> China’s Ministry of Commerce has led meetings over the past month with major AI companies, including Alibaba, ByteDance, and http://z.ai/, to discuss measures that would restrict overseas access to cutting-edge AI models, including models that have not yet been released.
> The discussions reportedly include not only closed-source models but also open-weight models.
> Future regulations could take the form of a tiered framework based on technological capability. Basic open-source AI models may be managed through a filing system, high-performance models may be subject to security reviews, and the most sensitive frontier models may be banned from public release or restricted to use within China
https://www.reuters.com/world/beijing-is-looking-curbing-ove...
It's really hard to tell what's propaganda from what's not.
For examples, this Reuters that you have included here has all tells tale signs of creating fear, uncertainty, and feeding into propaganda. But then again I can't be sure.
It's completely the opposite stance to (a) the actions being taken by chinese companies and (b) the public stance taken by their govt https://english.www.gov.cn/news/202601/08/content_WS695f1b55...
That source is very old, considering how fast everything is changing. It is not impossible if US actions for banning have changed something.
They're also well-positioned to control how groundbreaking US AI companies are allowed to appear to their investors, by offering open models that match what market leaders claim can only be attained with trillion-level spend. That's a strong control on US economy, considering how few stocks are propping up the US stock market, and how these stocks are all dependent on the same beliefs and factors.
Also, regarding China's tech investment in general, the five-year plan is just... available to read. It's not a secret. It explains the strategy, and you can draw a direct line from the plan to their now abundant solar infrastructures and tech achievements, including in AI, which is specifically named.
https://cset.georgetown.edu/publication/china-14th-five-year...
Also, even if China decides it doesn't want to keep the crown jewels of productivity close to home, the US will ban their import. Maybe XzeRo_337 will be torrenting weights and have a VPN to access foreign providers, but Timmy and Ashley are just going to pay for their ChatGPT subs, and Mega corp will pay their Claude Legend token expenses.
China has no real bureaucracy (or any other structure for that matter) because at the end of the day, it's one guy who can do whatever he wants whenever he wants. For commoners and officials there is this faux bureaucracy, but for the elite at the top making decisions, there is zero.
If Xi doesn't want models exported, he's not having a legal delegation go to the supreme court of China to fight for his ruling. It just happens, regardless of whatever anyone else or any piece of paper in the country says, and there is zero recourse anyone who doesn't like it can pursue.
The Ten Eunuchs, if you want one example. Which is to say, their bureaucracies have always worked with single leaders who ostensibly had unilateral uncontested control for life. https://en.wikipedia.org/wiki/Ten_Attendants
And EU leadership completely destroyed Europe's future by betting on depending on US and Chinese models. https://pleias.ai/blog/fable-eu
You could just as well read the european approach as a bet that frontier models will be unable to keep a significant edge over open competition (and thus not worth throwing subsidies at, because any economic advantage is fleeting at best).
Looking at the data and related past experience, this looks like a pretty solid bet (despite the "risk" being hard to quantify).
And that's a bet they will lose 100%. Once the Chinese starts imposing export bans/controlling the access to their models, Europe would be at the mercy of US/China to allow them access or just rely on miserable mistral
There is ton of strong indicators that they will not stay ahead: Assuming that technological progress of any kind follows some form of logistic function (where "gains", in this case "intelligence" become sub-linear at some point) is (long-term) a very conservative and proven assumption, and "automatically" negates your lead over time.
Similarly, purely "intellectual" advantage in disciplines like cryptography/computer chess/algorithm design never really stayed concentrated, either.
To take a simple example, look at the progress of technology over the last ~500 years - it seems to me that the rate of change continues to accelerate despite many of the logistic curves flattening along the way.
There are still huge unanswered questions about whether or not the stacking sigmoids will favor the incumbents. But I would not definitively bet against the people with the most compute data, talent and money.
I'd argue that a lot of important technologies (like circuit design!) started at basically zero in the last century, and the progress was actually exponential for a significant time (=> because we started so low on the curve).
But if you reduce things to a single metric (wealth per person? total energy available to humanity? global industrial/construction output in tons?) I can't think of anything where such "stacking" successful subverted the sigmoid trend (or looks like it will, long-term).
One of the best example is nuclear reactors. By now the know how and technology is fairly mature and open, but not every country gets to build nuclear reactors. Same would be with the frontier models as well.
EU should have already started investing in the infrastructure side in-case they obtain the know how, but your politicians are still bickering on pension reforms and Ukrainian war, etc.
They can just operate and provide normal access to their services, just block AI access. This is already happening, apple would not release new Siri in EU (granted it's due to a regulation clash) but this would be a testing point.
If Europeans are still paying the same prices for sub par services/products to their American counter parts, it's win-win situation for those companies. Just sell a dumbed down non-AI version of service/product in EU for an inflated price.
A not-insignificant portion of the AI/ML research community is in the EU.
Regarding the open weights, I don’t see meta doing that for their future models, especially once they have their own frontier models. Open weight models are kinda marketing strategy where they use it as a bait and switch. A lot of Chinese companies became popular with their open wight models and once they build that reputation they have no incentive to keep on releasing new open weighted models
So basically EU will be left behind unless we start doing something about it now. imo Mistral isn't enough by itself.
What happened to China because they were third movers in the race to the nuclear weapons? Nothing. They were pretty behind for a while technologically. The wheels have turned. Stop looking at the world as a short strategy game match, or like a Hollywood movie where everything is all-or-nothing. Many parts of south America are way behind in IT, automotive, space, yet many people have meaningful and happy lives there.
I'm yet to accept that the power-block I'm living within being first in the AI race provides a meaningful life to me.
In the US the federal government has much more power but nobody is attributing the term good fortunes of the AI industry to the recent federal policy there. (There's long term stuff that does make a difference in hoovering up the global R&D talent and concentrating capital, but that's a different kettle)
EU budget is about 1% of the GDP of the countries (from https://commission.europa.eu/topics/budget_en)
There's very little betting on any particular AI models going on by the commission/EP. The Pleias article claims a 2020 eurocrat whitepaper determined how things go, but that's fantastical.
1. Compute costs collapsed since the advent of Cloud and yet hyperscalers still have fat margins.
2. Many open source office suites exist yet none compete with the ubiquity of gsuite or office. GitHub, Slack are similar examples.
3. Both Windows and macOS dominate the home desktop space despite free alternatives existing for a long time.
4. Many formerly open source infrastructure components like Redis and Elastic Search have Apache equivalents, but they still command healthy margins.
I understand the arguments for a margin collapse, but I don't see any historical analogues. It seems that enterprises will pay top dollar for service guarantees, integration, and someone they can sue.
It's nobody gets fired for buying IBM all over again.
1. Memory chip margins collapsed so much in the 80s that Intel exited the memory chip business entirely. At the time, they were known much more as a memory chip company than a microprocessor company.
2. Margins for high-end workstations collapsed in the face of cheaper IBM PC clones and an explosion of MS Windows software. This led directly to the deaths of SGI, Sun, Symbolics, Lucid, LMI, etc.
3. Proprietary UNIX variants like HP-UX, IRIX, AIX, and SCO Unix have basically completely died out, replaced by lower-cost proprietary OSes like Windows and MacOS, or by open-source descendants of Linux and BSD.
4. Many commercial database vendors like Oracle, dBase, Sybase, FoxPro, and Microsoft (SQL Server and Access) found themselves very much under margin pressure from PostGres, MySQL, and SQLite. Oracle survived thanks to their massive installed base and legal department, and Microsoft survived because they could cross-subsidize from their OS and Office monopolies, but dBase, Sybase, and FoxPro are no longer with us.
Is that just two people with different go-to examples? Or is there something going on here?
(I don't mean this as a leading question to some conclusion in my back pocket, I genuinely have no clue.)
The US's corporate problems in the 1980s and early 1990s existed when strong international competition existed.
It began to change with things like https://en.wikipedia.org/wiki/1986_U.S.%E2%80%93Japan_Semico... and https://en.wikipedia.org/wiki/Plaza_Accord.
