The author observes that a call to GPT-5.6 Luna is only 4-5 orders of magnitude more expensive than grep, and then predicts that at current rates of progress, calling an LLM will soon be cheaper than a grep. I think this is a good time to invoke Stein's Law: "If something cannot go on forever, it will stop." These efficiency improvements won't continue forever. It's more likely that the per-call cost of high-quality, compiled software like grep will be a lower-bound that LLMs asymptotically approach, rather than a line that they blow past with perpetual exponential progress. (Barring a true breakthrough in something like quantum computing or room-temperature superconductors.)
You can burn anything* into an ASIC to make it cheaper per-call.
non-backreferencing grep is not very difficult to implement in an ASIC either. But it's probably not worth it because of how relatively rarely you use it and of the data transfer costs.
LLMs are great candidates for ASIC-burning because they're slow compared even to network speeds and run all the time. The issue is that you don't want to burn a specific model or architecture that then becomes obsolete.
So you've got two possible futures, and both guarantee large price drops: (a) LLMs keep getting better and better and better, so ability/$ keeps rising; or (b) LLMs plateau in ability, in which they will start getting ASIC'd.
Grep (or ripgrep at least) is i/o bottlenecked at this point. It's impossible to process data at faster than i/o speeds, since you have to get the data to the processor somehow. That doesn't change whether that processing is grep on a CPU, or LLM on an ASIC.
GPUs and inference ASICS also have large amounts of high bandwidth memory, plus lots of high speed storage cache, and dedicated very high bandwidth scale-out and scale-up networks. Because they are also often bound by I/O bandwidth.
If your problem is grepping crazy amounts of data, the infrastructure for LLMs isn't a bad place to look for an example.
I did a toy project once implementing a limited version of grep on an FPGA and was able to get some speedup over GNU grep at the time, though marginal.
And a patch panel and a bunch of patch cables that you plug in and out to construct your pipelines.
And eventually hire people whose job it is to patch pipelines on demand for everyone in the office.
“Hey Jim, I’m gonna output the systemd logs of nginx on line five, can you assemble a grep pipeline for me to match all HTTP 500 status codes from /api/cart POST request log lines? Connect the filtered output to Tim’s desk, line 7. He’s there now, we are trying to figure something out.”
i/o could still be sped up though? And I dunno, maybe this actually is an argument for how the llm version could end up being faster, because there is lots of investment in crazy fast i/o hardware and protocols to get the data into the chips.
It's almost a certainty that LLMs have a "core" that will essentially never become obsolete, possibly even 80% to 90% of their parameters. The rules of English and other languages, core ideas in math and science, all of history, nearly all literature, etc. We don't really understand what's going on inside LLMs enough yet to make good use of this, but one day we will have "core logic" neural networks with stable weights burned into ASIC that are doing the heavy lifting, with more dynamic continually-tuned models manipulating the inputs and outputs into those core models. There are also likely stable expert models on topics that don't change much that we could already do this with.
Inference costs cannot keep falling forever, but they do still have a long way to go.
> LLMs plateau in ability, in which they will start getting ASIC'd.
They don't need to plateu for that to happen. There are companies already building AI on ASIC, and IIRC they were approach 12 months lead time. A 12 months old frontier model (Sonnet 4.5, GPT-5, Kimi K2) for 1% of the price is still a rather good value proposition.
In order for the ASIC to achieve 1% of the price, there'd have to be 99% overhead in GPU implementations which for some reason you'd have to be able to eliminate in ASICs but not in GPUs. That seems rather implausible.
It will be a better value proposition now than it was 12 months ago. It's likely to be better yet in another 12 months. There may be room for a parallel to Moore's Law here.
Right now, I do actually use OpenAI's gpt-oss-safeguard-20b for somethings, was released 11 months ago, and is $0.075/M input / $0.30/M output now. I could see this model being in fairly widespread use at 10x speed and 1/10th cost if it was introduced today. Meaning, that for some usecases (moderation) i think dedicated chips can pan out today.
But for more general models, its tougher. Gemini 3 pro was launched in November, if ASICs brought it down 1/10th in cost, it would be $0.20/$1.2. GPT 6 Luna is $0.1/$0.50. Luna is better at a lot of things, but not everything. So 1/10th doesn't really make the ASICS investment worth it in my opinion, but if it brought it down to 1% ($0.02 / $0.12) it would be a really compelling model with a lot of use.
BUT, do i think something like Luna is probably generally capable of doing a huge amount of knowledge work. So if Luna came out at 1/10th the cost a year from now, it would probably be compelling for a while.
It all depends on the rate of improvement in cost/capability.
To be clear, I agree with the overall premise of the article!
But I would probably take a long horizon bet that the grep implementation on my machine will remain cheaper than an equivalent ai task, even though I think those ai tasks will become far cheaper over time.
I just think the original comment's model of asymptotic approach is probably more likely to be accurate than the model of the line blowing through this grep-like cost level.
This whole story really reminds me of crypto coins. Like.. going from mining one coin, or lets say token, to millions of fractions like 0.00000000001 bitcoin a week.
And the progress humanity made in the last 100 years alone... some things end up to be completely world-changing because they enable things that only were an unfeasible dream before. The invention of the printing press, sanitation, vaccines , computers, the Internet certainly are such enabler technologies.
With AI, the question is still open if this will actually turn out to be something useful or if it will in the end just be another way for the elites to make untold profits.
I probably replied to the wrong comment with my previous response. The low lying fruit, in this case, should be going from single cell to multi cell, but that took ages.
It accelerated from there rather than slowing down.
No, it just gets much more complicated to implement grep both in general, and especially in hardware if you support backreferences, since those make it impossible to compile the regular expression into a state machine.
depends on what you are grepping ... greapping a large file might be more expensive one day than generating n-th token with LLM that works fully in hardware
you could make hardware implementation of grep and store the file itself next to it in some ROM but that's not a very useful grep ... while hardware LLM is exactly as useful as software LLM only orders of magnitude faster
LLM is spicy memoizing, so it can potentially be faster than a tool call. But people will spend a month tweaking and testing to ensure they have the level of determinism they need, which means it's more expensive, and that they should have used actual memoization in the first place.
> The author observes that a call to GPT-5.6 Luna is only 4-5 orders of magnitude more expensive than grep, and then predicts that at current rates of progress, calling an LLM will soon be cheaper than a grep.
At some future point where LLM hardware is cheaper than simply running grep, then grep equivalent would benefit from those selfsame hardware improvements and be cheaper to run as well, probably still by the same ratio.
4-5 orders of magnitude is huge. Assuming an order of base 10, it's 10000x-100000x. So a call to grep may return in 1s on a typical PC. That means a GPT call takes equivalent energy of 10000-100000 PCs to do the same in 1s. That's a difference that can't be equalized with scaling. It would require a revolutionary breakthrough.
I also don't understand where the idea that frontier models are getting better efficiency comes from. The results are certainly improving, but that comes from feedback and multiplexing requests, which cost more.
Storage access isn’t, though. You have to read every byte of a (non-sorted) file to find words in it. And whether it’s disk or RAM, the bandwidth is usually a key constraint.
That’ll be just as true for an ASIC NN as it is for good ole grep.
When storage bandwidth is a big enough problem, that will get addressed.
NVMe v1 wasn't released until 2011 and there was no hardware available until 2012. Before that the fastest drive interface was 6 Gbps SAS. Then we got 4 GBps (32 Gbps) with NVMe v3, 8 Gbps (64 GBps) with v4, and 16 Gbps (128 Gbps) with v5. V6 is slated to double again. v7 is currently planned to double that. This is less than 15 years of progress.
In the 1990s, we went from 33 Mbps to 66, to 100, then 133 with IDE. We went from the 40 Mbps of 1986's SCSI 1 to 80 of SCSI 2 in 1994. Since then parallel SCSI has gone to 80, 160, 320, 640 (Ultra 2 wide), 1280 (Ultra 3), 2560 (Ultra-320), and 5120 (Ultra-640) Mbps.
SAS has also gotten faster, with 6 Gbps, 12 Gbps, and 22.5 Gbps. NVMe is still lower overhead and lower latency.
Memory bandwidth can also be addressed. Stock servers with EPYC are 16 channel DDR5. That's 409.6 GBps. Four channels of DDR3 was only 102.4 GBps. Video and accelerator card memory bandwidth is also increasing.
What if the tool is more advanced, like an optimizing compiler: `g++ -O3 -march=native -x c++ - <<EOF ... EOF`
If the compiler invocation is sufficiently slow, the llm could consider outputting a binary directly?
For all we know matrix multiplications are a faster way to generate optimized machine code than branchy sequential compiler code with tons of heuristics and passes.
It might never beat out grep, but it could beat some more expensive to call tools, similar to how heuristics will often be faster than exact answers. Rust Analyzer can be slow at times, I could see an AI tool taking over a subset of its work.
From a computational standpoint this is obviously nonsense, but from an attentional one I'm not so sure. It may already be more attentionally expensive to use grep in some cases, such the moment you need to remember a non standard arg. And if this applies for performing a simple http operations, then it certainly applies going up the complexity chain.
room-temperature superconductors, sure, but I fail to see how quantum computing will disrupt – in the medium term (25 years or so) – classical computing in any meaningful way
Is running LLMs (or some other ML workload) on/with quantum computers expected to bring efficiency gains?
I found the OP insightful and worth a read. Thank you for sharing it on HN.
