Key Takeaways

  • Mike Cannon-Brookes argues that startups burning venture capital on frontier models are ignoring the core muscle of software: technological cost discipline.
  • Atlassian routes and blends models across a trade-off triangle of speed, quality, and cost rather than defaulting to the largest model available.
  • Atlassian built an internal AI gateway team instead of outsourcing to third-party model aggregators to retain control over enterprise unit economics.
  • Optimization timing matters: Cannon-Brookes ignores unit costs for a five-person test feature, but aggressively refines costs before that feature hits five million users.
  • Inference economics will drop fast as massive talent and capital pour into custom silicon, data centers, and model distillation.

The Trap of Free-Spending Startups

Silicon Valley is full of early-stage startups running every prompt through the most expensive frontier model on the market. Their investors pay the cloud bills, their gross margins look disastrous, and nobody cares because growth numbers are up. Cannon-Brookes sees this as a fatal blind spot.

“I love all these private companies that have raised a bunch of money and don't have to care about cost cuz they're not building the discipline to care about cost,” Cannon-Brookes told Harry Stebbings. In his view, when the music stops and investors demand cash-flow positive unit economics, those startups will lack the technical systems required to survive. He insists that “a core strength and attribute of a great technology company will be its ability to manage technological cost.”

Atlassian built its business on enterprise discipline. When you serve hundreds of thousands of organizations, routing unoptimized calls to an external frontier model will destroy your gross margins overnight. Surviving at scale requires treating compute expenses as a primary engineering metric.

The Triangle: Speed, Quality, and Cost

Instead of treating artificial intelligence as a single vendor relationship, Atlassian runs an internal AI gateway. The gateway dynamically routes and blends queries across multiple foundation models, proprietary models, and smaller distilled open-weight models.

Every query gets evaluated across a strict trade-off triangle: speed, quality, and cost. A customer asking for a quick summary of a two-sentence comment in Jira does not need an expensive reasoning model that takes six seconds to generate an answer. A customer running complex analysis across an enterprise Teamwork Graph requires high quality and can tolerate a few extra seconds of latency.

By managing this gateway in-house rather than outsourcing routing to a third-party startup, Atlassian keeps direct control over its margin structure and latency profiles. They can swap out underlying providers the minute a cheaper or faster model ships without touching customer-facing application code.

Optimize for Five Million, Not Five

Cost discipline does not mean premature optimization. Cannon-Brookes separates prototype experiments from enterprise scale with a simple rule of thumb: “If five people using a feature, I don't care how much it costs. If 5 million people use a feature, I'm going to care a lot about cost. And in the middle, I usually have a fair bit of time to keep optimizing.”

Right now, technical supply outpaces actual customer absorption. As Cannon-Brookes notes: “The production of technology is faster than the consumption of technology, right? It is ahead. We've reached a tipping point. We can build features faster than customers can consume them in a realistic way that changes their business.”

Meanwhile, the price per watt of inference is collapsing. A massive wave of engineering talent and capital has left traditional fields to attack inference efficiency through custom silicon, dedicated data centers, and smaller specialized weights. Founders who build flexible routing layers today will capture those deflationary tailwinds directly into their gross margin.

What to Do With This

Audit your application prompts by token volume and model tier tomorrow morning. Identify the top 20% of repetitive, low-complexity agent tasks currently hitting expensive frontier models, and reroute them to a smaller distilled model or an open-weight alternative behind an internal proxy.