AI just fractured into four industries. Price your work accordingly.

Three numbers from this week, sitting on the same desk.

Google's Gemini Flash Live: $0.005 per input minute. A 24/7 voice agent for $25 a day, $9,460 a year — below minimum wage in every US state. Vercel: 30% of apps deployed on the platform, at roughly $340M ARR, are now generated by AI agents. ShinyHunters: breached the analytics vendor Anodot, harvested stored OAuth tokens, and pivoted into 12+ customer cloud environments including Rockstar Games, ransoming each one on the way out.

Those three numbers describe the same story from three angles. AI stopped being a software business this week. It fractured into four industries with four sets of economics, and most planning documents haven't caught up.

The four-layer split

Inference is a utility now. Google is pricing it like electricity, cross-subsidized by ads and custom silicon. No pure-play lab can match that structure, and OpenAI's response — acquiring Astral, the team behind uv and Ruff — is the tell. They didn't buy a better model. They bought the Python package manager, because their coding agents fail at dependency resolution and environment setup, not reasoning. NVIDIA agreed from the hardware side by shipping Vera, a CPU purpose-built for 22,500 concurrent agent environments per rack. Both the largest AI lab and the largest AI hardware company just invested in execution infrastructure, not model capability. That's a signal.

Hardware infrastructure is project finance. The Western AI buildout is running on $120B+ in leveraged financing, and the collateral is energy contracts, not revenue. NVIDIA put $2B into Nebius to guarantee demand for its own chips. Hyperscalers are debt-financing power grids. The US grid sits at 1.37 TW against China's 3.89 TW — a gap private capital cannot close on timelines that matter.

Workflow SaaS is still SaaS, but with commoditization pressure from below. Compliance and orchestration is the tollbooth layer — 70-85% margins, no dominant player, regulatory tailwind. That's where defensible margin now lives.

Yes, but — the counter-reading is that Google's pricing is predatory, not equilibrium. If enterprise ROI slips from 12 months to 24, the leveraged financing cracks and API prices correct 3-5x upward. That's real. But it argues for building your unit economics to survive both today's prices and 3x today's prices — not for ignoring the commodity signal. The floor moved either way.

The revenue war is a negotiation window

OpenAI's CRO leaked a memo accusing Anthropic of inflating ARR by $8B through gross revenue accounting that includes cloud partner pass-through. Normalized to net: OpenAI at ~$25B, Anthropic at ~$22B. Both accounting methods are GAAP-compliant. The headline dispute is theater. The signal underneath is that both companies are pre-IPO and desperate for enterprise logos, and OpenAI is breaking Azure exclusivity for AWS with demand its CRO calls "staggering."

That gives you one to two quarters of leverage. After the S-1s drop and lockups stabilize, it's gone.

Meanwhile, three senior OpenAI executives behind Stargate left for Meta. Meta committed another $21B to CoreWeave and is projected to pass Google in net ad revenue in 2026 — $243B to Google's $240B, driven by Meta's 22% growth against Google's 11% and the absence of Google's ~20% TAC drag. That's the first genuine three-platform ad market since mobile, and it's the same week Meta is building photorealistic Zuckerberg clones as a distinct product category. Two moats going up at once.

The security bill is coming due

The ShinyHunters breach is the pattern to internalize. Anodot did nothing exotic — it's a monitoring SaaS holding stored OAuth tokens with standing access to customer clouds. That's every observability, analytics, and CI/CD vendor in your stack. The tokens are the attack surface, and vendor-originated token usage looks legitimate to your monitoring. Twelve organizations found out this week.

Same week, OpenAI's internal tooling downloaded a compromised Axios update. Different vector, same architectural assumption: vendors you authorized will maintain the integrity of that authorization. They will not, reliably, forever.

Add Microsoft Copilot Cowork now routing M365 data to both OpenAI and Anthropic backends, OpenAI's Astral acquisition putting a package manager owned by an AI lab inside your CI/CD, and six mature local LLM families (four Chinese-origin) that don't traverse your CASB. Your third-party trust map tripled in complexity this quarter. Your DPAs, SBOMs, and vendor questionnaires assume a simpler world.

What to do this week

One inventory, one negotiation, one stress test.

Inventory every SaaS vendor holding delegated OAuth tokens or API keys to your cloud environments. Prioritize analytics, observability, and monitoring. Document access scope and last rotation date. Rotate anything older than 90 days and enforce IP allowlisting on the rest. This is what ShinyHunters exploited, and the fix is boring hygiene, not new tooling.

Renegotiate your primary AI vendor contract before the IPO window closes. If you're on OpenAI, benchmark against Anthropic and bring the numbers. If you're on Anthropic, do the reverse. Both providers are maximally motivated for another one to two quarters. After that, leverage shifts back.

Rebuild your AI unit economics against Google's floor — $0.005/min voice, $0.25/M tokens text — and stress-test at 3x those costs. If the model only works at today's subsidized prices, you don't have a business, you have a bet on somebody else's balance sheet.