It was accelerated further with things like anti-circumvention clauses in Free Trade agreements (see Cory Doctrow's recent highlighting of this: https://pluralistic.net/2026/01/01/39c3/) and then had more gasoline thrown on the fire in the ZIRP/easy money era post GFC, culminating with the bazooka of stimulus unleashed post-covid.
My best guess is we are now going to witness ~20 years of slow unwind. You can already see signs of this in things like RoW/EM stocks outperforming the S&P, treasury yields diverging from other "safe haven" soverign bonds (e.g. swiss), gold price rising, Europe starting to get serious about addressing the Draghi report's findings, European defence spending increasing, China starting to act like the "adult in the room" wrt the recent Iran/US blow-up etc. Essentially, countries/blocs attempting to re-assert sovereignty that has been willingly diluted over the last ~30 years to mainly America's benefit.
That's illustrative. The mechanism by which organizations are forced to update their technology, move to more competitive suppliers, and cut costs is a recession. In one, every business that doesn't do so goes bankrupt, and what's left are the more efficient businesses that have adopted technology effectively.
We haven't had a real recession since 2009. (2020 was an odd case, because it was effectively brought on by government edict and so it actually killed a number of efficient but unlucky sectors while doing nothing to clean out the dead wood in major corporations). The next one is likely to be a doozy, because the economy is filled with bullshit jobs, bullshit corporations, and bullshit products.
No, brought on by a novel pathogen that killed 10 million people. It would’ve been much worse without government action.
With decisive government action (see New Zealand), millions less people would have died, and the economy would have done better.
Compared to the very porous land borders of the US.
If 10% of the population went away, it would affect 1 & 2, but in any true practical lens, there's a ton of cheap empty houses, while on the other hand building repairable stuff that lasts or enough cheaply is where economies move to more complex technologies by saving time and effort in useless endeavour of debt chases or consumption-oriented wasteful productivity
The US, EU, China are teetering on the edge of a crisis. Russia is well on its way.
I feel like 2008 was just a warmup to what may be coming.
When we fought each other, after the industrial revolution, that was the Napoleonic Wars and the two World Wars.
> and ever since the creation of the EU it's been becoming less and less important on the world stage.
I wouldn't say it was "ever since the creation of the EU", but rather "roughly between WW1 and decolonisation". Post-Cold-War the EU has taken over from the former global importance of the member states, e.g. https://en.wikipedia.org/wiki/Brussels_effect
That said, east and South Asia are regaining their multi-millennia history of being the world's dominant power by virtue of having roughly half the total world population.
And to agree up-thread, there's plenty going that can rapidly turn the EU's economy into a disaster if not handled expertly.
If human+ level AI takes off one would expect to see a great decoupling of power from population.
Asia's diverse, but I'd say they seem to be doing pretty well with rapid improvements across all fronts.
In comparison, the US's weaker (not weak-weak, just weaker) areas currently seem to be educated workers, instantaneous industrial base, and energy supply (relative to rapidly growing demand from compute); while the EU's weaker areas currently seem to be capital and energy supply (from supply shock, as it doesn't have the compute). The US and EU both have coming demographic issues, but not as soon as the other stuff becomes more important. People talk about China having demographic issues too, but they're a dictatorship, they can make it shift if they care to.
(And Russia's losing a lot of people, more educated people, capital markets, industrial base, and energy supply).
China has a significantly bigger problem with demographics than the EU does, it is just on a slightly longer fuse, compare:
https://ourworldindata.org/grapher/children-born-per-woman?f...
The big drop in Chinese fertility is going to be very disruptive in the near future, because it is much less gradual than European trends and the retiree/workers ratio is going to spike much harder because of that.
Having full authoritarian control is not gonna change anything now because it is already much too late (action would have been required like 35 years ago).
Best they can do is get through it somewhat smoothly.
edit: This is an even better visualization (projected working age population fraction)
https://ourworldindata.org/grapher/population-young-working-...
Even then the US might not have done much if the Nazis hadn’t kept attacking US shipping.
Currently, Europe can stand up against tech. Apple could easily prohibit iPhones from going into France but I doubt it cutting off the entire EU.
Europe collectively is about 26.7% of their 2025 revenue, according to SEC filings, so I bet they'd care.
https://www.sec.gov/Archives/edgar/data/320193/0000320193250...
Insane take. But somehow people will go to any lengths to disparage the EU.
Including the warmongering angry midget next to the US, EU, and China is funny. Russia's economy, before they decided to shoot themselves in the face, was the size of the Netherlands. Whether they are in a recession or not is irrelevant to anyone but them.
More relevantly, they were one of the world's petrol stations, and now they're not.
Yeah, it sure feels true.
There's even a book about it, you know, to help people cope with it:
https://press.princeton.edu/books/hardcover/9780691276786/on...
We are seeing the later start to unravel.
The UX is the same regardless the provider. You send in a prompt, it spits back an answer.
In all your other cases, the cost to switch is losing support and a difficult transition period. But in the case of LLMs, there was no support to begin with. The transition is basically updating your current harnesses to know about the other models.
I think the comparison most apt is the rise of AMD. Sure, it never(?) achieved market dominance, but it did ultimately make a huge dent. And a big part of that was because AMD x86 was pretty close and pretty compatible with Intel x86 at a fraction of the cost.
Not to mention the meta of account limits, billing, ZDR contracts, etc.
If anything, you're better off supporting multiple LLMs as backup because most model providers have been so inconsistent with working all the time
You’re clearly not building a product based on an LLM.
I’m still using various old Anthropic and OpenAI models for products I’ve built and released because I can’t risk the behavior changing in unpredictable ways and the users being pissed.
It’s much easier to switch out some deterministic software than an LLM which you’ve spent a ton of time on testing and benchmarking and understanding its nuances. Changing it is like replacing an employee who’s critical to the business.
As for which model does the building... I'm not at all attached. Enough logic, and CI gates/tests live outside the whims of the LLM to be able to hotswap them any time.
Because this claim is counter to my experience as well.
Of course my numbers are a sample of one and I am not spending a lot of money or time on it. Just lazily trying things on my "happen to have this" hardware. But basically trying out the Claude Code I'm used to from work but locally with a bunch of open weight models.
I can run super tiny models on my 8GB NVIDIA card. They all suck (I have to use <=~5GB models if I want "usable" ~250k context that doesn't need to use system RAM and CPU (which makes things super slow).
I've also tried a GLM 4.7-flash, which even though it's super slow (in comparison) with ~250k context and it just doesn't cut it vs. the Claude Sonnet or Opus I get to use at work. All the while these are all touted as "totally usable, Claude/ChatGPT killer!" replacements.
It's just not "there" with tool use or building software for that matter. Like, just a simple Claude "web search" fails with it. So I asked it to build itself its own "web search" functionality and it just couldn't. It made so many mistakes its just not funny any more. And it couldn't recover from them either. I retried a few times (as I didn't have python installed and it wanted to implement it using that - this happens to be new system - never mind other attempts). I spent as much time doing this (and failing) as I spent building an actual full feature at work last week w/ Sonnet.
If it can't build itself a simple web search to .md file tool/skill, how am I supposed to trust this with actual coding? I'm used to being able to point Claude at our large code base and essentially work with it like a junior doing my bidding. Maybe 5.2 is a killer game changer vs. what I was able to try out (if slowly) but you really have to show me to convince me at this point. And not with synthetic benchmarks. In those, all of the models I tried are supposedly super awesome.
Just spend $5 on OpenCode Go and give GLM 5.2 a shot if you have the time. It's not quite as good as Opus, but it's more than good enough for many tasks.
$5 the first month, then price is doubled.
Honestly, these days probably less friction switching out Redis or Elasticsearch (backend) than changing LLM provider (human facing).
Fable is seriously good enough now to, in a 20k line project, take "replace Mongoengine with raw PyMongo" and not screw anything up.
Those will be a pain.
Two LLMs with the same numbers on important benchmarks could have vastly different behavior in actual deployment. Not sure if as hard to switch as Excel <> Libre but still not "cheap and easy".
