We just had a realization during a demo call the other day:
The companies that are entirely AI-dependent may need to raise prices dramatically as AI prices go up. Not being dependent on LLMs for your fundamental product’s value will be a major advantage, at least in pricing.
Yup. Also regardless of price they need to spend more and more as the project collapses under the inevitable incidental complexity of 30k lines of code a day.
It's similar to how if you know what you're doing you can manage a simple VPS and scale a lot more cost effectively than something like vercel.
In a saturated market margins are everything. You can't necessarily afford to be giving all your margins to anthropic and vercel.
> The companies that are entirely AI-dependent may need to raise prices dramatically as AI prices go up.
It's not that clear. Sure, hardware prices are going up due to the extremely tight supply, but AI models are also improving quickly to the point where a cheap mid-level model today does what the frontier model did a year ago. For the very largest models, I think the latter effect dominates quite easily.
There's only so far engineers can optimise the underlying transformer technique, which is and always has been doing all the heavy lifting in the recent ai boom. It's going to take another genius to move this forward. We might see improvements here and there but the magnitudes of the data and vram requirements I don't think will change significantly
State space models are already being combined with transformers to form new hybrid models. The state-space part of the architecture is weaker in retrieving information from context (can't find a needle in the haystack as context gets longer, the details effectively get compressed away as everything has to fit in a fixed size) but computationally it's quite strong, O(N) not O(N^2).
What's weird though is the bifurcation in pricing in the market: aka if your app can function on a non-frontier level AI you can use last years model at a fraction of the cost.
And I don't really mean new businesses that are entirely built around LLMs, rather existing ones that pivoted to be LLM-dependent – yet still have non-LLM-dependent competitors.
same as Uber… in the beginning everyone pretty much new that the cost of rides cannot possibly be that cheap and that it is subsudized. once you corner the market etc people just got used to “real” prices to the poibt that now there are often cheaper alternatives than Uber but people still Uber…
Its also quite interesting to read about Uber exploits their drivers and discriminating algorithms. Cory Doctorow mentioned it in his latest book, sadly cant link the direct sources.
Not really, the next move is to establish standards groups requiring the use of AI in product development. A mix of industry and governmental mandates. What you view are viewing as COGS instead becomes instead a barrier to entry.
How is that surprising? We've been taking that into account for any LLM related tooling for over a year now that we either can drop it, or have it designed in a way that we can switch to a selfhosted model when throwing money at hardware would pay for itself quickly.
It's just another instance of cloud dependency, and people should've learned something from that over the last two decades.
Not so much that it was surprising, rather that we looked at a competitor’s site and noticed that a) their prices went way up and b) their branding changed to be heavily AI-first.
So we thought, hmm, “wonder if they are increasing prices to deal with AI costs,” and then projected that into a future where costs go up.
We don’t have this dependence ourselves, so this seems to be a competitive advantage for us on pricing.
Absolutely. Pricing exposure is the quiet story under all the waves of AI hype. Build for convenience → subsidise for dependence → meter for margin is a well-worn playbook, and AI-dependent companies are about to find out what phase three feels like.
Hyperscalers are spending a fortune so we think AI = API, but renting intelligence is a business model, not a technical inevitability.
Indeed, it was clear from the beginning, "AI" companies want to become infrastructure and a critical dependency for businesses, so they can capture the market and charge whatever they want. They will have all the capital and data needed to eventually swallow those businesses too, or more likely sell it to anyone who wants the competitive advantage.
in fact I am betting opposite. frontier models are getting not THAT much better anymore at all, for common business needs at least. but the OSS models keep closing the gap. which means if trajectories hold there will be a near future moment probably where the big provider costs suddenly drop shaerply once the first viable local models consistently can take over tasks normally on reasonable hardware. Right now probably frontier providers rush for as much money as they possible can before LLMs become a true commodity for the 80% usecases outside of deep expert areas they will have an edge over as specialist juggernauts (iE a cybersecurity premium model).
So its all a house of cards now, and the moment the bubble bursts is when local open inference has closed the gap. looks like chinese and smaller players already go hard into this direction.
