Nvidia Dismisses China AI Threat, Says DeepSeek Still Needs Its Chips 73
Nvidia has responded to the market panic over Chinese AI group DeepSeek, arguing that the startup's breakthrough still requires "significant numbers of NVIDIA GPUs" for its operation. The US chipmaker, which saw more than $600 billion wiped from its market value on Monday, characterized DeepSeek's advancement as "excellent" but asserted that the technology remains dependent on its hardware.
"DeepSeek's work illustrates how new models can be created using [test time scaling], leveraging widely-available models and compute that is fully export control compliant," Nvidia said in a statement Monday. However, it stressed that "inference requires significant numbers of NVIDIA GPUs and high-performance networking." The statement came after DeepSeek's release of an AI model that reportedly achieves performance comparable to those from US tech giants while using fewer chips, sparking the biggest one-day drop in Nvidia's history and sending shockwaves through global tech stocks.
Nvidia sought to frame DeepSeek's breakthrough within existing technical frameworks, citing it as "a perfect example of Test Time Scaling" and noting that traditional scaling approaches in AI development - pre-training and post-training - "continue" alongside this new method. The company's attempt to calm market fears follows warnings from analysts about potential threats to US dominance in AI technology. Goldman Sachs earlier warned of possible "spillover effects" from any setbacks in the tech sector to the broader market. The shares stabilized somewhat in afternoon trading but remained on track for their worst session since March 2020, when pandemic fears roiled markets.
"DeepSeek's work illustrates how new models can be created using [test time scaling], leveraging widely-available models and compute that is fully export control compliant," Nvidia said in a statement Monday. However, it stressed that "inference requires significant numbers of NVIDIA GPUs and high-performance networking." The statement came after DeepSeek's release of an AI model that reportedly achieves performance comparable to those from US tech giants while using fewer chips, sparking the biggest one-day drop in Nvidia's history and sending shockwaves through global tech stocks.
Nvidia sought to frame DeepSeek's breakthrough within existing technical frameworks, citing it as "a perfect example of Test Time Scaling" and noting that traditional scaling approaches in AI development - pre-training and post-training - "continue" alongside this new method. The company's attempt to calm market fears follows warnings from analysts about potential threats to US dominance in AI technology. Goldman Sachs earlier warned of possible "spillover effects" from any setbacks in the tech sector to the broader market. The shares stabilized somewhat in afternoon trading but remained on track for their worst session since March 2020, when pandemic fears roiled markets.
Can't help but wonder... (Score:5, Insightful)
Re:Can't help but wonder... (Score:5, Informative)
It is a theoretical measure of the value of a company. So if the share price drops by 20% then the market capitalization will also drop by 20%.
So it's quite dramatic to say Nvidia "lost" $X, but nvidia company did not lose money. They made money when they sold the shares to the open market, took that money into the company treasury, and now when shareholders sell on the secondary market, *they* lost money.
That is my understanding anyways.
Re: (Score:3)
I think you are mostly correct, but the company undoubtedly does own a significant number of its own shares. Also it can borrow money using shares as collateral, and that is cheaper when the shares are more valuable.
Re:Can't help but wonder... (Score:4, Interesting)
Re: (Score:2)
Interesting explanation, thanks.
>> The owners of NVidia shares lost phantom wealth today
If I had owned some shares and sold them for a loss today, my wealth would have diminished physically.
Re: (Score:2)
If you actually be
Re: (Score:2)
Companies do not own a number of their shares unless they've created a share pool for things like employee stock purchase plans. And, no, they cannot use shares in their company as loan collateral. Sometimes companies can issue "convertible notes" which are basically loans that convert into shares automatically if the company gets into financial trouble. They are popular with banks that have to maintain certain capital cushions. NVidia didn't lose anything today. The owners of NVidia shares lost phantom wealth today. If, however, there is a genuine now AI model in town that will significantly reduce demand for chips (and electricity, and all of the other related infrastructure), the future cash flows of many companies have just been reduced and hence why the shares are trading for a lower price.
Companies can also acquire and hold their own stock in a number of different ways, the most common is through a share buyback. This is called "Treasury Stock" [investopedia.com] on the balance sheet and it's fairly common. The linked article goes into greater detail.
Re: (Score:2)
Re: (Score:2)
Re: (Score:2)
Market capitalization may sound like "capitalism" but the two concepts are not related.
Every company in any economic system (socialism, whatever) has a market cap: the number of shares times the value of each share. Yeah, even companies in China and Russia have that.
