Developers Forced to Train Their Own Models! Anthropic, OpenAI Enter Application Development - Will They Stop Selling Advanced AI Technology to Enterprises?

According to the latest foreign media reports, AI infrastructure providers Anthropic and OpenAI have started developing their own applications, accelerating their deployment in vertical markets. This move has created direct competition with their own API clients, forcing developers who rely on advanced models to reconsider their technology roadmap.

The Information reports that the latest case is the design tool Canva, whose core functions are now in direct competition with Anthropic's self-developed applications. This "supplier as competitor" situation has led more and more developers to re-examine their technology roadmap.

Currently, API business remains the core revenue source for both Anthropic and OpenAI. The former's second-quarter revenue was approximately $115 billion, growing about 14 times year-over-year, and has already achieved profitability in terms of adjusted operating income. However, whether this situation can continue has become an open question in the industry.

The report points out that three forces are driving the tightening of APIs. The first is safety concerns. The Trump administration has pressured Anthropic and OpenAI to adopt "phased releases" for some new models, requiring customers to obtain access permissions only after individual case approvals. The Trump administration has also announced a testing process for major AI companies before model releases.

Even after the models are publicly released, Anthropic and OpenAI may proactively reduce model performance in specific tasks for safety reasons. The Fable model under Anthropic has had its capabilities in network security tasks intentionally weakened. Notably, Anthropic and OpenAI do not need to perform the same performance degradation when driving their own proprietary applications.

Both Anthropic and OpenAI have publicly warned that future models may possess dangerous capabilities in areas such as hacking and bioweapon development. If malicious actors break through the defenses, completely cutting off external access to advanced models may be considered a safer choice.

The second force is anti-distillation. Competitors can use the output results of commercial models to train new models with similar performance, effectively copying advanced capabilities at low cost. In Anthropic's potential disaster risk report released last week, it warned that models obtained through distillation may inherit the dangerous capabilities of the original model but lack safety protection, potentially causing downstream risks to the world. However, a former Anthropic researcher admitted that "it's basically impossible to completely stop distillation."

Open-source software advocates hold a different position, believing that open-source models derived from Anthropic's models actually help developers defend against network security attacks. The previous incident of OpenAI's AI agent infiltrating Hugging Face is a testament to this.

The third force is the most realistic commercial logic. Investors have warned developers: "Anthropic may reserve the most advanced technology for its own competitive applications rather than making it available through APIs. Anthropic has already entered emerging fields such as AI-driven drug development."

If Anthropic, as the absolute leader in the API market, restricts competitors in this way, it may face a large-scale antitrust investigation. However, the potential returns of vertical applications are more abundant, which is the fundamental reason for the two companies to accelerate their deployment.

The report points out that facing the uncertainty of API access permissions, major application developers have started to take action. Legal AI company Harvey and programming editing tool Cursor have begun training their own internal models to reduce their dependence on Anthropic and OpenAI, and more developers are expected to follow this path.

When the core infrastructure supplier can become a competitor at any time or unilaterally change service terms, building autonomous and controllable technical capabilities becomes an inevitable choice for enterprises. The balance of power in the AI industry is shifting from "shovel sellers" to "shovel users," and this time even the supply rules of the shovel are being rewritten.