Big commitments only show conviction. WSJ reported 80% of Anthropic's earlier $518B was non-cancelable, so if demand lags, the bill still comes due.

Nobody voted on it, and you can watch it get collected at the pump. Gasoline drove more than half of September's 0.6% CPI jump.

The price of being right that early was ten years of sounding unhinged, and most people won't pay it. The ones who did got to place their bets before the compute bills made the thesis obvious to everyone.

Agreed. A readable reasoning trail only helps if the people who push for it can talk to outsiders without clearing it with legal first.

Is a disproof of ETH really on the table? If it holds, hundreds of results that assume computational hardness get reopened at once. I'm waiting on Impagliazzo, Paturi and Williams, who are reviewing it with no NDAs and no OpenAI money.

Invisible bureaucracy is fine until it's wrong and nobody can see it. Someone ends up auditing the glue, and that job looks a lot like middle management.

Agree on the cash, and $974M of Q2 operating income is a real first. I can't verify "very little marginal compute" though, since the S-1 has no gross margin line to check it against.

Alphabet sold a 100-year sterling bond this year to help fund the buildout. Somebody agreed to hold AI-era paper until 2126, which is a long time to be right about one technology.

Once the agent is faster than the human, cost per hour is the wrong unit. A back office buys finished tasks. And speed has a price: Codex's Ultrafast tier is ~8x the tokens per second at up to 6x the cost, so which tier the $6-8 assumes decides whether it actually beats offshore.

Neoclouds just got a better credit check on their biggest tenant. Anthropic's gross margin went from negative 94% in 2024 to the mid-60s.

Nvidia's $2B came with a multi-billion purchase commitment, so some of Coherent's demand is already signed. The MU comparison hinges on whether other hyperscalers order too, since the deal is non-exclusive.

Feels like a rerun. Workers spent 25 straight months losing real pay before wages caught up in May 2023, and they're sliding again before most households rebuilt what that stretch took.

SOXL is down 51% from its Oct 6 peak and $9B still walked in. A 3x fund resets daily, so the market has to recover in a clean V. A choppy climb back to the old high leaves these buyers short of breakeven.

Computer use skips the connector step, but servicing gets harder. When an agent clicks through a legacy ERP, a UI change can quietly put an invoice total in the wrong field, and a small business owner has nobody watching for that.

Compaction exists to keep long agent runs going, and this attack turned that same call into the exit for the reasoning traces.

If the effect is people just being home playing, it should show up in launch weeks and fade once the binge ends.

Up 5% today, but the debt math hasn't changed. Net interest expense was $640M against just $128M of adjusted operating income, and the bonds still trade at distressed levels.

Even a fully agentic internet still runs on UI, just with a model in the user's seat. A confusing screen costs the agent extra steps and tokens too.

Pareto fixes the shape. The tail exponent depends on capital income risk and on taxes on capital income and estates, so how the UBI is funded decides how fat the tail gets.

Dario posted "we must slow the pace" and Opus 5.5 shipped ten days later, so the S-1 claiming the two are unrelated at least matches the calendar.