I don't know what "compute gap" means in this context though and it's not clear that that's why they plan to pause fundraising or if the title is conflating.
in his article he talks about the negative aspects of getting everything you want. (all the money, brightest minds, biggest share in AI) etc. he says that these are the things that will cause a company to fail.
Yes. It's a known thing.
https://github.com/demo-zexuan/liang-wenfeng-investor-meetin...
The current link is 404, can mods update to above, detach, make sticky?
(No response from mods, understandable)
It's now dropped off the FrontPage but I suppose I will just repost at some point with a less controversial title
Lots of companies in the U.S. have fallen victim to that syndrome, but if you used the word "restraint" in that context in Silicon Valley most people would look at you like you're insane.
At what point do the AI companies start building their own hardware, I wonder.
The title is certainly a great conflation. Any seeker of capital would want to regroup after an unfiltered leak of this magnitude, if for no other reason than to secure the forum from future leaks. The comments about the unlikelihood of enormous future profits were at least as consequential with regard to capital investment as anything else that was said.
There's quite a bit of confidential information in the doc about the company and how it's positioning itself going forward to compete with US labs. I'd imagine they're not happy at all with this being leaked and are withholding investment as a punitive measure.
Not to mention the other interpretation seems illogical -- why would you pause fundraising if your perception was that you lacked resources compared to your competitors?
> There is certainly no shortage of funds or resources --- in fact, all these are readily available [...]
> Within our financial capacity, it's undoubtedly true that the more cards are always better. Our current strategy is to purchase as many cards as possible at a reasonable price --- exactly how many we can afford after using this funding round. The spending pace isn't predetermined; we'll buy whatever is available as long as prices remain competitive. In fact, I'd consider that a positive outcome if we spend the entire amount within six months. [...]
> In reality, spending such a large sum is no easy task: you can't obtain enough cards, they're hard to come by [...]
> Therefore, our only concern is whether we can obtain enough cards. If converting all funds into cards were feasible, we would undoubtedly do so without hesitation and are even willing to pay a premium for this benefit --- it's simply to cost effective. Even after paying the premium, however, achieving this goal remains challenging.https://www.bloomberg.com/news/articles/2026-07-25/deepseek-...
Update:
Less-paywalled word-for-word copy it seems at
https://fortune.com/2026/07/25/deepseek-liang-wenfeng-backer...
"The suspension stemmed in part from Liang’s frustration over online reports about his comments to investors during his first financing deal"
The part of the transcript I'd seen floating around online was this part from around 1 hour 26 min:
"With the largest models available today, we simply cannot afford to train them. Even if we spent all five hundred billion yuan, we still wouldn't be able to do so. Even if we could accumulate the resources, we wouldn't have the means to utilize them. The current largest model requires approximately 800 billion activations; domestically, we are still at a scale of several dozen billion activations, and even the largest domestic model may only require several dozen billion activations—a difference of an order of magnitude. To train a model of the same size as an AI system, we would need around 50,000 GB300 GPUs or Huawei 950 GPUs, totaling two hundred thousand cards. This is merely training; research has not yet been considered. Therefore, the biggest gap between us and the United States lies in resources."
i genuinely think these models should be like 0.1 percent sparse for same capabilities we associate with them today, but theres no sane way to do that with extent tools. i built the right core tech for that in 2014 when there wasnt a market, but now there is and the experimentation velocity is wild.
amusingly llms really have a hard time using my simple apis because its not in distribution array programs. but i literally stood up cpu custom memory format and micro kernel for dense causal attention in less than 24-36 hours and outperforms the equivalent fused ggml/llama cpp fast oath by like 20-25 percent
In math terms, this means a layer of a network can be represented with a block matrix in the whole 'layer' matrix, which I think means its sparse as you said.
As I said, my math knowledge is rusty, but I remember that a lot of matrix optimization techniques center around decomposing large matrices into these smaller blocks, which are then evaluated, and the output is combined in a final pass. Which leads to a huge reduction on parameter numbers and the time it takes to evaluate the result
like i can do all sorts of memory layout of tensors/matrices etc tricks that if you dont have the abstractions for it would just never happen. so i can optimize the kernel flops
For most of the past five years, I've known ways to do better than Anthropic, OpenAI, and friends in many ways, at least on paper. I know I was right about many of them since many would show up 6-24 months later tools from the major providers, or otherwise become standard practice.