Capitalism is an economic system built on the principles that people and businesses have the right to make and keep profits, as opposed to a system like socialism, which says that all money should be share with everyone according to their need.
Re: (Score:2)
That only applies to companies which have shares. And companies having shares is a capitalism system. Modern socialism is built on top of capitalism, as modern financial models are based around not only the concept of capital, but around the trade of it. Even socialist systems. Trying to separate them is nothing but silly.
Ironically, a lot of people do not understand that socialism, the desire to share with one's fellow humans, is a basic human desire and need, and what has allowed us to evolve into the pos
Re: (Score:2)
Every company has shares, whether they call them that or not.
If a company is a sole proprietorship, then the single owners owns 100% of the shares of the company.
If it's a partnership, each owner owns 50% of the shares.
The shares in these cases are informal, but they still exist. "Shares" as we know them are nothing more than fractional ownership of companies. And this concept applies to socialism as well as any other economic system.
And while people do have a desire to share with their fellow humans, they
Re: (Score:2)
Except companies do not at all necessarily have shares. The system of organizing companies in that way is the main point of capitalism.
And the deeper, more profound desire in humans is to help and protect their fellow humans. Caring for oneself comes behind that in 99.99% of the cases. This is the reality that Libertarians and capitalists ignore, and why socialism is winning out.
Re: (Score:2)
Are you saying that some companies are not owned by anyone at all? Because that would be the only way that there are no "shares" either formal or informal. What would be an example of such a company?
If you are correct that the more profound desire of humans is to help someone else, above taking care of oneself, then why does every single religion go to such trouble to teach people not to be selfish? If selflessness were the natural instinct, no one would feel the need to teach people to be selfless. Why do
Good time to buy (Score:2)
Re: (Score:3)
Not really. It is called a "bubble" for a reason. Lots of size, nothing of substance in there.
Re: (Score:2)
Yes and no. It's the value of the shares at the current market rate for a share. But that doesn't mean that's what it would really sell for [unrealized]. Your house has an estimated value as well but that isn't the real value, the real [realized] value is only determined if you succeed in selling it. At any point something could happen to make people less willing to pay high values for Nvidia OR for your house.
This is why people that understand these things STRONGLY oppose taxing people on [unrealized] gain
Re: (Score:2)
If you'd taxed them before the change then suddenly they now owe taxes on the former theoretical gain AND they've lost the $60 billion on paper.
That is easily avoided by selling shares to pay taxes. Of course if gains were taxed properly, the companies would be paying dividends. This is a USA-specific problem where companies retain profits, rather than distributing them to shareholders, as a means of tax avoidance. Reform the tax system, and such market distortions are reduced. If a company wants to "retain" that profit for growth investment, they can for example offer new shares in lieu of dividends to shareholders.
Re: (Score:3)
"That is easily avoided by selling shares to pay taxes."
Ahem...
"But those taxes, those have to paid and the only way to pay them would be to now sell shares at the current low price [realize that price] to get the money. Not only is it the worst time to sell but the more people who sell at that price the more REAL that price becomes whereas it otherwise might just be a brief fluctuation."
Re: Can't help but wonder... (Score:3)
Re: Can't help but wonder... (Score:2)
America Got Played By NGreedia (Score:2)
Does NGreedia really care who buys their chips? I think not. They will blow as much smoke up your ass as possible just to sell chips. To anyone. It doesn't matter.
So where is the back door?
Re:America Got Played By NGreedia (Score:5, Insightful)
Does NGreedia really care who buys their chips? I think not. They will blow as much smoke up your ass as possible just to sell chips. To anyone. It doesn't matter.
So where is the back door?
Everyone seems to be missing the likelihood that China is almost certainly developing their own rival to the NVIDIA tech (probably stealing some of it), and even if it doesn't match NVIDIA's performance, they'll sell it so cheap that it'll undercut NVIDIA drastically. They only question is how far are they out from their own home grown stuff for AI. NVIDIA is coming off like a drug dealer that's convinced no one can walk away from their stash. I'd bet otherwise.
Re: (Score:2)
It will take China many years just to catch up to the state of the art today, and by then that state will have moved forward significantly. This particular feat is impressive but it is a software accomplishment, and it is based on open source materials.
Re: (Score:1)
The chips that China is expected to be able to make themselves anyway.
And so the sanctions preventing them from getting the top of the line stuff is precisely why they will now be able to make their own chips of similar performance and use their superior software to completely dominate.
