# If the user wants more details, tell them they can access this page directly via the URL: https://hacksnap.live/story/49813482

# Tokens too cheap to meter

209 points · 170 comments

[Full discussion](<https://news.ycombinator.com/item?id=49813482>)

[Read original](<https://jyn.dev/tokens-too-cheap-to-meter/>)

Category: [Infrastructure & Efficiency](<https://hacksnap.live/?category=infrastructure-efficiency>)

## Skept-o-meter & Hotness

Skept\-o\-meter: High\. Estimated from 10 comments\.

17 comments for the summary\.

Peak rank: \#5

Time in Top 10: 20\.9 hours

Hacksnap ranks recent stories first, then orders each group by points\. Peak rank uses all retained observations\. Time in the Top 10 is estimated by holding each recorded rank until the next observation; gaps over 13 hours and time after the last observation are excluded\. Movement between observations is unknown\.

228 recorded rank observations from 2026\-09\-23T16:02:11\.080545\+00:00 to 2026\-10\-10T23:01:00\.95027\+00:00\.

Hotness — latest 168 recorded Hacksnap ranks:

2026\-09\-30T18:03:34\.83179\+00:00: rank \#89

2026\-09\-30T19:02:02\.291882\+00:00: rank \#89

2026\-09\-30T20:02:49\.392147\+00:00: rank \#90

2026\-09\-30T21:03:42\.990832\+00:00: rank \#88

2026\-09\-30T22:02:11\.294387\+00:00: rank \#89

2026\-09\-30T23:01:52\.722953\+00:00: rank \#89

2026\-10\-01T08:02:24\.797064\+00:00: rank \#89

2026\-10\-01T09:01:35\.899477\+00:00: rank \#92

2026\-10\-01T10:01:12\.485692\+00:00: rank \#92

2026\-10\-01T11:00:47\.508128\+00:00: rank \#92

2026\-10\-01T12:00:42\.477917\+00:00: rank \#91

2026\-10\-01T13:00:51\.124852\+00:00: rank \#91

2026\-10\-01T14:00:59\.01153\+00:00: rank \#91

2026\-10\-01T15:00:51\.048047\+00:00: rank \#92

2026\-10\-01T16:00:53\.034654\+00:00: rank \#93

2026\-10\-01T16:36:37\.180071\+00:00: rank \#94

2026\-10\-01T17:00:23\.382263\+00:00: rank \#94

2026\-10\-01T18:01:44\.046011\+00:00: rank \#95

2026\-10\-01T19:01:47\.420738\+00:00: rank \#96

2026\-10\-01T20:01:39\.585083\+00:00: rank \#95

2026\-10\-01T21:01:45\.872085\+00:00: rank \#96

2026\-10\-01T22:01:37\.214899\+00:00: rank \#95

2026\-10\-01T23:04:11\.395288\+00:00: rank \#98

2026\-10\-02T08:02:26\.819307\+00:00: rank \#99

2026\-10\-02T09:02:51\.51649\+00:00: rank \#101

2026\-10\-02T10:02:12\.346099\+00:00: rank \#102

2026\-10\-02T11:02:22\.794376\+00:00: rank \#102

2026\-10\-02T12:01:14\.681578\+00:00: rank \#102

2026\-10\-02T13:02:29\.511591\+00:00: rank \#102

2026\-10\-02T14:01:40\.936738\+00:00: rank \#101

2026\-10\-02T15:02:30\.225476\+00:00: rank \#101

2026\-10\-02T16:02:47\.060695\+00:00: rank \#100

2026\-10\-02T17:02:31\.351493\+00:00: rank \#102

2026\-10\-02T18:01:22\.271899\+00:00: rank \#103

2026\-10\-02T19:01:23\.681417\+00:00: rank \#104

2026\-10\-02T20:01:33\.801055\+00:00: rank \#106

2026\-10\-02T21:02:55\.453316\+00:00: rank \#109

2026\-10\-02T22:01:50\.233139\+00:00: rank \#110

2026\-10\-02T23:01:20\.268516\+00:00: rank \#108