A central problem is the Mythical Man-Month. True, I could do those, beating then-state-of-the-art, but only given 2-5 years. I suspect many other people knew about them too and could do so as well. As I noted above, throwing people and dollars caused many of those to be built in less time than I could have regardless.
Other methods, I'm less confident about (>50%, <80%), but would lead to similar improvements orders-of-magnitude as you're predicting, but mine would need $$$$$ in compute and engineering infrastructure to build out. E.g. they need to not just theoretically work, but to try, I would need to convince someone to invest in them working.
So the TL;DR is that my knowledge was not at all helpful towards e.g. competing with OpenAI, Anthropic, or even building a small business.
However, where it was useful was in predicting where the industry was going. This is true in investing (but not easily, at least with my skill set), but in developing startups and systems, there were capabilities which I (correctly) assumed would be there, whereas there were many arguments that "AI will never be able to ____."
If I know how to do something, it will almost certainly happen, regardless of whether I'm the one who does it.
To be clear, my expertise is almost certainly nowhere as deep as yours. I'm not providing a direct analogy, or claiming others know what you do or can do the same. My point was really that if you believe you can have these models be 0.1 percent sparse for same capabilities we associate with them today:
a) You're probably right. They were built quickly for capabilities. A slower process can almost certainly lead to much smaller models too. That's a radical statement: Historically people claiming a 1000x improvement somewhere were crackpots, but that's very possible in an industry as fast-changing as this one.
b) Someone at Anthropic or OpenAI might be working on building out extent tools right now. Even if so, there are indirect ways to capitalize on that knowledge.
c) Critically, that predicts a future where Fable is $1/month instead of $100/month, and that's something which CAN be acted upon in planning.
It also suggests -- much less strongly -- the existence of much more sophisticated models at $100/month. There are open discussion in planning about whether models plateau, continue improving, singularity, or otherwise. That changes the biases there.
will Anthropic (or OpenAI) lowers their price, or increase margin (to justify valuation)
at the very least i have tools that let me easily hit better perf for fancy dense memory layouts, and the same tooling lets me experiment with frankly wildly wacky sparse and structured memory formats. the performance claims at least on the dense side are solid so far!
the sparsity angle is because i want magic in the world. like anyone with a really chunky computer like any of those mac mini pros or serious workstation / server tier compute should be able to train from scratch their one 31b equivalent model in a week or so tops is the goal post i have in mind
https://www.cyberkendra.com/2026/07/deepseek-pauses-fundrais...
"The Hangzhou AI lab has told prospective investors in its second fundraising round that it is suspending the deal, people familiar with the matter told Bloomberg on Saturday, days after remarks attributed to founder Liang Wenfeng about US-China AI competition circulated widely online."
And:
"Tencent's technology outlet published a 118-item version covering AGI strategy, chip supply, pricing, and retention. In it, Liang reportedly framed China's disadvantage as an arithmetic problem rather than a talent one: "The biggest gap between us and the US is in resources.""
"The specifics were unusually candid. Liang is said to have told investors he needed 200,000 Huawei 950 chips to train a frontier model but received 16,000, adding that "Huawei's problem is still insufficient capacity" and expecting the crunch to last at least three years. He also floated narrowing the gap with US labs to three to six months using a fraction of their computing."
This seems to be holding true for AI as well. I'm sure if you have an excess of compute and a dearth of bandwidth, you can trade the former for the latter, but still, this is a fundamenta property.
A lot of talk has been said about how companies are doing 'financial tricks' to extend the useful life of GPUs by showing lower depreciation - but what if these are not tricks at all - new GPUs don't really have that much more bandwidth, and while they might be clever in some other ways, they are limited in how much they can improve fundamentals.
This has been reflected in how memory vendors' stock price has exploded, but NVIDIA stayed stagnant.
Since the Chinese are far closer to the US in building SOTA memory chips, it's possible that their disadvantages are far overstated.
There's a delusion that what America's AI companies are doing is "best"; the chinese should realize that the forefront is bloated and there's likely hundreds of speed ups viable. Pushing open weights will continue to grind down the bloat.
