SemiAnalysis presents GDPx benchmark for evaluating frontier AI models

Members of SemiAnalysis’s tokenomics team presented GDPx and described it, in their own view, as the best benchmark for measuring frontier-model capabilities. They criticize approaches based on an industry’s share of GDP or SOC and O*NET taxonomies: in their assessment, the first trillion dollars in AI revenue will be concentrated in fields that can adopt the technology quickly and earn high returns from frontier intelligence. They named chip design, biology, cybersecurity and RSI among GDPx’s focus areas.

A team representative said SemiAnalysis is working on a demand model to project OpenAI and Anthropic revenue in 2027 and 2028, and declined to detail the analysis at this stage. Another team member challenged the assumption that broad public adoption of AI will itself drive large AI spending: in his view, most people will struggle to use AI effectively across many tasks. He also said the cost of AI at a given capability level is falling by 100x to 1,000x a year, driven by progress in research, data and compute. He argued that the greatest demand will therefore come from users able to get rapid returns from large amounts of machine intelligence.

The team believes white-collar work will eventually be automated, but will not drive the first trillion dollars of AI spending: in their assessment, adoption in such industries will be slow because companies will need time to reorganize operating expenses.