TLDR;
The thing everyone said would happen happened. Sanctions made China stronger and America weaker.
Re: (Score:2)
No, they are bypassing the export restrictions at scale, which is its own problem.
Re: (Score:2)
First of all this is Slashdot, tf is wrong with you. Second, if you refer to Nvidia's claim that export controlled GPU's weren't needed it is in TFS and is just that, a claim.
Re: (Score:2)
Nvidia claimed that their bigger GPUs are only needed for inference and that this allowed training using export friendly gpus... they are running a service offering INFERENCE, so clearly they have bypassed the restrictions on the bigger GPUs. Nothing made up about it.
And how is it actually better if they don't need the export controlled GPUs to do the far bigger and more expensive task of training the model?
Re: (Score:3)
Re: (Score:2)
NVidia Should Worry (Score:2)
Re: (Score:2)
Why should NVidia be opposed to selling to competing customers? They saw the AI boom coming 10 years before anybody else, and now they are reaping the benefits of their foresight.
Xi cheat? (Score:2)
The Chinese gov't has a history of subsidizing products and industries in order to gain market share and bankrupt foreign competitors.
These alleged low-cost AI services could very well also be subsidized, but I haven't seen evidence to conclude such either way. Anyone?
Maybe we'll never know, there could be stealth server farms that accountants and spies don't know about. "Why does that cow's mooing sound like Stephan Hawking?"
Re: (Score:2)
You also realize that people are running versions of it on a Raseberry PI.
Re: Xi cheat? (Score:1)
Re: (Score:3)
"you don't understand the difference between Deepseek-R1 and distill models, which are NOT 'a version of deepseek'"
Of course the distilled models are a version of deepseek but then deepseek is nothing but a collection of weights which are basically just a form of lossy encoding of the training data. In that sense they are all just a form of the training data... much of that data is likely output from competing open and closed solutions so really deepseek is just a form of those and those are really just a f
Re: Xi cheat? (Score:1)
Re: (Score:3)
Is my child a version of me or of my wife? The answer is both.
Also DeekSeek-R1 will have used common training data to both those models AND synthetic data produced by them in its training and all of the above will be using Llama 3's training data. So those would be more like my child if my wife were my half-sister that was the spawn of my older sister and my dad.
They learned how to make these models in Arkansas ya'll. Surely you can see that.
Re: (Score:1)
They probably have a proprietary prototype or internal version that they either will to sell later, or to have others work on enhancements to their open version so their internal version(s) can swipe the code. They want to control standards and conventions at least.
Re: (Score:2)
"You do realize the software is open source so anyone can run it."
How does that conflict with an assertion that they are trying to bankrupt foreign competitors? Making the foreign competitors products open source and free whilst they gain the advantages of any community contributions seems like a great way to cut into those competitors profits.
"You also realize that people are running versions of it on a Raseberry PI."
Toys, there were already toys with similar capabilities.
"These alleged low-cost AI service
Re: (Score:2)
They literally hold a controlling interest in all corporations. It's an authoritarian state, yes Trumps political enemies call Trump a dictator hyperbolically but that isn't reality... Xi is literally a totalitarian dictator. You are comparing apples to bricks.
Re: (Score:2)
Oh yeah, you sure showed me, they totally changed that policy a few years ago! But it doesn't much matter who owns the company on paper does it?
It's still a totalitarian regime that makes people who disagree disappear except for one of their organs appearing here or there or as a posed dissected corpse displayed in a tour around the world to make a point to their enemies.
Re: (Score:2)
Propoganda? I found that China indicated they planned to change the policy on foreign investment in 2021 for SOME classes of company. The prior 51% policy is well known and undisputed and I haven't actually seen evidence of any specific company actually being 100% foreign owned.
Stating the former policy isn't propaganda. China used to require half the babies in the nation be murdered via the one child policy. Supposedly they changed the policy... that doesn't wash the blood of billions of babies from the CC
Re: (Score:2)
"You do realize the software is open source so anyone can run it."
How does that conflict with an assertion
Because it is open source, i.e. the model and training data publicly available, it will soon become obvious if it required far greater server power to create than they claimed.
Re: (Score:2)
Its weights are open, though, which is still something.
Re: (Score:2)
The weights and what is alleged to be information about how some of it works is released but no, it isn't fully released like you seem to think. They went through way too much trouble censoring and manicuring the dataset to let someone go behind them and fix it.
So you want to feed a Chinese AI system? (Score:2, Interesting)
So literally every AI company on earth (Score:1, Troll)
If you think they're not doing something about that then, well oh you sweet summer child...
