⚡️ The best engineers don't write the most code. They delete the most code. — Stay Sassy
TL;DR
AI spend is about to move from a flat perk to an individual budget fight — Stay SaaSy argues 2026 will force managers to decide who gets how much token budget, turning AI usage into something closer to department budgeting than buying everyone the same laptop.
Per-employee AI costs may get absurdly high, fast — Swyx cites an OpenAI team of three allegedly using 2.5 billion tokens a day, with one person alone at 1 billion, implying annual AI spend can look less like $50,000 and more like millions per engineer.
Build-vs-buy is getting more tempting, but mature software is still sneakily complex — Their practical heuristic: if the tool can basically be replaced by a spreadsheet, you may be able to build it internally; otherwise you're probably underestimating uptime, migrations, edge cases, and long-term ownership.
The first jobs to automate may be higher in the org chart than people expect — Stay SaaSy EM says many executive decisions are really just repeatable case statements, and companies should spend less time fantasizing about replacing junior staff and more time removing leadership bottlenecks.
AI works best today on automation, not the human core of management — They see real value in AI drafting, routing, and standard decision support, but trust-building, coaching, and handling messy people problems still look stubbornly human.
The scary reliability problem isn't AI writing code — it's tired humans pretending they reviewed it — Using Amazon's Jira-linked outage as a prompt, they argue companies are breaking a key software invariant: no single confused person, newbie, or fatigued engineer should be able to take down prod.
The Breakdown
How an anonymous management blog actually grew
Stay SaaSy started with the unglamorous answer: just write, distribute, and see what sticks. Early growth came from Hacker News, where five to ten posts a year would hit the front page, then Substack added a steadier readership, and X became the “turbo charge” layer with a much looser, more chaotic voice.
Serious blog, feral Twitter
They keep the same subject matter across platforms but tune the tone hard: Substack is “we’re going to school,” while X is where the shitposting lives. One example was the line about some jobs being a “cloud skill,” which came from hearing about a finance worker whose entire bottleneck was waiting for someone to send a one-line approval email.
The coming shock: token budgets per employee
The big management theme is that companies are moving from subsidized AI tools to consumption-based pricing, and that changes everything. EM compares it to a world where one employee’s laptop costs $500 a year and another’s costs $50,000 — except Swyx pushes the number way higher, citing an OpenAI example that could translate to roughly $2.5 million a year in model spend.
When building gets cheaper than distributing
They think cheaper AI-powered software creation will create weird new bottlenecks: teams that can build faster than they can sell, ship, or support. That means managers may have to tell high-performing engineers, “Yes, you could spend another $100,000 in tokens and build more, but the business can’t turn that into revenue fast enough.”
Build vs. buy still follows old truths
AI sharpens the question, but doesn’t erase the classic framework. PM’s test is blunt: describe the features clearly, and if your MVP is basically a spreadsheet, maybe build it; if not, you’re probably forgetting how much complexity hides in uptime, migrations, admin overhead, and all the ugly edge cases vendors already absorbed for you.
Stop only automating the leaves of the org chart
One of the spiciest takes is that companies are too focused on replacing junior work and not focused enough on automating executives. EM says a surprising amount of leadership work is standard, repetitive, and already close to a decision tree, so the better move may be using AI to reduce executive bottlenecks rather than fantasizing about replacing the bottom of the tree first.
AI can answer standard questions, but people are still the hard part
Their old post “You Know What To Do” comes up here: most management decisions aren’t mysteries, they’re just emotionally hard. AI might help by handling the standard cases without ego or hesitation, but human trust, morale, coaching, and the weirdness of actually managing people still look like the durable problems across every startup generation, including hot AI companies.
Amazon’s outage and the real code-review crisis
Prompted by Amazon going down for hours and blaming Jira-related software, they argue the industry’s real failure mode is cultural, not technical. Teams are acting like “armies of one,” shipping giant volumes of AI-assisted code without maintaining the old safety invariant — that no new hire, weak performer, or exhausted reviewer should ever be able to slip something into prod that can take the system down.