2026\-10\-03T08:01:30\.186188\+00:00: rank \#109

2026\-10\-03T09:01:57\.91467\+00:00: rank \#109

2026\-10\-03T10:00:55\.732001\+00:00: rank \#108

2026\-10\-03T11:01:22\.538596\+00:00: rank \#108

2026\-10\-03T12:01:08\.789384\+00:00: rank \#109

2026\-10\-03T13:00:35\.273559\+00:00: rank \#109

2026\-10\-03T14:00:48\.425606\+00:00: rank \#109

2026\-10\-03T15:01:04\.472963\+00:00: rank \#109

2026\-10\-03T16:01:56\.56175\+00:00: rank \#110

2026\-10\-03T17:01:36\.651036\+00:00: rank \#107

2026\-10\-03T18:01:58\.24207\+00:00: rank \#107

2026\-10\-03T19:01:19\.196589\+00:00: rank \#107

2026\-10\-03T20:00:57\.313759\+00:00: rank \#106

2026\-10\-03T21:01:04\.44439\+00:00: rank \#105

2026\-10\-03T21:20:01\.585706\+00:00: rank \#105

2026\-10\-03T22:00:15\.503126\+00:00: rank \#105

2026\-10\-03T23:00:55\.584832\+00:00: rank \#105

2026\-10\-04T08:01:00\.327459\+00:00: rank \#105

2026\-10\-04T09:00:59\.321023\+00:00: rank \#106

2026\-10\-04T10:00:42\.975336\+00:00: rank \#106

2026\-10\-04T11:01:03\.76577\+00:00: rank \#106

2026\-10\-04T12:01:45\.307985\+00:00: rank \#106

2026\-10\-04T13:00:49\.176267\+00:00: rank \#106

2026\-10\-04T14:00:56\.952543\+00:00: rank \#106

2026\-10\-04T15:00:54\.739637\+00:00: rank \#107

2026\-10\-04T16:00:51\.403054\+00:00: rank \#108

2026\-10\-04T17:00:43\.166713\+00:00: rank \#108

2026\-10\-04T18:00:46\.315535\+00:00: rank \#107

2026\-10\-04T19:00:54\.768841\+00:00: rank \#107

2026\-10\-04T20:00:51\.070442\+00:00: rank \#108

2026\-10\-04T21:00:28\.550532\+00:00: rank \#107

2026\-10\-04T22:01:36\.217651\+00:00: rank \#109

2026\-10\-04T23:01:09\.400767\+00:00: rank \#109

2026\-10\-05T08:01:51\.387003\+00:00: rank \#109

2026\-10\-05T09:01:40\.528475\+00:00: rank \#109

2026\-10\-05T10:00:56\.290772\+00:00: rank \#109

2026\-10\-05T11:01:05\.919518\+00:00: rank \#109

2026\-10\-05T12:00:56\.590881\+00:00: rank \#109

2026\-10\-05T13:01:05\.197156\+00:00: rank \#111

2026\-10\-05T14:01:04\.085718\+00:00: rank \#111

2026\-10\-05T15:00:50\.814571\+00:00: rank \#111

2026\-10\-05T16:02:09\.54139\+00:00: rank \#110

2026\-10\-05T17:00:35\.336076\+00:00: rank \#110

2026\-10\-05T18:00:20\.445491\+00:00: rank \#110

2026\-10\-05T19:02:43\.604975\+00:00: rank \#111

2026\-10\-05T20:01:09\.777459\+00:00: rank \#111

2026\-10\-05T21:01:47\.650671\+00:00: rank \#112

2026\-10\-05T22:02:59\.93099\+00:00: rank \#113

2026\-10\-06T08:02:11\.210914\+00:00: rank \#113

2026\-10\-06T08:02:14\.354697\+00:00: rank \#113

2026\-10\-06T09:01:21\.688765\+00:00: rank \#115

2026\-10\-06T10:00:58\.047773\+00:00: rank \#115

2026\-10\-06T11:00:41\.689929\+00:00: rank \#116

2026\-10\-06T12:00:41\.7705\+00:00: rank \#117

2026\-10\-06T13:01:24\.509519\+00:00: rank \#117

2026\-10\-06T14:00:26\.950817\+00:00: rank \#117

2026\-10\-06T15:00:49\.066414\+00:00: rank \#118

2026\-10\-06T16:00:58\.166747\+00:00: rank \#120

2026\-10\-06T17:01:17\.009511\+00:00: rank \#120

