Foreign media reports that a recent SemiAnalysis interview suggests a shift in the competitive landscape of generative AI. The competition, previously focused on pre-training and computing power, is gradually moving towards enterprise clients, API revenue, and model utilization efficiency. The article states that programming has become the primary scenario for token consumption, reshaping the competitive positions of OpenAI, Anthropic, Google, and xAI.
Programming becomes the core of income
The article states that current enterprise AI spending is not concentrated in general office work, but in programming and software engineering. SemiAnalysis estimates that over 70% of the API revenue of frontier labs comes from this scenario. Some heavy enterprise users are spending an average of $100,000 per year on AI.
This means that while companies are beginning to review their AI budgets, spending has not been reduced across the board; instead, it is concentrated on development phases with more tangible returns. The article argues that restricting employees' use of high-level models may not actually reduce costs and could even weaken efficiency.
Subscription subsidies under pressure, vendors turn to APIs
The article argues that fixed-month subscriptions are facing profit pressure. For example, a premium subscription of $200 per month, when converted to API pricing, corresponds to several thousand to tens of thousands of dollars in credit. Once user engagement increases even slightly, service providers' profits will be significantly squeezed.
SemiAnalysis believes this is also why OpenAI and Anthropic are more actively pushing enterprise customers to switch to API billing. The article mentions that Anthropic's enterprise solutions are now more pay-as-you-go, thereby improving gross margins and overall profitability.
- Programming is considered the largest token consumption scenario.
- Enterprise APIs are considered a higher-margin business.
- Subscription models place greater pressure on subsidies for frequent users.
OpenAI catches up, Google is accused of falling behind.
The article argues that OpenAI's growth slowed earlier this year, but with the release of models 5.5 and 5.6 and Codex, enterprise API revenue rebounded. SemiAnalysis concludes that the competition between OpenAI and Anthropic is approaching a duopoly.
In contrast, the article is more pessimistic about Google. Its core argument is that Gemini's performance fell short of expectations, and Google's previous long-term TPU rental agreements limited its own computing power allocation, affecting model training pace and exacerbating talent loss pressures.
xAI and Meta explore monetization of computing power
The article also mentions that xAI and Meta are experimenting with different business models for computing power. SemiAnalysis states that xAI leases Colossus computing power at above-market prices while retaining shorter payback periods to balance cash flow and flexibility for self-use. Meta, on the other hand, is said to be pursuing a Neocloud-like solution to find returns on its large capital expenditures.
According to the article, the relationship between cloud vendors and model companies is also changing. As enterprises place greater emphasis on compliance, model selection, and delivery capabilities, distribution channels around Token-as-a-Service are expanding, with AWS, Azure, and GCP still holding a clear advantage.
The RL data market is heating up.
Regarding the next stage of capability enhancement, the article argues that reinforcement learning is replacing pre-training as a more valued expansion direction. Budgets for purchasing high-quality RL environment data are also rapidly increasing.
SemiAnalysis estimates that frontier labs may have invested over $10 billion in this type of data this year. Particularly in software engineering tasks, the price of a single high-quality RL environment task has risen to five figures, indicating that the AI data supply chain is shifting from low-end annotation to services with higher technical barriers.
