Pricing has always been where the rubber meets the road in SaaS, but the numbers coming out of 2025–2026 are not normal. Cursor is losing roughly $32 a month on every Pro subscriber — that’s not a leak, it’s the public math from their July apology and an Anthropic rate card. The $20 plan was a tripwire, and the product they market as "unlimited Sonnet" is really a coupon for Sonnet plus an internal router that defaults to cheap models. That single stat, buried under dev Twitter outrage, is the canary for an entire generation of AI repackagers pretending to be SaaS companies. The margin isn't determined by Cursor; it's determined by whatever rate card Anthropic publishes next quarter.

The same logic ripples across the entire seat-based software layer. The AI Compression model, detailed in 12 SaaS Categories That Won’t Survive the AI Compression, puts $240B of ARR at risk inside 24 months because any workflow describable in fewer than 200 tokens is now a prompt-replacement candidate. Meeting transcription ($310M ARR), machine translation ($1.2B), scheduling ($280M) — every one of them has a thin product moat that a horizontal agent (ChatGPT Voice, Claude 4.5 translate) can cross for free. And the article points directly at Cursor as the clearest live case study: "even at scale, the gross margins are getting absorbed by the underlying model providers." The through-line is brutal: the seat-based layer was a UX bet that held from 2010 to 2024, and it no longer holds when Anthropic’s inference cost has collapsed roughly 1,000x in 36 months — the fastest computing cost decline in history.

The response from builders isn't to wait for the bubble to pop. It's to go to zero. The $0 Vercel-killer stack documents a $312.40 monthly invoice replaced with a $0/mo stack of Cloudflare Workers, Supabase, R2, and Resend — all in six hours. Vercel’s September 2025 credit-based billing and mandatory Turbo builds turned a predictable ceiling into a leaky meter, and the author notes that this connects directly to the AI Compression watch list: "the marginal value of managed Next.js is collapsing as the open-source adapters catch up." The migration isn't just about saving money — it's a vote that the premium tiers of hosting platforms are the next SaaS category to compress.

The macro debate — bubble or miracle — is a distraction. The AI bubble debate is wrong on both sides runs the numbers: $297 billion in Q1 2026 venture funding, 81% into AI, with four deals (OpenAI, Anthropic, xAI, plus one stealth round) consuming 63% of every venture dollar. Bears scream "1999" because OpenAI’s 34x revenue multiple matches Cisco’s peak. Bulls shout "productivity miracle" because Anthropic doubled revenue to $30B ARR and guided $10.9B in Q2 alone. But the article’s real insight is that neither side is reading the right chart: this is a margin reallocation, not a bubble. The dollars aren’t disappearing — they’re moving from Salesforce’s P&L to Anthropic’s. The model layer is commoditizing faster than the application layer can capture value, and whoever owns distribution wins.

Where these articles go quiet is on the human friction that slows the transfer. They assume rational economic actors, but enterprises don’t rip out HubSpot Starter or Trello overnight because ChatGPT can handle contacts — they have renewal cycles, compliance reviews, and the inertia of a thousand spreadsheets. The 12 Categories list is sharp but treats every $310M market as equally vulnerable, when in reality the switching cost for meeting transcription is near zero (just don’t hit record in Otter), while swapping CRM workflows is a six-month project. And the Vercel-killer story works because the author already knew the four services — for a team without that muscle memory, the Saturday isn’t six hours, it’s six weeks. The pricing play that wins might not be the one with the lowest unit cost, but the one that gives the enterprise executive an excuse not to move.

If you only read one, make it The AI bubble debate is wrong on both sides because it synthesizes the macro data — the $297B funding quarter, Anthropic’s revenue curve, the inference cost collapse — that makes the micro examples (Cursor’s $32 loss, Vercel’s $312 invoice) inevitable. The essay connects the dots between the chart and the spreadsheet, and it’s the only one that asks which side cracks first. The answer isn’t OpenAI or Salesforce; it’s the builder who thought they were running a SaaS business when they were really running a reseller that paid for Anthropic’s next training run.