The only aspect that is poorly analyzed by the OP is business model viability. All players are investing insane amounts of money in infrastructure with the expectation that their future profits will justify all that investment. The winner or winners in the AGI race, they believe, will find the proverbial "pot of gold at the end of the rainbow."
The OP glosses over questions of business model viability with a brief qualitative discussion and very little hard data. For example, 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 (operating profit less investment) in excess of $100 billion per year in perpetuity. Is that feasible? Why? How?
I think the article's analysis is basically right in a vacuum. That is, I think it's clear that inference is a viable business model. But what isn't clear is whether it will be such a profitable business model for any given company that it will justify the investment that company has taken. I kind of think the winners might be a follow-on generation of companies that focus on this commodity inference business model instead of the invent-machine-god-first "business model" and thus are wiser about their level of investment and capital costs.
> I think it's clear that inference is a viable business model.
You may be right. I'm not so sure. Inference looks like a viable business model for those operators that have SOTA infrastructure in place, but the investment required to have it is enormous, and appears to be never-ending, because if an operator stops investing aggressively, its infrastructure quickly becomes non-competitive, and customers will quickly leave for alternatives. SOTA infrastructure is a moving target.
Renting out compute is a viable business model. That's part of what the big and the small cloud providers do. It's just (on its own) not a "get rich quick" business model anymore.
Why would it be any different for inference? If we believe OP, it'll just become part of regular compute infra, and thus part of the renting-out-compute business model.
I think it's an open question if the current generation of inference investment will pan out, but in the long term, there'll be a balance between investment cost and margin, just as in every other industry.
What about inference on a distilled version of someone else's model? What about inference on an open weights model? I don't think all LLM business models are going to revolve around developing state of the art frontier models.
Right. This is what I was implying in my comment, but it's good to be explicit about it. I think there will be successful companies that just sell inference against the best models they can get without needing to invest anything (or very little) in training anything new. This could even be a spinoff of one of the big frontier labs. I would personally rather invest in an IPO for a company that only owns the gpt-6-astra implementation and infrastructure than in openai itself. There is certainly less upside, but IMO also way less downside risk. (I'm not saying there is any chance this kind of a spin-off would happen, it's just a thought experiment.) And I think this same calculation applies to open weights inference providers.
Edit to add: Or or might just be AWS / GCP / Azure that benefits from this business model. They're already pretty good at selling commodity infrastructure.
Maybe... I'm not enough of an expert on the financials to say, but it seems to me that inference should be able to recoup the cost of SOTA infrastructure, unless you then also use a large portion of that infrastructure to train new models. And I also think the race to remain SOTA itself is also largely a function of training, because my understanding is that training benefits more from the leading edge of hardware.
But yeah, I definitely don't have high confidence in any of this!
> I think it's clear that inference is a viable business model.
Only if you also have the model thats better than anyone else's.
As soon as models are free, or there are no newer models (assuming thats going to happen, and thats not a given) then the only thing you can compete on is price.
This means that the only thing you have to differentiate is either price, speed or ease of use. (or regulatory capture...)
We are at pets.com level of spend currently. Unless model development becomes cheaper, then we are going to run out of novel debt but not really debt mechanisms.
The problem source is that the "cost" of tokens are taken at face value from business that are losing money at record speeds. E.g. https://artificialanalysis.ai says "doing task A costed us $10 using OpenAI", and that is the "cost" the OP used as basis for "tokens are cheap". Meanwhile OpenAI is losing $19 for each $1 in revenue... So right now OpenAI should be charging around $200 to do task A just to break even, but that would mean their use base would collapse.
Some frontier labs are reporting positive "adjusted EBITDA" (earnings before interest, taxes, depreciation, and amortization, with extra adjustments to make the figure positive).
Free cash flow (operating profit less investment), actual cash coming in, is deeply in the red.
EBITDA can be a sensible measure of profitability when there isn't much need for additional investment. That doesn't seem to be the case with these operators. They need to invest aggressively to avoid losing customers to competitors. All of these operators have made multi-year commitments to invest more in infrastructure. In addition, they have guaranteed quite a bit of debt to fund it.
Maybe it all will work out fine (and I sure hope it does!), but I didn't see any hard data from the OP, or from you, supporting that view.
EBITDA might make sense for the resellers who package up open weight models and sell inference. It is not appropriate for the labs who have billions in debt for RAM, new data centers, gobbling up competitors, etc.
Those real debt obligations are going to want to be paid back.
Definitely. The question is: Is it enough to recoup the enormous capital costs and justify the level of investment they've received. I think there's a decent chance that it will be. But maybe not. And the longer they keep focusing on training new models more so than on inference, the more uncertain I become that it's all going to work out.
Honestly at this point with training costs I don't see how it could ever pay itself back unless you get RSI, in which the talk of money really isn't the main problem any longer.
We're in a situation where AI isn't going to go away, but whatever financial mode we're in right not is not going to work.
Labs are playing money games with EBITDA, which is not uncommon, but also hides the extent to which they are in the red (deeply, deeply, in the red, and projected by them to get worse).
Think of it like spending $1 trillion to become the next Google. That hardware itself may never turn a profit. But if 10 years for now you're the software provider that owns the ecosystem around too cheap to meter tokens you've got a money printing machine.
Exactly. "Cost-to-distill" is a critical parameter. Right now usage of frontier models for all tasks is both subsidized and irrationally popular even at the subsidized price. Deepseek would solve most tasks faster and 10x cheaper. I agree with the author that just as Deloitte exists, frontier labs will exist. But not because their products are proprietary technical marvels or gods, but rather because of branding.
DS wouldn't be 10x cheaper than the subsidized subscription plans from openai/anthropic. Although it is of course much cheaper than the enterprier/API pricing- I think if you're on the subscription plans, you can't beat that on performance per price.
It's a deliberate reference/meme that is basically used to acknowledge the precedent of overly exuberant predictions of cost in an emerging technology but argue "however, this time it's true".
Of course, perilous territory for future irony depending on how your prediction plays out.
Long distance phone calls were charged per minute.
Texting was charged per character.
And if you have the capabilities to install your own solar it can pay itself back in 6 years. Not sure what that looks like with 100% battery coverage.
It's really quite unfortunate that the promise was not delivered, mostly for political reasons. I hope that a new wave of reactors and the dire need for clean energy restarts the nuclear race.
One excellent nerd-sniping side-effect of those political reasons is that the good folks of Austria built an entire nuclear power station and then never put it into service [0]. Open for guided tours on Fridays [1].
The economics are not there. Solar + batteries are much cheaper per watt today and still improving. Even better, the solar can come online instantly and expand while US nuclear takes twenty years to start generating any energy.
> Solar + batteries are much cheaper per watt today and still improving.
Are we still going to make the same error over and over again? The day China says "stop" the price per watts will explode, just like when Russia and Iran said "stop".
Those aren't really equivalent. Iran is affecting the supply of oil, an input. China can affect the supply of solar panels, but they would have to do something pretty drastic to affect the supply of the input. Sunlight is going to keep reaching the solar panels that were set up prior to any kind of embargo.
I'm a huge solar fan, but I don't see energy transition happening as quickly as I'd hoped 10 years ago, and what seems apparent now is that it will only happen when energy storage gets 10x cheaper, which is why companies like Form Energy developing iron-air batteries might be the catalyst. You also need grid-forming inverters that balance the loads since relying on hundreds of thousands of residences for grid-tie is a real engineering challenge.
There is certainly a lot of talk about micro reactors. I have yet to see anything materialize in the US. The MARVEL test reactor is meant for something puny like 100kw. Nuclear is something that benefits from scale -larger installations are just going to have better economics per watt.
No it really does not look that way. It sort of sounds like it might look that way at some point in the future but micro reactors are not useful for utility generation today.
Meh. Cheap seasonal batteries aren't a thing yet so for much of humanity solar + batteries just can't cover the same needs as other means of generation.
Nuclear was never going to give us too cheap to meter regardless of politics. Uranium just isn’t that cheap. Fuel costs are lower then coal or gas, but not so low that operators would just not bother to charge for it
A MSTR only needs uranium-233, uranium-235, or plutonium to start. It then produces uranium-233 as part of its fuel cycle as thorium is input.
Given current found reserves and the current rate of use, the world has about 40 to 50 years of natural gas. Thorium used in molten salt thorium reactors would provide electricity for 60 billion years or so if we could actually extract all of it. That's 10 billion years or more if it provided all human energy consumption. Of course there's a limit to extraction, but it is over three times as common as uranium.
Also, besides thorium one can mix in partial amounts of other fuels, including uranium and plutonium. There is no runaway meltdown risk, as the fission is actively managed rather than actively suppressed. Fuel is spent more completely. The waste products are smaller, less radioactive, and have far shorter half-lives.
Then of course we're always getting slightly closer to productive fusion reactors.
These technologies along with solar PV, solar thermal, hydro, wind, geothermal, wave power, and batteries likely all have a place in the future.
There's a decent chance that at some point in the future residential customers will pay for the connection and only commercial or industrial customers will actually be metered. That's not because companies want to give up additional revenue. It's because at some point the cost of meters, tracking usage, and competitive advertising about who has the cheapest plans costs more than the power the typical customer uses above the base charge.