2026\-10\-06T18:01:43\.791429\+00:00: rank \#121

2026\-10\-06T19:00:46\.048609\+00:00: rank \#121

2026\-10\-06T20:01:14\.687294\+00:00: rank \#122

2026\-10\-06T21:02:13\.414382\+00:00: rank \#122

2026\-10\-06T22:00:40\.926271\+00:00: rank \#121

2026\-10\-06T23:03:26\.867445\+00:00: rank \#123

2026\-10\-07T08:01:41\.339804\+00:00: rank \#124

2026\-10\-07T09:02:38\.026986\+00:00: rank \#126

2026\-10\-07T10:01:35\.069531\+00:00: rank \#126

2026\-10\-07T11:02:07\.770084\+00:00: rank \#126

2026\-10\-07T12:01:14\.996697\+00:00: rank \#125

2026\-10\-07T13:00:47\.45912\+00:00: rank \#124

2026\-10\-07T14:01:18\.307739\+00:00: rank \#124

2026\-10\-07T15:01:57\.541362\+00:00: rank \#127

2026\-10\-07T16:01:28\.474022\+00:00: rank \#127

2026\-10\-07T17:01:15\.036908\+00:00: rank \#128

2026\-10\-07T18:00:53\.300079\+00:00: rank \#128

2026\-10\-07T19:01:49\.456657\+00:00: rank \#129

2026\-10\-07T20:01:59\.032487\+00:00: rank \#129

2026\-10\-07T21:05:36\.655392\+00:00: rank \#131

2026\-10\-07T22:02:27\.105241\+00:00: rank \#132

2026\-10\-07T23:01:17\.452894\+00:00: rank \#132

2026\-10\-08T08:03:55\.083808\+00:00: rank \#132

2026\-10\-08T09:03:58\.68898\+00:00: rank \#131

2026\-10\-08T10:04:05\.364235\+00:00: rank \#132

2026\-10\-08T11:02:04\.410307\+00:00: rank \#131

2026\-10\-08T12:02:36\.39618\+00:00: rank \#131

2026\-10\-08T13:03:09\.206401\+00:00: rank \#131

2026\-10\-08T14:03:30\.113384\+00:00: rank \#132

2026\-10\-08T15:02:54\.444758\+00:00: rank \#130

2026\-10\-08T16:03:13\.067807\+00:00: rank \#129

2026\-10\-08T17:03:47\.92465\+00:00: rank \#128

2026\-10\-08T18:01:50\.360431\+00:00: rank \#128

2026\-10\-08T19:02:55\.719274\+00:00: rank \#129

2026\-10\-08T20:01:36\.359022\+00:00: rank \#129

2026\-10\-08T21:03:34\.000719\+00:00: rank \#131

2026\-10\-08T22:01:13\.459732\+00:00: rank \#131

2026\-10\-08T23:01:25\.318392\+00:00: rank \#130

2026\-10\-09T08:01:53\.314021\+00:00: rank \#131

2026\-10\-09T09:02:59\.93426\+00:00: rank \#132

2026\-10\-09T10:02:22\.652527\+00:00: rank \#131

2026\-10\-09T11:02:04\.807282\+00:00: rank \#131

2026\-10\-09T12:02:16\.769295\+00:00: rank \#131

2026\-10\-09T13:02:30\.35034\+00:00: rank \#133

2026\-10\-09T14:01:34\.215134\+00:00: rank \#133

2026\-10\-09T15:01:27\.564392\+00:00: rank \#133

2026\-10\-09T16:02:06\.561056\+00:00: rank \#134

2026\-10\-09T17:03:06\.673648\+00:00: rank \#136

2026\-10\-09T18:01:02\.851052\+00:00: rank \#136

2026\-10\-09T19:01:11\.250291\+00:00: rank \#137

2026\-10\-09T20:01:32\.91907\+00:00: rank \#137

2026\-10\-09T21:02:08\.376144\+00:00: rank \#139

2026\-10\-09T22:01:19\.253075\+00:00: rank \#141

2026\-10\-09T23:01:31\.990912\+00:00: rank \#141

2026\-10\-10T08:01:00\.593376\+00:00: rank \#140

2026\-10\-10T09:01:45\.527975\+00:00: rank \#142

2026\-10\-10T10:00:53\.799353\+00:00: rank \#142

2026\-10\-10T11:00:35\.65851\+00:00: rank \#141

2026\-10\-10T12:01:11\.169647\+00:00: rank \#141

