Anthropic, Atlassian, and a Scale VP All Landed on the Same Answer: Build the Skills Layer, Not a New Stack

$688 billion in hyperscaler AI capex against $110 billion of revenue. No capex breakeven until 2031 or 2032.

Those are Scale Venture Partners' Rory O'Driscoll's numbers from SaaStr AI 2026, and they're the backdrop for the most useful strategy conversation I've read this month (SaaStr). O'Driscoll also says foundation models need to capture 15-17% of knowledge-worker wages to pencil out, 25%+ for developers. That's the hurdle rate the entire buildout is underwriting.

What makes the writeup worth your time is that three people with completely different incentives converged. Anthropic's Eleanor Dorfman: "not a new stack," be intentional about the existing one. Atlassian's Sharif Mansour, same conclusion from the incumbent side. O'Driscoll from the investor side. All three land on the harness and skills layer as the moat, not the model.

The counter-evidence to rip-and-replace is Anthropic's own stack. They thread Claude through Clay, LeanData, Salesforce, Gong, Slack, and Intercom. An AI lab with unlimited access to frontier models kept its SaaS vendors and built the connective layer instead. Their reported result: 54% of new enterprise logos arriving self-serve in 2026, with AEs previously burning 70% of their time on internal process. Atlassian reports 5M+ users on AI features, with juniors 19-30% more likely to use AI and seniors twice as likely to refine the output.

That junior/senior split is the finding I'd tattoo on a wall. Juniors reach for AI more. Seniors edit what comes back more. Both behaviors are correct for where those people are, and any org measuring "AI adoption" as a single number is averaging two opposite things.

The counterexample exists and it's loud. Health insurer Curative cancelled a $600,000/year Salesforce contract after building an internal replacement with AI-assisted coding in about two months, targets cutting ~80% of total SaaS spend this year, and moved visualization off Google Looker onto Snowflake directly (Business Insider). Their agent negotiates provider contracts for ~$70 each versus $1,500-$2,000 with human teams, at 10x the volume. But CEO Fred Turner supplies his own caveat, and it's the honest one: Curative still uses Slack, and its Anthropic bill went from tens of thousands to millions per month. The software spend moved to a model vendor. It didn't disappear.

Connect that to story two and the picture resolves. If margin is moving from the seat to the token, and tokens are getting rationed, then the layer that decides which tokens get spent is where the value concentrates. That's the harness. That's your skills layer, your routing logic, your context budget. Not the model, which you don't control and which will be commoditized on price within two release cycles.

Build the harness. I've spent more time this year on the orchestration layer of my own systems than on prompts, and it's the only investment that survived three model swaps.