"Between digital economics and the wholesale embrace of King's Gillette's experiment in price shifting, we are entering an era when free will be seen as the norm, not an anomaly. How big a deal is that? Well, consider this analogy: In 1954, at the dawn of nuclear power, Lewis Strauss, head of the Atomic Energy Commission, promised that we were entering an age when electricity would be "too cheap to meter." Needless to say, that didn't happen, mostly because the risks of nuclear energy hugely increased its costs. But what if he'd been right? What if electricity had in fact become virtually free? The answer is that everything electricity touched—which is to say just about everything—would have been transformed. Rather than balance electricity against other energy sources, we'd use electricity for as many things as we could—we'd waste it, in fact, because it would be too cheap to worry about."
... What Mead understood is that a psychological switch should flip as things head toward zero. Even though they may never become entirely free, as the price drops there is great advantage to be had in treating them as if they were free. Not too cheap to meter, as Atomic Energy Commission chief Lewis Strauss said in a different context, but too cheap to matter. Indeed, the history of technological innovation has been marked by people spotting such price and performance trends and getting ahead of them."
My issue is that datacenter energy costs are being prioritized for commerce over residential use, so the average consumer is paying more for electricity, because a datacenter needs more electricity and they're getting tax breaks. Assuming this all improves efficiency for new products like automated robotics, there is a debatable benefit. Jevon's Paradox has no ceiling, except the environment, and people's 401ks.
On the other hand, this did work out in other areas. I pay a flat monthly rate for all-I-care-to-eat internet access, for example. My email provider has limits on storage space but I don’t get charged per email sent or received. It’s not a crazy concept on its face, nuclear power just didn’t work out as well as was hoped.
I'm always reminded on Orwell's quote about the then new atomic bomb and his prescience on how it would all work out:
"Had the atomic bomb turned out to be something as cheap and easily manufactured as a bicycle or an alarm clock, it might well have plunged us back into barbarism, but it might, on the other hand, have meant the end of national sovereignty and of the highly-centralised police State. If, as seems to be the case, it is a rare and costly object as difficult to produce as a battleship, it is likelier to put an end to large-scale wars at the cost of prolonging indefinitely a “peace that is no peace”."
It seems, especially with open weights, that the AI is much more like the alarm clock and not the battleship. $20/mo would have been about $1 in 1944
I just want to rant about these Artificial Analysis charts that you see everywhere:
The "most attractive quadrant" is completely meaningless. The whole point of a Pareto curve is that each point on the curve is better than everything else on at least one dimension, and that you can make these comparisons without placing a value judgement on the relative importance of the different metrics. If you make a composite score of the two metrics (any monotonically non-decreasing function, e.g. a weighted sum with non-negative weights), that score will always be maximized by one of the points on the Pareto frontier.
So going by the numbers in the 2nd chart (1st AA chart) from TFA alone:
- there's no reason one would choose Deepseek V4 Pro 0813 (max) even though it's in the "most attractive quadrant", because GLM-5.3-Flash is both cheaper and scores better.
- Claude Fable 5.1 (max with fallback) on the top right* could be your most attractive option if you need the best scoring model and don't care about cost, even though it isn't in the "most attractive quadrant"
- The un-shown model off the left side of the chart could be your most attractive option if you just need lots of cheap tokens and don't care about quality.
(Obviously if you start including other factors in your score that aren't represented on the chart, then you might choose differently.)
* I also dislike the way they place the labels, and that grey line connecting the label to the point is way too subtle.
Sound critique. I'll add that the Artificial Analysis intelligence index is not considered a good metric for intelligence anymore. Most of the benchmarks that it comprises are saturated or considered low signal today.
It's true that LLMs "want" to be be local, but they won't shift broadly to being local until there's a sufficiently large supply of VRAM or (at least) "unified" memory from the manufacturers. (I'm also assuming here that radical regulatory changes like government bans of local models aren't going to happen.) So (AFAICS—I am no expert) the future of LLMs over the next few years comes down primarily to the nitty-gritty of how much memory fab capacity will be added and when, and to a lesser extent of what happens to future demand from LLM SaaS services (& maybe their existing stock of hardware if they get in trouble). (I'm also assuming no roughly-AGI-sized leap forward which makes the frontier models of the near future vastly more valuable than the near-fontier models of today.) For the incumbent manufacturers the high-margin business is selling to LLM SaaS providers who use VRAM efficiently, but the high-volume business is getting chips into millions of laptops which will use VRAM very inefficiently. I assume that they will want to move from high margins to high volumes as they build they physical capacity to ship higher volumes, but they seem to prefer to do it at a stately pace. Hopefully some jostling from Chinese competitors, and maybe a dropoff in demand from data centres, will speed things along.
I wouldn’t say so. Once local solutions pass a threshold of affordability consumers tend not to mind too much about their inefficient resource utilisation: see the many thousands of MacBooks which sit largely idle for most of the day. And efficient utilisation is actively against the interests of hardware manufacturers, at least while they’re not supply-constrained and looking to sell their limited supply to whomever can pay top dollar.
I think Nvidia is under the same pressure as Anthropic/OpenAI. Nvidia will dominate research and probably keep dominating training, but the real volume is in inference. And for inference Nvidia's lead is only a few months, similar to the lead frontier labs have over open source. Nvidia will sell a lot of Rubin CPX's, but their margin on that will be a lot smaller than B200 because there is so much more competition in that space.
A much deeper analysis on the falling price per task was published yesterday by Epoch AI [1]. It's a real statistical analysis and comes to more defensible and grounded conclusions. The headline takeaway is:
The cost of a given level of performance often falls fastest right after that level is first achieved, that is, when it is state of the art (SOTA). We see this pattern on three of our five main benchmarks of AI capability. Averaging across all five, cost falls 66% per quarter (75× per year) for performance that has just debuted as SOTA. Two years later, prices fall half as fast, at 32% per quarter (4.7× per year).
but the analysis itself has more nuance and is a quite interesting read.
There seems to be a mistake in the cost comparison between 2025 and 2026. The 2025 chart axis is the cost to run the entire "intelligence index", and the 2026 version is a weighted average cost per task.
I don't disagree with the thesis here, I just don't think costs are coming down quite that quickly.
GPU case seems very weak. The graph is impossible for me to reason about at least. You could draw basically any trend line through that GPU graph and it would look equally plausible to me. The main takeaway I get is that the NVidia H100 from four whole years ago is barely different in efficiency from the state of the art, which is surprising to me, and seems to indicate the exact opposite of what the article says.
If inference continues toward the trend of becoming a commodity, and inference hardware efficiency is doubling every two years, that will likely mean that the price of inference will continue to fall so long as the market is competitive. I don't see how companies investing billions in inference hardware today will see profit in the lifespan of the hardware.
for the first chart (sourced from https://epoch.ai/data/machine-learning-hardware?view=graph&y...), what is the audience supposed to think about that trend line? there's a step after you slap a regression on some points where you evaluate whether there's a real trend or noise, right? i don't see that either in the article or the linked source.
I agree with OP that we will continue to see improvements, but there are also some serious bottlenecks ahead of us:
- Energy is not infinite, neither energy efficiency is.
- Datacentres neither.
- Benchmarks are an abstraction of real world problems!
On top, there is an overall "economic" aspect that most of the people miss: every change carries a certain degree of risk (lose money, reputation, customers, death of people, ecc) that very few want to take and a lot of changes(e.g. rewrite some piece of SW in another Lang) don't produce a positive economic impact.
If we're all using distilled open-weight models in ASICs in our own systems the energy cost will come way down. The question is when that becomes a reasonable solution for a broad set of use cases.
I think this is a case where just drawing a "line goes up" extrapolation is incredibly misleading because there is _tremendous_ economic pressure to get costs down, and costs are very tightly tied to energy use. All of these systems are incredibly inefficient right now and have a lot of room to go down in energy use. I'd guess that the absolute _floor_ is burning model weights directly to silicon and that's like a 90+% reduction in energy use.
Not really related to the central point, by but I couldn't help but get caught up by
> Generally, models intended to be run locally will be much smaller, such as Muse Glimmer or Qwen3 Coder.
That is such an interesting set of models to use as examples here. One being essentially obsolete on release a month ago, and the other being completely ancient in LLM time. I really wonder how they landed on those two.
everything will be able to talk to anything else, for real this time
it will be like the internet of things only some asshole will call it "intelligence of things"
again, there will be no S for Security in this new IoT
your thermostat will one day start fucking with you. when you run a diagnostic llm on it, it turns out it's keeping around 5 different viral copies of personality files around, that were left behind by llm botnets/openclaw-like memetic replicators/your grandpa leaving behind easter eggs before his death.
the future will be pretty evenly distributed, and full of weird shit
I just stopped reading at that, for anyone else, Please find a better source and take everything in here with a grain of salt.
IMO
The number one improvement that mattered for local AI was llama.cpp, partial offloading to system cpu/ram. The next was quants, being able to take fp16 and turn it to q8, q4 etc. The next IMHO is unsloth dynamic quant, that have been able to do mixed precision so we have UDq1/q2 that is actually pretty damn coherent. Allowing individuals to drive K3 locally even if it's at Q1/Q2. Then MoE changed everything for everyone, cloud and local. The other is integrated GPU, Apple, Strix Halo, DGX Spark. Then all the extra improvements like MTP, DSpark, etc. Of course there's many other additional things that have mattered too
They're not secretly subsidizing, they're openly subsidizing.
