I think the way to parse the current title "DeepSeek pause fundraise after comments on compute gap to US leaked (transcript) [pdf]" is that there was a leak that DeepSeek will pause fundraising because they perceive there is a compute gap with the US.
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.
Maybe: "Leaked Deepseek transcripts reveal plan to pause fundraising due to compute gap"
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.
yep. the word they use is probably 克制 or self-restraint. no need to raise so much cash if you can't use it.
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. Once you're in such a comfortable position that you'll keep making lots of cash regardless of whether you do well or not, there is little incentive to make good decisions, let alone take risks. This is known as the "curse of Oil" and is also what Intel's decline is attributed to.
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.
I skimmed the doc and my impression is that your second listed interpretation -- DeepSeek is pausing investment because of a leak -- is the more correct one.
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?
Because money isn’t free and if you know there is a chokehold in supply, why raise money ar current valuation when things are getting better by the day?
All the Chinese reporting I see point to the second (majority) interpretation. Liang being furious about his private investor talk leaked online is the news here.
Those could be subsequent developments, but that's not what the linked transcript was about. The transcript was a discussion of the DeepSeek founder (Liang Wenfeng) with investors, and he does not mention any leaks, or any frustration. He simply says that he is constrained by the supply of cards, and he has no problem of getting funding, but has no reason to raise further funding because he can't transform the cash into cards.
> 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.
I wanted to post this which explains the wording but I thought the transcript was more interesting. Sorry. Maybe mods can help me to put what follows as auxiliary link. I don't know how.
Most of that is paywalled, but this one paragraph in the Bloomberg article suggests it might be more to do with investors leaking information:
"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."
amusingly ive been working on ultra sparse llm inference/ training/ model design because nature loaths a dense graph/matrix and cause i think it shoukd be possible. i actually stood up a 20-25 percent faster than sota causal fast attention kernel yesterday, will be standing up cuda/metal/armv8 kernels too and thats gonna be fun.
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
I am not much of a math person, but if we look at the how the brain is wired, we see that the dendrites (the inputs) of a neuron are hundreds of micrometers in length, and the axons (the outputs) are millimeters, and very rarely can stretch to tens of centimeters. So they can sample only a tiny amount of internal state, and affect a much larger, but usually still small output.
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
i definitely will be doing some drop of some faster attention kernels in the next few weeks.
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.
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
"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."
I wonder if it would be viable for the Chinese to pursue a huge buildout of less sophisticated logic fabs. My experience with GPGPU is that it seems to intensely skew towards memory bandwidth, and has a much lower compute intensity than graphics (by which I mean rasterization and shaders).
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.
this bodes well for continuing to refine smaller models and open sourcing them.
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.
I was gonna say, this just puts more pressure to deliver ground breaking research with limited resources. And if history teaches us anything it’s that scarcity produces ingenuity.
> 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.
Right, and if you come up with an efficiency gain that makes scaling better, e.g. a 50% reduction in required compute. Or even asymptotic improvements e.g. moving from quadratic to linear. Then you're much much better off.
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.
It reminds me a bit of the tyranny of the rocket equation. You can always scale your fuel to get a little more delta V, with ever diminishing returns. …but for something like a DEEP space/interstellar mission, it almost always pays to wait a few more years for a faster propulsion system because you’ll get there fastest by always delaying your launch and chasing better technology.
I’m not sure how well the analogy holds up, or if there’s anything to be learned from it though.
The implication here is that the only gains left to be had are from scale. That we are already maximally efficient. If that's true, then how has OpenAI repeatedly bragged about reducing the cost of their models by orders of magnitude? (And DeepSeek Flash even more so, of course.)
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?
> "There's a delusion that what America's AI companies are doing is "best""
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.
They can't get hardware because the US has put restrictions on how much can be sold to China. There is not a technical or know-how limitation, but political. Deepseek could otherwise write some checks to NVidia for what they want.
Thanks to the import restrictions, I expect Chinese GPU hardware to be competitive within a few years.
Here's something I really don't understand: If as alleged Chinese open weight models are catching up with US anyway, and the performance is near US frontier model level but Chinese can do it with a fraction of cost, and eventually AI model will be commodified, wouldn't that means that the billion or even trillion dollars that US labs spend have only diminishing returns and the lead is only temporary?
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?
U.S. policymakers believe that even if the gap is small—like six months to a year—whoever reaches AGI first (whatever that means) could gain such an overwhelming advantage over their perceived adversary that it would effectively kneecap them. (You can look at the kinds of things they mention—cyber, WMDs—to get a sense of what they mean.) Jensen Huang disagrees and has said AI is a marathon.
They'd have to use the gap though to actually kneecap them in that time, or else it is just shoveling money into the fire.
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.
> 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.
