Most legacy SaaS estates cut cloud cost in the wrong order: architecture first, bill second.
I do it the other way round. Year-one bill under current builds, then year-two architecture.
How I sequence it 🧵
@prasanth_kuna@GergelyOrosz Agree on judging by what merged. Caps help finance forecast, but a hard cap on your best engineers is the wrong place to save money
@mjovanovictech For a legacy .NET estate the runtime is rarely the cost line. The rewrite bill is testing and parallel running, and the compute saved seldom pays that back
@saastr $12B in 2021 for $3.68B now is 31 cents on the dollar. At 2.9x revenue and flat growth, this is a margin bet: Bending Spoons gets paid on what it can cut, not what it can grow
OpenAI now takes up to 40% off tokens if you can wait 2 to 8 hours. Most of what sits on the job server in an older SaaS estate could wait that long. The batch window was always the cheapest SKU, nobody priced it until now
@kimmonismus@kdaigle@fayfers The router is the part I'd watch. For anyone selling AI inside a per-seat SaaS product, gross margin gets decided by what share of calls the router sends to the cheap model
@MatthewBerman Usually means a long contract didn't land. A B200 depreciates and the debt accrues whether it's rented or not, so idle hours get sold at whatever covers the financing
@bindureddy Slower is fine for most agent work. OpenAI just priced that trade at 40% off for 2 to 8 hour windows, so detail over speed is a discount anywhere the work runs like a batch job
@saastr At 24% growth you only need a 16% margin to clear 40, and still only 6% do. In my view that's inference landing in COGS before pricing catches up
@asmartbear@r00k If the agent picks the deploy tool, the docs are the demo. Anything that needs a human onboarding call loses before a buyer ever sees it
@DanielLockyer If that holds, Microsoft renting AWS for its own runners says capacity beats loyalty. CI minutes are a commodity, the cheapest hour wins whoever owns the logo
@GergelyOrosz Priced for the sunny day, same as a single-region SaaS estate with no DR. The rent saving looks great right up to the one event that wipes out the whole booking
@XkaiiStudio Option 3 is $73 a month per tenant before backups. Fine for the few enterprise tenants who pay for isolation, a margin leak on the long tail
@vercel Permissions for agents is the bit enterprise buyers will pay for. Next ask will be a bill per agent, nobody signs off a swarm without knowing which one burnt the tokens
@haider1 $7.50 to $20 is under 3x across the whole line. With no cheap tier to route to, model choice stops being a cost lever and the saving has to come from sending fewer tokens
@theo Good model is half the story. For a SaaS bill the bigger question is whether Google ships a cheap tier under it, since the boring calls are where gross margin gets decided
@saastr A 281x gap between median and power users is what breaks per-seat AI pricing. One flat price per head and the top tenants eat the margin the median ones make
@GergelyOrosz And each shift moves the bill too. Review time was salary, CI minutes and observability are metered, so faster code ends up as a bigger cloud invoice