AI just fractured into four industries with four different margins

Three numbers from this week, in the order they should change your thinking.

Google dropped Gemini Flash Live to $0.005 per minute of input. A 24/7 voice agent now costs $25/day — $9,460/year. Below minimum wage in every US state.

OpenAI bought Astral. Not a model lab, not a chip company. The team behind uv and Ruff — the Python package manager and linter sitting in your CI pipeline. The acquisition thesis, barely hidden: coding agents fail at dependency resolution and environment bootstrapping, not reasoning. Buying smarter models was the obvious move. They bought the plumbing instead.

Vercel said 30% of apps deployed on its platform are now agent-generated. At ~$340M ARR, that's not a demo metric. That's production traffic.

Line those up and the picture is hard to miss. Inference is collapsing into a utility. Software is getting cheap to produce. And the company with the most leverage in the market just decided the fight isn't about the model anymore — it's about who owns the surface where work happens.

The four-layer map

If you're still pricing AI companies on SaaS multiples, you're going to misprice most of them. The industry has split into four economic layers, and each one is governed by the economics of the industry it now resembles, not the industry it came from.

Inference is a utility. 20–30% gross margins on a good day. Google can sustain $0.005/min because it's vertically integrated from custom silicon through ad-subsidized cloud. No pure-play model lab can match that cost structure, and OpenAI implicitly conceded by spending acquisition dollars on developer tooling instead of price-matching.

Hardware and infrastructure is project finance. NVIDIA put $2B into Nebius targeting 5 GW by 2030. Hyperscalers are building dispatchable grid assets that can curtail 25% of load in under a minute — engineering around a permitting problem they can't solve politically. The US grid sits at 1.37 TW; China's at 3.89 TW and added 500 GW in a single year. Power, not compute, is the binding constraint.

Workflow SaaS is software, but with compressed margins because the agent-generated 30% on Vercel is going to keep growing. If your moat is the difficulty of building software, the moat is shrinking monthly.

Compliance and orchestration is a tollbooth. 70–85% margins, recurring, and no hyperscaler dominates yet. This is where defensible margin actually lives, and it's the least crowded layer in the stack.

The trap most companies are walking into: they think they're a workflow SaaS company, but 40% of their COGS is inference at Google's below-cost pricing. That's not SaaS economics. That's a blended utility/SaaS business with margin leakage they aren't modeling.

The leverage underneath

Roughly $120B in Western AI financing is cross-collateralized against energy contracts, not revenue. OpenAI's $122B round, Meta's $21B to CoreWeave, NVIDIA backstopping demand into Nebius. Each player's bet validates the next player's collateral.

The scenario nobody's pricing: enterprise AI ROI takes 24 months instead of 12. Plausible — Gallup's productivity numbers are coming in marginal, lawyers report AI-generated emails are increasing downstream review time, and the agent execution problems OpenAI just paid to fix are real. If ROI slips, debt servicing cracks first, and today's artificially cheap API prices correct upward 3–5x with little warning.

If your unit economics work at $0.005/min but break at $0.015/min, you don't have a business. You have a position in someone else's leveraged trade.

What the OpenAI–Anthropic spat actually tells you

OpenAI's CRO leaked a memo accusing Anthropic of inflating ARR by $8B through gross revenue accounting. Both methods are GAAP-compliant. Normalized to net, OpenAI leads by ~$3B regardless of which way you measure.

The interesting part isn't the accounting. It's that both companies are knife-fighting for enterprise logos ahead of 2026 IPOs while OpenAI is simultaneously breaking Azure exclusivity for AWS, losing three Stargate executives to Meta, and acquiring a Python toolchain. That's not a company confident about model-layer dominance. That's a company hedging into developer surface area, distribution channels, and vertical absorption (Hiro, personal finance) because the model layer alone won't carry the valuation.

For anyone procuring AI right now, this is a 1–2 quarter negotiation window. After IPO lockups stabilize, leverage shifts back to the vendors. Use it.

What to do this week

One specific thing: take your three highest-volume AI features and rebuild their unit economics in two columns — one at today's API prices, one at 3x. If both columns work, ship faster; the inference price floor is your tailwind. If only the cheap column works, you're not running a product, you're running a subsidy arbitrage, and you need to either move the workload to the workflow or compliance layer where margin is durable, or get to self-hosting capability before the financing structure decides for you.

The teams that figure out which of the four layers they're actually competing in — and reprice accordingly — own the next eighteen months. The teams still benchmarking models against each other are fighting over the lowest-margin segment of a fracturing industry.