Three new cards from Tomasz Tunguz across two posts, covering model economics, AI product retention, and enterprise software disruption. A single thread runs through all three: switching costs, and who actually holds them.
Theme 1, The frontier model supremacy window has a measurable lifespan
Tunguz's July 13 post measured something most AI vendors would rather buyers not track. The average frontier model holds its performance crown for 41 days. After that, a competing model overtakes it on benchmarks, typically at lower cost. The average frontier model holds its performance crown for 41 days, resetting customer negotiating power on the same cycle captures the mechanism: model release cadence now runs faster than any enterprise procurement cycle, so the pricing premium attached to top-of-benchmark status expires before most customers finish their evaluation.
The implication is structural. A vendor selling on raw capability has eight to nine regime changes per year to defend its price. Customers who understand the cycle can time renegotiations to each reset. This is why the most durable AI moats have nothing to do with current benchmark position. Tunguz's earlier card (The AI deployment bottleneck has shifted from model capability to customer adoption, making embedded engineering teams the institutional moat) describes the countermeasure: embedded engineering teams that convert capability advantage into institutional relationships and learned workflows. The 41-day window is the threat; the FDE layer is the defense.
"The average frontier model maintains its crown for 41 days."
"Customers gain negotiating leverage every 41 days."
- Tomasz Tunguz, The average frontier model holds its performance crown for 41 days, resetting customer negotiating power on the same cycle. The average frontier AI model holds its performance lead for 41 days; customers gain negotiating power on the same cycle.
Theme 2, AI product retention is a colander problem, not a capability problem
The same July 13 post supplies the product-level counterpart. AI products retain customers at rates between mobile games and social networks: high single digits to about 40% by month five, with the stickiest cohorts near the top. AI products retain customers between mobile-game and social-network benchmarks, with workflow integration separating the top cohort from the rest names the dynamic: customers pour in, and many pour out. What keeps the top cohort is not better performance. It is workflow embedding deep enough that switching requires re-learning a daily operational habit, not just swapping a tool.
Social networks hold 80% at month five through bidirectional relationships and content gravity. SaaS retains 90% through sunk data and contract friction. AI products that sit in the high single digits are substitutable; they have not become part of a customer routine. The 40% ceiling is achievable, but only by becoming part of how the customer does the work, not just a tool the customer reaches for occasionally.
"AI is like a colander. Customers pour in, & many pour out."
- Tomasz Tunguz, AI products retain customers between mobile-game and social-network benchmarks, with workflow integration separating the top cohort from the rest. AI products retain between high single digits and 40% at month five; workflow integration, not capability, determines which cohort a product lands in.
Theme 3, Agents are displacing ERP systems one workflow at a time
The July 10 "Three Years In" post supplies the enterprise disruption angle. AI agents bypass ERP switching costs by automating operational workflows from within rather than displacing the system at once describes how AI agents are reaching ERP-class systems without a rip-and-replace migration. Traditional displacement required a full data migration, a project large enough to justify the incumbent's switching-cost moat. Agents skip the migration. They automate individual subprocesses, order approvals, inventory reconciliation, pipeline routing, while the ERP stays nominally in place. The operational value migrates to the agent layer gradually. By the time the organization recognizes the ERP has become a shell, the displacement is mostly complete.
This is the same switching-cost dynamic from the other side. Where The AI deployment bottleneck has shifted from model capability to customer adoption, making embedded engineering teams the institutional moat describes how AI companies build new institutional lock-in, the ERP bypass describes how agents dissolve existing lock-in without the friction that used to protect it.
- Tomasz Tunguz, AI agents bypass ERP switching costs by automating operational workflows from within rather than displacing the system at once. AI agents displace ERP systems by automating individual operational workflows from within, gradually transferring value out of systems whose switching-cost moat had previously made displacement prohibitive.