Local open inference can address hardware scarcity by repurposing the existing hardware that users need anyway for their other purposes. But since that hardware is a lot weaker than a proper datacenter setup, it will mostly be useful for running non-time-critical inference as a batch task.
Many users will also seek to go local as insurance against rug pulls from the proprietary models side (We're not quite sure if the third-party inference market will grow enough to provide robust competition), but ultimately if you want to make good utilization of your hardware as a single user you'll also be pushed towards mostly running long batch tasks, not realtime chat (except tiny models) or human-assisted coding.
one graph, One graph and the author is pinning an entire theory on it?
Infra is always limited, even at hyper scalers. This leads to a bunch of tools dfofr caching, profiling and generally getting performance up, not to mention binpacking and all sorts of other "obvious" things.
As a recent example in AI space itself. China had scarce GPU resources, quite obvious why => DeepSeek training team had to invent some wheels and jump through some hoops => some of those methods have since become 'industry standard' and adopted by western labs who are now jumping through the same hoops despite enjoying massive computeresources, for the sake of added efficiency.
China already operates like this. Low cost specialized models are the name of the game. Cheaper to train, easy to deploy.
The US has a problem of too much money leading to wasteful spending.
If we go back to the 80s/90s, remember OS/2 vs Windows. OS/2 had more resources, more money behind it, more developers, and they built a bigger system that took more resources to run.
Mac vs Lisa. Mac team had constraints, Lisa team didn't.
Though I do agree with you, I just came back from a trip to China (Shanghai more specifically) and while attending a couple AI events, the overwhelming majority of people there were using VPNs to access Claude code and codex :-/
Harness is a big one, Claude Code still has trouble editing files with tabs. I wonder how many tokens per day are wasted on Claude attempting multiple times to edit a file.
It’s the tool that calls the model, give it access to the local file system, calls the actual tools and commands for the model, etc, and provide the initial system prompt.
Basically a clever wrapper around the Anthropic / OpenAI / whatever provider api or local inference calls.
I'm having an hard time getting my mind to see this.
> Users should re-tune their prompts and harnesses accordingly.
I read this in the press release and my mind thought it meant test harness. Then there was a blog post about long running harnesses with a section about testing which lead me to a little more confusion.
Yes, the word 'harness' is consistently used in the context as a wrapper around the LLM model not as 'test harness'.
This field is chock full of people using terms incorrectly, defining new words for things that already had well known names, overloading terms already in use. E.g. shard vs partition. TUI which already meant "telephony user interface ". "Client" to mean "server" in blockchain.
pi vs. claude code vs. codex
These are all agent harnesses which run a model (in pi's case, any model) with a system prompt and their own default set of tools.
Why is written with an assumption that we have finite hardware production capacity? Industrial processes can scale up, new factories can come online… it will take a while but the whole point of economics is that supply will scale to meet demand. The shortage is a temporary, point-in-time metric.
And that’s not considering the software innovation that can happen in the meantime.
The economic hypothesis that has dominated the past hundred years is that economic growth is infinite because resources are infinite and (almost) free. We all know this is unrealistic and disconnected from our human condition.
Regarding "innovation", I agree with your idea. I even think that the major innovation will be to transpose models locally, using reduced infrastructures that will still be sufficient for the majority of use cases.
... and I have this little idea in the back of my mind: when companies can no longer keep up with demand and people have (albeit more limited and reduced) local capacity, minds will start focusing on techniques (more humble and modest ones) to keep part of the system running locally, without dependency.
I know it may sound ridiculous, but it could actually become a way to break away from the business models that have been developed over the past few decades. Broadly speaking, this even amounts to saying that the biggest victims of AI could be the companies that bet on AI as a service.
Yet I know my vision is way too idealistic but I'm coming to imagine that a human brain, although less efficient in the long run, remains a reliable way to control the resulting costs and could even turn out to be more advantageous and more readily available than its silicon-based counterpart.
The human brain is incredibly efficient (Approximately 20W of energy consumption¹). These AI systems use many orders of magnitude more energy than human equivalents.
Whoever running and selling their own models with inference is invested into the last dime available in the market.
Those valuations are already ridiculously high be it Anthropic or OpenAI to the tune of couple of trillion dollars easily if combind.