Token pricing was a small minority of customers up until this year, when all the labs started trying to force customers onto token-based billing. Within the last week, Anthropic repriced my team's plan from a temporary "50% extra tokens" to 25%: https://support.claude.com/en/articles/15910845-claude-code-...
The fact that all this is ongoing within such a short timeframe should make you suspicious of any analysis that claims to be observing "statistical trends" like they've discovered a new Moore's Law out of 6 months of pricing data from 2 companies.
your repricing has nothing to do with subsidising which means selling at a loss. Within this year, the real prices have gone down more than 10x on average which is way more than the teeny 25% you are fighting for.
The cost is decreasing quickly, mostly because the labs stopped competing on quality. At the same time, the costs are enormous, and all labs are very openly subsidizing usage hoping to get enough scale to be profitable (while 1 of them has suspicious unity numbers and can be hiding negative marginal income).
Also, it's impossible that they become cheaper than specialized software. Or even as cheap as them. It's still possible that they become cheap enough that it doesn't matter.
how is this nonsense still so persistent? There's piles and piles of evidence that inference has massive gross margins at api pricing. what are you actually talking about?
I find it difficult to believe the inference only providers (Baseten, Fireworks, Digitalocean, etc) are all selling tokens at a loss.
Asking Claude for a rough estimate based on publicly available throughput and cost data for open weight models on modern GPUs suggests serverless, pay-as-you-go inference is profitable on owned GPUs with reasonable utilization (30-50%).
This is the core of my belief that data center construction is a huge bubble.
AI is not a bubble, IMO, though we may see a retrench and some companies with sky-high valuations will crash to more reasonable ones. But data center demand is probably a bubble, and the main driver will be reduction in the actual amount of power and data center space required to serve escalating demand.
I think hardware and model improvements will pace or maybe outrun demand and then when demand starts to saturate will keep going and leave a lot of orphaned data centers.
I disagree. I own over 1TB of vram at home. I can tell you that it's not a bubble. From my builds, I would rather have cloud, cloud is easier. From running small models like Qwen3.8-27B to large models like Qwen3.8-2.4T. I can tell you that small models will never be enough or match up. Everyone will want the smartest model, not just a good enough model.
> Everyone will want the smartest model, not just a good enough model.
Not so sure about this. There’s always a potential threshold. After all, we don’t all use the most powerful computers, the latest phones, the highest resolution cameras, the fastest or best cars.
I am already not interested in cloud LLMs and I don’t even use the best (on paper) model that I can run locally. I prefer a model that people insisted (here) was “dead on arrival” but appears to work better for me.
I think the difference is that AI as an edge. That edge will turn into more money, better quality of life, etc. Of course, with serious skills, you might be able to use use a not so smart model to keep up with folks with smart models. People are lazy tho, and will prefer for AI to do all the work if it means they do none.
Jevon's paradox isn't a physical law, it doesn't magically apply to everything. Millions more copies of Atari's ET game didn't cause everyone to pickup a cheap copy, and cause extra demand for a garbage video game. Some times (actually, usually, I'd argue) things are made that will sell for less than the cost of construction because of irrationality, and they don't induce extra demand and they don't change the negative profit margins.
You can't simply wave Jevon's paradox at things. Thousands of miles of canals were dug in the UK that couldn't be sustained and were abandoned. Thousands of miles of railways were laid that could be sustained and were abandoned. And those are potentially durable investments, unlike cheap walls, pillars and roofs laid over a levelled concrete slab full of fast depreciating IT equipment.
I'm so glad the tide here is turning on this talking point, brought on by exactly the same people beating us over the head with it for months while no progress is made towards it materializing.
Many, many people who post here are capable neither of real analysis nor distinguishing real analysis from memes. They aren't hackers, they are adherents of a cult that happens to focus on the same subject matter as hackers.
Jevon's paradox only applies to products with near infinite demand. Energy being the most famous example. I dont think its difficult to argue that compute/intelligence is also a base input into the economy and theres almost no limit to the amount of intelligence the world will want.
The article observes that the cost of frontier intelligence from 2025 has fallen 100x in the last year. It also notes that the energy to run models is also collapsing. Consumer hardware is borked right now because these new algorithms are revolutionizing the utility of a computer. Computing is technology who's cost has been collapsing for 90 years, and its a safe prediction that it will decrease again.
> Yeah okay bud, anyone checked in with the state of consumer hardware recently? Not the author, evidently.
RAM prices will crash when demand drops even a little. They'll probably crash to a lower (inflation adjusted) level than before. This has happened before.
Industrial scaling in general often looks like a sawtooth: price spike, capacity investment, crash, repeat.
Part of what's keeping prices high a little longer is that everyone knows this and is a little reluctant to plow resources into chip fabs for fear of having the bottom fall out before they recoup or sell that to someone else to hold that bag.
Graph the average compute and RAM in a mid-high end laptop at an inflation adjusted price point for the past 40 years. It's very exponential and hasn't slowed down much.
No. Prices will crash when supply side expands to meet the increased demand. Because demand won't go down to pre-bubble times any time soon. Unfortunately the supply side has been very slow in increasing production, partly because most steps of the production chain are all maxed out.
On a long enough scale you are right that prices will likely normalize to a better level, but before 2030? That would mean the factories are built quickly once they begin.
The entire reason why hardware prices are so absurd right now is because manufacturers across the board are doing everything to prevent that crash.
They all collectively chose NOT to increase supply with increased demand. So if the bubble pops, they just go back to previous prices without oversupply driving the prices to rock bottom.
> Tokens become cheaper than tool calls
The author observes that a call to GPT-5.6 Luna is only 4-5 orders of magnitude more expensive than grep, and then predicts that at current rates of progress, calling an LLM will soon be cheaper than a grep. I think this is a good time to invoke Stein's Law: "If something cannot go on forever, it will stop." These efficiency improvements won't continue forever. It's more likely that the per-call cost of high-quality, compiled software like grep will be a lower-bound that LLMs asymptotically approach, rather than a line that they blow past with perpetual exponential progress. (Barring a true breakthrough in something like quantum computing or room-temperature superconductors.)
Yeah I bumped on that too. If it's possible to make llms cheaper than current grep, then it is also almost certainly possible to make grep cheaper.
You can burn anything* into an ASIC to make it cheaper per-call.
non-backreferencing grep is not very difficult to implement in an ASIC either. But it's probably not worth it because of how relatively rarely you use it and of the data transfer costs.
LLMs are great candidates for ASIC-burning because they're slow compared even to network speeds and run all the time. The issue is that you don't want to burn a specific model or architecture that then becomes obsolete.
So you've got two possible futures, and both guarantee large price drops: (a) LLMs keep getting better and better and better, so ability/$ keeps rising; or (b) LLMs plateau in ability, in which they will start getting ASIC'd.
Grep (or ripgrep at least) is i/o bottlenecked at this point. It's impossible to process data at faster than i/o speeds, since you have to get the data to the processor somehow. That doesn't change whether that processing is grep on a CPU, or LLM on an ASIC.
GPUs and inference ASICS also have large amounts of high bandwidth memory, plus lots of high speed storage cache, and dedicated very high bandwidth scale-out and scale-up networks. Because they are also often bound by I/O bandwidth.
If your problem is grepping crazy amounts of data, the infrastructure for LLMs isn't a bad place to look for an example.
ripgrep may be.
I did a toy project once implementing a limited version of grep on an FPGA and was able to get some speedup over GNU grep at the time, though marginal.
In any case, LLMs aren't IO bound :))
I am now imagining a future where developers buy fancy "grep cards" for their machines. I don't hate it.
It would be cool if they could be daisy-chained, so you can have a hardware implementation of |
The next logical step would be to have a separate card for each Unix utility.
And a patch panel and a bunch of patch cables that you plug in and out to construct your pipelines.
And eventually hire people whose job it is to patch pipelines on demand for everyone in the office.
“Hey Jim, I’m gonna output the systemd logs of nginx on line five, can you assemble a grep pipeline for me to match all HTTP 500 status codes from /api/cart POST request log lines? Connect the filtered output to Tim’s desk, line 7. He’s there now, we are trying to figure something out.”
“Sure thing Bob, give me a moment.”
It’s OK: you can just tell us you have a financially and relationally crippling Eurorack addiction. You’re amongst friends here.
Patching cables is so primitive... Use punched cards to define the connection patterns or something!
Transputerpunk.
At some point Intel was experimenting a Xeon with built-in FPGA, maybe they were just a bit too early to the game :)
Universal Basic ASICs
If your BASIC is compiled that's just a CPU!
I've evaluated some of these and in almost all cases they failed to outperform Intel's ridiculously optimized CPU regex library, hyperscan.
Right, but I read the comment's point as: being i/o bound is the eventual limit of the asymptote and it's the same one grep has.
i/o could still be sped up though? And I dunno, maybe this actually is an argument for how the llm version could end up being faster, because there is lots of investment in crazy fast i/o hardware and protocols to get the data into the chips.
It's almost a certainty that LLMs have a "core" that will essentially never become obsolete, possibly even 80% to 90% of their parameters. The rules of English and other languages, core ideas in math and science, all of history, nearly all literature, etc. We don't really understand what's going on inside LLMs enough yet to make good use of this, but one day we will have "core logic" neural networks with stable weights burned into ASIC that are doing the heavy lifting, with more dynamic continually-tuned models manipulating the inputs and outputs into those core models. There are also likely stable expert models on topics that don't change much that we could already do this with.