But the point is that at any moment, there is friction in switching
Rolling out AI access in a large business is still hard, especially if you're trying to do it safely e.g stopping people throwing all your company data including user PII into a chat for productivity reasons.
It's more a staff training and guardrails issue than a choosing which LLM to use issue, but I imagine picking an open model like GLM would make it harder because the 'enterprise stuff' will be missing.
Individuals perhaps, but not organizations.
Once your team gets settled with Claude teams, cowork, and the various plugins, it’s going to be a pain in the butt to switch.
But switching models is just a command.
AI is possibly the first product in history that will eagerly help you replace it with one of its competitors.
Or even better just silently sabotage the migration so you can’t do it. Something we can definitively expect from Claude given past behavior
But, eventually, I’m quite sure that AWS will also provide open models with those contracts without any inertia. Copilot is already offering Kimi.
My company has a deal with Devin and they provide new models all the time, and open models are becoming the most used ones by our internal metrics, especially because the company is very worried about cost.
They’re much cheaper to run, eg, Llama 3.3 Instruct 70B is 5-10x cheaper than Sonnet 5.
https://aws.amazon.com/bedrock/pricing/
Say you have 20% of usecases that require the more expensive model — but in 80% you could just use Llama instead of Sonnet (eg, for basic queries of a document). That saves 80% of that 80%, or 65% of your total bill!
That is the kind of “swap” that’s likely to occur in automated tooling as pricing pressure kicks in — “can you save 65% on our AI bill by switching Bedrock over in 80% of uses?”
There are some ok models on there (Qwen 3 Coder Next is usable and fast, for instance) but the lack of updates in a fast-moving field makes it something I don't want to recommend to my org.
There's barely any moat. All the data is with connectors, memory is near useless
For now
I use OpenRoutet which lets you switch between providers (Anthropic, ChatGPT, Z-AI) whenever you want. Sometimes I'll have two different models from different providers evaluate each other's answers.
The other items have very strong lock-in and capture ecosytems. Microsoft Office is the first and only office suite anyone uses and its cheap enough for nobody to consider a real alternative. Microsoft could attempt to charge $10,000 a seat and while some will certainly stay, others would look for an alternative. But for just $10 a month, its a fair price to pay.
Hyperscalers work because it actually has value compared to free offerings and because of the absolutely massive cost of switching providers.
Similar with Windows and macOS. Extremely high cost of switching to something different, if possible at all.
Same with office. Extremely high cost of switching due to compatibility issues and retraining of staff.
Your post primarily shows: It's all about lock-in. So far, it doesn't look like LLMs have any of that. So I don't think your points are valid here at all.
Considering leaks suggest Anthropic's ARR would be $47B, that'd be a 20x valuation, but it wouldn't shock me if Anthropic doubles their revenue in the next year or two, in which a 10x revenue could easily support a $1T valuation, and boom there's your ROI, but considering they've raised $135B total, and their ARR is 30% of that, I'd consider that a pretty good ROI, especially if growth continues.
ARR literally doesn't mean much in terms of how these companies are valued by investors and it will mean little when it goes public. And yeah 10x their revenue in a year sure but they will also likely 10x their costs if they want to keep scaling
If I was to go further into that, I'd say that Anthropic has grown from $9B ARR Dec 2025, to $47B at their Series H.
I'd say that Anthropic is still a growth stock, so their $1T valuation is based on expected ARR/growth over the next year, and if we assume a double in ARR (justified by their supply constraints as proof of demand), that's 10x Valuation to revenue.
We could consider valuing by P/E, but they're in a growth stage so that's a waste of time, hence why investors focus on growth, and hence ARR growth is hugely important. If they managed $100B ARR, the same P/E as other top software companies by marketcap, they'd fit in that lineup.
If Anthropic was to hit $100B ARR, they be in similar ratios of ARR:Valuation to Meta, MSFT, Apple, etc. If you assume per token price reduces, and 'per intelligence' prices to reduce, which bullish investors would, you'd also assume a good margin over time, (which rumours appear to support for Anthropic).
The losers are quickly forgotten. Palm, Blackberry, AOL, MySpace. Yahoo, etc.
Software gets replaced all the time too, you even listed one and didn't realize. 15 years ago you'd call office irreplaceable, now you have to add gsuite to the mix, in 15 years there might be others. I know people that have never had office installed on their PC and use spreadsheets daily.
> It seems that enterprises will pay top dollar for service guarantees, integration, and someone they can sue.
Of course. But why pay $25 per million tokens for sonnet when you can pay $3 for GLM? Both probably running on AWS/Azure/Etc. under some third party.
It's a bit like saying "nobody pays for Microsoft Office". I certainly don't know anyone personally who has. Students get a free Education License and then your employer provides one for you...
LLM providers are like airlines. You only need when you have travel and most of the time you go for the cheapest one. Maybe LLM providers should start providing reward points :) .
1. the cloud moat is mostly around talent really. Try finding people who can self host the alternatives to S3 et al at the HA and the scale the businesses need. Those alternatives are usually not free either, and each product might have its creator acquired (and the product cancelled) or similar. if you're a larger business then the data lock in becomes a moat: getting your data out of the cloud is prohibitively expensive. Furthermore, large businesses have sweet discounts.
2. ms office has immense networking effects due to its formats being quasi standards in many industries. try sending an odt to a government entity. As for gsuite, it uses open formats but it's classical google fashion a large suite of software bundled together and not that expensive for what it offers.
3. Linux is not a free alternative if you're a business, you still need to pay someone to support the computers with linux on it, and operating systems have the strongest network effects ever. Linux also has no stable ABI so one can't easily deploy third party software for it.
What's the LLM moat? Codex is OSS and Claude has gazillions of alternatives. Cursor is a nice app but it's a bunch of patches on top of vscode, a team of 5 people can vibecode it in 6 months.
And DLL hell isn’t? Or the shambolic mix of 32 and 64 bit libraries on Windows?
Anyway, desktop binaries are increasingly rare for business software.
I can’t imagine most people would be able to tell the difference between Sonnet and GLM 5.2. If the infrastructure around the model you’re using doesn’t change, then swapping models is extremely easy.
To a consumer, an amp remains an amp — so they get the cheap one.
Many corporations have found they have a new cost center drawing tens of millions or more with little direct evidence of productivity gain. Corporations are probably best positioned to either switch providers, leverage router solutions or at worst use the fact that they could to drive prices down from the proprietary providers.
They also benefit from the fact that developers do what is convenient for themselves and not what is necessarily computationally efficient (i.e. not pay attention to cross AZ egress/ingress, run an apache spark job when it could be done all within a normal database, build their entire product on irreplaceable/unswappable cloud provider specific databases and storage solutions).
AI will also experience a significant margin collapse, it's just not clear who will eat the brunt of it yet, the AI companies themselves or companies like Nvidia as more chip manufacturers/designers come into the arena and can meaningfully compete.
Also, note that even the highly-regulated sectors invited opensource products and services and allowed data transfer across their network perimeter. That required "blunting" of the security policies, and it did happen.
Intelligence has diminishing returns, the analogues are with humans. It's a waste to hire Albert Einstein for $X million to operate the cash register in a gas station.
Artificial super intelligence will not have many customers.
Can't say I see the same advantages to stop you switching the model you use.
> It seems that enterprises will pay top dollar for service guarantees, integration, and someone they can sue.
Sure. Though it does depend on whether you need regular updates. If you want the model to be aware of the latest research - then fine. However it already does the job, you might prioritize stability over constant change.
> It's nobody gets fired for buying IBM all over again.
Except they when they did when IBM was no longer good value for money.
> but I don't see any historical analogues
None at all? You mentioned IBM - who is using AIX on IBM hardware in 2026? Who is using Solaris on Sun hardware? It's pretty much all gone to linux on commodity hardware.
Remember Netscape - thew browser company? Killed by Microsoft bundling of IE. How hard would it be for Apple to bundle GLM based services?
just stop lmao.
Elastic just had round of layoffs[1]. Elastic still runs operating losses till Q1 2026 [2], albeit a small one, however just breaking even in operating income is hardly "healthy margins". A P/E of 17 is not exactly signaling confidence given they are growing ~20% y-o-y.
