SemiAnalysis, the semiconductor and AI research shop that industry insiders actually read, just dropped its own internal numbers on AI agent productivity — and they're absurd. The firm tracked nine real analyst workflows, from deep company research to financial data pulls, and compared the token costs against what a human would bill. Every single task delivered an ROI above 10x, most landed between 60x and 90x, and one standout hit 93.8x.
https://twitter.com/SemiAnalysis_/status/2056480113357729977
Here's the kicker: one deep-dive on Hewlett Packard Enterprise cost just $21.33 in tokens and replaced 20 hours of analyst time worth $2,000. That's a 93.8x return. Pulling financial data and emailing an Excel sheet? $1.87 for 3 hours' worth of work, 80.2x ROI. A quick earnings summary on Marvell? $2.82, 70.9x. Even building a Slack bot for tokenomics disclosures cost $58.02 in tokens vs. $1,000 in labor — still a 17.2x multiplier.
- Deep company research on HPE: $21.33 token cost, $2,000 labor (20 hours), 93.8x ROI
- Financial data pull & Excel email: $1.87 token cost, $150 labor (3 hours), 80.2x ROI
- Marvell earnings summary: $2.82 token cost, $200 labor (4 hours), 70.9x ROI
- Tokenomics disclosure Slack bot: $58.02 token cost, $1,000 labor (20 hours), 17.2x ROI
The internal benchmark is one thing, but SemiAnalysis also points to external data that screams how fast this is scaling. The firm tracks Claude Code's GitHub commit activity daily. In February 2026, the AI coding agent hit 134,000 commits per day — about 4% of all public GitHub commits. That's a 42,896x increase in 13 months, and their internal tracking shows the curve still climbing.
Once you've compressed 20 hours of grunt work into $21 in tokens, there's no going back. Still, banks and other large institutions are held back by compliance and IT restrictions. The real enterprise explosion hasn't even started yet.