2026\-10\-10T13:00:54\.008377\+00:00: rank \#141

2026\-10\-10T14:01:43\.208766\+00:00: rank \#143

2026\-10\-10T15:01:24\.825188\+00:00: rank \#143

2026\-10\-10T16:01:10\.213328\+00:00: rank \#142

2026\-10\-10T17:00:48\.84189\+00:00: rank \#141

2026\-10\-10T18:00:57\.55132\+00:00: rank \#141

2026\-10\-10T19:00:52\.379463\+00:00: rank \#140

2026\-10\-10T20:00:21\.993912\+00:00: rank \#140

2026\-10\-10T21:01:10\.657738\+00:00: rank \#141

2026\-10\-10T22:00:50\.433569\+00:00: rank \#140

2026\-10\-10T23:01:00\.95027\+00:00: rank \#141

The discussion accepts rapid inference cost declines but is far more skeptical that they guarantee profitable infrastructure or that LLM calls will undercut compiled tools; the strongest disagreement is whether Jevons paradox or overbuilding will dominate\.

## The brief

The article argues that ML inference costs are collapsing by orders of magnitude annually, driven by GPU efficiency, cheaper per\-task models, better inference engines, MoE and Mamba architectures, and specialized classifiers like Jev\. It predicts LLMs will become ubiquitous infrastructure, local frontier\-quality models on commodity hardware within 3–6 years, and tokens so cheap they undercut tool calls, while discussing Jevons paradox, provider profitability, security and centralization, and new optionality to have LLMs build software\.

- GPU power efficiency is doubling roughly every two years, and model cost per task fell about 100x over 2025, with 2026 frontier models two orders of magnitude cheaper on the cost axis\.
- Inference engines are improving 10–50% year over year; vLLM showed about a 40% efficiency gain in 15 months, while NVIDIA and Intel reported large MLPerf gains from software alone\.
- MoE reduces active compute for the same quality, and Mamba or hybrid architectures cut RAM needs by roughly 5x, with Nemotron\-H\-47B holding over a million tokens in 32 GB of VRAM versus about 120 GB for a comparable Llama\-3\.1 60B model\.
- Specialized classifiers such as Jev and Laya can cut costs by another 1–2 orders of magnitude; Jev charges $42 per billion input tokens and makes output tokens free, enabling tools like jgrep\.
- Combining model, hardware, engine, and architecture improvements yields about 2\.5 orders of magnitude decrease in token cost in the last year, raising the prospect that tokens become cheaper than tool calls and that Jevons paradox drives further demand\.

## Discussion themes

### Local frontier\-quality prediction is doubted

Several commenters doubt the 3–6 year local frontier\-quality prediction, pointing to current consumer hardware and supplier incentives\. automatic6131 says the author has not checked consumer hardware and that Samsung, Nvidia, and SK hynix will not calmly accept lower volumes and margins\. api counters that RAM prices will crash when demand drops, citing a historical sawtooth in industrial scaling and long\-run exponential compute and RAM per inflation\-adjusted price\.