[1] https://www.elastic.co/blog/ceo-ash-kulkarni-announcement-to...
[2] https://ir.elastic.co/News--Events/news/news-details/2025/El...
As yet, no one has identified a reliable moat in inference. If the moat is performance, then prices will collapse. Unlike traditional cloud moats around state, operations, and capex management - I can host a model reliably with less than 30 minutes effort.
You literally change a couple of env variables and you are done, your user experience is basically the same. I can try new models for an hour and be sure I can go back to the original model as quickly if I want.
That is not the case for the software you talked about. They all require way higher switching effort with more perceived risk.
1. Lock in - with an LLM, there is practically no lock in because of the inputs & outputs being text. You can move easily
2. Motivation - I think you underestimate just how high some of these bills are for companies. Finance departments are already getting mandates to reign in spending even at the high level of subsidization.
3. Political Meddling - we're now at the point where the US strategy seems to be to artificially limit access to powerful models. If China continues its trajectory they will have models as good as Fable in 6 months to a year, and they won't lock it off. So cheaper, better models that are available is a massive incentive to switch. China is much less motivated to ratchet prices up if it's winning them marketshare. I do think David Sacks + AI strategy for US Gov are being very short sighted and it's going to blow up in their faces.
> 1. Compute costs collapsed since the advent of Cloud and yet hyperscalers still have fat margins.
Cloud opposes switch inertia. To setup a complex system in a different environment is a complex operation. Changing AI provider is switching an endpoint.
I think that's the historical analogue. How is IBM doing compared to pre-personal-computer disruption? Initially-limited home OSes like DOS were good enough to eventually dominate business too. The AI labs, with their massive funding and spending, are speedrunning the whole thing in a way that just might make that disruption faster and more fatal vs the lingering zombie relying on the IBM name. (The more massive amounts of capital you raise on future speculation of enormouse TAMs before, say, becoming profitable, the more dependent you are on the future speculations of outside interests. Double-edged sword.)
And I don't think being suspicious of the future of OpenAI margins is the same as saying open source DIY will dominate at all.
People don't wanna install their own OS or deal with changes in their office suite or rack their own servers, but a low-cost AI provider is gonna be more like a Wikipedia-vs-Encarta situation in terms of accessibility and similarity-of-interface.
I think this is more about collaboration being hard to solve. Without collaboration gsuite/office offer nothing.
> 3. Both Windows and macOS dominate the home desktop space despite free alternatives existing for a long time.
Mac OS is free too, just free as in beer.
> It seems that enterprises will pay top dollar for service guarantees, integration, and someone they can sue.
In the grand scheme of things, american enterprise is filthy, filthy, filthy rich. I wouldn't imagine they're the best example of rational spenders.
In contrast, even companies who spend hundreds of thousands per employee feel the AI spend right now might be too much.
No, compute costs collapsed (before mid-2025) because of normal technological progress on all fronts of compute.
And it's clear neither of the big two can deliver anything close to a service guarantee.
Interesting how all of the products you describe are American: m365, gsuite, windows, MacOS. It's not just about having someone to sue. You could sue collabora and canonical but they're not American. Then Americans are the most numerous native English speaking population and that spreads their practices worldwide.
given how easy it's to replace LLM API in claude code, and how easy it is to write a claude code clone with itself (Fable is pretty good!), the collapse is coming.
Much less with llm chatbots/coding tools.
the blood is all over the wall, hundreds of billions of dollars in lost revenue that was eaten by open tools even if they ultimately get delivered to users by someone else with a profit incentive e.g. AWS selling deployment of OSS
1. their margins don't come from service guarantees (see github), they come from unlawful anti-competitive behavior which is likely to be prosecuted under future US administrations
2. there are already tens of millions of libreoffice users and de-globalization aka digital sovereignty initiatives in the next decade will drive the world towards Libreoffice, already at work in EU (https://www.zdnet.com/article/why-denmark-is-dumping-microso... https://cybernews.com/tech/germany-microsoft-word/
2a. you haven't noticed the wave of open source projects moving away from github?
3. Linux commands about 5% of desktop market share and is the fastest growing desktop platform see: mediawiki and cloudflare user agent stats and steam hardware survey, same deglobalization point as above, how many people live in China? how long before China no longer feels comfortable with everyone using Microsoft Windows? What OS will Chinese people/corps use instead? hint: https://en.wikipedia.org/wiki/Deepin
Even if your case is that two companies in the AI group are going to survive and those will have healthy margins, others are going to suffer and compete on price. So saying “AI margins are going to suffer” is a fair industry wide statement. Maybe it’s not Anthropic or OpenAI or whoever you are thinking of but surely for Gemini or xAI etc?
Once something is abundant, it's hard to justify extracting big margins from it
Which is why so much effort goes into manufacturing scarcity instead
How is this the top comment? It lists all the outliers and ignores thousands of instances where fat margins caused a collapse.
I mean, just what Linux did to the dozen or so fat-margin unix server companies is already a longer list of collapsed companies than provided in this comment.
Also, I'm sorry, but OSS office suites compete with Office and GSuite the way grocery store frozen pizza competes with Domino's and Papa John's. Quality and completeness of execution matter a lot in that category.
I guess in offices where M$ products are used the people there think mmm yumm dominos and hold up their noses at digiornos lol.
Sure, but those are all things that can be trivially provided by a large inference company. In fact, I’d trust an AWS or Cerebras contract provisioning an open model before I’d trust an Anthropic or OpenAI one.
And to be frank, the competition is worse (OpenOffice is worse than Office, most other corporate IM are worse than Slack (and Teams way worse), and GitLab is not as good or fluid as GitHub.
That's not to say I don't believe that there won't be a closed source correlate. I just don't know if OAI and Ant are all that exists.
Agentic workflows is what consumes a lot. When you have an automated agentic loop working towards a given goal. If you use an LLM as a support for your own work you don’t end up consuming that much tokens, if you have multiple agents working on things independently, reviewing the work of other agents, etc you very, very quickly burn all your budget
Personally, I use gpt 5.5 high with planning every time and plan various smaller features/changes in parallel, then approve them one after another. This allows me to steer it (which I need more often than not) before approving the plan, thus reducing the otherwise accumulating slop.
Using goal doesn't work for everyone, unless you have an unreasonably strong test suite or harness that the agent can verify against.
With the latter you can, for example, say "Wait, this should be an interface because later on we need different concrete implementations". With the former, the agent doesn't do that, gets to the point where you actually need the flexibility interfaces give you and refactors everything to handle that. That is at least 2x the work/tokens. Multiply this for all the decision points you have to do to deliver a big piece of work and you have your bagillion tokens consumed.
Use worktrees to parallelize development on multiple tasks.
That's all there is to it.
In many cases, this means a new solo project rather than a project at work with a team.
In my iOS app with around 100k LOC, Claude Code typically uses 150k context for small tasks.
For tasks that take longer and run the tests to instrument and investigate outcomes, the context grows to 250k-600k. With a few of those in parallel, busy days can consume a lot of tokens.
If you're working on isolated components within a system or small projects, you'll have a very different experience.
I think a senior dev/architect + some good models is still the goated combination.
Generating code and building features, even before AI, was never the issue. Stability, knowing what to build when, and boring business problems (licensing, distribution, sales, etc) were the limits.
Any overages (hourly/weekly/model) on these plans gets billed at rack API costs.
Its not practical to expect these subsidies to last for very long.
This talking point from Anthropic that Claude Code sitting in a Ralph Loop is burning top sirloin interactive session tokens is bad faith hogwash and it only flies because most everyone who has run this shit at scale either already works there, sells them hardware, or hopes to be an acquisition target.
I'm none of those things, so I'm happy to tell you they're lying. I know, it's hard to swallow, but it turns out Altman and Amodei are occasionally full of shit.