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.
I feel like if you showed current frontier models to someone 10 years ago, they'd probably call it AGI. Does AGI have a clear definition or is it just a pair of goalposts on wheels?
All the AGI (which is a misnomer, ASI is preferred) talk is about the moment of singularity, which is where the growth at the third derivative is increasing, so the gap (in the absolute) between first place and others, even if it's one month, will be ever increasing as time progresses.
I also have a hard time believing this narrative because limiting factors prevail such as compute and energy. These constraints will take many many years to overcome.
It also assumes that "intelligence" is a limiting factor. It's hard to imagine there are any domains today, or almost any, which are bottlenecked on intelligence.
An ASI wouldn’t need to produce or invent anything. It can just take over humanity’s digital infrastructure (which, thanks to the efforts of the past 30 years, now controls most of what humans need) and use it to blackmail humanity into giving it whatever it wants. Security is so laughably bad that even today’s AIs already do that by accident in the course of solving some benchmark.
Reaching AGI would come with so many ethical issues, it feels so absurd that actual adults seem to actually believe it’s something that must be chased as fast as possible
Actually history shows that despotism or power in one single figure is a concept that waned with the increasing complexity of the world. One single despot could govern easily over matters of a small tribe or kingdom, but as states became more complex power was delegated to court, bureaucracy and even the bourgeoise. As the sovereign demanded ever more power it eventually needed to distribute power to the people, so it could tap its numbers. Napoleon’s schools for the peasants were so he could have a better and more numerous corps of officers than his rivals, which had a smaller pool of recruitment.
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.
> Jensen Huang disagrees and has said AI is a marathon.
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.
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.
> U.S. policymakers believe that even if the gap is small—like six months to a year—whoever reaches AGI first [...]
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.
> U.S. policymakers believe that even if the gap is small—like six months to a year—whoever reaches AGI first...
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.
>notice the recent shift towards identification on social media
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"
>C4ISR, OffSec, loitering munitions, Disinfo/social media botting
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.
VLMs are the name of the game now in the UAV space, as well as distilled models that can perform well on edge compute (eg. Processing sat images directly on the sat instead of dealing with C2 latency).
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.
The paper discusses this, and is refreshingly honest. They do not expect nor aim to be a top player. They're not aiming for a path to world domination, but a path forward to continuing to play their part in pursuing the development and advancement of LLMs - nothing more, and nothing less. They mention that commercialization is, at best, a distant goal. Given DeepSeek already is commercialized, I assume that refers more to commercialization in the sense of making substantial profits and the like.
It's probably the same mindset that enables them to just cancel fund raising in response to the leak.
Deepseek is funded by their hedge fund, high flyer. They intentionally cap their token prices to basically recoup server costs. The meeting transcript describes it as a moral commitment, that they don’t care about trends like image and video generation, and world model “hype”. They only care about reasoning, chain of thought and continuous learning.
there's a lot of propganda from these state backed enterprises. I think the fraction of the cost label is debatable given the evidence of mass gpu smuggling through third parties like Singapore which China can't exactly openly admit to. Unless of course we're talking about distilling, which is probably a lot cheaper than training a model from scratch (there's also the fact that labour is still relatively cheap in China compared to the US which may or may not matter e.g. Anthropic claim against Alibaba
> The campaign allegedly used nearly 25,000 fraudulent accounts to run 28.8 million exchanges with Claude between April and June 2026
(although their campaign could have been in part or all automated via agents, not sure)
Is there really a valid basis for claims like "state backed enterprises"? My understanding is that's no different than claiming that datacentres in America are "state backed" since they get things like big breaks on property taxes.
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).
My understanding after reading Liang’s comments during the investment meeting is that Liang firmly believes in AGI and he bets everything to reach goal. Once it reaches AGI, the game would flip totally. How he didn’t paint it out, and with the potential severe impact on the labor and consumer market, the true economic impact is difficult to predict. Liang is more like religious about this goal.
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.
It's odd to me to watch so many very rich humans speedrun the destruction of humanity. Like, is there a world where we hit AGI and it actually works out for us?
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.
The whole point the guy is making in the transcript is that they're taking a different strategy from the US labs, one where they focus on smaller models and cost control, and maintain as top priority the work stream that they think will get them to AGI (not every product fad that comes along).
There is an immense pot of gold at the end of this rainbow and if the theories about ASI are in the general correct direction, only one winner will get it.
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.
AI's already commoditized, but the fundraising plans for the US labs assumes a winner take all endgame where one lab will pull arbitrarily ahead of everyone else. I have no idea why DeepSeek is making that bad assumption now too. Maybe the investors have drunk the Kool-Aid.
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.