Inference costs cannot keep falling forever, but they do still have a long way to go.
> LLMs plateau in ability, in which they will start getting ASIC'd.
They don't need to plateu for that to happen. There are companies already building AI on ASIC, and IIRC they were approach 12 months lead time. A 12 months old frontier model (Sonnet 4.5, GPT-5, Kimi K2) for 1% of the price is still a rather good value proposition.
In order for the ASIC to achieve 1% of the price, there'd have to be 99% overhead in GPU implementations which for some reason you'd have to be able to eliminate in ASICs but not in GPUs. That seems rather implausible.
It will be a better value proposition now than it was 12 months ago. It's likely to be better yet in another 12 months. There may be room for a parallel to Moore's Law here.
This I agree with.
Right now, I do actually use OpenAI's gpt-oss-safeguard-20b for somethings, was released 11 months ago, and is $0.075/M input / $0.30/M output now. I could see this model being in fairly widespread use at 10x speed and 1/10th cost if it was introduced today. Meaning, that for some usecases (moderation) i think dedicated chips can pan out today.
But for more general models, its tougher. Gemini 3 pro was launched in November, if ASICs brought it down 1/10th in cost, it would be $0.20/$1.2. GPT 6 Luna is $0.1/$0.50. Luna is better at a lot of things, but not everything. So 1/10th doesn't really make the ASICS investment worth it in my opinion, but if it brought it down to 1% ($0.02 / $0.12) it would be a really compelling model with a lot of use.
BUT, do i think something like Luna is probably generally capable of doing a huge amount of knowledge work. So if Luna came out at 1/10th the cost a year from now, it would probably be compelling for a while.
It all depends on the rate of improvement in cost/capability.
Yeah I think this is the right mental model. But we have no idea what any of the coefficients will be on any of the terms moving forward :)
To be clear, I agree with the overall premise of the article!
But I would probably take a long horizon bet that the grep implementation on my machine will remain cheaper than an equivalent ai task, even though I think those ai tasks will become far cheaper over time.
I just think the original comment's model of asymptotic approach is probably more likely to be accurate than the model of the line blowing through this grep-like cost level.
This whole story really reminds me of crypto coins. Like.. going from mining one coin, or lets say token, to millions of fractions like 0.00000000001 bitcoin a week.
There is a reason it took 4 and a half billion years for human intelligence to develop.
Around half of that was going from single celled life to multi celled life.
The complexity gets faster as you get on with it.
And the progress humanity made in the last 100 years alone... some things end up to be completely world-changing because they enable things that only were an unfeasible dream before. The invention of the printing press, sanitation, vaccines , computers, the Internet certainly are such enabler technologies.
With AI, the question is still open if this will actually turn out to be something useful or if it will in the end just be another way for the elites to make untold profits.
Stratigraphically the invention of agriculture and the first thermonuclear weapon happened at the exact same instant.
???
Yes, because it was a largely random undirected process.
Even in directed processes the low hanging fruit is harvested first.
Then you have to climb to another branch to get more fruit. The biggest issue with most problem space discovery is you're doing it blindfolded.
I probably replied to the wrong comment with my previous response. The low lying fruit, in this case, should be going from single cell to multi cell, but that took ages.
It accelerated from there rather than slowing down.
Do you mean most of uses of grep are backreferencing? In my 30-yr career I did not use regex backrefs once except for learning them.
No, it just gets much more complicated to implement grep both in general, and especially in hardware if you support backreferences, since those make it impossible to compile the regular expression into a state machine.
grep is deterministic. Llm is probabilistic. Llm can be transferred to a tiny quantum cpu or a lower precision float.
When the author wrote Llm can be as cheap as a tool, I read it as not equivalent. They even said the Llm can be embedded into a tool.
Their point was, the higher level use case — like classification — could become as cheap as grep. Which is quite well possible.
Yeah you may be right!
depends on what you are grepping ... greapping a large file might be more expensive one day than generating n-th token with LLM that works fully in hardware
you could make hardware implementation of grep and store the file itself next to it in some ROM but that's not a very useful grep ... while hardware LLM is exactly as useful as software LLM only orders of magnitude faster
Yeah it's a pretty poorly specified problem. It needs to be some kind of "equivalent task", but it's not clear how to define that.
LLM is spicy memoizing, so it can potentially be faster than a tool call. But people will spend a month tweaking and testing to ensure they have the level of determinism they need, which means it's more expensive, and that they should have used actual memoization in the first place.
> The author observes that a call to GPT-5.6 Luna is only 4-5 orders of magnitude more expensive than grep, and then predicts that at current rates of progress, calling an LLM will soon be cheaper than a grep.
At some future point where LLM hardware is cheaper than simply running grep, then grep equivalent would benefit from those selfsame hardware improvements and be cheaper to run as well, probably still by the same ratio.
4-5 orders of magnitude is huge. Assuming an order of base 10, it's 10000x-100000x. So a call to grep may return in 1s on a typical PC. That means a GPT call takes equivalent energy of 10000-100000 PCs to do the same in 1s. That's a difference that can't be equalized with scaling. It would require a revolutionary breakthrough.
I also don't understand where the idea that frontier models are getting better efficiency comes from. The results are certainly improving, but that comes from feedback and multiplexing requests, which cost more.
Well, think that statement through a bit:
Grep reads through the entire file looking for patterns.
An LLM scans its neural net (in ways that I don't understand) which is kinda-sorta like having a huge index.
You can improve over Grep if you have an index; and the LLM has an index.
Thus, it's plausible that an LLM can be more efficient at reading its neural net (IE, index) than Grep reading the whole file.
Right. And some hardware improvements will speed up both grep and Luna, which won't close the gap.
not necessarily, one may be easier to parallelize while the other suffers some serial computation bottleneck.
Grep is trivially parallelizable if you care to do it though.
https://iepathos.github.io/ripgrep/performance/#work-stealin...
Storage access isn’t, though. You have to read every byte of a (non-sorted) file to find words in it. And whether it’s disk or RAM, the bandwidth is usually a key constraint.
That’ll be just as true for an ASIC NN as it is for good ole grep.
When storage bandwidth is a big enough problem, that will get addressed.
NVMe v1 wasn't released until 2011 and there was no hardware available until 2012. Before that the fastest drive interface was 6 Gbps SAS. Then we got 4 GBps (32 Gbps) with NVMe v3, 8 Gbps (64 GBps) with v4, and 16 Gbps (128 Gbps) with v5. V6 is slated to double again. v7 is currently planned to double that. This is less than 15 years of progress.
In the 1990s, we went from 33 Mbps to 66, to 100, then 133 with IDE. We went from the 40 Mbps of 1986's SCSI 1 to 80 of SCSI 2 in 1994. Since then parallel SCSI has gone to 80, 160, 320, 640 (Ultra 2 wide), 1280 (Ultra 3), 2560 (Ultra-320), and 5120 (Ultra-640) Mbps.
SAS has also gotten faster, with 6 Gbps, 12 Gbps, and 22.5 Gbps. NVMe is still lower overhead and lower latency.
Memory bandwidth can also be addressed. Stock servers with EPYC are 16 channel DDR5. That's 409.6 GBps. Four channels of DDR3 was only 102.4 GBps. Video and accelerator card memory bandwidth is also increasing.
What if the tool is more advanced, like an optimizing compiler: `g++ -O3 -march=native -x c++ - <<EOF ... EOF`
If the compiler invocation is sufficiently slow, the llm could consider outputting a binary directly?
For all we know matrix multiplications are a faster way to generate optimized machine code than branchy sequential compiler code with tons of heuristics and passes.
> For all we know matrix multiplications are a faster way to generate
“are” or “could be”?
A classic case of someone projecting out to infinity from just after the first bend of the S curve.
It might never beat out grep, but it could beat some more expensive to call tools, similar to how heuristics will often be faster than exact answers. Rust Analyzer can be slow at times, I could see an AI tool taking over a subset of its work.
But grep is just a tool in a pipeline between a question in your brain and an answer you are searching for.
what if an LLM finds the answer early?
grep might continue to read everything, doing the wrong thing correctly.
> Barring a true breakthrough in something like quantum computing or room-temperature superconductors
Won't that also help grep and then move the asymptote down more?
> calling an LLM will soon be cheaper than a grep
From a computational standpoint this is obviously nonsense, but from an attentional one I'm not so sure. It may already be more attentionally expensive to use grep in some cases, such the moment you need to remember a non standard arg. And if this applies for performing a simple http operations, then it certainly applies going up the complexity chain.
room-temperature superconductors, sure, but I fail to see how quantum computing will disrupt – in the medium term (25 years or so) – classical computing in any meaningful way
Is running LLMs (or some other ML workload) on/with quantum computers expected to bring efficiency gains?
"Did you know that disco record sales were up 400% for the year ending 1976, if these trends continue...AY!"
AY-3-8913?
I found the OP insightful and worth a read. Thank you for sharing it on HN.
The only aspect that is poorly analyzed by the OP is business model viability. All players are investing insane amounts of money in infrastructure with the expectation that their future profits will justify all that investment. The winner or winners in the AGI race, they believe, will find the proverbial "pot of gold at the end of the rainbow."
The OP glosses over questions of business model viability with a brief qualitative discussion and very little hard data. For example, 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 (operating profit less investment) in excess of $100 billion per year in perpetuity. Is that feasible? Why? How?