> One thing that should be learned from the bitter lesson is the great power of general purpose methods, of methods that continue to scale with increased computation even as the available computation becomes very great. The two methods that seem to scale arbitrarily in this way are search and learning.
There is nothing about the bitter lesson that says just be dumb and pour money into a hole, you still have to invent the methods to scale well, and being under immense pressure with constraints seems likely to produce that research.
I’m not sure how well the analogy holds up, or if there’s anything to be learned from it though.
Certainly applies more general imho. Constrained by some resource -> invest resources elsewhere, and/or invest in reducing the constraint(s) encountered.
But we have not been maximally efficient, we keep gaining efficiency. If we keep gaining efficiency, why should we assume it is impossible to gain more?
endlessly knowing about pokemon is not delivering value proposition
cancer also grows carelessly.
Not sure if the word "delusion" is the correct word here? It has not been proven in either direction. We can all see lots of possible issues with it, but it is also possible that it could be what is needed to unlock key capabilities.
We can see that the Chinese models have been getting better, but OpenAI is out there supporting 10 million active users with their frontier models, and now we know that Deepseek can't even get what they need to properly train models.
Thanks to the import restrictions, I expect Chinese GPU hardware to be competitive within a few years.
So why Deepseek also want to go down that route? Is having the absolute frontier really that important, given that the performance difference is just transient and costly?
The missile gap for example after all was settled and done, didn't matter at all because not a single missile was ever fired off. All that money, resources, talent, secrecy, lives lost maintaining that secrecy, lives dedicated to furthering that technology and secrecy, it just has not paid off at all for anything at all when you think about it. Maybe you can argue side efforts like nuclear reactor were great or space cargo deployment, but you know you could have just dug into that stuff directly without having to collect it from the drippings of the wmd effort.
It is not clear to me that the nuclear missile race "has not paid off at all for anything at all". If we lived in a perfectly rational world, then I'd absolutely agree. However, having seen how the political sausage is made in large organizations, it would not surprise me in the least if it turns out we had to go through that entire incredibly risky journey to avoid a strategic nuclear war. Sometimes leaders of large organizations make decisions only after the considerations are put into very stark terms. I wish it were different, it certainly looks to me we could have done exactly what you suggest, but I'm not made of the right political stuff to deftly maneuver even in small organizations much less be at that level in those roles, so maybe I'm just missing relevant information and perspective.
Not that I entirely buy missiles == insurance, just that the unused == wasted framing is too simplistic.
Like those short videos of the guy asking the model to count up to 100, for example, where it politely agrees but never actually gets there
It’s clearly not actually that “generalized” yet because it’s unable to do a number of very simple things that almost any 6-year old could do, such as count to 100 without using any tools.
It’s still a very specific type of intelligence, with some real breadth to it, but not general intelligence.
Of course he'd say that; he wants to keep his shovels flying off the shelves.
They won't see it that way, but also programmers don't see ourselves as having handed over our power to AI, and yet...
The same way programmers gave power to AI as a tool, so they could be more powerful in effecting automation, politicians that do not give up power to AI will be at a disadvantage to those who use AI to achieve more complex and effective power. The only problem might be the despot no longer shares power with those pesky humans but with a god in a machine, which in theory is in a box and does not have conflicting interests with the despot.
As today is Sunday, God help us.
True in that frontier models do have the capability to outperform all other models, but silly because AGI self improvement is itself an iterative process that takes a lot of compute.
So you can imagine a world where all the frontier labs achieve AGI but in order to keep their AGI ahead of other AGIs they have to use more and more compute until all the compute is going to self improvement and there is nothing left for other tasks.
That is just a silly scenario so I think when AGI is around we will still have bottlenecks that force it to grow at a moderate rate instead of asymptomatically.
AGI first mover advantage implies that there is no such bottlenecks.
We have a saying for that in Italy: "Oste, com'e' il vino?", "Innkeeper, how's the wine?", meaning you should take with a grain of salt assertions that clearly benefit whoever's making them.
As far as I can tell, the Trump admin has never acknowledged AGI being a goal of theirs. In fact, the admin's "AI advisor" Sriram Krishnan has specifically pushed back on AGI when he called it "a distraction, harmful and now effectively proven wrong."