All that investment is seeking return. Correct me if I'm wrong.
Developers and software companies are the only serious users because they (mostly) review output of these models out of both culture and necessity.
Anywhere else? Other fields? There these models aren't any useful or as useful while revenue from software companies by no means going to bring returns to the trillion dollar valuations. Correct me if I'm wrong.
To make the matter worst, there's a hole in the bucket in form of open weight models. When squeezed further, software companies would either deploy open weight models or would resort to writing code by hand because that's a very skilled and hardworking tribe they've been doing this all their lives, whole careers are built on that. Correct me if I'm wrong.
Eventually - ROI might not be what VCs expect and constant losses might lead to bankruptcies and all that build out of data centers all of sudden would be looking for someone to rent that compute capacity result of which would be dime a dozen open weight model providers with generous usage tiers to capitalize on that available compute capacity owners of which have gone bankrupt and can't use it any more wanting to liquidate it as much as possible to recoup as much investment as possible.
No matter how low and reasonably Anthropic is valued, don't think $200 Max plans are going to recoup the investment + some return on top because size of the software industry is not that huge and profit margins for AI inference aren't very high either.
> because size of the software industry is not that huge
I onboarded marketing on a premium team Claude seat yesterday. And one of our sales vibecoded an internal tool in the last three weeks using Claude Code that they now use every day. I wouldn’t have imagined it a month ago. We still had to take care of deployment for him, but things are moving fast.
Seems like everybody an their mothers are using max plans these days. I wouldn't be surprised if LTV of each customer was big enough to justify spending.
Assuming there are 10 million developers and everyone is at $200 max plan, that would be $2 billion/month or $24 billion/year maximum.
Note - this is just the revenue not the profit. No salaries, no compute paid for. Just plain revenue. Profit would be way less.
But even that - if we take it to $24 billion/year and we take a 10x multiple, the company is barely valued at $240 billon dollar, lets be generous and make it double at $480 billion and then round it up to $500 billion for a nice round number.
Far far from the $800 billion valuation Anthropic is looking at.
Companies are spending far more than $200/month/developer. The $200 Max plan is a great value but you hit limits far too soon, and it also doesn't cover any of the other styles of integrations and tools that you can build and use to help your developers, like code review suggestions, which at the very least would come from additional Max plans, and not from the individual developers' plans.
It feels like a repeat of the dot com infrastructure buildup that spurred the whole 2005 explosion in affordable hosting and new companies. This will probably leave us massive access to affordable compute in a couple of years.
AKA, the beginning of big companies being able to roll over small companies with moar money
(note: I don't expect this to actually happen until the AI gets good enough to either nearly entirely replace humans or solve cooperation, but the long term trend of scarce AI will go towards that direction)
China is installing something like 500 GW of wind and solar per year now. Even if they're only able to build and otherwise access chips that have half the SoTA performance per watt, they will win.
A dollar is an entirely fictional unit and trillions of it can be manufactured at no cost, while watts are constrained by the laws of physics, photons/electrons, supply chain of electricity and all that fun stuff in the real world.
I wouldn’t agree. Even at national scale, these projects cost resources. And the resources of all agents (org, countries) are constrained.
While we could reason in "performance / watt" and "performance / people", "performance / whatever other resource involved", and "performance / opportunity cost of allocating these resources to this use case and not another", "performance / whatever unit of stable-ish currency" is a convenient and often "good enough" approximation that somewhat encapsulates them all.
A simplification, like any model, but still useful.
US energy is constrained by the utility monopolies/oligopolies which have to extract more rents, specifically by increasing costs. Their profit is a percentage of cost, these perverse incentives + oligopolies will make it increasingly expensive to make anything (including AI) in US.
The dynamics vastly favor China, part of the reason the US sprinting towards "ASI" isn't totally boneheaded is that the US and its industry needs a hail mary play to "win" the game, if they play it safe they lose for sure.
I'd be fine with a world without AI, honestly. Nobody really wins this race except the very wealthy. And I don't think it's really going to play out the way the wealthy think it will. It's more like a dog catching a car than it is a race.
