AI spending is becoming extremely concentrated: The proportion of US software-spending businesses using paid GPU infrastructure is up to a record ~10%, from less than 4% two years ago. Model serving and inference, or spending on the computing power needed to run AI models and process AI requests, account for the vast majority of this increase. The proportion of firms using model serving and inference is up to a record 8.7% from 2.4% over this period. Meanwhile, the top 10% of firms by spending account for 99.5% of model-serving spending and 99.0% of neocloud spending, or specialized AI cloud services. The top 10% also represent 95.7% of cloud computing spending and 95.6% of spending on OpenAI and Anthropic. AI adoption is broadening, but spending remains heavily concentrated.