They are not catching up to US models. The only Chinese models that attain a modicum of competence are all, sooner or later, are discovered to be trained by exploiting US models (in fact Deepseek itself admitted so about 1 year back).
Chinese models are not innovating anything, they are just doing what China does everywhere else: copying the West… poorly but cheaper.
Chinese models most likely are distillations of frontier models with tricks for subpar hardware. If you want to be ahead of the us labs you need to spend billions for pretraining from scratch.
If that is the case, it means one thing only - US labs don't have moat whatsoever and their expectation to have trillion dollar valuation is just laughable.
Thank you. These documents are a particularly valuable insight into the kind of thinking going on at DeepSeek. The part about how inference should be priced at a level that's enough to return capex in 10 months, but no higher, is really interesting. Liang Wenfeng simply has different motivations than we're used to over here in the West.
Did you also read the transcripts of meetings of Anthropic/OpenAI's investors?
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"
Not sure why people keep lumping OAI and Anthropic together. Really, Anthropic are the evil ones. You can make the case OAI are evil too if you want, but Anthropic are very clearly significantly worse and they aren't even in the same ballpark.
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.
You think OAI are doing all that out of the goodness of their hearts?
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.
The behavior of the two companies and their culture are entirely different.
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.
> Objectively speaking, if I can spend two billion this year, it would indicate that our procurement department has achieved outstanding performance. The main gap between us and the United States lies in resources, while the disparity in personnel is minimal—there is virtually no difference, as we are essentially the same team of people, possibly from China.
> 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
Wow, that is fascinating, I didn't realize China was now blocking foreign chips, lol. It's not a definitive indicator, but I feel that doesn't bode well for US dominance in this area -- when your competitor thinks they'd be helping _you_ by using your resources, that's not great.
I don't know that it's clear that the motivation is that it's "helping" their competitors directly. Maybe the motivation is "if we rely on these, then the next time a US president arbitrarily decides to block us from buying them, we won't have the infrastructure already in place to be able to work around it". It seems more betting on a shorter-term cost with less uncertainty in the long term rather than a shorter-term win with a lot harder to quantify risks in the long term.
This was the Chinese government burning the boats[0]. Beyond the symbolism and ensuring everyone's commitment, they want to direct the firehose of AI money towards a local champion (Huawei). Money, and experience bourne of being in the trenches developing, manufacturing, deploying and debugging on actual workloads will supercharge how quickly Huawei get good, compared to when they had to fairly compete against a well-resourced Nvidia
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.
The money and resources aren't going to Huawei exclusively, but to a whole range of companies. This selection changes over time as new companies show promise, or previously selected companies prove themselves to be incompetent.
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.
They've achieved self-sufficiency in >14nm chips in remarkable timing. Unfortunately for DeepSeek, it's the <14nm chips that are needed for massive training tasks. I wouldn't be surprised if China backs down and lets them purchase the chips given that they are still years away from being able to make them themselves. Either that or the gov't steps in and forces them to share resources
And even if Huawei's Ascend 910C can compete with NVIDIA's H200, CUDA is still a large moat
Bypassing the CUDA moat is, in fact, one of the major tasks Deepseek set for itself. Their efforts in this area are likely one of the main reasons for their slow release cadence, culminating in their v4 inference setup that runs on Huawei Ascend chips.
I think the point is more than by banning the NVIDIA hardware they are forcing local development of potentially competitive hardware, basically giving Huawei a subsidy or leg-up.
I don't think his pitch when asking money from investors should mean too much for us. He wants the funds, and he needs to point to a deficiency that those funds should cover. We cannot know for sure but he may be exaggerating, or let's just say, talking strategically.
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.
Curious what the fundamental limit on Huawei's capacity is. China has shown if nothing else they know how to scale when they want to. If it came down to just building more of what they know how to do, it would be happening. Is there more to it?
Their yields on high performance chips that could do training is really bad, and they aren’t getting more of the outdated ASML machines that they could use to scale up even with bad yields. It will still take China a few years or a decade to build out the tech needed to fab high performance chips economically on their own.
They can already make chips economically. Yields are worse than TSCM but good enough since they no longer need to pay the "Qualcomm tax". Huawei's phones are profitable. The issue is rather that Chinese capacity comes from a low quantity, and scaling capacity while simultaneously indigineousizing parts takes time — years. Foreign capacity was too good and too cheap so they never succeeded in scaling capacity, because the demand for Chinese fabs wasn't there. Now the demand for domestic capacity is there and they're scaling like crazy, like 100-200% growth per year. But demand still far outpaces capacity. Fabs are hard to build. You need many more years of 200% growth to even approach the demand.
And yes, I made use of emdash. It's a legit grammar tool. Sue me.
They can make chips economically or they can make high performance chips. They just can’t do both at the same time yet.
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.