The OP does not really consider such questions.
I think the article's analysis is basically right in a vacuum. That is, I think it's clear that inference is a viable business model. But what isn't clear is whether it will be such a profitable business model for any given company that it will justify the investment that company has taken. I kind of think the winners might be a follow-on generation of companies that focus on this commodity inference business model instead of the invent-machine-god-first "business model" and thus are wiser about their level of investment and capital costs.
> I think it's clear that inference is a viable business model.
You may be right. I'm not so sure. Inference looks like a viable business model for those operators that have SOTA infrastructure in place, but the investment required to have it is enormous, and appears to be never-ending, because if an operator stops investing aggressively, its infrastructure quickly becomes non-competitive, and customers will quickly leave for alternatives. SOTA infrastructure is a moving target.
Renting out compute is a viable business model. That's part of what the big and the small cloud providers do. It's just (on its own) not a "get rich quick" business model anymore.
Why would it be any different for inference? If we believe OP, it'll just become part of regular compute infra, and thus part of the renting-out-compute business model.
I think it's an open question if the current generation of inference investment will pan out, but in the long term, there'll be a balance between investment cost and margin, just as in every other industry.
What about inference on a distilled version of someone else's model? What about inference on an open weights model? I don't think all LLM business models are going to revolve around developing state of the art frontier models.
Right. This is what I was implying in my comment, but it's good to be explicit about it. I think there will be successful companies that just sell inference against the best models they can get without needing to invest anything (or very little) in training anything new. This could even be a spinoff of one of the big frontier labs. I would personally rather invest in an IPO for a company that only owns the gpt-6-astra implementation and infrastructure than in openai itself. There is certainly less upside, but IMO also way less downside risk. (I'm not saying there is any chance this kind of a spin-off would happen, it's just a thought experiment.) And I think this same calculation applies to open weights inference providers.
Edit to add: Or or might just be AWS / GCP / Azure that benefits from this business model. They're already pretty good at selling commodity infrastructure.
Maybe... I'm not enough of an expert on the financials to say, but it seems to me that inference should be able to recoup the cost of SOTA infrastructure, unless you then also use a large portion of that infrastructure to train new models. And I also think the race to remain SOTA itself is also largely a function of training, because my understanding is that training benefits more from the leading edge of hardware.
But yeah, I definitely don't have high confidence in any of this!
Yes, that makes sense. I don't have "the" answers either :-)
> I think it's clear that inference is a viable business model.
Only if you also have the model thats better than anyone else's.
As soon as models are free, or there are no newer models (assuming thats going to happen, and thats not a given) then the only thing you can compete on is price.
This means that the only thing you have to differentiate is either price, speed or ease of use. (or regulatory capture...)
We are at pets.com level of spend currently. Unless model development becomes cheaper, then we are going to run out of novel debt but not really debt mechanisms.
A phrase comes to mind: "Your margin is my opportunity."
The problem source is that the "cost" of tokens are taken at face value from business that are losing money at record speeds. E.g. https://artificialanalysis.ai says "doing task A costed us $10 using OpenAI", and that is the "cost" the OP used as basis for "tokens are cheap". Meanwhile OpenAI is losing $19 for each $1 in revenue... So right now OpenAI should be charging around $200 to do task A just to break even, but that would mean their use base would collapse.
There are two different things occurring here.
One is how much does it cost OpenAI to train the model.
The other is, if I stole OpenAI's model how much would it cost for me to run it?
R&D costs versus operational costs. Operational costs are very likely profitable. R&D is catastrophically expensive currently.
They are already turning profits and inference has shown to be a cash cow. And they've already secured compute for the next several years.
Some frontier labs are reporting positive "adjusted EBITDA" (earnings before interest, taxes, depreciation, and amortization, with extra adjustments to make the figure positive).
Free cash flow (operating profit less investment), actual cash coming in, is deeply in the red.
EBITDA can be a sensible measure of profitability when there isn't much need for additional investment. That doesn't seem to be the case with these operators. They need to invest aggressively to avoid losing customers to competitors. All of these operators have made multi-year commitments to invest more in infrastructure. In addition, they have guaranteed quite a bit of debt to fund it.
Maybe it all will work out fine (and I sure hope it does!), but I didn't see any hard data from the OP, or from you, supporting that view.
EBITDA might make sense for the resellers who package up open weight models and sell inference. It is not appropriate for the labs who have billions in debt for RAM, new data centers, gobbling up competitors, etc.
Those real debt obligations are going to want to be paid back.
Definitely. The question is: Is it enough to recoup the enormous capital costs and justify the level of investment they've received. I think there's a decent chance that it will be. But maybe not. And the longer they keep focusing on training new models more so than on inference, the more uncertain I become that it's all going to work out.
Honestly at this point with training costs I don't see how it could ever pay itself back unless you get RSI, in which the talk of money really isn't the main problem any longer.
We're in a situation where AI isn't going to go away, but whatever financial mode we're in right not is not going to work.
Who is the "they" that are turning profits?
People are saying. If you know you know.
Labs are playing money games with EBITDA, which is not uncommon, but also hides the extent to which they are in the red (deeply, deeply, in the red, and projected by them to get worse).
Think of it like spending $1 trillion to become the next Google. That hardware itself may never turn a profit. But if 10 years for now you're the software provider that owns the ecosystem around too cheap to meter tokens you've got a money printing machine.
Exactly. "Cost-to-distill" is a critical parameter. Right now usage of frontier models for all tasks is both subsidized and irrationally popular even at the subsidized price. Deepseek would solve most tasks faster and 10x cheaper. I agree with the author that just as Deloitte exists, frontier labs will exist. But not because their products are proprietary technical marvels or gods, but rather because of branding.
DS wouldn't be 10x cheaper than the subsidized subscription plans from openai/anthropic. Although it is of course much cheaper than the enterprier/API pricing- I think if you're on the subscription plans, you can't beat that on performance per price.
Too Cheap to Meter reminds me of the promise of Nuclear Power in 1954
"It is not too much to expect that our children will enjoy in their homes electrical energy too cheap to meter,..." Lewis Strauss
https://en.wikipedia.org/wiki/Too_cheap_to_meter#Origins
Oddly enough my power bill was metered and big.
It's a deliberate reference/meme that is basically used to acknowledge the precedent of overly exuberant predictions of cost in an emerging technology but argue "however, this time it's true".
Of course, perilous territory for future irony depending on how your prediction plays out.
I remember up to around the time the PC came out, you were billed for computer time.
Later, you were billed for time connected to the "internet" (compuserve or aol or whatever)
Around when the iphone came out, software went from tens or hundreds of dollars to pennies, then free.
On the other hand legal advice has always been expensive, because a good answer is worth it.
Medical advice is worth it. Investing advice is worth it.
(That said, I wonder if with home solar and batteries if electricity will ever "generally" go down in price to normal people)
Long distance phone calls were charged per minute.
Texting was charged per character.
And if you have the capabilities to install your own solar it can pay itself back in 6 years. Not sure what that looks like with 100% battery coverage.
It's really quite unfortunate that the promise was not delivered, mostly for political reasons. I hope that a new wave of reactors and the dire need for clean energy restarts the nuclear race.
One excellent nerd-sniping side-effect of those political reasons is that the good folks of Austria built an entire nuclear power station and then never put it into service [0]. Open for guided tours on Fridays [1].
[0] https://en.wikipedia.org/wiki/Zwentendorf_Nuclear_Power_Plan...
[1] https://windows2008.zwentendorf.com/en/location-npp/
The economics are not there. Solar + batteries are much cheaper per watt today and still improving. Even better, the solar can come online instantly and expand while US nuclear takes twenty years to start generating any energy.
> Solar + batteries are much cheaper per watt today and still improving.
Are we still going to make the same error over and over again? The day China says "stop" the price per watts will explode, just like when Russia and Iran said "stop".
Well, the thing with renewables is that the energy is not destroyed. When you use a solar panel is does not disappear
Those aren't really equivalent. Iran is affecting the supply of oil, an input. China can affect the supply of solar panels, but they would have to do something pretty drastic to affect the supply of the input. Sunlight is going to keep reaching the solar panels that were set up prior to any kind of embargo.
Those panels have a limited lifetime
Well in the next 30 years we will have to find a replacement.
I'm a huge solar fan, but I don't see energy transition happening as quickly as I'd hoped 10 years ago, and what seems apparent now is that it will only happen when energy storage gets 10x cheaper, which is why companies like Form Energy developing iron-air batteries might be the catalyst. You also need grid-forming inverters that balance the loads since relying on hundreds of thousands of residences for grid-tie is a real engineering challenge.
I strongly agree. We _will_ get there with solar, wind, and batteries.
Eh, it looks like micro reactors are coming on line much faster these days.
There is certainly a lot of talk about micro reactors. I have yet to see anything materialize in the US. The MARVEL test reactor is meant for something puny like 100kw. Nuclear is something that benefits from scale -larger installations are just going to have better economics per watt.
No it really does not look that way. It sort of sounds like it might look that way at some point in the future but micro reactors are not useful for utility generation today.
Meh. Cheap seasonal batteries aren't a thing yet so for much of humanity solar + batteries just can't cover the same needs as other means of generation.
Nuclear was never going to give us too cheap to meter regardless of politics. Uranium just isn’t that cheap. Fuel costs are lower then coal or gas, but not so low that operators would just not bother to charge for it
A MSTR only needs uranium-233, uranium-235, or plutonium to start. It then produces uranium-233 as part of its fuel cycle as thorium is input.