"The dog that caught the car" refers to how dogs sometimes chase cars. Suppose the car stops and the dog catches up - what is it going to do? It has no plan, it has no purpose, it isn't going to bite the car, it isn't going to get anything out of catching the car. The car may even run it over. I intended it basically as "play stupid games, win stupid prizes", or "be careful what you wish for".
My observation is that the dog sniffs all the tires, picks one tire, lifts one leg and does the deed. I don't know if its a way of marking territory or domination. We need a dogatologist to explain what it means.
Initially I thought "Well... good for AI companies because they can then charge more" but IMHO that's a very tricky position because it means the cheap wave is behind us.
It's one thing to "sell" free or symbolically cheap stuff, it's another to have an actual client who will do the math and compare expenditure vs actually delivered value.
> and compare expenditure vs actually delivered value
Which means that the hype production will be driven up another few notches to make people doubt their rational findings and keep them in irrational territory just a tad longer. Every minute converts to dollars spent on tokens.
It seems very possible that we have at least five years of real limitations on compute coming up. Maybe ten, depending on ASML. I wonder what an overshoot looks like. I also wonder if there might be room for new entrants in a compute-scarce environment.
For instance, at some point, could Coreweave field a frontier team as it holds back 10% of its allocations over time? Pretty unusual situation.
Well it's in the books. O(n^2) algorithms are bad in the long run, transformers algorithm has such complexity, so not a big surprise we hit the limits.
Open Weight models are 6 months to a year behind SOTA. If you were building a company a year ago based on what AI could do then, you can build a company today with models that run locally on a user's computer. Yes that may mean requiring your customers to buy Macbooks or desktops with Nvidia GPUs, but if your product actually improves productivity by any reasonable amount, that purchase cost is quickly made up for.
I'll argue that for anything short of full computer control or writing code, the latest Qwen model will do fine. Heck you can get a customer service voice chat bot running in 8GB of VRAM + a couple gigs more for the ASR and TTS engine, and it'll be more powerful than the hundreds of millions spent on chat bots that were powered by GPT 4.x.
This is like arguing the age of personal computing was over because there weren't enough mainframes for people to telnet into.
It misses the point. Yes deployment and management of personal PCs was a lot harder than dumb terminal + mainframe, but the future was obvious.
I've seen this claimed, but I'm not sure it's been true for my use cases? I should try a more involved analysis but so far open models seem much less even in their skills. I think this makes sense if a lot of them are built based on distillations of larger models. It seems likely that with task specific fine tuning this is true?
> I've seen this claimed, but I'm not sure it's been true for my use cases?
I'd be surprised if it isn't true for your use cases. If you give GLM-5.1 and Optus 4.6 the same coding task, they will both produce code that passes all the tests. In both cases the code will be crap, as no model I've seen produces good code. GLM-5.1 is actually slightly better at following instructions exactly than Optus 4.6 (but maybe not 4.7 - as that's an area they addressed).
I've asked GLM-5.1 and Opus 4.6 to find a bug caused by a subtle race condition (the race condition leads to a number being 15172580 instead of 15172579 after about 3 months of CPU time). Both found it, in a similar amount of time. Several senior engineers had stared at the code for literally days and didn't find it.
There is no doubt the models do vary in performance at various tasks, but we are talking the difference between Ferrari vs Mercedes in F1. While the differences are undeniable, this isn't the F1. Things take a year to change there. The performance of the models from Anthropic and OpenAI literally change day by day, often not due to the model itself but because of the horsepower those companies choose to give them on the day, or them tweaking their own system prompts. You can find no end of posts here from people screaming in frustration the thing that worked yesterday doesn't work today, or suddenly they find themselves running out of tokens, or their favoured tool is blocked. It's not at all obvious the differences between the open-source models and the proprietary ones are worse than those day to day ones the proprietary companies inflict on us.
If you don't know C, in older versions that can be a catastrophic failure. (The issue is so serious in modern C `free(NULL)` is a no-op.) If it's difficult to get a `FOO == NULL` without extensive mocking (this is often the case) most programmers won't do it, so it won't be caught by unit tests. The LLMs almost never get unit test coverage up high enough to catch issues like this without heavy prompting.