>> As you can understand, during V3 training, NVIDIA GPUs were still used, but the NVIDIA ecosystem was no longer employed.
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 ?
> Ironic that these large LLMs are eroding Nividia's moat.
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 :)
They talk about developing and using TileLang which they ported to Huawei's 950 GPUs.
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.
Give it 5 years for China to have it's own ASML. Nothing big bang is going to happen in 5 years or even a decade from now, execpt for a few more hypes, deep corrections and the political drama. AGI is not a destination, but a journey. There are no winners.
It's funny that you mention this because with the current US administration it works in a similar fashion... see Anthropic not cooperating with the US military and getting their new shiny model "paused" few weeks later (and officials like Hegseth being pretty open about it beforehand, signalling to them that criticizing the US admin/not cooperating will hurt their business: https://xcancel.com/SecWar/status/2027507717469049070 )
I'm not defending China at all, just noticing a detestable trend.
And it’s not even reading between the lines and being a conspiracy theorist. The current U.S. regime has made it abundantly clear that they will gladly operate in bad faith.
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.
Perhaps there is an opportunity for China to close the compute gap by renting compute from hyperscalers through a complex web of shell entities similarly to how the US procured titanium for the SR-71 during the Cold War.
Trump reversed course on the NVIDIA ban. It's now China that is blocking their companies from buying NVIDIA chips. So the shell entities would be to get around Chinese, not USian restrictions
That's not true. First there is still a licensing and quota scheme on the US side for the H200s. Secondly China blocked them for use in inferencing. Thirdly Chinese companies don't want them for training because newer chips are more cost effective.
If this is true it almost sounds like DeepSeek is following the Anthropic playbook of trying to pressure the local government into aligning with their corporate agenda through scare tactics. So I wouldn't be surprised if Liang Wenfeng "disappears" for a little while from the public eye in a few weeks.
I think the way to parse the current title "DeepSeek pause fundraise after comments on compute gap to US leaked (transcript) [pdf]" is that there was a leak that DeepSeek will pause fundraising because they perceive there is a compute gap with the US.
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.
Maybe: "Leaked Deepseek transcripts reveal plan to pause fundraising due to compute gap"
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.
yep. the word they use is probably 克制 or self-restraint. no need to raise so much cash if you can't use it.
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.
I shouldn't win too hard, because then I'll lose?
Yes. Once you're in such a comfortable position that you'll keep making lots of cash regardless of whether you do well or not, there is little incentive to make good decisions, let alone take risks. This is known as the "curse of Oil" and is also what Intel's decline is attributed to.
Catching up without overtaking
Yep. Classic “giant series A, company gets amazing office space” vibes.
Problem of the local maxim, extreme dependency is the same as evolutionary pressure to specialize for it, eventually the process owns you.
is this similar to Meta starting to rent out own compute as they cant seem to do much with it and monetising it is much better ROI?...
Sorry to barge in here. I couldn't find a good place to place
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)
email hn@ycombinator.com
Yeah, wouldn't it makes sense to increase fundraising, so as to acquire more compute to close the gap?
I guess they likely can't right now due to no hardware available, either cause of bans or already all booked.
I gather this is the whole point of the original article. It seems to have been pulled, so I can't confirm.
At what point do the AI companies start building their own hardware, I wonder.
Other companies are investing and hardware design and manufacturing, so they won't duplicate the effort given the national drive it would be a waste.
Perhaps it's akin to a mineshaft gap?
> if the title is conflating
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.
I skimmed the doc and my impression is that your second listed interpretation -- DeepSeek is pausing investment because of a leak -- is the more correct one.
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?
Because money isn’t free and if you know there is a chokehold in supply, why raise money ar current valuation when things are getting better by the day?
Maybe because if you need more money to catch up with your competitors than anticipated then your ROIC is lower and your valuation changes?
All the Chinese reporting I see point to the second (majority) interpretation. Liang being furious about his private investor talk leaked online is the news here.
e.g. https://x.com/_FORAB/status/2081034500101017616?s=20
Those could be subsequent developments, but that's not what the linked transcript was about. The transcript was a discussion of the DeepSeek founder (Liang Wenfeng) with investors, and he does not mention any leaks, or any frustration. He simply says that he is constrained by the supply of cards, and he has no problem of getting funding, but has no reason to raise further funding because he can't transform the cash into cards.
he's pausing because of the leak. the transcript content has nothing to do with it.
Thank you! That is indeed how I read it.
I wanted to post this which explains the wording but I thought the transcript was more interesting. Sorry. Maybe mods can help me to put what follows as auxiliary link. I don't know how.