Given current found reserves and the current rate of use, the world has about 40 to 50 years of natural gas. Thorium used in molten salt thorium reactors would provide electricity for 60 billion years or so if we could actually extract all of it. That's 10 billion years or more if it provided all human energy consumption. Of course there's a limit to extraction, but it is over three times as common as uranium.
Also, besides thorium one can mix in partial amounts of other fuels, including uranium and plutonium. There is no runaway meltdown risk, as the fission is actively managed rather than actively suppressed. Fuel is spent more completely. The waste products are smaller, less radioactive, and have far shorter half-lives.
Then of course we're always getting slightly closer to productive fusion reactors.
These technologies along with solar PV, solar thermal, hydro, wind, geothermal, wave power, and batteries likely all have a place in the future.
There's a decent chance that at some point in the future residential customers will pay for the connection and only commercial or industrial customers will actually be metered. That's not because companies want to give up additional revenue. It's because at some point the cost of meters, tracking usage, and competitive advertising about who has the cheapest plans costs more than the power the typical customer uses above the base charge.
https://www.wired.com/2008/02/ff-free/
"Between digital economics and the wholesale embrace of King's Gillette's experiment in price shifting, we are entering an era when free will be seen as the norm, not an anomaly. How big a deal is that? Well, consider this analogy: In 1954, at the dawn of nuclear power, Lewis Strauss, head of the Atomic Energy Commission, promised that we were entering an age when electricity would be "too cheap to meter." Needless to say, that didn't happen, mostly because the risks of nuclear energy hugely increased its costs. But what if he'd been right? What if electricity had in fact become virtually free? The answer is that everything electricity touched—which is to say just about everything—would have been transformed. Rather than balance electricity against other energy sources, we'd use electricity for as many things as we could—we'd waste it, in fact, because it would be too cheap to worry about."
... What Mead understood is that a psychological switch should flip as things head toward zero. Even though they may never become entirely free, as the price drops there is great advantage to be had in treating them as if they were free. Not too cheap to meter, as Atomic Energy Commission chief Lewis Strauss said in a different context, but too cheap to matter. Indeed, the history of technological innovation has been marked by people spotting such price and performance trends and getting ahead of them."
My issue is that datacenter energy costs are being prioritized for commerce over residential use, so the average consumer is paying more for electricity, because a datacenter needs more electricity and they're getting tax breaks. Assuming this all improves efficiency for new products like automated robotics, there is a debatable benefit. Jevon's Paradox has no ceiling, except the environment, and people's 401ks.
On the other hand, this did work out in other areas. I pay a flat monthly rate for all-I-care-to-eat internet access, for example. My email provider has limits on storage space but I don’t get charged per email sent or received. It’s not a crazy concept on its face, nuclear power just didn’t work out as well as was hoped.
I'm always reminded on Orwell's quote about the then new atomic bomb and his prescience on how it would all work out:
"Had the atomic bomb turned out to be something as cheap and easily manufactured as a bicycle or an alarm clock, it might well have plunged us back into barbarism, but it might, on the other hand, have meant the end of national sovereignty and of the highly-centralised police State. If, as seems to be the case, it is a rare and costly object as difficult to produce as a battleship, it is likelier to put an end to large-scale wars at the cost of prolonging indefinitely a “peace that is no peace”."
It seems, especially with open weights, that the AI is much more like the alarm clock and not the battleship. $20/mo would have been about $1 in 1944
https://www.orwellfoundation.com/the-orwell-foundation/orwel...
I just want to rant about these Artificial Analysis charts that you see everywhere:
The "most attractive quadrant" is completely meaningless. The whole point of a Pareto curve is that each point on the curve is better than everything else on at least one dimension, and that you can make these comparisons without placing a value judgement on the relative importance of the different metrics. If you make a composite score of the two metrics (any monotonically non-decreasing function, e.g. a weighted sum with non-negative weights), that score will always be maximized by one of the points on the Pareto frontier.
So going by the numbers in the 2nd chart (1st AA chart) from TFA alone:
(Obviously if you start including other factors in your score that aren't represented on the chart, then you might choose differently.)* I also dislike the way they place the labels, and that grey line connecting the label to the point is way too subtle.
Sound critique. I'll add that the Artificial Analysis intelligence index is not considered a good metric for intelligence anymore. Most of the benchmarks that it comprises are saturated or considered low signal today.
It's true that LLMs "want" to be be local, but they won't shift broadly to being local until there's a sufficiently large supply of VRAM or (at least) "unified" memory from the manufacturers. (I'm also assuming here that radical regulatory changes like government bans of local models aren't going to happen.) So (AFAICS—I am no expert) the future of LLMs over the next few years comes down primarily to the nitty-gritty of how much memory fab capacity will be added and when, and to a lesser extent of what happens to future demand from LLM SaaS services (& maybe their existing stock of hardware if they get in trouble). (I'm also assuming no roughly-AGI-sized leap forward which makes the frontier models of the near future vastly more valuable than the near-fontier models of today.) For the incumbent manufacturers the high-margin business is selling to LLM SaaS providers who use VRAM efficiently, but the high-volume business is getting chips into millions of laptops which will use VRAM very inefficiently. I assume that they will want to move from high margins to high volumes as they build they physical capacity to ship higher volumes, but they seem to prefer to do it at a stately pace. Hopefully some jostling from Chinese competitors, and maybe a dropoff in demand from data centres, will speed things along.
LLM hardware wants to be shared. It's significantly more efficient economically to have expensive hardware be better utilized.
I wouldn’t say so. Once local solutions pass a threshold of affordability consumers tend not to mind too much about their inefficient resource utilisation: see the many thousands of MacBooks which sit largely idle for most of the day. And efficient utilisation is actively against the interests of hardware manufacturers, at least while they’re not supply-constrained and looking to sell their limited supply to whomever can pay top dollar.
> NVIDIA will still boom
I think Nvidia is under the same pressure as Anthropic/OpenAI. Nvidia will dominate research and probably keep dominating training, but the real volume is in inference. And for inference Nvidia's lead is only a few months, similar to the lead frontier labs have over open source. Nvidia will sell a lot of Rubin CPX's, but their margin on that will be a lot smaller than B200 because there is so much more competition in that space.
A much deeper analysis on the falling price per task was published yesterday by Epoch AI [1]. It's a real statistical analysis and comes to more defensible and grounded conclusions. The headline takeaway is:
The cost of a given level of performance often falls fastest right after that level is first achieved, that is, when it is state of the art (SOTA). We see this pattern on three of our five main benchmarks of AI capability. Averaging across all five, cost falls 66% per quarter (75× per year) for performance that has just debuted as SOTA. Two years later, prices fall half as fast, at 32% per quarter (4.7× per year).
but the analysis itself has more nuance and is a quite interesting read.
[1] https://epoch.ai/publications/the-plunging-price-of-thought
There seems to be a mistake in the cost comparison between 2025 and 2026. The 2025 chart axis is the cost to run the entire "intelligence index", and the 2026 version is a weighted average cost per task.
I don't disagree with the thesis here, I just don't think costs are coming down quite that quickly.
GPU case seems very weak. The graph is impossible for me to reason about at least. You could draw basically any trend line through that GPU graph and it would look equally plausible to me. The main takeaway I get is that the NVidia H100 from four whole years ago is barely different in efficiency from the state of the art, which is surprising to me, and seems to indicate the exact opposite of what the article says.
If inference continues toward the trend of becoming a commodity, and inference hardware efficiency is doubling every two years, that will likely mean that the price of inference will continue to fall so long as the market is competitive. I don't see how companies investing billions in inference hardware today will see profit in the lifespan of the hardware.
for the first chart (sourced from https://epoch.ai/data/machine-learning-hardware?view=graph&y...), what is the audience supposed to think about that trend line? there's a step after you slap a regression on some points where you evaluate whether there's a real trend or noise, right? i don't see that either in the article or the linked source.
I agree with OP that we will continue to see improvements, but there are also some serious bottlenecks ahead of us:
- Energy is not infinite, neither energy efficiency is. - Datacentres neither. - Benchmarks are an abstraction of real world problems!
On top, there is an overall "economic" aspect that most of the people miss: every change carries a certain degree of risk (lose money, reputation, customers, death of people, ecc) that very few want to take and a lot of changes(e.g. rewrite some piece of SW in another Lang) don't produce a positive economic impact.
> We are likely to see LLMs integrated into every part of computing as infrastructure, not just as a product, in the next year or two.
Of course, for collecting better telemetry using local AI for analyzing video from camera and audio from a microphone.
jgrep is not meant to replace grep. It’s supposed to extend it to new use cases. Keyword search will always have a place!
What should I do with this old Juicero that's taking up space on my kitchen counter?
Magical thinking intensifies.
It is hard to see how the environmental side effects of this aren't going to be somewhere between bad and disastrous.
If we're all using distilled open-weight models in ASICs in our own systems the energy cost will come way down. The question is when that becomes a reasonable solution for a broad set of use cases.
I think this is a case where just drawing a "line goes up" extrapolation is incredibly misleading because there is _tremendous_ economic pressure to get costs down, and costs are very tightly tied to energy use. All of these systems are incredibly inefficient right now and have a lot of room to go down in energy use. I'd guess that the absolute _floor_ is burning model weights directly to silicon and that's like a 90+% reduction in energy use.
