Key Takeaways

  • Ephemeral virtual machines for tools like Replit and Lovable cost $4 to $5 per active user each month by spooling up only when triggered.
  • Persistent consumer agents like Dots, Grokbot, and Muse cost between $20 and $30 per active user each month because they run continuous background loops.
  • Charging flat $20 consumer subscription fees on 24/7 autonomous agents destroys unit economics unless products cap background compute or charge for metered usage.
  • Rory O'Driscoll argues that hardware supply cycles and Moore's law will compress inference costs within two to five years, while Jason Lemkin warns startups cannot survive the cash burn required to reach that window.

The Disagreement

Jason Lemkin started out believing that consumer AI agents followed the same cost profile as modern developer sandboxes. He was wrong, and he admitted it openly.

“I thought back when we started talking about this that the cost to run dots or grock or muse or instinct was like four to five bucks,” (play from 32:07) Lemkin said. “That's what it cost to run lovable or replet. Okay. And it's it's VMs. It's it's non-persistent VMs that spool up and spool down for your website.” (play from 32:15)

The math breaks when software moves from on-demand execution to continuous autonomy. Once an agent tracks data, watches feeds, and runs background logic around the clock, compute bills explode.

“This is like up to 20 bucks or 30 bucks a month to run these things,” (play from 32:32) Lemkin noted. “And the learning from Grokbot and Dots is we want them to do even more. My Muse is running all day all day long for 20 bucks a month.” (play from 32:37) For Lemkin, turning these agents into intermittent, batch-triggered tools defeats their entire purpose: “If you can't run an agent as often as you want 24/7 doing your work, it's just not that interesting. It's just an automated workflow.” (play from 33:10)

Rory O'Driscoll sees this friction as a standard hardware cycle that founders should look past. Infrastructure shortages inflate costs today, but historical trends always drive compute toward zero.

"Costs go down," O'Driscoll said. “Short-term compute prices go up because of supply and demand. Long-term compute prices go down because of Mo's law. In the end, it'll get cheaper. It'll be fine. The question is, are you two years away or five?” (play from 33:25)

Lemkin snapped back with the hard operational reality for early-stage companies: “The boys at Meta can carry the cost.” (play from 33:40)

Who's Right (and When They're Wrong)

O'Driscoll is right on the physics of computing, but Lemkin is right on the survival of startups.

If you run a balance sheet like Meta or Microsoft, you can absorb negative gross margins for five years while waiting for cheaper chips and surplus data center capacity. If you raised a $4 million seed round, you will run out of cash before hardware efficiency catches up. Selling a $20 monthly subscription while paying AWS or cloud providers $30 per user in raw compute means every new customer shortens your company's life.

O'Driscoll's logic holds only if you sell high-contract enterprise software with gross margins wide enough to swallow temporary compute spikes. Lemkin's warning applies directly to consumer apps and prosumer tools. Users love unmetered, always-on agents precisely because they push the financial pain onto the software vendor. When you offer an infinite loop for a fixed price, your power users bankrupt you.

What to Do With This

Audit your active user compute logs tomorrow morning and separate non-persistent sandbox calls from background agent tasks. If any user generates background inference that costs more than 40 percent of their monthly subscription price, push a hard usage cap or transition them to metered credits before the end of the quarter.