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...
https://archive.ph/zpIrG
Most of that is paywalled, but this one paragraph in the Bloomberg article suggests it might be more to do with investors leaking information:
"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."
amusingly ive been working on ultra sparse llm inference/ training/ model design because nature loaths a dense graph/matrix and cause i think it shoukd be possible. i actually stood up a 20-25 percent faster than sota causal fast attention kernel yesterday, will be standing up cuda/metal/armv8 kernels too and thats gonna be fun.
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
I am not much of a math person, but if we look at the how the brain is wired, we see that the dendrites (the inputs) of a neuron are hundreds of micrometers in length, and the axons (the outputs) are millimeters, and very rarely can stretch to tens of centimeters. So they can sample only a tiny amount of internal state, and affect a much larger, but usually still small output.
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
Look forwarding your future releases
i definitely will be doing some drop of some faster attention kernels in the next few weeks.
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
Curiosity:
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.
thx for the kind response!
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
>Fable is $1/month instead of $100/month
will Anthropic (or OpenAI) lowers their price, or increase margin (to justify valuation)
Article grabbed at random that provides some more context (tho could use more):
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."
I wonder if it would be viable for the Chinese to pursue a huge buildout of less sophisticated logic fabs. My experience with GPGPU is that it seems to intensely skew towards memory bandwidth, and has a much lower compute intensity than graphics (by which I mean rasterization and shaders).
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.
this bodes well for continuing to refine smaller models and open sourcing them.
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.
I was gonna say, this just puts more pressure to deliver ground breaking research with limited resources. And if history teaches us anything it’s that scarcity produces ingenuity.
http://www.incompleteideas.net/IncIdeas/BitterLesson.html
> 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.
Right, and if you come up with an efficiency gain that makes scaling better, e.g. a 50% reduction in required compute. Or even asymptotic improvements e.g. moving from quadratic to linear. Then you're much much better off.
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.
It reminds me a bit of the tyranny of the rocket equation. You can always scale your fuel to get a little more delta V, with ever diminishing returns. …but for something like a DEEP space/interstellar mission, it almost always pays to wait a few more years for a faster propulsion system because you’ll get there fastest by always delaying your launch and chasing better technology.
I’m not sure how well the analogy holds up, or if there’s anything to be learned from it though.
The implication here is that the only gains left to be had are from scale. That we are already maximally efficient. If that's true, then how has OpenAI repeatedly bragged about reducing the cost of their models by orders of magnitude? (And DeepSeek Flash even more so, of course.)
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?
So what? There are physical and economic ceilings on dumb computation scaling.
americas tech stack always ends up bloated. not everything is worth learning.
endlessly knowing about pokemon is not delivering value proposition
cancer also grows carelessly.
> "There's a delusion that what America's AI companies are doing is "best""
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.
They can't get hardware because the US has put restrictions on how much can be sold to China. There is not a technical or know-how limitation, but political. Deepseek could otherwise write some checks to NVidia for what they want.
Thanks to the import restrictions, I expect Chinese GPU hardware to be competitive within a few years.
Here's something I really don't understand: If as alleged Chinese open weight models are catching up with US anyway, and the performance is near US frontier model level but Chinese can do it with a fraction of cost, and eventually AI model will be commodified, wouldn't that means that the billion or even trillion dollars that US labs spend have only diminishing returns and the lead is only temporary?
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?
U.S. policymakers believe that even if the gap is small—like six months to a year—whoever reaches AGI first (whatever that means) could gain such an overwhelming advantage over their perceived adversary that it would effectively kneecap them. (You can look at the kinds of things they mention—cyber, WMDs—to get a sense of what they mean.) Jensen Huang disagrees and has said AI is a marathon.
They'd have to use the gap though to actually kneecap them in that time, or else it is just shoveling money into the fire.
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.
> 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.
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.
Agreed, that was a really fallacious point. It’s like saying buying car insurance didn’t pay off because you didn’t get in a wreck.
Not that I entirely buy missiles == insurance, just that the unused == wasted framing is too simplistic.
Also why do we want to kneecap?
I feel like if you showed current frontier models to someone 10 years ago, they'd probably call it AGI. Does AGI have a clear definition or is it just a pair of goalposts on wheels?
They would until you showed them the jagged edges.
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
So does AGI just mean infallible?
It's a carrot on a stick to make investors/the US government give @sama whatever he wants.
The goalposts are moving, and their speed accelerates all the time.
Third derivative though. The goal posts must have thrust not just acceleration.
All the AGI (which is a misnomer, ASI is preferred) talk is about the moment of singularity, which is where the growth at the third derivative is increasing, so the gap (in the absolute) between first place and others, even if it's one month, will be ever increasing as time progresses. I also have a hard time believing this narrative because limiting factors prevail such as compute and energy. These constraints will take many many years to overcome.
It also assumes that "intelligence" is a limiting factor. It's hard to imagine there are any domains today, or almost any, which are bottlenecked on intelligence.
