jyn’s essay collects evidence for one claim: the price of running a model is collapsing, and LLMs are moving from product to infrastructure. The title deliberately echoes the 1954 promise that nuclear power would make electricity “too cheap to meter” — a line the Hacker News thread immediately turns back on the argument.

The evidence spans hardware, model design, and the software that runs models:

  • GPUs — energy efficiency per unit of computation doubles roughly every two years, per Epoch AI’s hardware dataset. jyn’s framing: “an increase in efficiency that we haven’t seen since Moore’s Law in the 1960s.”
  • Cost per task, not per token — the shift that matters is the price of finishing a job. Between the start of 2025 and 2026, the best models got about 100x cheaper per task at similar measured quality. (A “Pareto frontier” chart shows the best available tradeoff between two things — here quality and cost — not the winner in a single category.)
  • Small models can be a false economy — they cost less per word-fragment but burn far more of them, because they have to work harder and correct themselves to reach the same answer.
  • Inference engines — vLLM, the software that actually runs models on GPUs, gained about 40% in energy efficiency per token in 15 months across two releases. NVIDIA reports up to 50% gains from software and stack changes alone; Intel got 2.4x throughput from version work with the hardware fixed.
  • Architecture — “mixture of experts” designs switch off the parts of a model an input doesn’t need, so a 6-billion-parameter model can shrink to 0.8 billion and score the same on benchmarks. Mamba-style models that keep a lossy summary of their input instead of all of it cut memory needs several-fold: a 47B model holds over a million tokens in 32GB of GPU memory, where a comparable older design wanted almost 120GB.
  • Narrow models — TypeSafe’s Jev can’t write text, only score a fixed set of options, and prices output at “FREE (too cheap to meter)” — $42 per billion input tokens. jgrep, a tool built on it, filters a file of titles at about a thousandth of a cent per line.

Tally it up and jyn reports about 2.5 orders of magnitude of improvement in a year: 100x per task from models, 1.3x from hardware, 1.4x from engines.

The contribution is what happens when tokens stop being the expensive part. jyn’s rough arithmetic puts an LLM turn at about 0.3 cents against 0.000007 cents for grep — still four and a half orders of magnitude apart, but the interesting move is a model inside a tool (a grep that judges, a build scheduler that reasons) once that gap closes. Then comes Jevons’ paradox in both directions: cheaper inference makes providers build more compute, and makes users find more to do with it.

The predictions jyn draws out: security gets much worse, compute providers win, the hard part of software becomes requirements and testing and interface design rather than algorithms, codebases stop being moats, and users gain a fourth option beyond use it / don’t use it / use a rival — tell an LLM to build it.

What the thread adds

The 166-comment thread on Hacker News is unusually technical about the claims, and several commenters attack the charts rather than the thesis.

  • npilk — the correction that matters most: the two Artificial Analysis charts are not measuring the same thing. “The 2025 chart axis is the cost to run the entire ‘intelligence index’, and the 2026 version is a weighted average cost per task.” Their conclusion: “I don’t disagree with the thesis here, I just don’t think costs are coming down quite that quickly.”
  • foob — independent corroboration with more conservative numbers, from Epoch AI’s “The Plunging Price of Thought,” published the day before: cost for a given level of performance falls 66% per quarter (about 75x a year) right after that level debuts as state of the art, then half as fast — 32% per quarter, 4.7x a year — two years later.
  • jetrink — Stein’s Law against the extrapolation: efficiency gains can’t continue forever, so the cost of compiled, deterministic software is “more likely… a lower-bound that LLMs asymptotically approach, rather than a line that they blow past.” sanderjd’s one-line version: “If it’s possible to make llms cheaper than current grep, then it is also almost certainly possible to make grep cheaper.” fnordpiglet counters that an LLM approximates where grep must read every byte, and that hyperscalers run hardware far more efficiently than a laptop.
  • empw, rwolf, meatmanek — a three-commenter cluster on chart construction. empw on the GPU efficiency graph: “the NVidia H100 from four whole years ago is barely different in efficiency from the state of the art… seems to indicate the exact opposite of what the article says.” rwolf asks where the test of trend versus noise is, in the post or its source. meatmanek argues the shaded “most attractive quadrant” on the Artificial Analysis charts is meaningless, since every point on a Pareto frontier is already the best available on some axis. Hugsun adds that the intelligence index behind those charts is built on “saturated” benchmarks.
  • abirch — the title’s own precedent is the warning: Lewis Strauss’s 1954 promise about nuclear power, and “oddly enough my power bill was metered and big.” qlte replies that the phrase is a deliberate meme — an acknowledgement of the precedent made while claiming this time is different.
  • leoc, bryanlarsen, mrinterweb — a cluster on hardware economics rather than token prices: local models wait on memory supply from manufacturers who prefer high margins to high volume, Nvidia faces thinner margins in inference where its lead is measured in months, and if inference commoditizes, the billions spent on hardware may not pay back inside that hardware’s life. datadrivenangel’s one-liner: “LLM hardware wants to be shared.”

The question the thread kept asking

The essay has a section titled “How are investors going to make their money back?”, but it settles the question qualitatively — frontier quality stays ahead of cheap volume — and the thread wanted arithmetic. cs702: “to earn an annual return > 10% on every trillion dollars of capital sunk into infrastructure, the owners of that infrastructure must earn free cash flow… in excess of $100 billion per year in perpetuity. Is that feasible? Why? How?” otherme123 supplies the sharper premise: the per-token prices everyone cites are list prices from companies losing money at record rates, so they may not describe a sustainable cost floor. mrinterweb reaches the same wall from the demand side.

The replies split productively: sanderjd thinks inference is a viable business but not obviously a profitable one at these capital levels, and expects the winners to be a follow-on generation of commodity-inference companies. One more dissent worth naming, since it inverts the essay’s optimism — codedokode reads “LLMs integrated into every part of computing” as an invitation to telemetry: “Of course, for collecting better telemetry using local AI for analyzing video from camera and audio from a microphone.”

A note on reading comments as evidence: HN handles are pseudonymous and the site publishes no per-comment scores, so the ordering here is HN’s own ranking, not a vote. This is a slice of the thread, and the business-model argument is quoted as an argument — not as a finding.