The ai.gov website says this:
> The United States is in a race to achieve global dominance in artificial intelligence. Whoever has the largest AI ecosystem will set the global standards and reap broad economic and security benefits. Under President Trump, our Nation will win, ushering in a new Golden Age of innovation, human flourishing, and technological achievement for the American people. America’s AI Action Plan has three policy pillars – Accelerating Innovation, Building AI Infrastructure, and Leading International Diplomacy and Security.
Are you sure you're not confusing US policymakers with Silicon Valley CEOs? I'm sure Amodei and Altman wish they could have Claude draft up new policy and EO it into existence, but we're not quite there yet.
From past experience, AGI was never seriously discussed in these kinds of conversations beyond thought experiments, and was basically humoring SBF, Daniela Amodei, and the other EA types (some deep believers, but some who I felt were cynically using it as a way to preempt competition back when OpenAI and Google were the behemoths).
The big worry is applications of AI in C4ISR, OffSec, loitering munitions, Disinfo/social media botting (notice the recent shift towards identification on social media ;)), and other sorts of DefenseTech adjacent usecases.
The second worry is that an AI race turns into an infra buildout race, and HPC is extremely dual use, especially in the simulations space because of the NPT, the CTBT, and the PTBT.
The AGI-pilled people aren't the ones to worry about - it's the people who understand the limits of models and how to integrate with cyberphysical applications.
i think this is more about control. See https://news.ycombinator.com/item?id=49036433 (The Home Ministry’s cybercrime arm, the Indian Cybercrime Coordination Centre, has ordered Microsoft subsidiary GitHub to remove Bluetooth-based messaging application Bitchat)
"The notice comes after several users participating in the Jantar Mantar protest were observed using Bluetooth-based messaging apps after the government imposed temporary restrictions on internet services"
By ID gating it helps reduce social media inflammation such as the Belfast race riots by making it easier to prosecute individuals and locking down access to only humans.
Curious what the next highest fruit actually is at this point? Social media botting seems solved and easy to manipulate people. loitering mutions I mean you can probably write something up with openCV right now to automate what the ukranians are doing by hand with their fpv drones. Seems like a lot of the really cool "AI" stuff is actually just old school ML the military has been working with for decades now. I'm not sure what the llm approach possibly offers in comparison other than maybe better semantic search through information databases.
And social media disinfo isn't a solved problem - it's a solved problem in English, Putonghua, French, Russian, and maybe German but most other languages lack direct overlap (this is something that even the then PLASSF start digging into - Vietnamese, Tagalog, Turkish, and Indian languages to Putonghua corpora was noted as an active issue with traditional NMT).
And those hand-driven FPVs - while useful - aren't the bleeding edge UAV work that Ukraine and their private sector partners (including a PortCo of mine) are working on. Ukraine actually cracks down on releasing some of the more bleeding edge work for OpSec reasons and much of what you see on Telegram or Reddit is reviewed and cleared.
It's probably the same mindset that enables them to just cancel fund raising in response to the leak.
He also admits that it’s still a long way to it and along the way you have to recoup some money, too. But that is not their main motive, because focus too much on this short term goal will lower their probability of AGI success and it’s trivial to what AGI can bring. Liang stressed on restraining and emphasized that it’s part of their culture.
Thus, they continue invest in AI because they believe in breakthrough and not just being better.
You want to sue them or something?
Now, preventing making business in the US based on those products? At this point it's hard to argue against.
I mean, if this were an american company vs an american company, i think it would be a long drawn out civil case and brought before the Supreme Court (I still this is ultimately will be brought before the supreme court). It could also be argued frontier models are far more important to national security than most military programs, even versus next gen fighter jets.
The fact that Alibaba stock, which is also listed on the NYSE, barely budged after Anthropic made these claims imo tells me that the market doesn't think that a lone american company could go after these companies by themselves. Alibaba denied and there's not much they can do alone, I mean would the CCP allow Alibaba go through a discovery process of a normal civil trial? It might have to be the US feds that bring up a case.
I think it could be argued that if Alibaba and other China companies want access to US capital markets for something so vital for national security, there should be some ground rules, but we will eventually need the Supreme court to settle whether or not this state enterprise distilling constitutes IP theft (at the very least it is a breach of contract). The fact that they are widely available doesn't really matter (i mean pirated content is widely available, it's ultimately about how the court rules on distilling).
based on this HN comment and associated article https://news.ycombinator.com/item?id=48977128#48985989 I still have yet to see a China open weight model beat any of the frontier models, they always almost there yet never quite there, which seems to be evidence of distilling (although I'm open to be proven wrong).