An ASI wouldn’t need to produce or invent anything. It can just take over humanity’s digital infrastructure (which, thanks to the efforts of the past 30 years, now controls most of what humans need) and use it to blackmail humanity into giving it whatever it wants. Security is so laughably bad that even today’s AIs already do that by accident in the course of solving some benchmark.
> Jensen Huang disagrees and has said AI is a marathon.
Of course he'd say that; he wants to keep his shovels flying off the shelves.
Reaching AGI would come with so many ethical issues, it feels so absurd that actual adults seem to actually believe it’s something that must be chased as fast as possible
I don't forsee politicians in either country handing over their power to AIs, ever. Unless nukes are dropped, "the other side" will catch up.
> I don't forsee politicians in either country handing over their power to AI
They won't see it that way, but also programmers don't see ourselves as having handed over our power to AI, and yet...
Politicians have control over their power. Programmers don't. Compare with how politicians never vote to reduce their income.
Actually history shows that despotism or power in one single figure is a concept that waned with the increasing complexity of the world. One single despot could govern easily over matters of a small tribe or kingdom, but as states became more complex power was delegated to court, bureaucracy and even the bourgeoise. As the sovereign demanded ever more power it eventually needed to distribute power to the people, so it could tap its numbers. Napoleon’s schools for the peasants were so he could have a better and more numerous corps of officers than his rivals, which had a smaller pool of recruitment.
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.
> Jensen Huang disagrees and has said AI is a marathon.
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.
It’s kind of true but also kind of silly.
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.
> U.S. policymakers believe that even if the gap is small—like six months to a year—whoever reaches AGI first [...]
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.
> U.S. policymakers believe that even if the gap is small—like six months to a year—whoever reaches AGI first...
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.
>notice the recent shift towards identification on social media
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"
>C4ISR, OffSec, loitering munitions, Disinfo/social media botting
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.
VLMs are the name of the game now in the UAV space, as well as distilled models that can perform well on edge compute (eg. Processing sat images directly on the sat instead of dealing with C2 latency).
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.
The paper discusses this, and is refreshingly honest. They do not expect nor aim to be a top player. They're not aiming for a path to world domination, but a path forward to continuing to play their part in pursuing the development and advancement of LLMs - nothing more, and nothing less. They mention that commercialization is, at best, a distant goal. Given DeepSeek already is commercialized, I assume that refers more to commercialization in the sense of making substantial profits and the like.
It's probably the same mindset that enables them to just cancel fund raising in response to the leak.
Deepseek is funded by their hedge fund, high flyer. They intentionally cap their token prices to basically recoup server costs. The meeting transcript describes it as a moral commitment, that they don’t care about trends like image and video generation, and world model “hype”. They only care about reasoning, chain of thought and continuous learning.
One word: Focus.
there's a lot of propganda from these state backed enterprises. I think the fraction of the cost label is debatable given the evidence of mass gpu smuggling through third parties like Singapore which China can't exactly openly admit to. Unless of course we're talking about distilling, which is probably a lot cheaper than training a model from scratch (there's also the fact that labour is still relatively cheap in China compared to the US which may or may not matter e.g. Anthropic claim against Alibaba > The campaign allegedly used nearly 25,000 fraudulent accounts to run 28.8 million exchanges with Claude between April and June 2026 (although their campaign could have been in part or all automated via agents, not sure)
Is there really a valid basis for claims like "state backed enterprises"? My understanding is that's no different than claiming that datacentres in America are "state backed" since they get things like big breaks on property taxes.
Ownership and corporate governance is much more state-led via GGIFs as well as mandated party oversight depending on the size of company.
All that and so what? Fact is the Chinese have several near peer models, they've released the weights and they are widely available.
You want to sue them or something?
> You want to sue them or something?
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).
My understanding after reading Liang’s comments during the investment meeting is that Liang firmly believes in AGI and he bets everything to reach goal. Once it reaches AGI, the game would flip totally. How he didn’t paint it out, and with the potential severe impact on the labor and consumer market, the true economic impact is difficult to predict. Liang is more like religious about this goal.
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.
They want to achieve AGI first because, once it is achieved, no one knows what the world will look like.
It's odd to me to watch so many very rich humans speedrun the destruction of humanity. Like, is there a world where we hit AGI and it actually works out for us?
I doubt this is the case. It should be common knowledge at least among the people building these things that a true AGI isn’t possible with LLMs.
Unless I’ve missed some advancement?
It's understood that LLMs have limitations and people are working on "the next thing" to try and make it to real AGI, e.g. Yann LeCun.
Many people have tried before, but the bitter lesson has come for them all.
>Unless I’ve missed some advancement?
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.
