The AI industry split into two economies this week

Three data points landed in the same 48 hours and they tell the whole story.

Nvidia posted $68.1B in a single quarter, 73% growth, 55.6% net margins, $96.6B in annual free cash flow — second only to Apple. Salesforce's Agentforce hit $800M ARR growing 60% quarter-over-quarter, and organic growth still decelerated to 8%. And an NBER survey of 6,000 executives found 80% of firms report zero AI productivity impact from any of it.

The AI economy just bifurcated in public. Infrastructure captures monopoly economics. The application layer is running a cannibalization treadmill. If you ship software for a living, both of those facts change your next planning cycle.

The cannibalization trap is real and it's already in the numbers

Salesforce's own CFO said it out loud on the earnings call: Agentforce growth is being offset by weakness in marketing, commerce, and Tableau. At 1.7% of projected FY2027 revenue, the fastest-growing product in the portfolio can't outrun the erosion in the products it's replacing. Snowflake's CFO added the punchline — AI product margins are lower than legacy products, and they just laid off 200 people. Workday's returning CEO called third-party AI agent providers "parasites." HubSpot's CEO said the plan is to "monitor, meter, and monetize" agent access to customer data.

That's four incumbents in the same week saying the quiet part. AI features are growing revenue and destroying margin at the same time, and the defensive move is to wall off the data layer before agents route around you entirely.

Meanwhile, in a single week: Perplexity shipped Computer as a general-purpose digital worker, Anthropic shipped Cowork with scheduled recurring tasks and a plugin architecture, Cursor put its agents in dedicated cloud VMs producing merge-ready PRs, Notion launched Custom Agents with 24/7 org access. Five agent platforms went to production in seven days. Jason Lemkin testified on the Salesforce call that he went from 15 humans to 2.5 humans plus 20 agents while closing $3M in deals. The TAM shifts from software budget to labor budget, and the pricing metric has to shift with it.

Yes, but — you could read the 80% zero-impact number as an early-innings story and expect the cannibalization to convert to net-new value as workflows mature. Claude Code hitting $1B run-rate in six months is the exhibit for that reading. It might be right for coding. It's not right for the enterprise SaaS book, where the seat-based revenue is being subtracted in real time and the AWU-style replacement metric doesn't exist yet. Benioff told investors they're "still trying to figure out what these numbers mean." A $300B company inventing pricing on the fly is not a signal of a mature transition.

The infrastructure side is a different sport

Amazon is negotiating up to $50B into OpenAI at a $730B pre-money — $15B firm, $35B contingent on IPO or AGI. Nvidia is considering $30B alongside. Google just closed a multibillion-dollar TPU deal with Meta, the first real crack in the Nvidia monopoly at hyperscaler scale, while Meta simultaneously scrapped its most advanced custom training chip. Alphabet, Amazon, and Meta collectively hold roughly $600B in untapped borrowing capacity without touching their credit ratings.

The AGI clause in Amazon's deal deserves its own paragraph. Microsoft's Azure exclusivity with OpenAI ends when AGI is declared. Amazon's additional $35B unlocks precisely when that exclusivity dissolves. This is a call option on the shape of the cloud market, structured as an investment. And "AGI" is described as loosely defined — one party wants to declare it, the other wants to delay. If you're signing anything with an AGI-linked trigger this quarter, demand benchmark-based definitions or don't sign.

OpenPipe's ART framework is the counter-current worth watching. A 14B Qwen model trained with GRPO for under $80 on a single GPU beat o3 on an email search benchmark at 96% accuracy — $0.85 per 1,000 runs versus $55.19. That's a 64x cost reduction on a specific agentic task. OpenAI's own VP of Science says the internal architecture is orchestrator plus cheap specialized models, not single frontier calls, and considers the ensemble mistake the most common costly one teams make on their APIs. Three MIT-licensed Chinese frontier models — GLM-5, Qwen 3.5, DeepSeek V4 — dropped in the same week. GLM-5 is #1 on open leaderboards.

The frontier API tax is optional for well-defined tasks. That's the leverage point.

And your coding tools are actively under attack

Socket documented a self-propagating npm worm — SANDWORM_MODE — that specifically targets Claude Code, Cursor, VS Code Continue, and Windsurf by injecting malicious MCP servers. It steals from Bitwarden, 1Password, and LastPass, propagates via stolen npm and GitHub tokens, and includes a polymorphic engine using local Ollama to rewrite itself. Claude Code itself has two confirmed RCE CVEs (CVE-2025-59536, CVE-2026-21852) triggered by opening a cloned repository — no code execution required. Meta's Manus agent had a CVSS 9.8 zero-click prompt injection allowing Gmail exfiltration and passwordless sudo shells. Microsoft Semantic Kernel's Python SDK has a CVSS 9.9 RCE in the InMemoryVectorStore filter — the exact component every RAG prototype ships with. Fix is 1.39.4.

The threat model moved from "don't run untrusted code" to "don't open untrusted projects." Your MCP server allowlist is now a production security control.

What to do this week

One concrete move, three parts. Sit down with your CFO and pick the single feature in your product that consumes the most frontier API spend. Instrument its per-task cost as a first-class metric on a dashboard your finance team sees. Then build the replacement path: prototype an ART-style GRPO run on a 7B or 14B open model against that exact task, and stand up an MCP server allowlist across every AI coding tool your engineers touch before you ship the replacement. If the infrastructure economy is where the margin lives, your job is to stop renting it by the token on the workflows you can own — and to make sure the tools you're renting today aren't the reason your credentials end up on a Russian broker's shelf next week.