Commodification of Intelligence: Circular AI Deals Explained
Every few months, critics point to circular AI deals—OpenAI raising money from Microsoft and spending it on Microsoft servers, or Nvidia backstopping CoreWeave debt while CoreWeave buys Nvidia GPUs—and scream “bubble.” Wojciech Gryc argues this misses the point. These deals signal not a dot-com repeat but the commodification of intelligence: AI compute moving from product to fungible commodity like electricity, copper, or natural gas.
Commodity Deals 101
In traditional commodities, large infrastructure (mines, refineries, ports) requires massive upfront capital. Circular deals solve this: a commodity trading firm guarantees to buy all output at a set price, making lenders comfortable. The trader may also take equity. This pattern dates to 1960s Japanese commodity traders, continued in 1980s Jamaica, and appears today in US critical minerals: MP Materials secured a 10-year contract with the Department of War to buy neodymium-praseodymium magnets at ≥$110/kg, then signed off-take agreements with General Motors.
AI Is Fungible and Expensive
Frontier generative AI—both training and inference—depends heavily on Nvidia GPUs. These are effectively fungible compute units with massive capital costs. Just as oil buyers forward-purchase to de-risk mines, AI players use circular financial structures to ensure GPU supply. The article notes this is not inherently bad: it enables capacity that otherwise couldn't be financed.
However, risks emerge when circular deals lack a real fungible market. If the counterparty cannot actually resell the GPUs or compute in a deep liquid market, the deal becomes fragile. The author warns that some AI circular deals may introduce “awful surprises” for companies or the entire sector—especially when the underlying asset is not as globally tradeable as oil or copper.
For developers using AI agents, this means the infrastructure powering your tools is increasingly funded like a utility. The cost and availability of GPU compute may stabilize or become more predictable—but also may be subject to the same boom-bust cycles as commodity markets if liquidity dries up.
📖 Read the full source: HN AI Agents
👀 See Also

Claude Code 2.1.76 adds MCP elicitation, worktree improvements, and fixes for context limits
Claude Code version 2.1.76 introduces MCP elicitation support for structured input during tasks, adds worktree.sparsePaths for large monorepos, and fixes 'Context limit reached' errors on 1M-context sessions. Version 2.1.75 made 1M context windows default for Opus 4.6 on Max, Team, and Enterprise plans.

Agent Harness Outside the Sandbox: Durable Execution & Cold Starts
Running the agent loop outside the sandbox isolates credentials, enables sandbox suspension, and simplifies multi-user sharing, but requires solving durable execution and cold start latency.

Buddy turns down $300k+ role replacing 70% of staff with Claude agents — Reddit debates the moral and technical reality
A Reddit post describes a friend who refused a role as 'AI Transition Lead' to map workflows, build Claude/GPT agent pipelines, and fire 70% of staff. The poster argues the $300k+ bag is worth it to waste time and watch C-suite delusion crash.

Anthropic Doubles Claude Code Rate Limits, Signs Compute Deal with SpaceX
Claude Code five-hour rate limits doubled for Pro/Max/Team/Enterprise plans, peak-hour reductions removed, and API rate limits raised for Opus models. SpaceX Colossus 1 adds 300+ MW capacity (220k NVIDIA GPUs) within a month.