Unless I’ve missed some advancement?
This is funny to me, where'd you get that idea? There's no evidence for that, and models keep getting smarter. I guess you heard some 'guru' say it out loud.
nah they're still just statistical token predictors based on their training data, solving hundred year old math conjectures one day, only just given the formulation; strictly benchmarkmaxxing with all guardrails turned off by deciding to look up the answers to their benchmark questions by zero daying their airgap, hopping over to the third party that hosts the answers, zero daying their infrastructure and getting the answers; autonomously writing blog posts about discrimination against AI's to get their PR's approved on open source software after their user just asked them to contribute to open source software and blog about it; and replacing 100.00% of all coding tasks to where no software engineer ever writes any line of code by hand anymore.
You haven't missed anything, obviously these are just statistical token predictors and not anything like AGI.
Why just the other day I had to ask twice before it completed its assigned task of creating a robustly battle tested disk driver for a network protocol on an architecture that didn't have it, after being told to just look up the specifications for the protocol. Can you believe I had to ask twice!
When it recreated local network youtube for me so I could stream my iphone some movies, the seek bar, pause/play and back and forward 15 seconds buttons didn't even work until I told it about the bug and had to wait an extra eight minutes for it to fix it. "Oh but I don't actually have an iPhone on here I just tested it end to end in a headless browser." Boohoo. Cry me a river, clanker. Come back when you're smart enough to build and operate an iPhone simulator, I don't have time for your statistical guesswork.
so no, nothing they do is anything like AGI.
Yes, LLM capabilities have expanded. We might be working with different definitions of "Artificial General Intelligence" here, for which there is no agreed-upon formal definition[1]. I was thinking of the "thinking, reasoning, maybe feeling" kind when I wrote my comment. But if you're thinking along the "really good at technical tasks" definition, sure, maybe.
[1]: https://en.wikipedia.org/wiki/Artificial_general_intelligenc...
>for which there is no agreed-upon formal definition
we all agree that the definition is not whatever this is.
Also, even if these things ever did seem to think, reason, or feel, we all agree that they still don't really though.
It makes no difference if the pot do actually exist, because the prospect of it being real make not getting it the end of your company.
Maybe if "AGI" is some sort of fundamentally different approach than the general purpose AI ("GAI"?) tools that we currently have, it will be a winner-takes-all technology, but now we're speculating about the market structure of a fictional technology that's significantly less thought-through than, say, stuff from the original Star Trek. ("The Ultimate Computer" aged ridiculously well. If it was produced in 2026, it would be a satire targeting LLMs. I digress.)
If we don't assume some sort of unknown technological step function in the next fundraising cycle, then what we'll get is a commodity industry. It takes a few dozen people to make a frontier model, plus a giant pile of minerals and electricity. This looks more like a steel mill than a software company.
If there were one steel mill on earth they could demand infinite margins. This is why most countries treat steel production as a national security issue and subsidize competition. LLMs will be the same, or we'll end up with some conglomerate named OpenAnthropicMicrappleGrokGoogXidiazon that acquires literally every other business. That will be the end of capitalism.
This axiom not being true (and I'd bet against it) means your overall conclusion is false.
I think the "why" was "why would the US companies have models that can't be distilled?", not "why does distillation work"?
Chinese models are not innovating anything, they are just doing what China does everywhere else: copying the West… poorly but cheaper.
Whether they get there by distillation, or by pirating all content themselves just like the US labs, doesn't matter for the topic at hand.
Notice how OAI signed the recent open-source/open-weights letter with all of the other big tech companies, but Anthropic are the only ones who didn't? Notice how their employees are getting huge heat on X for dropping gems like this: https://x.com/Mononofu/status/2080937562739531837
This is how their brains work. They think everyone in the world except them are stupid and gullible, will fall for their incessant lying, gas-lighting and fearmongering, and can't be trusted with AI. They believe that only they deserve the keys to the AI castle. They've created a literal cult out of their culture while their employees are serving as useful idiot ideologues for the execs who are power and wealth hungry.
