Every few months a Chinese lab ships a model that matches a US frontier model at a fraction of the training cost, and the internet reacts like someone got caught cheating. DeepSeek did it first. Then Alibaba’s Qwen. Then Moonshot’s Kimi. Each time, the headlines read the same way: the labs that spent billions are getting outrun by labs that spent a lot less.
I went and read what the US CEOs actually said, not what got said about them. It’s a lot calmer than the discourse around it.
Building good things takes the same three ingredients
Building something good, physical or digital, takes a good idea, some experience, and a good workshop. That’s true whether you’re building a cabinet or a piece of software. None of those three things are unique to you. Other people have good ideas. Other people gain experience. Other people build good workshops, sometimes for less money than you spent on yours. I feel that last one personally.
Software doesn’t get an exception to that. It never has.
No production software escapes commoditization
Thinking you’ve built a piece of technology that won’t eventually become a commodity is the height of hubris. Databases were a moat once. Then they became infrastructure. Web browsers were a competitive battleground. Now they’re free and mostly interchangeable. Search was a research problem few companies could solve. Cloud compute was scarce and expensive right up until it wasn’t. VMs. Containers. Kubernetes. All had their moment. All are pipes inside the walls now.
It’s never happened that a piece of production software stayed uniquely valuable forever. It’s not going to start happening with LLMs. A model is software. Software commoditizes. That’s the whole pattern, and it was always going to apply here too.
What the CEOs actually said
Here’s the part that surprised me. Go read the actual statements from US AI leaders when DeepSeek’s R1 dropped in January 2025, and none of it sounds like panic.
Sam Altman posted on X that R1 was “an impressive model, particularly around what they’re able to deliver for the price,” and called it “legit invigorating to have a new competitor.”
Dario Amodei wrote on his own blog to correct the record, not to sound the alarm. DeepSeek hadn’t matched frontier US models for pennies on the dollar. It had produced something “close to the performance of US models 7-10 months older, for a good deal less cost.” He used that as an argument for tighter export controls, not a reason to panic about being behind.
Demis Hassabis, per CNBC’s reporting on internal remarks, called DeepSeek’s work “probably the best” to come out of China, then said Google has “all the ingredients” to stay ahead.
None of that is alarmed. It’s measured, a little competitive, and mostly focused on what to do next. The people closest to the actual technology were the calmest people in the entire story. The zeitgeist did the panicking for them: the “China is beating us” framing came out of media coverage and policy debate, not out of the labs building the models.
There’s one real exception, and it’s worth naming because it’s the opposite of a benchmark panic. In June 2026, Anthropic told the Senate Banking Committee that Alibaba-affiliated operators had run 28.8 million exchanges with Claude through roughly 25,000 fraudulent accounts, calling it the largest distillation attack they’d seen. That’s a specific grievance about IP theft. It’s not “we lost the race.” It’s “someone stole the answer key,” and that’s a legitimately different complaint.
Thompson called it before the panic did
Ben Thompson wrote Stratechery’s DeepSeek FAQ on January 27, 2025, the same week all of this broke. His line has aged well: “If models are commodities… then long-term differentiation comes from having a superior cost structure.”
That’s the actual shift, and he said it before most of the internet caught up. If the model itself stops being the moat, the advantage moves down the stack, to whoever can serve inference cheapest, and up the stack, to whoever builds the best product on top of it. That’s bad news for the companies whose whole pitch was “our model is better.” It’s good news for the cloud platforms and for anyone building an actual product instead of just a model.
They were never magic
LLMs are a commodity now, and they were always going to become one. That’s not a knock on the technology. It’s just what happens to every piece of software good enough to matter. I’ve made a version of this argument before: it was never really about the model, it’s about what surrounds it.
Interacting with them well is a UX problem. Getting quality, reliable output out of them is a hygiene and testing rigor problem: know what you’re asking for, check what you got back, build the harness that catches it when the model gets it wrong.
They’re not magic. They never were. They’re just sharper tools than what we had before, and sharper tools still need someone who knows how to use them. I have scars to prove that one too.