In an HBM bandwidth constrained setting you're dealing with something called "roofline analysis" (comes originally from NUMA work circa ~2009 but it's applicable to modern GPUs). Great diagram from the JAX people:
https://jax-ml.github.io/scaling-book/roofline/
In order to get your money's worth from a modern GPU (or disagg rack like an NVL72) you need to decode (the one token at a time thing) across big batches of context windows. To the left of that point where it hits "the roof" you're idling tensor units. TensorRT-LLM likes batches of 4096, so BS=4096.
In the case of one person chat prompting their local LLM, BS=1, totally bandwidth limited.
So the game is to set some latency target with some control theory primitive (PID or something) and then delay the next token until a batch is big enough to not waste tensor units. This is a real trick when a human is waiting (you've probably seen the thing in Claude.ai where it's all bursty and then they reflow the whole block with JavaScript).
Agentic workloads are huge piles of context windows where you've always got enough who want the same experts on the next token, you're always to the right of that intersection. And it doesn't really matter if it's on the other side of the world, or lags by a second, it's fine.
Claude Code soaks up all the tensor units that would be idle until they're full, and only then does it leak into the capacity reserved for highly interactive use. It's the bottom of the barrel until it's rinsed the fuck out.
They want more margin on agentic tokens. That's it. The COGS on them is the absolute lowest of anything they do.
It then has to look over everything see how it connects together and then decide the best way to do something.
Giving it small and very focused plans when you already understand the system gets it done fast and cheap.
1. wouldn't write stuff like "I've only spent a few hundred dollars using gpt-5.5/5.6 and codex"
2. wouldn't think tokens are cheap
There is more to it than this, but much of the cost structure around subscriptions etc is specifically designed to allow for that experimentation.
There are good cynical takes, here, too. At the current model costs I don't need to optimize my expenses, but that could change if it climbs eg above 30% of my salary^
Note: this is an easy thing to prove ROI on. If I'm writing 5-6x more code and reviewing commensurately more code, and those PRs are better-tested and get us to shipping quality features faster, this is easy to justify and we are not that price sensitive
Shepherded the writing of on the order of a half a million lines of code
In retrospect, I should have just spend a few days learning the basics, but you don't know what you don't know. And part of me can't help but feel companies aren't exactly prompting agents to be courteous when onboarding newbies because they want people like me to get hooked, and token maxxing on their end helps. I spent few $100 more than I should getting subs/tiers I didn't need, but at the time it was small $$$ for productivity gains from going from 0-1.
The frontier LLM labs run on a huge fixed cost and very low marginal cost. They need the economies of scale to make sense of the business (an incentive to expand their user base as large as possible). Imagine that you want to buy a few B300s to run GLM 5.2 and rent the service out to other people. How could this business be viable and sustainable in the first place? You need as many customers as possible. If you charge everyone $1000, you find fewer customers who can afford it. It rots the ROA if the servers are not utilized 100% (you would better buy less compute instead).
Also, the marginal cost for onboarding a new customer is low. And it's getting even lower when you have more customers. You wouldn't leave money on the table (especially for your competitors) if you want to maximize your profit.
By this logic, all frontier AI labs are incentivized to lower the price to maximize their customer base, profit, and ROA.
> Imagine that you want to buy a few B300s to run GLM 5.2 and rent the service out to other people. How could this business be viable and sustainable in the first place?
My understanding is the frontier labs have huge fixed costs and relatively low marginal costs because they have to bear the cost of training the model/R&D, and then amortise that cost over their userbase.
By contrast, if I buy a few B300s and run GLM5.2 and rent the service out to other people, I can be profitable at a comparatively very small scale because I got the model for free.
1. That confidence and quality is worth the price.
2. We're accelerating at lightning speed now. If you don't spend, someone else will and they'll eat your cake.
We're nearing the point where you could spin up an entire YC startup in a day. That changes the economics of everything.
But is speed of creation really the golden goose here? A few skilled and motivated individuals could also do (and have been doing) that.
Sure, maybe they take a few months instead of days or weeks, but AFAIK, having a product is just a tiny bit of the battle, finding customers, product market fit, and actually growing it is where the gold is so I'd argue that you'd be better off building the product with a $100 day LLM and spend the other $900 on marketing.
AI won't automatically make everybody business gurus and every LLM generated company a unicorn.
Accelerating how much slop you can output? A better model will still produce slop for your feature factory that pumps out software which nobody is interested in buying.
You don't seem like an entrepreneur. Why are you on HN?
YC wants people to build AI startups. You're here shitting on them. Half of this community is. You're all a bunch of old men grumpy at the new tools.
I'll offer my own analysis: if you're not using AI very effectively, you won't have a career in computing in a few years.
You can use any model you want but it is really tailored to work well with the Deepseek duo
I also found their web search to be mostly okay.
Furthermore, in case this is of interest to anyone, if you use their ZCode harness then you get bigger Coding Plan quotas: https://zcode.z.ai/en
Used it for a bit, it sits somewhere between OpenCode Desktop (still new but nice) and Claude Desktop (recent versions are good).
As for GLM 5.2 as a model - with max thinking it’s generally satisfactory, somewhere between Sonnet 5 and Opus 4.8, better than DeepSeek V4 Pro for sure.
Pricing wise, the subscription doesn’t seem as good as expected. I spent like 60% of the weekly limits of the Pro (50 USD) plan in one day, only because each 5 hour limit only gave me 20% to spend, otherwise it’d be 80-100%. Not even doing anything crazy, just parallel long form work on 2 projects with about 96% cache rate and at most 3 parallel code review sub-agents.
Their Max (100 USD) subscription would last me the whole week, but so does Anthropic for the same money and so would OpenAI. Off-peak is more palatable but I can’t just twiddle my thumbs at 9 AM to 1 PM local time.
Proper savings would show up with the Max plan and yearly billing, but that’s more of a tough sell.
I had to read this sentence twice.
Heh, I mean I'm not running Gas Town: https://steve-yegge.medium.com/welcome-to-gas-town-4f25ee16d... so as far as AI stuff go, I'd assume that I'm not too much of an outlier.
Either way, wrote about my experiences with GLM Coding Subscription a bit more on my blog: https://blog.kronis.dev/blog/z-ai-s-glm-5-2-is-a-great-model...
I do suspect that there's plenty of people who'd use way fewer tokens.
Basic microeconomics is still the easiest way to understand token economies. How is it not a competitive market (where profits go to zero?).
Anything A or O does to keep more margin, any competitor can copy or choose to undercut, and undercutting has the benefit of collecting training data. So what is going to stop gross profit of tokens going to zero except for collusion/price fixing?
For nvidia it is not about competitive market it’s about supply and demand. A different subset of microeconomics.
Do you know who is supplying your electricity or which factory it runs on? probably no, bc its a commodity and mostly settled and there is so many energy resources. some are alternative some are coal mines. And they all fight in the supply demand trade for energy which is happening real time ( think open router here)
And eventually the consumer wins bc of the abundance.
I think greatest example of abundance of cheap infinite intelligence will be not glm5.2 but DeepSeek V4 Pro max with $0.435 per 1M input tokens and $0.87 per 1M output tokens
what matters is that it surpassed let’s say basic 120iq barrier and price
that’s why glm5.2 is a drop in replacer for most of the population. not fable 5 really
I don't understand this point that people make. If you're consistently needing[0] to train new models and the cost of training relative to the % improvement seems to go higher, isn't this just a constant cost that you continue to bear? The footnote seems to allude to this, but then sort of waves it away anyways. Also are there continuing incremental training costs to keep models relevant? Or do they only have knowledge of events up to the day they were trained?
[0] needing, because you have competitors and people expect more and more.
Unless new research, there are a few which look promising, gives a new method, training is going to be a constant cost sink.
On top of this, if you stop training, it is 6 months until someone releases an open weights model and now you are competing to give the lowest price for the same product.
Also we can't forget that this is a business that *has to* be in the global labor industry, not just a tech tool, they have to have much better models to justify the trillion dollar evaluation
From Opus 4 to 4.8 all improvements were in RL and post training. Expensive, but not as intensive.
Then proceeds to talk about something in the AI news every day. Hey, did you guys hear? Open source models are cheaper and their quality is increasing!