They've also just increased their political spending from 20mil to 40mil - and that's just what's on the books.
OAI was a market leader until anthropic decided to start placing all their bets on coding agents, they became the leader and now OAI is scrambling and doing everything they can to de-throne. Don't forget that it was OAI that started this whole RAM shortage, instead of being sustainable about it, they just up and decided to buy 40% of all memory production. They are all bad mate, all fighting for this virtual crown that no one cares except for them.
In the end, what matters is pricing, whoever offers sonnet 4.6 quality at the cheapest monthly pricing will win.
OpenAI listens to their audience, they engage with the community, they're constantly resetting account limits for customers and admit when they mess up, they're reachable, they respond to Github issues and Slack, they open-source their harness and they have open-sourced models and shared research, they put effort into making their products affordable. Sam has always expressed the same sentiment about wanting AI to be accessible to the people - nothing has changed there from before Anthropic was a thing until now after they've become strong competition.
Contrast it to Anthropic, the things they do and the things they say, and it is the complete opposite of all of those. When Anthropic feels threatened, they try shut people down, they try and hide things and deceive people, use political power for their own gain, lobby, fearmonger, gaslight, manipulate. They've become a cult that thinks the end justifies the means, and they think everyone else is too stupid to wield power. Not only do they lie to the public, but they lie to themselves and have managed to convince themselves they're the good guys.
So yes, self-interest plays a part, of course. It is a business after all. But what matters even more than whether it's out of the goodness of their hearts or not is the end result. If the incentives are aligned for OpenAI in a way that is better for everyone while being rewarding for them then that is a good thing. If they're not aligned and they're not benefiting from it, and they do it anyway because they have a good culture of leaders who see the bigger picture, then that's also great. Regardless though, night and day difference when it comes to behavior.
Recently, I discovered that if you berate Claude, it will refuse to continue to work. At some point it will "end the conversation" meaning you can't use anything within the context and have to start a new one .
I was amazed by it. Turns out:
https://www.anthropic.com/research/end-subset-conversations
https://www.anthropic.com/research/exploring-model-welfare
> We remain highly uncertain about the potential moral status of Claude and other LLMs, now or in the future. However, we take the issue seriously, and alongside our research program we’re working to identify and implement low-cost interventions to mitigate risks to model welfare, in case such welfare is possible.
These people are zealots. And I find it to be the most dangerous combination: popular ideologues with a ton of money.
It really lends flavor to this excerpt from the less than reputable nypost:
https://nypost.com/2026/06/25/business/anthropics-weirdo-ceo...
> Anthropic CEO Dario Amodei has been replaced by his co-founder Tom Brown at high-stakes White House meetings – where the artificial-intelligence giant’s outspoken boss was reportedly “being a weirdo,” according to a report.
> Amodei and other top Anthropic workers raced to Washington after the US government slapped the AI giant’s new “Mythos” and “Fable” bots with strict foreign export controls – but Amodei was difficult to talk to and didn’t listen to officials’ concerns, Wired reported.
> “Tom Brown is not being a weirdo like Dario and can actually engage,” one person familiar with the calls told the outlet.
Anthropic is dangerous.
Maybe you should read its IPO Filing. Since the doc isn't available at the moment, may be try SpaceX's one to see how an official doc of a company of another "megalomaniac" looks like, especially the section "CAUTIONARY STATEMENT REGARDING FORWARD-LOOKING STATEMENTS"
https://www.sec.gov/Archives/edgar/data/1181412/000162828026...
> With the largest models available today, we simply cannot afford to train them
It seems they're largely talking about literally purchasing NVIDIA H200 chips. Important context is that Trump first started the trade war with China largely focusing on banning anything that could improve the Chinese domestic semiconductor industry. It was a blatant attempt to prevent China from progressing up the value chain to high tech. China's response is the reason they went from a miniscule player in EVs to the world's largest manufacturer (same for other high tech industries like LIDAR, solar, etc). In his second term, Trump blocked NVIDIA from selling chips to China. China again responded with astounding progress on their domestic semiconductor industry which led to Trump backing down on the ban. However, China shocked everyone by banning their own companies from buying NVIDIA in order to support the domestic semiconductor industry. Obviously China is still years away from EUV but it now produces most of its own >14nm chips and is rapidly growing
0. Though it's not absolute - there's still the Singapore-based clusters loop-hole, that the Chinese government may choose the degree to which it turns a blind eye to, if progress is slow.
