Skip to content
John D. Clay

On X

Articles

Essays John D. Clay publishes on X about Super Intelligence, the power behind it, and what it really costs.

  • October 16, 2026

    Your SI Bill Isn't Set by the Price Sheet. It's Set by the Harness.

    Companies haggle over price per token. Meanwhile the coding tool their engineers use picks the model, launches the helper agents and sets the defaults. A September paper calls that layer the harness, and the fix is simple: put someone in charge of the defaults, because in most companies nobody is.

    Read on X

  • October 14, 2026

    They're Trucking Natural Gas to Keep the SI Campuses Online.

    Oracle couldn't wait for gas pipelines, so it has been trucking compressed natural gas to its data centers at roughly four times the usual fuel cost. It's a bridge, not a fix. It also shows that the ship date for SI is set by steel and permits, not chips.

    Read on X

  • October 9, 2026

    “Just Self-Host the Open Weights” Is Usually the Wrong Bill.

    The weights are free. The GPUs aren't, and they bill whether they're working or sitting dark. Whether self-hosting beats an API comes down to utilization and the people who keep the thing running, not open vs. closed.

    Read on X

  • October 7, 2026

    Synthetic Data Isn’t Free. You Still Owe a Human Floor.

    Labs say human text is running out, so they train on their own outputs. The research is blunt: replace real data with synthetic and models degrade, but keep the real data and add synthetic on top and the damage can be contained. You still owe a human floor, so meter it every training cycle.

    Read on X

  • October 2, 2026

    Your Agents Aren’t Paying for Answers. They’re Paying to Re-Read Themselves.

    Output tokens get the blame, but in an agent's tool loop most of the bill is the same instructions, tools and history getting sent back on every turn. Keep the front of the prompt stable, use the provider's cache on purpose, and measure whether it's actually hitting.

    Read on X

  • September 30, 2026

    The Real AI Bottleneck Isn’t the Model. It’s the Grid.

    Data centers can go up in two to five years, but the transmission lines to feed them can take more than ten. Add power spikes that hit several sites at once, and who actually scales SI comes down to firm power and the wires, not the next model launch.

    Read on X

  • September 23, 2026

    What Happens When Quantum Computers and AI Meet

    They've already met, and it isn't a quantum chatbot. It's a hybrid stack: SI already helps build and correct quantum machines today, while quantum's help to SI is real in theory but still narrow in practice. The real prize is computing pieces of nature that classical machines can only approximate.

    Read on X