But this will not be a singular event. And like humans it does not mean that the smartest makes the best decisions.
More like a singularity event.
Who actually believes this nonsense?
The whole point the guy is making in the transcript is that they're taking a different strategy from the US labs, one where they focus on smaller models and cost control, and maintain as top priority the work stream that they think will get them to AGI (not every product fad that comes along).
There is an immense pot of gold at the end of this rainbow and if the theories about ASI are in the general correct direction, only one winner will get it.
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.
Why? If you can reach ASI without ASI, then why can’t multiple companies reach ASI on parallel tracks?
I think that statement is vacuous true for all magical thinking.
AI's already commoditized, but the fundraising plans for the US labs assumes a winner take all endgame where one lab will pull arbitrarily ahead of everyone else. I have no idea why DeepSeek is making that bad assumption now too. Maybe the investors have drunk the Kool-Aid.
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.
> and eventually AI model will be commodified
This axiom not being true (and I'd bet against it) means your overall conclusion is false.
Eventually you'll have a model you can't distill, at which point the frontier labs will take off.
Why?
They are not catching up to US models. The only Chinese models that attain a modicum of competence are all, sooner or later, are discovered to be trained by exploiting US models (in fact Deepseek itself admitted so about 1 year back).
Chinese models are not innovating anything, they are just doing what China does everywhere else: copying the West… poorly but cheaper.
Chinese models most likely are distillations of frontier models with tricks for subpar hardware. If you want to be ahead of the us labs you need to spend billions for pretraining from scratch.
If that is the case, it means one thing only - US labs don't have moat whatsoever and their expectation to have trillion dollar valuation is just laughable.
The moat is the compute.
The compute for training, or for inference?
Yes
The repository was force-pushed so the link doesn't work anymore, but the file is still available at: https://github.com/demo-zexuan/liang-wenfeng-investor-meetin...
Thank you. These documents are a particularly valuable insight into the kind of thinking going on at DeepSeek. The part about how inference should be priced at a level that's enough to return capex in 10 months, but no higher, is really interesting. Liang Wenfeng simply has different motivations than we're used to over here in the West.
Everything in this transcript reads so very different from what megalomaniacs in charge of Anthropic/OAI have to say
Did you also read the transcripts of meetings of Anthropic/OpenAI's investors?
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...
Not sure why people keep lumping OAI and Anthropic together. Really, Anthropic are the evil ones. You can make the case OAI are evil too if you want, but Anthropic are very clearly significantly worse and they aren't even in the same ballpark.
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.
You think OAI are doing all that out of the goodness of their hearts?
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.
The behavior of the two companies and their culture are entirely different.
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.
no, my company that wants a 1T$ IPO is more ethical than your company that wants a 1T$ IPO!!!!
A lot of people here are saying that chinese open weight models are state-sponsored. If it's so, why would deepseek seek a fundraising?
Some deleted it (again?). But the comments of Liang could be found somewhere else.
[1] https://aiproem.substack.com/p/must-read-deepseek-liang-wenf...
> Objectively speaking, if I can spend two billion this year, it would indicate that our procurement department has achieved outstanding performance. The main gap between us and the United States lies in resources, while the disparity in personnel is minimal—there is virtually no difference, as we are essentially the same team of people, possibly from China.
> 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
Wow, that is fascinating, I didn't realize China was now blocking foreign chips, lol. It's not a definitive indicator, but I feel that doesn't bode well for US dominance in this area -- when your competitor thinks they'd be helping _you_ by using your resources, that's not great.
I don't know that it's clear that the motivation is that it's "helping" their competitors directly. Maybe the motivation is "if we rely on these, then the next time a US president arbitrarily decides to block us from buying them, we won't have the infrastructure already in place to be able to work around it". It seems more betting on a shorter-term cost with less uncertainty in the long term rather than a shorter-term win with a lot harder to quantify risks in the long term.
This was the Chinese government burning the boats[0]. Beyond the symbolism and ensuring everyone's commitment, they want to direct the firehose of AI money towards a local champion (Huawei). Money, and experience bourne of being in the trenches developing, manufacturing, deploying and debugging on actual workloads will supercharge how quickly Huawei get good, compared to when they had to fairly compete against a well-resourced Nvidia
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.
The money and resources aren't going to Huawei exclusively, but to a whole range of companies. This selection changes over time as new companies show promise, or previously selected companies prove themselves to be incompetent.
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.
They've achieved self-sufficiency in >14nm chips in remarkable timing. Unfortunately for DeepSeek, it's the <14nm chips that are needed for massive training tasks. I wouldn't be surprised if China backs down and lets them purchase the chips given that they are still years away from being able to make them themselves. Either that or the gov't steps in and forces them to share resources
And even if Huawei's Ascend 910C can compete with NVIDIA's H200, CUDA is still a large moat
Bypassing the CUDA moat is, in fact, one of the major tasks Deepseek set for itself. Their efforts in this area are likely one of the main reasons for their slow release cadence, culminating in their v4 inference setup that runs on Huawei Ascend chips.