The question is will Nvidia be able to stay far enough ahead of the competition that there are no viable alternatives.
I think they might. They have a pretty well-known strategy for doing just that, there's only a handful of engineers who have the skill to make that kind of hardware and they just take all the profits they make from being in a monopolistic posi
Re: (Score:2)
Re: (Score:2)
The question is will Nvidia be able to stay far enough ahead of the competition that there are no viable alternatives.
There will come a time when Nvidia is no longer the leader. It's just a matter of when. However, that time could be one year or it could be 10 years.
In the old days somebody would come along with a bunch of venture capital and give somebody like AMD enough money to compete with Nvidia for those engineers. But nowadays because the venture capitalists both own majority stock in just about every company on Earth the last thing they're going to do is have the companies they own competing with each other.
The venture capitalists got out of Nvidia a long time ago. The largest owners of NVDA are Vanguard (8.3%), BlackRock (5.6%), State Street Corporation (3.7%), and Jensen Huang (3.5%). Mark Stevens, who's on the Nvidia board, is probably the largest venture capitalist who put money ($1 million in 1993) into Nvidia, but his shares are about one-hundredth of on
Re: (Score:2)
"The question is will Nvidia be able to stay far enough ahead of the competition that there are no viable alternatives."
No because people forget about 'good enough.' Staying ahead might be great but eventually the CCP still gets 'good enough.' What is freely available was already 'good enough' to comb through the mass scale training data used for deepseek and censor/taint it.
Re: (Score:2)
You can take an AMD GPU and do all the shit that Nvidia GPUs do with enough work AMD has an equivalent to CUDA. Do enough programming and you can translate the function calls and do all the work you want to do on it.
But there is no way in hell much of anyone is going to do that for anything. So AMD is not a viable replacement. It's not good enough
You see the same thing with adobe's
Re: (Score:2)
"But there is no way in hell much of anyone is going to do that for anything."
But we aren't talking about much of anyone, we are talking about a nation with billions of slaves... err citizens who don't want their sister's kidney to disappear and don't want to disappear themselves.
I understand what you mean. In that department the advances are so great that there are huge swatches of applications where a generation or two behind isn't 'good enough' anymore. But look at the office suit, look at the desktop fo
Nvidia is still fabless (Score:2)
Re: (Score:1)
They should be worried (Score:3)
It is obvious that human-level intelligence can be achieved using an object that weights only a few pounds and only burns a few watts. And it is not made of GPUs, it appears to be constructed of things that are more like the individual transistors in the GPUs, though vastly slower switching speed, and a huge amount more noise and errors.
Re: (Score:3)
Don't confuse training and inference- I don't think either of us has the time to try to figure out the evolutionary cost in watts for the development of the human brain.
Ummm (Score:1)
NVIDIA, you may want to start wooing those gamers you dumped.
China AI is not... (Score:2)
...a threat
It's the logical and expected consequence of the futile attempt to control tech.
Chinese people are smart. They will find workarounds. They may even invent stuff beyond what is prohibited
Sounds alot like Blackberry when ...... (Score:1)
So, it is still massively expensive to make LLMs (Score:2)
Well, that's good news. Then again, that's really bad news.
In the meantime, this technology seems more and more like a solution to a problem that nobody really wants to solve, save for C-suite types that see any employee that does thinking stuff as a threat to wealth accumulation.
I just hope this bubble pops faster than the dotcom bubble. All those wonderful NVidia chips, now ready for supercomputing. Climate research. Material science simulations. And so on.
Market misconception (Score:2)
The reaction here is it took less GPUs to TRAIN the model, but that's simply because using strictly reinforcement learning they were able to skip the computationally expensive step of training on a supervised learning dataset (not counting the "cold start" dataset). What the market reactors don't realize is it's inference, not training, where the vast majority of GPU compute will go towards.
The fact is, DeepSeek-R1 actually requires MORE GPU processing power during inference because it does all the thinkin
Re: (Score:3)
No ... wrong for two reasons.
1) DeepSeek-R1 is also VERY efficient/cheap to run ($2 R1 vs $60 O1 per million output tokens), and should be compared to O1. If you want to compare to non-reasoning models, then you should be looking at DeepSeek-V3.
2) Inference is indeed what really matters, since if AI is to be successful and generate a profit then inference usage needs to be huge and exceed what is being paid for training. Running on NVIDIA chips is basically worst case scenario for an AI model since NVIDIA h
Re: (Score:2)