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.
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."
"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.
The real question is whether it's easier to improve the software side instead. There are likely a lot more optimizations possible in terms of model architecture, and if there is a compute bottleneck, then it's going to put a lot of pressure on Chinese labs to address the problem using more efficient designs.
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.
> "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.
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.
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?
> 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.
This sounds like someone has swallowed too many sci-fi novels. Both the US and China can already nuke each other, they don't need AGI to that. The reason they don't is because it's war is bad for everyone involved.
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.
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
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.
> 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.
> 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.
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.
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.
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.
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.
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?
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.
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.
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.
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.
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.
I don't disagree with you, but also sometimes I feel like Anthropic really are huffing their own gas.
Recently, I discovered that if you berate Claude, it will refuse to continue to work. At some point it will "end the conversation" meaning you can't use anything within the context and have to start a new one .
> We remain highly uncertain about the potential moral status of Claude and other LLMs, now or in the future. However, we take the issue seriously, and alongside our research program we’re working to identify and implement low-cost interventions to mitigate risks to model welfare, in case such welfare is possible.
These people are zealots. And I find it to be the most dangerous combination: popular ideologues with a ton of money.
It really lends flavor to this excerpt from the less than reputable nypost:
> Anthropic CEO Dario Amodei has been replaced by his co-founder Tom Brown at high-stakes White House meetings – where the artificial-intelligence giant’s outspoken boss was reportedly “being a weirdo,” according to a report.
> Amodei and other top Anthropic workers raced to Washington after the US government slapped the AI giant’s new “Mythos” and “Fable” bots with strict foreign export controls – but Amodei was difficult to talk to and didn’t listen to officials’ concerns, Wired reported.
> “Tom Brown is not being a weirdo like Dario and can actually engage,” one person familiar with the calls told the outlet.
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"
> 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.
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
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.
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.
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.
The production capacity constraint seems to come from SMIC who make the Ascend processors for Huawei. Huawei's memory comes from CXMT who seem to have plenty of capacity, with Apple looking to buy memory from them. Huawei then combines processors and memory into chiplets similar to what NVIDIA does with their GPUs.
The reason SMIC are capacity constrained is at least in part because they've been blocked from buying ASML's EUV machines, and are therefore having to make do with previous generation lower resolution DUV machines. These DUV machines can be coaxed into making surprisingly competitive 5-7nm chips, but at the expense of using many more production steps ("multi patterning") which limits productivity.
The main goal is AGI, and the underlying prerequisite theme is continuous learning.
Everyone is trying to figure out how to achieve this prerequisite. I'm thinking of agent harnesses. That's what everyone is trying to do at this point.
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.
>> 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.
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.
17 replies →
Yeah, wouldn't it makes sense to increase fundraising, so as to acquire more compute to close the gap?
3 replies →
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?
[dead]
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.
1 reply →
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."
9 replies →
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.
The real question is whether it's easier to improve the software side instead. There are likely a lot more optimizations possible in terms of model architecture, and if there is a compute bottleneck, then it's going to put a lot of pressure on Chinese labs to address the problem using more efficient designs.
2 replies →
And that's why OpenAI bought all the future ram contracts
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.
7 replies →
> "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.
1 reply →
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.
6 replies →
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?
7 replies →
> 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.
> 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.
4 replies →
This sounds like someone has swallowed too many sci-fi novels. Both the US and China can already nuke each other, they don't need AGI to that. The reason they don't is because it's war is bad for everyone involved.
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.
2 replies →
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
1 reply →
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.
3 replies →
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.
1 reply →
> 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.
> 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.
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.
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.
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.
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.
1 reply →
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?
2 replies →
They want to achieve AGI first because, once it is achieved, no one knows what the world will look like.
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?
8 replies →
But this will not be a singular event. And like humans it does not mean that the smartest makes the best decisions.
1 reply →
Who actually believes this nonsense?
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?
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?
2 replies →
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?
3 replies →
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.
3 replies →
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.
For all practical uses of the word, that's exactly what we mean by "catch up".
Whether they get there by distillation, or by pirating all content themselves just like the US labs, doesn't matter for the topic at hand.
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
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.
2 replies →
I don't disagree with you, but also sometimes I feel like Anthropic really are huffing their own gas.
Recently, I discovered that if you berate Claude, it will refuse to continue to work. At some point it will "end the conversation" meaning you can't use anything within the context and have to start a new one .
I was amazed by it. Turns out:
https://www.anthropic.com/research/end-subset-conversations
https://www.anthropic.com/research/exploring-model-welfare
> We remain highly uncertain about the potential moral status of Claude and other LLMs, now or in the future. However, we take the issue seriously, and alongside our research program we’re working to identify and implement low-cost interventions to mitigate risks to model welfare, in case such welfare is possible.
These people are zealots. And I find it to be the most dangerous combination: popular ideologues with a ton of money.
It really lends flavor to this excerpt from the less than reputable nypost:
https://nypost.com/2026/06/25/business/anthropics-weirdo-ceo...
> Anthropic CEO Dario Amodei has been replaced by his co-founder Tom Brown at high-stakes White House meetings – where the artificial-intelligence giant’s outspoken boss was reportedly “being a weirdo,” according to a report.
> Amodei and other top Anthropic workers raced to Washington after the US government slapped the AI giant’s new “Mythos” and “Fable” bots with strict foreign export controls – but Amodei was difficult to talk to and didn’t listen to officials’ concerns, Wired reported.
> “Tom Brown is not being a weirdo like Dario and can actually engage,” one person familiar with the calls told the outlet.
Anthropic is dangerous.
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...
[dead]
> 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.
2 replies →
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
3 replies →
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.
1 reply →
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...
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.
3 replies →
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.
1 reply →
Can't they use GlobalFoundries, Intel or TSMC in addition to their own fabs?
4 replies →
The production capacity constraint seems to come from SMIC who make the Ascend processors for Huawei. Huawei's memory comes from CXMT who seem to have plenty of capacity, with Apple looking to buy memory from them. Huawei then combines processors and memory into chiplets similar to what NVIDIA does with their GPUs.
The reason SMIC are capacity constrained is at least in part because they've been blocked from buying ASML's EUV machines, and are therefore having to make do with previous generation lower resolution DUV machines. These DUV machines can be coaxed into making surprisingly competitive 5-7nm chips, but at the expense of using many more production steps ("multi patterning") which limits productivity.
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.
The main goal is AGI, and the underlying prerequisite theme is continuous learning.
Everyone is trying to figure out how to achieve this prerequisite. I'm thinking of agent harnesses. That's what everyone is trying to do at this point.
That's the same problem I'm trying to solve: https://github.com/rush86999/atom
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.
>> 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?
No. I was referring to the global AI industry.
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?
Idiots say idiotic things.
> when you're the Chinese CEO, but all your phraseology comes from reading fortune cookies
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.
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.
1 reply →
Yes the Trump administration acts a lot like the CCP and it is despicable.
There is no war in Ba Sing Se
Is there a mirror? Getting 404...
From another comment: https://github.com/demo-zexuan/liang-wenfeng-investor-meetin...
link broken!
so this is why they still not release deepseek r2 yet
there is just not enough resources right now, US sales block is working
[dead]
[dead]
[dead]
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.