanced chips and physical hardware above all. Querying a model through its public interface does not count as exporting it, which leaves a wide gap between what the law forbids and what actually transfers know how.
Anthropic has urged Washington to close that gap, pressing for controls on access to advanced American computing power rather than just the machines themselves. Lawmakers have floated measures that would limit how foreign users reach American technology through cloud services, and officials have condemned unauthorised copying. For now, though, the policy is racing to catch up with a method that treats a chatbot's open door as a way in.
What it means for the AI boom
The stakes reach well beyond a single listing. If the most advanced models can be partly cloned simply by using them, then the enormous sums pouring into frontier AI rest on a shakier foundation than the headline valuations suggest. The race may reward not whoever builds the best model once, but whoever can keep building the next one fastest, staying just ahead of the copies.
That is a different kind of business from the one investors usually pay a premium for. It looks less like owning a fortress and more like running on a treadmill that never slows. Anthropic may yet prove it can stay in front, and a successful listing would suggest the market believes it can. But the quarrel with Alibaba has put a sharp question at the centre of the AI gold rush, namely whether anyone can truly build a wall around genius that answers to anyone who asks.
