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.
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?
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.
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 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
"The Hangzhou AI lab has told prospective investors in its second fundraising round that it is suspending the deal, people familiar with the matter told Bloomberg on Saturday, days after remarks attributed to founder Liang Wenfeng about US-China AI competition circulated widely online."
And:
"Tencent's technology outlet published a 118-item version covering AGI strategy, chip supply, pricing, and retention. In it, Liang reportedly framed China's disadvantage as an arithmetic problem rather than a talent one: "The biggest gap between us and the US is in resources.""
"The specifics were unusually candid. Liang is said to have told investors he needed 200,000 Huawei 950 chips to train a frontier model but received 16,000, adding that "Huawei's problem is still insufficient capacity" and expecting the crunch to last at least three years. He also floated narrowing the gap with US labs to three to six months using a fraction of their computing."
this 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.
> "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.
> U.S. policymakers believe that even if the gap is small—like six months to a year—whoever reaches AGI first [...]
Who are these policymakers? 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.
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)
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.
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).
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.
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.
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.
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.
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.
> 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
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.
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.
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.
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.
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.
The title is certainly a great conflation. Any seeker of capital would want to regroup after an unfiltered leak of this magnitude, if for no other reason than to secure the forum from future leaks. The comments about the unlikelihood of enormous future profits were at least as consequential with regard to capital investment as anything else that was said.
There's quite a bit of confidential information in the doc about the company and how it's positioning itself going forward to compete with US labs. I'd imagine they're not happy at all with this being leaked and are withholding investment as a punitive measure.
Not to mention the other interpretation seems illogical -- why would you pause fundraising if your perception was that you lacked resources compared to your competitors?
e.g. https://x.com/_FORAB/status/2081034500101017616?s=20
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
"The suspension stemmed in part from Liang’s frustration over online reports about his comments to investors during his first financing deal"
The part of the transcript I'd seen floating around online was this part from around 1 hour 26 min:
"With the largest models available today, we simply cannot afford to train them. Even if we spent all five hundred billion yuan, we still wouldn't be able to do so. Even if we could accumulate the resources, we wouldn't have the means to utilize them. The current largest model requires approximately 800 billion activations; domestically, we are still at a scale of several dozen billion activations, and even the largest domestic model may only require several dozen billion activations—a difference of an order of magnitude. To train a model of the same size as an AI system, we would need around 50,000 GB300 GPUs or Huawei 950 GPUs, totaling two hundred thousand cards. This is merely training; research has not yet been considered. Therefore, the biggest gap between us and the United States lies in resources."
i genuinely think these models should be like 0.1 percent sparse for same capabilities we associate with them today, but theres no sane way to do that with extent tools. i built the right core tech for that in 2014 when there wasnt a market, but now there is and the experimentation velocity is wild.
amusingly llms really have a hard time using my simple apis because its not in distribution array programs. but i literally stood up cpu custom memory format and micro kernel for dense causal attention in less than 24-36 hours and outperforms the equivalent fused ggml/llama cpp fast oath by like 20-25 percent
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
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."
There's a delusion that what America's AI companies are doing is "best"; the chinese should realize that the forefront is bloated and there's likely hundreds of speed ups viable. Pushing open weights will continue to grind down the bloat.
> One thing that should be learned from the bitter lesson is the great power of general purpose methods, of methods that continue to scale with increased computation even as the available computation becomes very great. The two methods that seem to scale arbitrarily in this way are search and learning.
Not sure if the word "delusion" is the correct word here? It has not been proven in either direction. We can all see lots of possible issues with it, but it is also possible that it could be what is needed to unlock key capabilities.
We can see that the Chinese models have been getting better, but OpenAI is out there supporting 10 million active users with their frontier models, and now we know that Deepseek can't even get what they need to properly train models.
Thanks to the import restrictions, I expect Chinese GPU hardware to be competitive within a few years.
So why Deepseek also want to go down that route? Is having the absolute frontier really that important, given that the performance difference is just transient and costly?
They won't see it that way, but also programmers don't see ourselves as having handed over our power to AI, and yet...
Who are these policymakers? 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.
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.
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.
Maybe if "AGI" is some sort of fundamentally different approach than the general purpose AI ("GAI"?) tools that we currently have, it will be a winner-takes-all technology, but now we're speculating about the market structure of a fictional technology that's significantly less thought-through than, say, stuff from the original Star Trek. ("The Ultimate Computer" aged ridiculously well. If it was produced in 2026, it would be a satire targeting LLMs. I digress.)
If we don't assume some sort of unknown technological step function in the next fundraising cycle, then what we'll get is a commodity industry. It takes a few dozen people to make a frontier model, plus a giant pile of minerals and electricity. This looks more like a steel mill than a software company.
If there were one steel mill on earth they could demand infinite margins. This is why most countries treat steel production as a national security issue and subsidize competition. LLMs will be the same, or we'll end up with some conglomerate named OpenAnthropicMicrappleGrokGoogXidiazon that acquires literally every other business. That will be the end of capitalism.
This axiom not being true (and I'd bet against it) means your overall conclusion is false.
Not to mention there is a lot of demand from various factors, not deepseek only. Huawei itself is a major consumer.
> 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
And even if Huawei's Ascend 910C can compete with NVIDIA's H200, CUDA is still a large moat
I'm not defending China at all, just noticing a detestable trend.
Being rational and predictable is likely a more important quality than ideology now that the Americans are threatening everyone and forcing us all to pick sides.
US incumbent party criticism is nothing like CCP criticism.
https://theaviationgeekclub.com/in-1960s-russia-sold-titaniu...
https://nationalinterest.org/blog/buzz/titanium-russia-was-s...