I think the point is more than by banning the NVIDIA hardware they are forcing local development of potentially competitive hardware, basically giving Huawei a subsidy or leg-up.
I don't think his pitch when asking money from investors should mean too much for us. He wants the funds, and he needs to point to a deficiency that those funds should cover. We cannot know for sure but he may be exaggerating, or let's just say, talking strategically.
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.
Curious what the fundamental limit on Huawei's capacity is. China has shown if nothing else they know how to scale when they want to. If it came down to just building more of what they know how to do, it would be happening. Is there more to it?
Their yields on high performance chips that could do training is really bad, and they aren’t getting more of the outdated ASML machines that they could use to scale up even with bad yields. It will still take China a few years or a decade to build out the tech needed to fab high performance chips economically on their own.
Can't they use GlobalFoundries, Intel or TSMC in addition to their own fabs?
Chinese companies are effectively embargoed from using them for <14nm.
Sounds like an invasion of Taiwan could fix that problem for them.
They can already make chips economically. Yields are worse than TSCM but good enough since they no longer need to pay the "Qualcomm tax". Huawei's phones are profitable. The issue is rather that Chinese capacity comes from a low quantity, and scaling capacity while simultaneously indigineousizing parts takes time — years. Foreign capacity was too good and too cheap so they never succeeded in scaling capacity, because the demand for Chinese fabs wasn't there. Now the demand for domestic capacity is there and they're scaling like crazy, like 100-200% growth per year. But demand still far outpaces capacity. Fabs are hard to build. You need many more years of 200% growth to even approach the demand.
And yes, I made use of emdash. It's a legit grammar tool. Sue me.
They can make chips economically or they can make high performance chips. They just can’t do both at the same time yet.
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.
Huawei chips need advanced 3d packaging in order to keep up. Since the process is too complex the yield is still bad.
Not to mention there is a lot of demand from various factors, not deepseek only. Huawei itself is a major consumer.
>> As you can understand, during V3 training, NVIDIA GPUs were still used, but the NVIDIA ecosystem was no longer employed.
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 ?
> Ironic that these large LLMs are eroding Nividia's moat.
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 :)
They talk about developing and using TileLang which they ported to Huawei's 950 GPUs.
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.
Give it 5 years for China to have it's own ASML. Nothing big bang is going to happen in 5 years or even a decade from now, execpt for a few more hypes, deep corrections and the political drama. AGI is not a destination, but a journey. There are no winners.
Are you saying there are no deep corrections and the political drama in China?
One has to be careful when pointing out problems in China, lest such criticism be confused with criticism of the Party's policies.
It's funny that you mention this because with the current US administration it works in a similar fashion... see Anthropic not cooperating with the US military and getting their new shiny model "paused" few weeks later (and officials like Hegseth being pretty open about it beforehand, signalling to them that criticizing the US admin/not cooperating will hurt their business: https://xcancel.com/SecWar/status/2027507717469049070 )
I'm not defending China at all, just noticing a detestable trend.
And it’s not even reading between the lines and being a conspiracy theorist. The current U.S. regime has made it abundantly clear that they will gladly operate in bad faith.
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.
Yes the Trump administration acts a lot like the CCP and it is despicable.
This seems like when Indians were super excited about China having castes too. Then, it turned out that they were egregiously incomparable.
US incumbent party criticism is nothing like CCP criticism.
It is really sad how India doesn't try to be better but tries to prove China is just as bad.
There is no war in Ba Sing Se
Perhaps there is an opportunity for China to close the compute gap by renting compute from hyperscalers through a complex web of shell entities similarly to how the US procured titanium for the SR-71 during the Cold War.
https://theaviationgeekclub.com/in-1960s-russia-sold-titaniu...
https://nationalinterest.org/blog/buzz/titanium-russia-was-s...
Trump reversed course on the NVIDIA ban. It's now China that is blocking their companies from buying NVIDIA chips. So the shell entities would be to get around Chinese, not USian restrictions
That's not true. First there is still a licensing and quota scheme on the US side for the H200s. Secondly China blocked them for use in inferencing. Thirdly Chinese companies don't want them for training because newer chips are more cost effective.
> Thirdly Chinese companies don't want them for training because newer chips are more cost effective.
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.
If this is true it almost sounds like DeepSeek is following the Anthropic playbook of trying to pressure the local government into aligning with their corporate agenda through scare tactics. So I wouldn't be surprised if Liang Wenfeng "disappears" for a little while from the public eye in a few weeks.