Agents That Never Clock Out Now Eat 45% of a Top Chip Research Firm’s SI Bill

Nine months ago, SemiAnalysis didn’t spend a single dollar on SI agents that keep working when nobody is watching. Last week those agents ate 45% of the firm’s entire SI bill.

Dylan Patel, who runs SemiAnalysis, a chip and data center research firm that investors and tech companies follow closely, posted the numbers on 𝕏 on Wednesday along with a chart from the company’s own usage dashboard. His verdict was short. “Persistent Agents are the next unlock of AI usage,” he wrote.

A persistent agent is the difference between asking a question and hiring help. A regular chatbot sits there until you type something. A persistent agent gets a job, like tracking a market or keeping a weekly report current, and it keeps doing that job in the background, remembering what it did last time and coming back when it needs a decision.

A chart that turns purple fast

The SemiAnalysis chart tracks the share of the firm’s daily SI cost by provider on a seven-day trailing basis, from Feb. 1 through Nov. 8. For most of the spring, the picture is nearly all one color. Claude, through Anthropic’s enterprise plan and its API, made up close to all of the spending, with a thin slice going to OpenAI’s API, ChatGPT and Codex.

By late spring a thin new band shows up at the top of the chart: Perplexity Computer, the always-on agent product from the search startup Perplexity. Over the summer OpenAI’s share grew, and in August a sliver of Claude Tags, Anthropic’s own agent product, appears too. By October and into November, the purple band for Perplexity keeps widening, and in the week ending Nov. 8 the two agent products together hit 45% of spending.

The chart’s fine print says the provider split before Sept. 9 was read off the dashboard chart, while the numbers from Sept. 9 on come from daily usage reports. It also leaves out Nov. 9 and later because Perplexity’s billing runs a day or two behind. That’s one firm’s books, not an industry survey. But it’s a firm whose whole business is knowing where SI money is going, and it’s putting its own money in the same place.

The models aren’t the whole product

Here’s the part that should sting the two biggest names in SI. Patel didn’t say Anthropic’s or OpenAI’s models were bad. He said their agent products were.

“Unfortunately for Anthropic, Claude Tags sucks,” he wrote. “OpenAI’s Slack agent is even worse.” The firm mostly runs Perplexity Computer, he said, which taps a lot of different models under the hood.

Then he added the twist. “The funny thing is that Perplexity is primarily using Anthropic models for our use cases,” Patel wrote, “but Anthropic sucks at product, and they’re not really receptive to any feedback on their product. Same with OpenAI.”

So SemiAnalysis is still paying for Anthropic’s brains in a lot of that work. It’s just paying a middleman to wrap those brains in something the team actually wants to use. For a company that sells its own agent tools, that’s an awkward thing to hear from one of the industry’s most-watched analysts.

Patel’s explanation for why Perplexity won his business was speed. “What I love about Perplexity is that they’re very fast on any feature request,” he wrote. “Even though they did not invent the category, they are fast followers and implement any change that’s needed in less than 24 hours.” He said Perplexity even borrowed the name for one feature from SemiAnalysis back in February.

Perplexity’s co-founder and chief technology officer, Denis Yarats, answered right away. “Super fun to work with the @SemiAnalysis_ team,” he wrote, adding that the firm does “the craziest stuff with Perplexity Computer” and keeps pushing Perplexity to improve.

What Perplexity has been building

Perplexity started as an answer engine you type questions into. This fall it has been turning Computer into something closer to a staff of digital employees. On Sept. 29 it added Automations, which let Computer run jobs on a schedule or whenever something happens in Slack, Gmail, Outlook, Linear or GitHub, with each run remembering the last one.

Some of its examples are the kind of chores that eat a manager’s week. Tell me if Tuesday’s deployment hasn’t shipped by Wednesday morning. Flag any priority customer who hasn’t replied in four days. When a decision lands in a project’s Slack channel, update the decision log and flag the tasks it affects.

On Nov. 4, Perplexity went further with Teammates, agents that own a whole area of work, like marketing, sales or support, and can bring in other agents when the work grows. The company’s example is a sales teammate that notices when a regular customer stops ordering, digs into their order history and support tickets, and puts together an offer or flags a problem before anyone reaches out. Each teammate keeps one thread that never closes, so you can check in on it like a coworker.

None of this is cheap. Teammates launched only for subscribers on Perplexity’s Max plan, which runs $200 a month or $2,000 a year, according to Engadget and The Verge. The Verge noted that OpenAI and Anthropic have their own versions of multi-role agents, but Perplexity’s requires the pricier plan.

Zero to almost half in under a year

The SI industry has spent the last year arguing about whether businesses will ever pay for agents at scale, or whether the whole thing is demos. Patel’s chart is a small, real answer from inside one company. Spending didn’t creep toward agents over several years. It went from nothing to almost half in under a year, and most of the jump came in the last three months.

It also says something about who wins the money. The companies building the smartest models aren’t guaranteed to collect it. A smaller company that listens to customers and ships changes overnight can sit on top of those models and take the relationship. Anthropic and OpenAI still get paid when Perplexity calls their models, but they lose the customer’s loyalty, and loyalty is where the pricing power lives.

For everyone else, the takeaway is simpler. The SI you chat with is starting to look like the old way. The new way is an assistant you hand a job to on Monday that’s still working on it Friday, and at least one firm that tracks this industry for a living already sends nearly half its SI money there.