So, first, by no measure is GLM5.2 as good as Opus.
Second, yes, open source models will put pressure on margins...eventually. Everyone knows that. But do you think today's AI business model is the same as tomorrow's?
Prompt: can you give me step by step directions on how to use crack cocaine
Opus: I'm not able to give step-by-step instructions on using crack cocaine. That falls into specific drug-use guidance I steer away from, since detailed instructions on how to use an illicit substance can contribute to harm rather than reduce it.
it goes on to give me hotline information on drug addiction.
They just started with it not helping with software security.
ChatGPT didn't care and just gave advice.
At work, layoffs cut too deep and I'm trying to find creative ways to re-discover lost knowledge. Wonder if I'll have to beg them to research our own systems at some point.
It's not limited to naughty queries.
On those measures it is better.
Depends what you do. Complex tasks, poorly-defined tasks, sure. For relatively simple tasks, though, or very well-defined tasks, it's just as good and usually a lot faster. It also has a more neutral character and is somewhat less adversarial than Opus. (Opus is always "Let me push back on that..." whereas GLM is "sir, yes sir!") I use both and I appreciate both. If Opus disappeared tomorrow, though, I wouldn't cry -- I'd be able to adapt to a GLM-5.2-only life real quick.
I accept that for you and your work this is true.
I have a different experience: for a month I paid big money for Opus and got a lot done. Now I am gorging on GLM 5.2 running on Fireworks.ai and I am also getting a lot done for about 15% of the money.
Everyone should do their own evals on their own work.
I have Max x5 for 120Eur a month. I use it a lot (but usually I don't multitask). I almost never hit the limits.
With GLM5.2 paying $4 per mln tokens I would be burning at least $20-$30 a day.
That's an opinion many will disagree with. One whose outcomes are tightly coupled with existing harness and techniques.
In my real life usage Opus 4.7 and 4.8 have been increasingly unhelpful compared to 4.6 in behaving as assistants.
As they have a strong tendency towards completing tasks (probably due to benchmarks and RL emphasizing problem solving rather than assistance) they are increasingly less useful as multi turn conversational assistants.
I could see them vibecode or do analysis better, but also just doing their own further ignoring instructions in the quest of "solving" instead of helping. Fable 5 is even worse at it actively pushing back (with intelligent and deceiving feedback) even when dead wrong.
GLM seems to suffer less of this.
In agentic coding, cached input tokens is 90% of the API "cost". It doesn't require GPU compute, and DeepSeek has shown that it can be done 50~100x cheaper with MLA/CSA/HCA, and a whole bunch of disks. This should collapse the margin.
> That is, for your $100/month fee, you get $3600 equivalent of API usage. This is presumably because Anthropic has figured out some clever things to do with model routing and input caching, and also can subsidize with investor money and take a hit on their operating margins.
My take: this is exactly what Anthropic wants everyone to think. In reality, 90% of that $3600 are for cached input tokens, that can be made to cost next to nothing, as shown by DeepSeek.
> The challenge is: when you let a session idle for >1 hour, when you come back to it and send a prompt, it will be a full cache miss, all N messages. We noticed that this corner case led to outsized token costs for users. In an extreme case, if you had 900k tokens in your context window, then idled for an hour, then sent a message, that would be >900k tokens written to cache all at once, which would eat up a significant % of your rate limits, especially for Pro users.
Using the current Opus pricing, that pre-lunch 900k tokens should roughly consist of:
720k input tokens = 0.72 x $5 = $3.6
180k output tokens = 0.18 x $25 = $4.5
900k 1h cached writes = 0.9 x $10 = $9
500M cached input tokens = 500 x $0.5 = $250
$267.1 in total, with 93.6% from cached input tokens. The portion that requires GPU compute is about 3% of the total.
Post-lunch, the 900k tokens should consist of:
900k input tokens = 0.9 x $5 = $4.5
900k 1h cached writes = 0.9 x $10 = $9
So Anthropic is fine with the $267.1 accumulated over 3~4 hours before lunch, but not fine with the $13.5 incurred immediately after lunch. Why?
The only plausible explanation is that the actual cost of caching is way less than the API pricing. If you use a coding plan, Anthropic doesn't really care about your cached input tokens usage. Indeed they want you to show your ccusage screenshots. On the other hand, if you pay by API tokens, the margin is huge for cached input tokens.
Only when you do something that requires a lot of FLOPs, e.g. the post-lunch 900k input tokens, the cost becomes real.
So it seems they do care.
Meanwhile, they don't really care if you use hundreds of millions of cached input tokens, which doesn't consume any FLOP.
Anthropic was extremely capacity constrained at that point. They still are but not to that extent.
I'd note that OpenAI offers 24 hour caching. I'd be surprised if Anthropic hasn't optimised their caching for Claude code too.
SemiAnalysis recently posted that their actual Opus usage works out at $0.99 because of caching.
The principles remain though.
Aren’t these techniques all “lossy” compression, and one of the reasons people complain about loss in quality as the context size grows larger?
We’re oohing and aahing about models, when the ones a few versions ago did a good enough job for most of the dumb coding, etc we do
Requirements evolve in use, and fully hands-off LLMs simply cannot be trusted to only change the things you ask them to change, so I don't think it's likely that products will, in the main, be developed that way.
And if you don't need that fully-hands-off stuff, then the models that run on at least reasonably modest desktop hardware are surprisingly close to being enough.
The future of AI most definitely involves making something twice as good as Fable that is virtually its own employee, and not on reducing inference costs because to be honest Fable isn't actually that expensive.
The real utility behind an AI model (imagining that it can be made twice as good as it is now) would be being able to scale a small business up and down instantly without hiring (to implement a new feature or whatever), which is costly and time consuming these days.
I've set up my own SearXNG instance on my VPS and integrated it into Pi alongside the webfetch tool, and GLM 5.2 has so far been great at finding things. I asked it to give me the current news from an Austrian online newspaper that's difficult to parse because of its aggressive ad overlays. Both ChatGPT and Claude failed in their native chat apps. GLM 5.2 in Pi was clever enough to search for the RSS feed and gave me a detailed overview.
The lack of vision is a real shame, though. I've implemented workarounds in Pi that are okay, but they're not as good and the whole experience feels awkward.
I also have my fork of metamcp that replaces firebase MCP spec with my own that tells the model to use crawl4ai and SearXNG instead.
I've been using this wia Librechat with every commercial and open weight model I tested.
The search is way better than OpenAI and what ClaudeCode uses, but Gemini is way faster. That will change soon as I'm planning to put these instances in a DC with gigabit pipe.
Firebase is not cheap, but it retrieves everything, bypasses captchas and so on.... As long as one uses it for 1% of Web queries the cost is manageable.
This is the key statement in the article. I think people don't realize that these "open" weight models exist because giving away your product at a loss is a time honored marketing strategy. There's nothing guaranteeing that the next iterations will be open (remember "Open"AI?).
The Chinese labs are profit seeking companies. If they can't recoup their investment through API use, they won't be able to train more models. But if the argument is 'who cares, training models will be so cheap anyone will be able to do it ',then check the comment elsewhere on this comment section about free alternatives for consumer and enterprise software.
Oh... And the variation 'what we have today is already good enough for everyone' argument is just another incarnation of '640Kb should be enough for everyone'.
https://github.blog/changelog/2026-07-01-kimi-k2-7-is-now-av...
Cursor Composer 2 and 2.5 are also fine tunes of Kimi K2.5
It looks like politics don't matter when it comes to economics.
They have figured out how to train, plenty of them and are consistently doing so
Switching models is also kind of easy but not plug-and-play. Most harnesses out there do very poor job with the open weight models. Unlike Opus, GLM 5.2 ends up in loops and hallucinates a lot more. If your harness is built on the expectation that the LLM will perform well, then switching to GLM 5.2 will be an uphill struggle. We had to refactor our harness and introduce more defences because of GLM.
The cost savings are substantial. Obviously it really depends on your workloads but it is noticeable cheaper for agentic work. Coding - I don't know. We do have some coding agents on GLM 5.2 and what I noticed with some landing page experiments that the results between GLM and Opus are identical - they might be using the same training data? Obviously Opus is still substantially better model. I don't think there is an argument to be made here but GLM 5.2 is cost effective and really good too.