But that's the least of it. The models (all of them) are absolutely hopeless at DRY'ing out the code, and when they do turn it into spaghetti because they seem almost oblivious to isolation boundaries, even when they are spelt out to them.
None of this is a problem if you are vibe coding, but you can only do that when you're targeting a pretty low quality level. That's entirely appropriate in some cases of course, but when it isn't you need heavy reviews from skilled programmers. No senior engineer is going to stomach the repeated stretches of almost the "same but not quite" code they churn out.
You don't have to take my word for it. Try asking Google "do llm's produce verbose code".
`free(NULL)` is harmless in C89 onwards. As I said, programmers freeing NULL caused so many issues they changed the API. It doesn't help that `malloc(0)` returns NULL on some platforms.
If you are writing code for an embedded platform with some random C compiler, all bets on what `free(NULL)` does are off. That means a cautious C programmer who doesn't know who will be using their code never allows NULL to be passed to `free()`.
In general, most good C programmers are good because they suffer a sort of PTSD from the injuries the language has inflicted on them in the past. If they aren't avoiding passing NULL to `free()`, they haven't suffered long enough to be good.
> That means a cautious C programmer who doesn't know who will be using their code never allows NULL to be passed to `free()`.
If your compiler chokes on `free(NULL)` you have bigger problems that no LLM (or human) can solve for you: you are using a compiler that was last maintained in the 80s!
If your C compiler doesn't adhere to the very first C standard published, the problem is not the quality of the code that is written.
> If they aren't avoiding passing NULL to `free()`, they haven't suffered long enough to be good.
I dunno; I've "suffered" since the mid-90s, and I will free NULL, because it is legal in the standard, and because I have not come across a compiler that does the wrong thing on `free(NULL)`.
So what would be the best practice in a situation like that? I would (naively?) imagine that a null pointer would mostly result from a malloc() or some other parts of the program failing, in which case would you not expect to see errors elsewhere?
> imagine that a null pointer would mostly result from a malloc() or some other parts of the program failing, in which case would you not expect to see errors elsewhere?
Oh yes, you probably will see errors elsewhere. If you are lucky it will happen immediately. But often enough millions of executed instructions later, in some unrelated routine that had its memory smashed. It's not "fun" figuring out what happened. It could be nothing - bit flips are a thing, and once you get the error rate low enough the frequency of bit flips and bugs starts to converge. You could waste days of your time chasing an alpha particle.
I saw the author of curl post some of this code here a while back. I immediately recognised the symptoms. Things like:
if (NULL == foo) { ... }
Every 2nd line was code like that. If you are wondering, he wrote `(NULL == foo)` in case he dropped an `=`, so it became `(NULL = foo)`. The second version is a syntax error, whereas `(foo = NULL)` is a runtime disaster. Most of it was unjustified, but he could not help himself. After years of dealing with C, he wrote code defensively - even if it wasn't needed. C is so fast and the compilers so good the coding style imposes little overhead.
Rust is popular because it gives you a similar result to C, but you don't need to have been beaten by 10 years of pain in order to produce safe Rust code. Sadly, it has other issues. Despite them, it's still the best C we have right now.
C is fundamentally a bad target for LLMs. Humans get C wrong all the time, so we can not hope the nascent LLM, which has been trained on 95% code that does automatic memory management, to excel here.
I always found myself writing verbose copypasta code first, then compress it down based on the emerging commonalities. I think doing it the other way around is likely to lead to a worse design. Can you not tell the LLM to do the same? Honest question.
> I always found myself writing verbose copypasta code first, then compress it down based on the emerging commonalities. I think doing it the other way around is likely to lead to a worse design.
I do pretty much the same thing, which is to say I "write code using a brain dump", "look for commonalities that tickle the neurons", then "refactor". Lather, rinse, and repeat until I'm happy.
> Can you not tell the LLM to do the same?
You can tell them until you're blue in the face. They ignore you.
I'm sure this is a temporary phase. Once they solve the problem, coding will suffer the same fate as blacksmiths making nails. [0] To solve it they need to satisfy two conflicting goals - DRY the code out, while keeping interconnections between modules to a minimum. That isn't easy. In fact it's so hard people who do it well and can do it across scales are called senior software engineers. Once models master that trick, they won't be needed any more.
