Key Takeaways

  • Rental rates on older chips continue climbing: the Silicon Data H100 rental tracking index recently hit some of its highest levels yet rather than dropping.
  • Scarcity reaches the highest levels of the industry; according to Daniel Saeedi, OpenAI president Greg Brockman asked Elon Musk directly for compute while waiting to receive medals at the White House.
  • The United States currently has between 5 million and 10 million leading-edge chips deployed against an estimated domestic base of 100 million knowledge workers.
  • Workflows where autonomous agents run around the clock require continuous token generation rather than intermittent search queries, multiplying the hardware baseline per seat.
  • Builders can forecast macro infrastructure deficits and evaluate capacity risk using Saeedi's One GPU Per Knowledge Worker Heuristic.

The Saeedi One GPU Per Knowledge Worker Heuristic

Market commentators regularly debate whether hyperscale datacenter spending is running too far ahead of demand. The pricing reality tells a different story. As the Odd Lots hosts pointed out, “I'm looking at the Silicon Data H100 rental tracking index. This is an old chip, it's an H100, it's still going up, it's at the highest in a while.” (play from 42:42)

Hardware shortages remain acute even for the most well-funded frontier operations. Saeedi recalled an encounter among industry leaders: “The most telling thing about how short compute we are is allegedly when Elon and Jensen were waiting to receive their medals from the White House, apparently Greg Brockman asked Elon for compute there.” (play from 45:49)

To determine whether the datacenter buildout is an overbuilt bubble or a structural deficit, Saeedi models token consumption from the worker level up:

  • Component 1: Individual Consumption Baseline

A fully AI-native employee consumes approximately one leading-edge GPU's worth of tokens per year to power continuous autonomous background agents.

  • Component 2: Total Addressable Demand

Multiply the national knowledge-worker labor force (approx. 100 million in the US) by one GPU to establish aggregate structural computational demand.

  • Component 3: Capacity Deficit Assessment

Compare aggregate target demand against current deployed leading-edge chips (e.g., 5 to 10 million Blackwell GPUs deployed in the US) to assess long-term infrastructure overbuild versus underbuild.

As Saeedi summarizes: “It won't be a surprise that we have kind of one B300 worth of computational capacity for every knowledge-working employee in the economy. So in my mind, it's like if everyone's going to be using AI in the way that I do it, we're still massively short compute.” (play from 45:07)

When This Works (and When It Doesn't)

This rule works well when forecasting macro datacenter capacity, evaluating CapEx cycles, and identifying whether hardware investments are running into cyclical oversupply or facing severe structural deficits. It helps founders avoid the trap of assuming token prices will collapse to zero in the near term.

The heuristic fails if your workforce relies strictly on chat interfaces. If a human engineer merely queries a model four times an afternoon, that worker consumes a tiny fraction of a GPU annually. The math only holds when workers run continuous autonomous agents in the background that constantly ingest code, audit data, and draft communications. If model efficiency or quantization outpaces agent complexity, individual token appetites could also drop below a dedicated chip equivalent.

What to Do With This

If you are planning your infrastructure budget for the next twelve months, audit your team's actual token usage by employee role.

First, divide your current annual token spend by the cost of running a dedicated instance like an H100 or B300 for a year. Second, count how many employees actively maintain persistent background agents versus standard chat prompts. If your engineers run automated test suites and continuous code review loops, allocate for roughly one dedicated chip equivalent per engineer in your vendor contracts. If your team only uses AI for quick copy edits, do not overbuy capacity; stick to shared serverless endpoints until you deploy autonomous workflows that run while your staff sleeps.