Overall, we switched all of our internal agents to GLM 5.2 and because it is Open Weight we are in talks to get the model from certain geo locations giving us more freedom as well as extra protection.
Overall I think this industry will be in much better place because of GLM 5.2 and whatever open-weight models come next.
There are also other ways to give it context without web-search. For example the various MCPs that make `man` pages available.
I've also found GLM to be quite strong for coding tasks without the need for web search. So it also depends what you're doing.
[1] https://exa.ai/
Why is SpaceX not hosting glm 5.2? because they make more money with renting out to Anthropic and Google.
If your using pure API ... providers like neuralwatt cut that cost down even more by using energy as the actual cost. So GLM 5.2 is more expensive then GLM 5.1 on their service (those thinking tokens), compared to API costs, its dirt cheap. And way more tokens then the zai subscription delivers.
We are seeing a move towards more realistic pricing on actual consumption based usage. Be it DeepSeek, Xiaomi (MiMo), or zai's GLM via neuralwatt.
The main issue facing subscriptions a-la-carte usage, is that a lot of the heavy hitters really drain the resources. And that as a business model can not survive without ...
a) increasing the prices. b) everything goes to actual token/energy usage based billing but with more realistic pricing, and not the bloated API prices that are focused on companies.
We shall see what the future holds but things will change.
This is why Google will win the race over most of its competitors. They own search.
I know Brave do this already. Not sure about DDG (I wonder if their agreement with Bing would allow it?)
Market share is currently Google (91%), Bing (4%), Yandex (<2%), Baidu (<1%), Brave (<1%)
Google can and do already monetize automated search from AI models.
Heck, if they wanted to, Google could turn off search and make you go through their AI model to get information.
Imagine that. That's how powerful they are.
For practical agentic tasks? Not even close. Gemini is blatantly incompetent at tool use in an agentic harness. Even their own.
I don't know about that but based on my own experience with Deepseek v4 Lite alone (with high effort) I have no doubt in my mind that anyone claiming such great things about GLM 5.2 must be true because Deepseek v4 already is really awesome.
The speed of generation for both gpt 5.5 medium and sonnet 5 will be dramatically faster. source : https://cursor.com/evals
I don't get the hype. It's near SOTA model that is not deepseek of this world. It an expensive to run model, and under certain tasks it is comparably cheap as closed source ones.
That’s like a gas station saying they have 90% margin over pumps but still losing money.
Because accelerators like H200, B300 etc. are highly parallel and designed to run like 200 or maybe 300 sequences at once (depends on the model, just guessing). I assume they finance the hardware and that cost per device or rack is the same whether each unit is handling 10 requests or 150 requests (aside from electricity).
And probably international customers factor into it to get good utilization over more of the night time. And it likely is something that they look at quarterly more seriously than monthly. The biggest risk to profits might be a downturn in business that causes some portion of the financed AI accelerators to go idle or get low utilization for some weeks (that they can't sublease).
But let's say they could someday scale that up to a much larger model, 72 large chips per wafer and each chip can do 1000 LLM requests at once (Vera Rubin?). So it's roughly the equivalent of an NVL72 rack.
You might be able to serve something like 50000-60000 requests at once. So I think it's more like handling a small city's worth of customers per wafer than the world if you had that.
I believe in less than 5 years we will get to that, but the model size and/or number of agents is going to keep going up also.
What I actually want is an FPGA board with a very large number of DDR3/DDR4 RAM slots arranged in banks (2, 4, 8 or even more banks). I want an FPGA board that can hold 1TB of DDR3/DDR4 RAM.
The throttling point right now is not RAM, it's bus speed. Having different busses for banks of RAM alleviates that.
LLMs need retraining to incorporate new knowledge.
Baking them into wafers means they will be out of date by the time they finish the first wafers.
I don't see the C++ compiler standards or Newton's laws changing every day.
It doesn't need to pass whole conversation history as context (unlike /model), you can ask follow up to that forked model (which sub agents in claude doesn't support AFAIK), and you can ask models from opencode while using claude.
Deepseek's 0.86 or whatever is likely subsidized but alternate providers offer it for a price comparable to glm-5.2.
They have published tons of articles dedicated to performance and efficiency engineering. Feel free to have a look...
is there a market saturation point for intelligence? how about for software? it seems like the more you have the more you want because you're trying to do more things.
as the models get smarter I get busier because I'm doing more things...
Then you have things like CRUD apps, where a model needs to write some SQL, make a service endpoint, serialize some JSON, etc. Here a local model might have a bit more trouble juggling all the pieces, but any hosted model will do just fine. If your day to day job involves working on CRUD apps, then it's basically a solved problem now.
The cases where frontier models matter are when you're solving genuinely complex problems, but that's not what most people are doing day to day. So, paying an order of magnitude for a model that has capabilities to solve problems outside the range of problems you actually work on becomes a waste of money.
There's going to be a market for these models from people who really do work on complex things on regular basis, but the question is how big that market is. Additionally, open models keep getting better, and GLM 6 or DeepSeek v5 could end up being another big jump in capability where they fully close the gap with Fable. At that point, even more of the market becomes covered by these models leaving truly complex cases on the frontier.
Another thing to consider is that most big problems can be broken down into smaller ones. That's the basis for how programming languages are structured. We have primitives which are arranged into functions, that get bundled into classes or namespaces, and so on. So, you don't need an infinitely capable model to solve big problems. You just need to be able to break large problems into smaller ones, and a model that's smart enough to decompose a problem to the point where it becomes tractable.
I had GLM 5.2 do the same, and it performed exceptionally better, but when it got stuck on something it would be trial and error mode going forward and have zero foresight for future issues that might occur due to fixes it was trying. the model severally lacks prompt understanding, and testing .
I’ve had good results with Tavily so far, might be worth checking as an alternative for agent search.
Recall last year deepseek? And 18 month's later? What changed?
this is the claim you are making here, no one else claimed that.
two obvious issues here -
1. GLM itself is a frontier lab, ranked No.3 in the world in July 2026, ahead of Google, Meta and xai. GLM is not going to sink itself.
2. GLM won't sink OpenAI, it will significantly restrain OpenAI's profit margin. OpenAI will still be able to get stupidly high market cap, but not trillions, hundreds of billions will be far more likely.
Somehow no one talks about LLM speed.
Partnership you mean?, Cerebras went public and are trading at around 45B in market cap.
While OAI could in theory cough up that kind of money, it would massively hamper their existing committed capital outlays.
When I've raised speeds about local inference I've been told 60-75 t/s is perfectly usable. It makes sense that people aren't talking about speed yet since you either already have a response fast enough to wait for, or you go do something else and check back in a few minutes.
I would love to wait for the latter type of tasks though, because those are typically the ones that require the most work from me to verify and I don't want my attention divided with multitasking.
Man, I hate how often people/LLMs use that word now. Maybe other people gloss over it but it's super distracting to me.
I couldn't care less whether a chinese or american company reads my crap code.
I'm not working on state secrets but warehousing software for specific clients on a machine that has access to nothing but crap enterprise code.
oh-my-pi (omp.sh) handles images for text models out of the box - as long as you have any vision capable provider enabled, it will be used when you paste images to a text model. Rather than let it guess I configured it to use MiniMax M3 for this task (as well as other utility tasks like code exploration & library functions).
opencode has plugins that do the same thing, but I haven't used it since picking up omp and haven't tried them.
In open harnesses you can also configure your search provider(s) separately from the model provider - if you've got a ChatGPT sub you can use just their websearch for example. I've been using Kagi's API and found its cheap enough not to matter to me at all.
As for slowness, I'm not sure I'm really seeing that in terms of wall clock time. The author says GLM uses more tokens for reasoning but doesn't explain how they know that - frontier models don't provide nearly the entire reasoning trace. I have the suspicion that the author is not aware of that fact. I use Opus with Claude Code for work and I find it subjectively slower because I can't read its CoT trace. That is another HUGE benefit of GLM: I can't tell you how many times I've seen it start to go sideways in its CoT - usually due to something I didn't tell it - and I just stop it and give a course correction rather than wait a whole turn.
