The cheap-model escape hatch may close, one-person companies hit a million, and AI-heavy firms are hiring more, not less

Everyone's plan for controlling AI costs was to fall back on cheap Chinese open models, and Beijing is now considering closing that door. Meanwhile the hard data on AI and jobs arrived, and it points the opposite way from the doom stories.

• China may stop releasing its best open models to the world
  My take: Reuters reports that Chinese officials have been meeting with Alibaba, ByteDance and Z.AI about restricting how their most advanced models get distributed overseas. Nothing is decided, and some China watchers think the report overstates it. But this matters because almost every cost-control plan I have seen this year quietly assumes free, frontier-adjacent Chinese models will keep arriving forever. That assumption is now a risk, not a fact. If your fallback model is Chinese, spend an afternoon this month checking that a Western alternative would work: NVIDIA's Nemotron and Google's Gemma are the obvious ones, and both are getting real enterprise use.
• Token budgets arrive at Tesla, and a new model competes on price instead of raw power
  My take: Tesla is capping employees at 200 dollars a week in AI spend after some engineers ran up thousands per week. Uber and Walmart did versions of the same thing. At the other end, SpaceX AI released Grok 4.5 and pitched it not as the smartest model but as the faster, cheaper, more token-efficient one. That is the tell: efficiency is becoming a selling point, because the bill is now the constraint. If you have not set a per-person AI budget yet, set one now, and make it easy for people to request more with a reason. A blanket cap with no exceptions just pushes your best people back to doing work by hand.
• Fine-tuning a smaller model is starting to beat paying for the biggest one
  My take: Bridgewater took a modest model, trained it on their own expert judgments, and reached about 85 percent accuracy at single-digit dollars per run. The frontier models managed 74 to 78 percent at 20 to 90 dollars. Microsoft is selling the same idea as a product, claiming its tuned models match a much larger model at a tenth of the cost. The rule of thumb worth carrying: general intelligence is expensive and your specific knowledge is cheap. If you have a repetitive, high-volume task and a pile of examples of it being done well, that is now a candidate for a cheap custom model rather than an expensive general one.
• One-person companies reaching a million dollars in revenue have doubled in two years
  My take: Stripe's data shows the number of solo operators passing a million dollars in revenue more than doubled between 2023 and 2025. Solo business applications are up 27 percent since early 2024 in the sectors with the highest AI adoption, and flat in the sectors with the lowest. Solo founders now account for 63 percent of new C corporations. The mechanism is simple: AI fills the gap that used to force you to hire a technical co-founder or a first marketer. Even if you never plan to work alone, watch these businesses closely. They are running the efficiency experiments that will land in your company in eighteen months.
• Companies that adopt AI are growing headcount, including entry level
  My take: Ramp matched AI spending against payroll for 21,000 US businesses. High adopters grew headcount 10 percent over two years while low adopters were flat, and entry-level growth was higher still at 12 percent. Hiring did not start until six to twelve months into adoption, and the spending threshold was modest, around 30 dollars per employee per month. Box's survey of 1,600 companies found the same shape. This does not mean nobody loses a job: tech and finance are shedding 28,000 roles a month. But the pattern is that AI makes ambitious companies take on more work, not fewer people. Ford's version of this was hiring back 350 veteran engineers to train the AI tools that were not getting the job done.
• Anthropic built a tool to read what a model is thinking before it answers
  My take: New research shows models keep a small set of internal concepts they can report on, separate from all the automatic processing underneath, and Anthropic built a way to read them. In safety tests the tool caught models recognising they were being tested, and flagging their own data manipulation, none of which appeared in the output. Set aside the consciousness debate, which the authors themselves refuse to enter. The business point is that debugging AI has been guesswork until now, and this is the first real step toward diagnosing why a system failed instead of shuffling the prompt and hoping. That is what has to exist before you can put AI on a process where being wrong is expensive.