The moat you shipped in Q1 has a two-quarter half-life

The week the SaaS vintage got its death certificate

T. Rowe Price marked Airtable down more than 60% from its $11B peak. DataRobot went to near-zero from $6B. Gusto is off about 30%. Databricks tripled its way toward $175B in the same window. HubSpot is down 75% since early 2025 and spent four days last week reversing an opt-out AI data pooling policy after customers revolted — the third time in three years a mid-market SaaS vendor has tried the same pooled-data play and been forced to unwind it (Zoom 2023, Slack 2024, HubSpot 2026).

The part that should end the debate about whether this is a rate story or a structural one: a non-technical Atlanta real estate manager rebuilt his Salesforce workflow on Replit and Claude Code for $100K/year in savings. Five enterprises confirmed doing the same in the past six months. These customers are not swapping vendors. They are removing the vendor from the transaction.

Battery Ventures' Brandon Gleklen said it plainly: product-market fit is now ephemeral. Bookings spike on launch, then foundation models catch up within a quarter or two. If you're planning a roadmap on the assumption that a feature moat buys you a year of growth, you are planning against evidence the market has already priced in.

Yes, but — the counter-reading is that Databricks and infrastructure comps are compounding, so this is a rotation inside software, not an extinction. That's true, and it's also the point. The line between AI-eaten and AI-compounded runs through defensibility against foundation models themselves. DataRobot sold ML tooling to non-experts. Foundation models now do that job for free. Proximity to AI is not a moat. Selling AI is not a moat.

The agent economics nobody modeled correctly

Uber deployed Claude Code across engineering and burned its full 2026 AI budget in four months. Gartner is forecasting 40% of AI agent projects cancelled by 2027 on cost overruns alone. This is not a startup scaling problem — Uber has one of the more sophisticated infrastructure orgs on the planet, and they still got the math wrong.

The math is not subtle once you look at it. Per-token prices dropped 80% between 2023 and 2026. Agentic workflows consume 5–30x more tokens per task than a chatbot query. Stanford measured that 62% of what gets sent to agents on every call is redundant repeated context — system prompts, tool definitions, documents the model already saw two turns ago. Even at the new cheaper rates, the effective cost multiplier is roughly 7–8x. And the best agent in AutomationBench-AA (Claude Fable 5) still only completes 48.6% of 657 real SaaS automation tasks. Every model tested breaks business rules. You are paying 8x per task for a coin flip.

The fix exists. LMCache runs as a disaggregated process with multi-tier GPU/CPU/SSD storage and reports 14x TTFT improvement on H200 with Qwen3-235B at 50 concurrent users. Tencent's fused FP8 MoE kernels landed in vLLM main this week — 2.95x throughput, 24% TTFT reduction, no code changes beyond a version bump. A small model as a relevance filter in front of your RAG pipeline cuts 68% of chunks at 96% recall. Anthropic's prompt caching gives you 90% off on the cached prefix if you restructure prompts so the stable parts come first.

The numbers are single-config lab wins and won't survive contact with your traffic untouched. But the direction is unambiguous: this quarter's lever isn't a smarter model, it's cheaper orchestration around the model you have.

And the perimeter is on fire this week

While you rework agent economics, six critical CVEs are under active exploitation with windows measured in hours. Adobe ColdFusion (CVE-2026-48282) was exploited within hours of disclosure. Oracle EBS (CVE-2026-46817) was exploited before any public PoC — Defused counted six honeypot hits in a two-hour window. SimpleHelp (CVE-2026-48558) is delivering Djinn Stealer, and Djinn specifically targets AI integration credentials alongside cloud keys and SSH sessions. BeyondTrust's PAM tool — the thing that exists to control privileged access — shipped an emergency auth-bypass patch. FortiBleed credentials are trading at $60K on the dark web. KVM's Januscape (CVE-2026-53359) breaks 16 years of assumed guest-to-host isolation on every Intel/AMD x86 host running nested virtualization.

The pattern that matters for anyone shipping AI: LLM API keys are now a named, targeted asset class. If your .env files sit in Git history or an RMM tool touched them, assume compromise. Rotate to a secrets manager with short-lived tokens. Do it this week.

What to do before Friday

Three things, verb-first, no hedging.

Run a moat durability audit on your top five features. For each one, answer honestly: could a non-technical user replicate 80% of the value in a weekend with Claude Code and Replit? If yes, that feature is not a roadmap item — it's a churn risk. Reallocate the engineering headcount toward compounding moats: proprietary data that gets better with usage, network effects between users, compliance certifications, integration ecosystems expensive to replicate.

Instrument prompt caching hit rate and redundant-token ratio per request in production before you scale another agent feature. You cannot manage 62% waste you haven't measured. If you find it, LMCache and prompt restructuring are two-week wins. If you don't, you have a different problem worth knowing about.

Rotate every LLM and cloud API key that has ever touched a .env, a notebook, or an RMM tool. Move them behind a secrets manager with short-lived tokens today. Djinn Stealer is not a hypothesis this week — it's a named payload being delivered through infrastructure many of you still run.

The application layer is being repriced. The compute layer is loosening. Power is the new scarcity. Data is the new compute. Your team's AI identity predicts retention better than comp does. Pick the two of those you can act on before your next sprint planning and start there.