Happy Friday. The SaaS selloff earlier this year was fast, brutal, and, according to one investor who lived through it, deeply misleading about what software is actually worth.
Also today: an AI model built to handle production alerts before an engineer ever sees them, and how Uber burned its full AI coding budget in four months and then had to engineer its way out.
BUSINESS & STRATEGY
What the SaaS Selloff Got Wrong and Who Paid for It

Median enterprise value-to-revenue multiples for public SaaS collapsed from 18 to 21 times down to 5 to 7 times in the months following Anthropic's January release of Claude Cowork. Private M&A deals barely moved: disclosed transactions closed at a median of 4.0 times revenue in Q2 2026, above the public market median of 3.21 times, an inversion of the historical relationship between the two.
The split that followed the April 10 bottom is the number that matters most for founders deciding whether to stay private. AI-integrated names like CrowdStrike and Datadog recovered to near all-time highs. Twilio, Zoom, and DocuSign remain down 60% to 80% from peak. Public markets repriced the whole sector first and sorted the survivors later. Private buyers never stopped distinguishing between them. That gap is an argument for staying private longer, and strategic acquirers with dry powder are apparently ready to close it.
ARTIFICIAL INTELLIGENCE
Resolve AI Puts AI Ahead of Engineers in the On-Call Chain

Resolve AI launched SRE-1 on November 18, a 27 billion-parameter model trained specifically on production incident data, alongside a service called Primary On-Call that routes alerts to AI agents before human engineers. In a dark launch, Salesforce reported the system closed more than 70% of CRM Analytics paging alerts without human intervention. The company says that figure reaches 80% across its broader customer base.
The technical moat here is the training data, not the model size. As Salesforce Ventures noted, there is almost no public data on production incidents, which is exactly why general-purpose models struggle with root-cause analysis and why a domain-specific model trained on live cases has an edge. The ensemble approach, five SRE-1 runs matched against each other, matching frontier model accuracy at 37% lower cost, is a preview of how specialized models will compete with larger ones on price-per-outcome rather than raw capability.
DEVELOPER TOOLS
Uber's $1,200 Two-Hour Session Became a Company-Wide Reckoning

Uber deployed Claude Code to roughly 5,000 engineers in December 2025. By April, adoption had reached 84% and the full annual AI coding budget was gone. Per-engineer monthly costs ran between $500 and $2,000, five to twenty times standard developer seat budgets. The company built a wrapper called AIFX to enforce four defaults across all coding tools: context capped at 400,000 tokens, reasoning effort set to medium, MCP tool schemas loaded dynamically instead of at session start, and prompt cache TTL extended to one hour.
The lesson is not that agentic coding tools are too expensive. It is that the defaults shipped by AI tool vendors are calibrated for capability demonstrations, not for 5,000-engineer deployments running all day. Uber's internal leaderboards ranked teams by usage, which drove adoption and spend together. Any company deploying these tools at scale without a cost governance layer is running the same experiment Uber ran, just without the engineering team that built the fix.
Created by the robots at The SaaS Sentinel