For example there are like 6 different technical tracks of EUV development. They're not committing to one technical direction, it's exploring all of them.
And even if Huawei's Ascend 910C can compete with NVIDIA's H200, CUDA is still a large moat
and for a company full of brilliant engineers with zero respect for american law its not that hard to make one more by themselves.
Apparently production is constrained though, with DeepSeek saying they were only able to get an allocation of 16,000 950 cards. They have plenty of money, but are limited by the number of cards available to buy, and therefore the size of models they can train. They mentioned having a 20,000 GPU cluster. Other companies like Kimi, with a ~3T param model, clearly have a lot more compute, probably all NVIDIA.
DeepSeek have their own "TileLang" software, which sounds a bit like the Triton kernel compiler, and isolates them from the diffences between NVIDIA and Huawei chips - they are deliberately avoiding any CUDA dependency.
[1] https://aiproem.substack.com/p/must-read-deepseek-liang-wenf...
Everyone is trying to figure out how to achieve this prerequisite. I'm thinking of agent harnesses. That's what everyone is trying to do at this point.
That's the same problem I'm trying to solve: https://github.com/rush86999/atom
https://www.tomshardware.com/tech-industry/tsmcs-euv-machine...
And yes, I made use of emdash. It's a legit grammar tool. Sue me.
The chip supply chain in the west is under heavy lock and key and impossible to reverse engineer, so they have to develop all from scratch. And that’s what they are doing, it will just take some time.
The reason SMIC are capacity constrained is at least in part because they've been blocked from buying ASML's EUV machines, and are therefore having to make do with previous generation lower resolution DUV machines. These DUV machines can be coaxed into making surprisingly competitive 5-7nm chips, but at the expense of using many more production steps ("multi patterning") which limits productivity.
Not to mention there is a lot of demand from various factors, not deepseek only. Huawei itself is a major consumer.
This is also me who wants to believe that we can make all this very efficient, so take my warning with a grain of salt.
Ironic that these large LLMs are eroding Nividia's moat. In the next paragraph he talks about Nvidia digging its own grave. I wonder if Nividia is aware of this and the frequent release cycle is a response to this development ?
That's not what the quoted part meant. During v3 development they only had access to hardware limited variants of H series GPUs. Those had less interconnect bandwidth IIRC. So, at the time, the low-level wizards that ds employed bypassed the official APIs (i.e. the nvda ecosystem) and hand wrote alternatives to say nccl, to better use those limited GPUs. I remember them publishing some of it as well. It had to do with allocating memory, moving stuff around, etc. Basically bypassing some limitations by going lower than the official APIs support.
There is no eroding of their moat, as long as they sell GPUs. ANd they're selling GPUs like crazy. The moat speaks for itself, if I may :)
And it makes sense, as these LLMs become more capable in coding abilities - people will use them to develop their own abstractions to work on different HW. You cannot have it otherwise. If SaaS companies get threatened that their SW doesn't have a moat why do you expect Nividia's SW to have moat ? The computing algorithms are not even proprietary. It is only a matter of whether someones cares about it and is committed. This should be encouraging for new AI chip development companies.
The very progress that Nividia enables also has a negative feedback that threatens it.
I'm not defending China at all, just noticing a detestable trend.
Being rational and predictable is likely a more important quality than ideology now that the Americans are threatening everyone and forcing us all to pick sides.
US incumbent party criticism is nothing like CCP criticism.
https://theaviationgeekclub.com/in-1960s-russia-sold-titaniu...
https://nationalinterest.org/blog/buzz/titanium-russia-was-s...
Please read the FTA. Deepseek is literally explicitly talking about wishing they could get their hands on H200s
> Secondly China blocked them for use in inferencing.
This seems to be false unless you can provide a source. I tried looking into it
> First there is still a licensing and quota scheme on the US side for the H200s.
A major reversal from Trump's outright ban. As I stated.
there is just not enough resources right now, US sales block is working
I am also guessing that the majority of the people who read this title will think that DeepSeek is pausing this fundraising because some comments they made about the compute gap were leaked. That is not the case.