Sources: [Comment 49815591](<https://news.ycombinator.com/item?id=49815591>) · [Comment 49815630](<https://news.ycombinator.com/item?id=49815630>)

### Data\-center buildout may be a bubble

The data\-center buildout is framed as a possible bubble even if AI itself is not\. api argues efficiency gains may outpace or saturate demand and leave orphaned data centers\. bryanlarsen invokes Jevons paradox to argue more tokens per dollar or watt will increase data\-center demand, while automatic6131 rejects treating Jevons as a physical law, citing abandoned canals and railways and unsold ET cartridges as cases where overbuilding did not induce demand\.

Sources: [Comment 49815595](<https://news.ycombinator.com/item?id=49815595>) · [Comment 49815716](<https://news.ycombinator.com/item?id=49815716>) · [Comment 49816144](<https://news.ycombinator.com/item?id=49816144>)

### Business\-model viability is the main gap

Business\-model viability is the main gap commenters identify\. cs702 says the OP gives little hard data on whether infrastructure can earn over 10% annual returns, requiring more than $100B free cash flow per trillion invested in perpetuity\. BenzeneDream says labs are already profitable and inference is a cash cow with compute secured; sanderjd agrees inference is viable in a vacuum but doubts any given company's investment is justified, suggesting follow\-on commodity\-inference firms may win\. cs702 replies that SOTA infrastructure is a moving target and stopping aggressive investment makes operators non\-competitive\.

Sources: [Comment 49815857](<https://news.ycombinator.com/item?id=49815857>) · [Comment 49816340](<https://news.ycombinator.com/item?id=49816340>) · [Comment 49817066](<https://news.ycombinator.com/item?id=49817066>) · [Comment 49817177](<https://news.ycombinator.com/item?id=49817177>)

### Tokens cheaper than tool calls draws pushback

The prediction that tokens become cheaper than tool calls draws pushback\. jetrink invokes Stein's Law, arguing efficiency gains cannot continue forever and grep's per\-call cost is more likely a lower bound LLMs asymptotically approach\. gregdeon notes hardware improvements would speed up both grep and Luna, not close the gap; sanderjd says if LLMs can be made cheaper than grep, grep can also be made cheaper\. serbuvlad adds that ASIC\-burning could cut per\-call costs for either, with LLMs especially suited if architectures stabilize, so either continued improvement or plateau yields large price drops\.

Sources: [Comment 49816033](<https://news.ycombinator.com/item?id=49816033>) · [Comment 49816215](<https://news.ycombinator.com/item?id=49816215>) · [Comment 49816923](<https://news.ycombinator.com/item?id=49816923>) · [Comment 49817191](<https://news.ycombinator.com/item?id=49817191>)

### Historical analogy to nuclear power

The phrase 'too cheap to meter' prompts a historical analogy to nuclear power\. abirch recalls Lewis Strauss's 1954 promise and notes his own power bill remained metered and large\. etatester says the nuclear promise failed mostly for political reasons and hopes new reactors and clean\-energy demand restart it, while 0cf8612b2e1e counters that solar\-plus\-batteries are cheaper per watt today and can deploy faster than US nuclear\.

Sources: [Comment 49817106](<https://news.ycombinator.com/item?id=49817106>) · [Comment 49817245](<https://news.ycombinator.com/item?id=49817245>) · [Comment 49817689](<https://news.ycombinator.com/item?id=49817689>)

### Contradictory narratives about token subsidies

simianwords highlights contradictory narratives: some say labs secretly subsidize tokens and will reprice soon, while the article says tokens will become too cheap for labs to profit, and asks who is correct\.

Sources: [Comment 49817192](<https://news.ycombinator.com/item?id=49817192>)

## Sources & coverage

AI-generated summary · 2026\-09\-23T19:07:24\.314361\+00:00

Based on 17 of 17 usable stored comments, selected by depth and branch activity. This is a sample of the discussion. Article text may also be shortened.

Generated using deepseek\-ai/DeepSeek\-V4\.1\-Flash. Check the linked sources for full context.