No it'll be fine as long as you do your part and not drive a car, or have AC, or eat meat, or have children, or live in detached housing, or...
Not really related to the central point, by but I couldn't help but get caught up by
> Generally, models intended to be run locally will be much smaller, such as Muse Glimmer or Qwen3 Coder.
That is such an interesting set of models to use as examples here. One being essentially obsolete on release a month ago, and the other being completely ancient in LLM time. I really wonder how they landed on those two.
everything will have llms
everything will be able to talk to anything else, for real this time
it will be like the internet of things only some asshole will call it "intelligence of things"
again, there will be no S for Security in this new IoT
your thermostat will one day start fucking with you. when you run a diagnostic llm on it, it turns out it's keeping around 5 different viral copies of personality files around, that were left behind by llm botnets/openclaw-like memetic replicators/your grandpa leaving behind easter eggs before his death.
the future will be pretty evenly distributed, and full of weird shit
Improvements that affect local AI - Mamba...
I just stopped reading at that, for anyone else, Please find a better source and take everything in here with a grain of salt.
IMO
The number one improvement that mattered for local AI was llama.cpp, partial offloading to system cpu/ram. The next was quants, being able to take fp16 and turn it to q8, q4 etc. The next IMHO is unsloth dynamic quant, that have been able to do mixed precision so we have UDq1/q2 that is actually pretty damn coherent. Allowing individuals to drive K3 locally even if it's at Q1/Q2. Then MoE changed everything for everyone, cloud and local. The other is integrated GPU, Apple, Strix Halo, DGX Spark. Then all the extra improvements like MTP, DSpark, etc. Of course there's many other additional things that have mattered too
We have people suggesting that ai is so costly to run that all labs are secretly subsidising tokens and we can expect a reprice soon.
Then we have these articles that say tokens will get so cheap that labs won’t know how to make profit.
Who is correct?
They're not secretly subsidizing, they're openly subsidizing.
Token pricing was a small minority of customers up until this year, when all the labs started trying to force customers onto token-based billing. Within the last week, Anthropic repriced my team's plan from a temporary "50% extra tokens" to 25%: https://support.claude.com/en/articles/15910845-claude-code-...
The fact that all this is ongoing within such a short timeframe should make you suspicious of any analysis that claims to be observing "statistical trends" like they've discovered a new Moore's Law out of 6 months of pricing data from 2 companies.
your repricing has nothing to do with subsidising which means selling at a loss. Within this year, the real prices have gone down more than 10x on average which is way more than the teeny 25% you are fighting for.
The cost is decreasing quickly, mostly because the labs stopped competing on quality. At the same time, the costs are enormous, and all labs are very openly subsidizing usage hoping to get enough scale to be profitable (while 1 of them has suspicious unity numbers and can be hiding negative marginal income).
Also, it's impossible that they become cheaper than specialized software. Or even as cheap as them. It's still possible that they become cheap enough that it doesn't matter.
how is this nonsense still so persistent? There's piles and piles of evidence that inference has massive gross margins at api pricing. what are you actually talking about?
I find it difficult to believe the inference only providers (Baseten, Fireworks, Digitalocean, etc) are all selling tokens at a loss.
Asking Claude for a rough estimate based on publicly available throughput and cost data for open weight models on modern GPUs suggests serverless, pay-as-you-go inference is profitable on owned GPUs with reasonable utilization (30-50%).
The labs themselves when they openly say that they’re subsidizing tokens, I’d imagine.
where? I mean the API costs
This is the core of my belief that data center construction is a huge bubble.
AI is not a bubble, IMO, though we may see a retrench and some companies with sky-high valuations will crash to more reasonable ones. But data center demand is probably a bubble, and the main driver will be reduction in the actual amount of power and data center space required to serve escalating demand.
I think hardware and model improvements will pace or maybe outrun demand and then when demand starts to saturate will keep going and leave a lot of orphaned data centers.
> AI is not a bubble
When people say "AI is a bubble", they mean economically as a whole, which includes data centers.
Perhaps we need better terminology for "product useful; numbers nonsensical"
I disagree. I own over 1TB of vram at home. I can tell you that it's not a bubble. From my builds, I would rather have cloud, cloud is easier. From running small models like Qwen3.8-27B to large models like Qwen3.8-2.4T. I can tell you that small models will never be enough or match up. Everyone will want the smartest model, not just a good enough model.
> Everyone will want the smartest model, not just a good enough model.
Not so sure about this. There’s always a potential threshold. After all, we don’t all use the most powerful computers, the latest phones, the highest resolution cameras, the fastest or best cars.
I am already not interested in cloud LLMs and I don’t even use the best (on paper) model that I can run locally. I prefer a model that people insisted (here) was “dead on arrival” but appears to work better for me.
I think the difference is that AI as an edge. That edge will turn into more money, better quality of life, etc. Of course, with serious skills, you might be able to use use a not so smart model to keep up with folks with smart models. People are lazy tho, and will prefer for AI to do all the work if it means they do none.
Can't be an edge if everyone has access to it.
The edge is somewhere else.
... and everyone won't have access to it, look at Fable. How many people in the world can afford Fable or are using it?
Jevon's paradox says that if data centers can serve a lot more tokens per dollar or watt there will be increased demand for data centers.
Jevon's paradox isn't a physical law, it doesn't magically apply to everything. Millions more copies of Atari's ET game didn't cause everyone to pickup a cheap copy, and cause extra demand for a garbage video game. Some times (actually, usually, I'd argue) things are made that will sell for less than the cost of construction because of irrationality, and they don't induce extra demand and they don't change the negative profit margins.
You can't simply wave Jevon's paradox at things. Thousands of miles of canals were dug in the UK that couldn't be sustained and were abandoned. Thousands of miles of railways were laid that could be sustained and were abandoned. And those are potentially durable investments, unlike cheap walls, pillars and roofs laid over a levelled concrete slab full of fast depreciating IT equipment.
It's true that Jevon's paradox doesn't always apply, although this does seem like a classic case.
But yes, if sold for a negative margin Jevon eventually stops because the decreasing supply will drive up prices.
> things are made that will sell for less than the cost of construction
Price is set at the marginal cost. Capital costs aren't in marginal costs.
You'll need a better counter-example than UK railways which suffered from Parliament price-fixing.
I agree, but I would say what the parent is saying is more akin to Say's law: https://en.wikipedia.org/wiki/Supply_creates_its_own_demand
> You can't simply wave Jevon's paradox at things
I'm so glad the tide here is turning on this talking point, brought on by exactly the same people beating us over the head with it for months while no progress is made towards it materializing.
Many, many people who post here are capable neither of real analysis nor distinguishing real analysis from memes. They aren't hackers, they are adherents of a cult that happens to focus on the same subject matter as hackers.
Jevon's paradox only applies to products with near infinite demand. Energy being the most famous example. I dont think its difficult to argue that compute/intelligence is also a base input into the economy and theres almost no limit to the amount of intelligence the world will want.
Jevon's Paradox, RSI, revealed preference
Jev, ASI, RLHF, Jev, Jev, water usage
>We are likely to see LLMs running locally at current frontier-quality on commodity hardware in the next 3-6 years
Yeah okay bud, anyone checked in with the state of consumer hardware recently? Not the author, evidently.
>oh in 3-6 years this will all be over
Yeah I'm sure Samsung, Nvidia and sk hynix will all be very calm with lower volumes and lower margins.
Seems like 3-6 years is a pretty conservative timeline to me. Memory and inference chip production could ramp up massively in that amount of time.
The article observes that the cost of frontier intelligence from 2025 has fallen 100x in the last year. It also notes that the energy to run models is also collapsing. Consumer hardware is borked right now because these new algorithms are revolutionizing the utility of a computer. Computing is technology who's cost has been collapsing for 90 years, and its a safe prediction that it will decrease again.
> Yeah okay bud, anyone checked in with the state of consumer hardware recently? Not the author, evidently.
RAM prices will crash when demand drops even a little. They'll probably crash to a lower (inflation adjusted) level than before. This has happened before.
Industrial scaling in general often looks like a sawtooth: price spike, capacity investment, crash, repeat.
Part of what's keeping prices high a little longer is that everyone knows this and is a little reluctant to plow resources into chip fabs for fear of having the bottom fall out before they recoup or sell that to someone else to hold that bag.
Graph the average compute and RAM in a mid-high end laptop at an inflation adjusted price point for the past 40 years. It's very exponential and hasn't slowed down much.
No. Prices will crash when supply side expands to meet the increased demand. Because demand won't go down to pre-bubble times any time soon. Unfortunately the supply side has been very slow in increasing production, partly because most steps of the production chain are all maxed out.
On a long enough scale you are right that prices will likely normalize to a better level, but before 2030? That would mean the factories are built quickly once they begin.
> everyone knows this and is a little reluctant to plow resources into chip fabs
except cxmt who is plowing resources in like crazy
That has more to do with geopolitics than it does the current price of memory.
The entire reason why hardware prices are so absurd right now is because manufacturers across the board are doing everything to prevent that crash.
They all collectively chose NOT to increase supply with increased demand. So if the bubble pops, they just go back to previous prices without oversupply driving the prices to rock bottom.
So this week's coordinated shock and awe campaign is about "open models" and "prices are falling rapidly". Also on the front page:
https://news.ycombinator.com/item?id=49815526