By "they" I mean "me".
[0] Blacksmiths could produce 1,000 or so a day, but it must have been a mind-numbing day even if it paid the bills. Then automation came along, and produced them at over a nail per second.
a) The agent doesn't need to read the implementation of anything - you can stuff the entire projects headers into the context and the LLM can have a better birds-eye view of what is there and what is not, and what goes where, etc.
and
b) Enforcing Parse, don't Validate using opaque types - the LLM writing a function that uses a user-defined composite datatype has no knowledge of the implementation, because it read only headers.
Write code? No. Use frontier models. They are subsidized and amazing and they get noticably better ever few months.
Literally anything else? Smaller models are fine. Classifiers, sentiment analysis, editing blog posts, tool calling, whatever. They go can through documents and extract information, summarize, etc. When making a voice chat system awhile back I used a cheap open weight model and just asked it "is the user done speaking yet" by passing transcripts of what had been spoken so far, and this was 2 years ago and a crappy cheap low weight model. Be creative.
I wouldn't trust them to do math, but you can tool call out to a calculator for that.
They are perfectly fine at holding conversations. Their weights aren't large enough to have every book ever written contained in them, or the details of every movie ever made, but unless you need that depth and breadth of knowledge, you'll be fine.
I just mean is the claim that the open source models where the closed models were 12 to 6 months ago true? They do seem to be for some specific tasks which is cool, but they seem even more uneven in skills than the frontier model. They're definitely useful tools, but I'm not sure if they're a match for frontier models from a year ago?
Frontier models from a year ago had issues with consistent tool calling, instruction following was pretty good but could still go off the rails from time to time.
Open weight models have those same issues. They are otherwise fine.
You can hook them up to a vector DB and build a RAG system. They can answer simple questions and converse back and forth. They have thinking modes that solve more complex problems.
They aren't going to discover new math theorems but they'll control a smart home and manage your calendar.
ASML only makes a certain number of machines a year that can do extreme ultra-violet lithography.
Also - turbine blades limit power, according to Elon.
Between them - we cannot chip fabs past a certain rate, and we cannot stand up the datacenter to run these desired chips past a certain rate. Different people believe one or the other is the 'true' current bottleneck. The turbine supply chain scaling looks much more tractable -- EUV is essentially the most complicated production process humans have ever devised.
Is ASML really the bottleneck? Do you believe anybody but TSMC and few fabs could really use and acquire those machines? I don't know the throughput of a EUV device from ASML but I imagine you need :
- clean room, itself needing the infrastructure for it (size, airCo, filtering, electricity) and the staff to run and maintain that basically empty space
- wafers to "print" on, so that's a lot of water and logistic to manipulate them (so infrastructure for clean water and all chemicals) also with dedicated staff
- finally staff who would be able to design something significantly better than NVIDIA, Intel, Broadcom, IBM, etc while (and arguably that's the trickiest part IMHO) being able to get it good enough as at a scale that can be manufactured from their own fab.
so I'm wondering who can afford this kind of setup that can only then make use of ASML machines.
> (so infrastructure for clean water and all chemicals)
Fabs are some of the most complex chemical engineering sites (dealing with some of the most dangerous substances) in the world. So don't underestimate the complexity of this part.
Yes. At least, the manufacturing of compute is. And a lot of the chain has been bitten hard by increasing capacity prematurely in the past so they're reticent to increase bandwidth at vast cost.
Presumably ASML can increase production if demand is high enough the question is over what time frame. 5 years seems plausible to me but I honestly don't know what that number is.
Yes. And the fab companies and their suppliers are deliberately and wisely slow to scale up production to meet short term changes in demand. They've seen the history of the semiconductor industry, it's constant boom and bust cycles. But they have the highest op-ex costs of anyone. So when the party's over they are the ones who pay for it the most.
If only there were some form of cheap, widely manufactured power generation technology that didn't use turbines... Are they really going to wait until 2030 to get more turbines rather than invest in solar?
