• by Laurie Sullivan , Staff Writer, Yesterday

Analysts at Madison & Wall, an advisory firm focused on the advertising industry, led a conversation with Gil Luria, head of technology research at D.A. Davidson, connecting the dots between the advertising industry and some of hardware’s biggest challenges related to AI.

They spoke about data centers, chip design, manufacturing, memory, servers, large language models (LLMs) and the capital required to support innovation. 

This knowledge can help media companies better anticipate businesses across the advertising industry.

The connection seems natural to this reporter, who formerly wrote about semiconductors and other electronic components, 

Luria said OpenAI and Anthropic are operating at a run rate of about $120 billion -- up from zero dollars three years ago and perhaps $20 billion a year ago, which represents the cost demand for AI tools.

“They are probably at their peak power,” Luria said, adding that he foresees a future in which their scope will narrow in two ways.

For simple queries like “summarize my emails” or “book me a reservation,” users will not need a deluxe model, but more likely open source.

For more complex queries, there is an entire software infrastructure that today will route queries to companies like Anthropic Claude or OpenAI ChatGPT.

The companies can route the query to the least expensive model and the person making the query will receive the answers needed. But they will take all the information.

The process begins with enterprise software infrastructure, which uses semantic routing to automatically direct queries to the least expensive adequate model, while privacy advocates confirm that users willingly give their conversation data to providers like OpenAI and Anthropic.

Companies manage this through dedicated AI gateways to optimize infrastructure costs and handle complex enterprise AI agents.

Luria cautioned that this is "okay if it’s me and my email, but if you’re a company, I can not warn you enough."

He added: “Do not have unprotected sex with Anthropic or OpenAI. You are giving them the keys to the kingdom when you put your information into the model unfettered.”

"Unprotected sex" is a metaphorical warning -- the high risk of sharing confidential information or granting AI-based chatbots unvetted access to local systems without digital protective barriers.

Luria said that in the future, people will realize this and will want a buffer -- a routing layer like software from Microsoft and other companies.

This is where the person’s information gets retained, and the AI query gets routed by software manufacturers to the “optimal” model, and that’s when the information layer will be routed away from Anthropic and OpenAI.

For people using consumer apps with the free, public web interfaces of ChatGPT or Claude, those companies retain user data by default to train future models, although users can opt out through user settings.

If a business uses the "software infrastructure" to route queries via developer APIs, both OpenAI and Anthropic have strict legal agreements. They do not log the user data for training, and they delete the data after a short retention window, usually 30 days for abuse monitoring.