Two things happened in the last two weeks that should have every enterprise CTO redrawing their AI architecture diagrams.
First, Anthropic crossed $30 billion in annualized revenue, up from $1 billion in January 2025. That's 30x in fifteen months. They quietly passed OpenAI in enterprise revenue specifically — not total revenue, but the segment that matters for anyone running a real business.
Second, Google announced it's putting up to $40 billion into Anthropic, including 5 gigawatts of dedicated compute. That's not a partnership. That's infrastructure-grade capital. For context, Google's entire net income last quarter was about $35 billion. They're spending more than a quarter on a single AI bet that isn't even their own model.
And in the same window, Microsoft and OpenAI restructured their partnership — meaning OpenAI can now ship through AWS and Google Cloud, not just Azure. The exclusive marriage that defined the last three years of enterprise AI is over.
The "Safe" Choice Just Became the Risky One
Walk into any Fortune 1000 IT shop in 2024 and ask why they picked OpenAI through Azure. The answer was almost always the same: "It's the safe choice. Microsoft is the enterprise standard. We already have the contract. Compliance approves it."
That logic was correct in 2024. It's broken in 2026.
The architecture decisions made on that "safe" assumption now have a problem. If you built your AI stack assuming you'd always live inside Azure OpenAI Service, you're sitting on:
- A vendor lock-in nobody promised you
- A model that's no longer the enterprise revenue leader
- A set of contracts that didn't anticipate multi-cloud, multi-model
- An engineering team that only knows one provider's tools, prompts, and quirks
That last one is the staffing story. And it's the part nobody's talking about yet.
What Multi-Model Actually Costs
When the AI duopoly was "OpenAI on Azure" for enterprise and "everything else" for hobbyists, you could staff for it cheaply. One AI engineer who knew the OpenAI API. Maybe a prompt engineer. Done.
The new reality is different. Real enterprises are now running three or four models — Claude for long-context document work, GPT for general agents, Gemini for the Google Workspace tie-in, open models for anything that needs to live inside the firewall. Each has its own SDK, its own prompt patterns, its own failure modes, its own pricing model, its own evals.
The job description "AI engineer" doesn't mean what it meant a year ago. It now means: someone who can architect a routing layer across multiple model families, build evals that compare them apples-to-apples, and rebuild that layer every six months as the leaderboard shuffles.
Those engineers exist. There aren't many of them. And the wage premium is no longer 20% — it's closer to 60-90% over a single-model engineer.
What This Means If You're Hiring
Three things, in priority order:
Stop hiring "OpenAI engineers." That title made sense in 2024. In 2026 it's a yellow flag — it usually means someone who learned one API and never branched out. Hire for model-agnostic architecture skills.
Audit your existing AI team for vendor concentration risk. If your entire AI org has only ever shipped against one provider, you have the same exposure as a company whose entire payments stack runs through one bank. Diversify the skill base before the architecture forces you to.
Get serious about multi-model evals. This is the new core competency. Companies that can A/B test Claude vs. GPT vs. Gemini on their actual workloads will pick the right model for each job. Companies that can't will keep paying premium prices for whichever provider their procurement team signed first.
The Bigger Pattern
Every five or six years in enterprise tech, the "safe choice" stops being safe. It happened with mainframes, with Oracle databases, with on-prem data centers, with single-cloud strategies. The pattern is always the same: a dominant vendor becomes the default because of inertia, the inertia becomes a contract, the contract becomes an architecture, the architecture becomes a hiring profile, and then the market moves and the whole stack is suddenly a liability.
We're at the architecture-becomes-hiring-profile stage of the AI cycle right now. The next eighteen months will sort companies into two groups: the ones who recognize that and start hiring for flexibility, and the ones who keep hiring for the old default and wonder why their AI projects feel slower and more expensive every quarter.
The duopoly is over. The single-vendor bet was never as safe as it looked. And the talent market is already pricing that in — whether your hiring team has noticed yet or not.