This notion that "we don't have enough compute" does not cleanly reconcile with the fact that labs are burning cash faster than any cohort of companies in history.
If I am a grocery store that pays $1 for oranges and sells them for $0.50, I can't say, "I don't have enough oranges."
There is a major logic flaw in what you're saying.
'If I am a grocery store that pays $1 for oranges and sells them for $0.50, I can't say, "I don't have enough oranges."'
How about 'if I'm a grocery store and I see no limit on demand for oranges at $.50 but they are currently $1, I can say 'if oranges were cheaper I could sell orders of magnitude more of them'.
Buying oranges for $1 and selling for $0.5 is an investment into acquiring market share and customer relationships and a gamble on the price of oranges falling in the future.
"I built a ship to go to the Indies and bring back tea."
"Bro, the ship cost 100,000 pounds sterling and only brought back 50,000 pounds of tea. I don't care if you paid 12,500 pounds for the tea itself, you're losing money."
There is a very rational reason labs are spending everything they can get for more compute right now. The tea (inference) pays 60%+ margins. And that is rising. And that number is AFTER hyper scalars make their margins. There is an immense amount of profit floating around this system, and strategics at the edge believing they can build and control the demand through combined spend on training and inference in the proper ratios.
60%+ margins according to numbers which are not published publicly and have not AFAICT been audited.
Could they be accurate? Sure, I think people who claim this is impossible are overconfident. But I would encourage anyone who assumes they must be right to read a history of the Worldcom scandal. It's really quite easy for a person who wants to be making money (or an LLM who's been instructed to "run the accounts make no mistakes"!) to incorrectly categorize costs as capital investments when nobody's watching carefully.
The problem with this idea is that someone can, and likely will, come up with the next best architecture that leapfrogs the current frontier models at least once a year, likely faster, for the foreseeable future. This means by the time you've manufactured your LLM on an ASIC, it's 4-5 generations behind, and probably much less efficient than current SOTA model at scale.
It won't make sense for ASIC LLMs to manifest until things start to plateau, otherwise it'll be cheaper to get smarter tokens on the cloud for almost all use cases.
That said, a 10 trillion parameter model on a bespoke compute platform overcomes a lot of efficiency and FOOM aspects of the market fit, so the angle is "when will models that can be run on an asic be good enough that people will still want them for various things even if the frontier models are 10x smarter and more efficient"
I think we're probably a decade of iteration on LLMs out, at least, and the entire market could pivot if the right breakthrough happens - some GPT-2 moment demonstrating some novel architecture that convinces the industry to make the move could happen any time now.
I don't think so. GB200 prices are GOING UP. A100s are still expensive. This implies massive utilization and demand, no? These machines are not sitting idle, or prices would drop in the very competitive hyperscaler environment.
Hard to say at this point. I'm sure you can run your LLM chips 24/7 for training and for the public to make weird thirst-trap videos about Judy Hopps but how real is the utilization and demand, really? Maybe very real, maybe not, I don't think we can know yet.
Its like being back in 1850 and you build the world's first amusement park where the rides are free or very cheap. People are like Amusement parks are the next big thing since Steam Boats! And tons of other rich people start to build huge amusement parks everywhere. The people who are skilled at making amusement park rides will increase their prices, and since the first amusement parks are free so they can get the public going to them demand will be huge.
But how sustainable is that? - well obviously we know from history that amusement parks did, in fact, take over the world and most people spent virtually all their time and money at amusement parks - I think the Crimean War was even fought over some religious-based theme park in Israel - until moving pictures came out, so it worked out for them, but for AI?
1. Supply can scale. You can point to COVID/supply-chain shocks, but the problem there is temporary changes. No one spins up a whole fab to address a 3 month spike. Whereas AI is not a temporary demand change.
2. Models are getting more efficient. DeepSeek V3 was 1/10th the cost of contemporary ChatGPT. Open weight models get more runnable or smarter every month. Cutting edge is always cutting edge, but if scarcity is real, model selection will adjust to fit it.
The companies that are entirely AI-dependent may need to raise prices dramatically as AI prices go up. Not being dependent on LLMs for your fundamental product’s value will be a major advantage, at least in pricing.
reply