Where AI Is Headed

I see two major trends in the AI industry going forward. One is addressing capacity shortages; the other is improving cost efficiency.

The capacity shortage comes from the compute bottleneck. Relative to a world of effectively unlimited compute, the algorithmic ceiling of industrial-scale language models is still far from exhausted. In that sense, the AI bubble has not truly arrived yet. At the same time, however, demand currently exceeds supply, so companies are under pressure to expand capacity and improve token efficiency in order to reduce the cost borne by users.

For example, Antigravity has argued that the token loss introduced by third-party agent frameworks needs to be reduced, and as a result it cut off third-party agent harnesses from using Claude subscriptions.

At bottom, the model layer has become strong enough while the agent layer remains too weak, creating a mismatch between subscription pricing and real inference cost. That alone tells us a great deal about the direction in which compute utilization is heading.

The second trend, cost reduction and efficiency gains, is driven mainly by the financial bubble around AI. Right now the field is still in its summer, but when autumn and winter arrive, the bubble will contract. OpenAI reportedly lost 12 billion dollars last year, and every major player is still burning enormous amounts of cash while monetization remains insufficient.

I think this will lead to three consequences. First, the bubble may burst overseas before it does in China, weakening confidence among domestic investors and forcing companies to cut headcount, trim expenses, and restructure their business chains. Second, the industry will have to move toward models that burn less money while generating more revenue. Spending less means improving inference efficiency and infrastructure; earning more means pushing further into the application layer, including vertical products and consumer-facing content, to monetize more effectively. On the consumer side, I am especially bullish on Physical AI. Many current products are still little more than VLM wrappers, but I believe this area will become extremely important. Third, competition will intensify and weaker players will be eliminated, with both large and small companies building their own moats and operating more independently. OpenAI, for instance, is now investing in AI communities and AI news delivery, using social features to increase stickiness, turn weekly active users into daily active users, and broaden its monetization channels.