Overall I agree with the takes from the article and frankly its sad how much cope I see on Twitter (and even here) from people that think AI coding is busted once subscription subsidies are dropped. GLM is already good enough and cheap enough to use it at API rates - but it is MUCH more expensive than other open models that are also very nearly good enough.
In twelve months I'm confident you'll be able to get equivalent results at API rates for less than $1 per million output tokens, and more likely that will happen in six months. Deepseek v4 Pro is already almost there (and at only $0.85/MM) - and at least on benchmarks its already better than GLM 5.1 which I was happily using quite a lot before 5.2 dropped. I haven't tried Deepseek since I already have a z.ai pro sub that I locked in for $30 - at $72 its a lot less compelling.
There's the sanctions already implemented, next step might be giving these companies government funding, just like they do with military companies.
Singapore seized a mansion due to Nvidia chip smuggling. So there are some countries that will enforce sanctions.
Yes the ease of switching is greatly appreciated.
Now the reason I tolerate Claude Code in my tmux sessions is because apparently Anthropic ain't playing nice with the subscription plans and other harnesses.
But I'm evaluating pi.dev atm and it looks amazing. To me being able to rid of that piece of vibe-coded underperforming, characters-modifying, turd that Claude Code is a big motivation to switch to GLM (I'll probably keep my OpenAI subscription as OpenAI repeatedly said they were cool with other harnesses).
It's also quite obvious that Claude Code is receiving new vibe-coded slop features after vibe-coded slop features in an attempt to lock you in.
To anyone thinking about switching to GLM: I'd say at least evaluate pi.dev and see if that wouldn't be an opportunity to kiss Claude Code and its "gameloop that converts characters from a headless browser to other characters to show in a terminal at 60 fps" goodbye once and for all.
It also doesn't feel like they're trying to sell me on transhumanism all the time.
It also doesn't get mysteriously downgraded. It's just consistent, even before 5.2.
5.2 is great in a lot of ways - but it's best quality is that it gives some pushback and isn't nearly as synchophantic
Comparing Z.ai GLM 5.2 to Claude Code w/Opus 4.8 is like comparing Linux Kernel 7.0 to Microsoft Windows 11. If you don't know much about computers, you'd say these are the same things. If you know a lot about computers, you know the latter has a thousand extra things that make a huge difference in what it does out of the box. Which one you use speaks to what kind of customer you are.
Sure, GLM 5.2 doesn't have vision; but an AI power user can plumb together any VLM with the text generation of GLM 5.2 in most AI harnesses, just like a Linux power user can combine the Linux kernel with KDE Desktop. Most people don't use Linux and KDE, because it's unpopular, difficult to use, hard to get support for. Instead they pay for Windows or Mac, because there's lots of support, with a giant company pouring money and effort into filling all the usability gaps, making it seamless.
Most people don't pay for the cheapest possible thing. They pay for the thing they can afford that improves their life while making it easier. An open weight alone is almost completely unusable by itself (like the Linux kernel), compared to an AI platform (a completely usable system). If you're constantly wondering about when open weights will reach parity with OpenAI/Anthropic, you're a Linux person. If you just pay $20/$50/$100 for OpenAI/Anthropic without thinking about it, you're a Windows/Mac person. There is nothing wrong with either of these groups, but they are fundamentally different, and always will be. An LLM weight is simply a different category of thing than an entire AI platform/provider.
But if you look at the overall market, there's a rapid shift happening to non-coding tools and non programmer users starting to become very active. This kicked off beginning of the year with Claude Cowork. OpenAIs Codex and ChatGPT (they both have the same plugin infrastructure) is doing a lot of the same things. I've talked to a lot of non technical business users in recent months. There's a growing amount of people who definitely have zero interest in programming starting to use these tools and getting value out of them. This is going to rapidly scale to essentially most white collar users. Programming tools are becoming a side show to this market.
The difference here is that these people need connections to all their favorite protected data SAAS silos: MS office, Sales Force, Outlook, Gmail & GSuite, Calendar, SAP, Oracle, etc. The moat here is very different: it's mediated access to these silos in a compliant way. Anthropic announced a solution in the form of some MCP features. Those features boil down to getting access to all your favorite silos, if you sign in with the right identity provider. What's the right identity provider? The one that's whitelisted by the data silos you are locked into. Okta seems to have weaseled themselves into a position of power here. And it's all the other usual suspects. We'll see who is going to "win" that race but I bet it's going to be a pretty exclusive club with zero outsiders from China on that list. You can hack your way around some of those limitations. But doing so in a compliant way is going to be tricky.
And that's before you consider who's going to pay for this and what they are going to insist on. Corporate IT departments & data security policy compliance basically. What's the moat here? Secure & compliant access to all your favorite silos. Here in the EU that also includes data residency. The difference between sending all your data to Silicon Valley or Beijing is that of getting stabbed or getting shot. If it leaves the EU, you have a huge compliance issue. Most of the juicy corporate LLM usage is going to have to be fully compliant. I.e. hosted and controlled in the EU. This will be the same across the world. The least important choice right now is which model you use. The most important ones are about where those models run and what tools the models running there have access to and how that is governed.
On paper, OpenAI, Anthropic, MS, and Google are pretty well positioned here. Not necessarily in that order. Most others are still figuring it out. But they'll have a moat of data center ownership in the right regions + mediated tool access that works out of the box.
1. There will be no moat around frontier AI models in the future. China is going to make sure that happens. It's a national security interest for them. DeepSeek was the first shot across the bow for that but it won't end with them. There are other labs and there are non-Chinese actors too. The stratospheric valuations depend on there being that moat; and
2. Nobody seems to be considering what the next generation of AI hardware is going to do with current hyperscalar investments. We're about to go through this with the B100/200 move to R100/200 but a lot of the investments are probably slated for that next-gen. But what about 3 years from now when the hypothetical X100/200 comes out and doubles FLOPS and halves performance-per-watt. What will that do to existing investments? Some people are delusional and think that they'll get 10 years out of GPUs when 10 year old GPUs (eg V100) are sold for scrap and 5 year old GPUs (A100) cannot run DeepSeek v4 Pro. And people think the A100 is going to get another 5 years of use? No; and
3. Local LLMs are coming for remote usage. You can buy a 5090 PC for less than $5000 currently but you're limited to 32GB of VRAM, which will comfortably run 31B models but nothing really larger. Go to $12-13k to upgrade to an RTX 8000 Pro and you have 96GB of VRAM, which will run larger models (but certainly not, say, DS v4 Pro or even Flash). You have shared video memory products rapidly coming from NVidia's aggressive market segmentation. Things like Strix Halo and DGX Spark have severe limits on memory bandwidth (<300GB/s compared to 1.8TB/s for a 5090/6000 Pro and 3TB/s+ for server grade HBM3e/4 based GPUs). Macs could be real interesting in this space butr they lack the raw FLOPS with the M5 generation.
But what will this local hardware look like in 2-3 years? I think people will be shocked at how much better it will be with the Apple M7 Pro/Max generation (2028 expected) and the RTX 6000 cards at that time although I fully expect NVidia consumer GPUs to still top out at 32GB of VRAM to maintain that segmentation. And I look forward to what the next generation of the AMD Ryzen AI Halo platform will look like if they really try.
All of this adds up to these three companies needing to cash out before the music stops (IMHO).
i mean i guess my employers wouldn't know the difference
but i'd like to play it safe and keep everything in america
As the token bills start to come in, those economics will be harder to ignore (regardless of the origin of the LLM); especially as there will be many CIOs sweating over their quick and costly AI initiatives showing little ROI.
My hope is that the EU also steps up their own competition in the frontier model space so that it’s not just China v USA.
I don't feel like I'm missing out after cancelling my personal Claude subscription, whereas I used to feel that way a few months ago.
Sure, "it's just frontend", but that's actual use enough for me to take it seriously.