Ask HN: Confused about how DeepSeek hurts Nvidia

36 points by prng2021 ↗ HN
I’m genuinely confused about why people think Deepseeks results will mean fewer GPUs being needed in the future. DeepSeek won’t be top dog forever. At some point, all their big competitors will figure out how they created their model, copy the approach, and get the same efficiencies. After that, why wouldn’t every competitor add more compute to go beyond DeepSeek’s capabilities and each other? Is there some experimental evidence out there that having 10X or 100X the compute DeepSeek used for training wouldn’t result in a much more advanced model?

49 comments

[ 3.7 ms ] story [ 114 ms ] thread
If work can be done with fewer GPUs then people will use less. This is what leading to Nvidia fall.
But there isn’t a finite amount of work. If we can do something with less then we should be able to do significantly more with more by applying the same optimizations.
I'm out here waiting for model compute costs to drop so I can run model ensembles and immediately improve accuracy.
I agree as such. Do note that there's finite amount of training data though and humankind is already close to the limits.
only if deepseek-r1 has achieved 100% AGI and the marginal benefit of more compute decreases, otherwise there is no reason to use less compute.
AI being more available and locally runnable will induce demand for GPUs.
I don't think this is true, there's a limit to how "good" the average person needs their AI to be — assuming the average person purposefully uses much AI at all (thinking of people like my mother here).
(comment deleted)
> DeepSeek won’t be top dog forever.

I agree with this in the sense that no model will be top dog forever. However, it's important to note their contributions to open source. They're raising the bottom bar, and that is important.

yeah the FOSS angle is big. it's not just good, it's good and it's out there for anyone.
Market reactions go to immediacy usually. You don't see people stonking in or out of something for an outcome in 5 years time: It's immediacy which makes a wave happen.

So noting your long term investment ideas seem plausible, what do you think is the immediate short term impact on this kind of spend? Do you think Nvidia will sell more or less units in the next reporting interval? Because thats what most people are reacting to.

It would not surprise me if there are plenty of willing buyers, looking to buy in a dip and sell on the inevitable upward swing.

I am not a direct investor. I have no idea what my pension fund did, if anything.

I too can't understand why? won't a cheaper model make people use AI more? For example, the current $200 chatgpt plan is too expensive for me, but making it $4, I will become a customer.

Many small companies, which would never think about training models in house, could now do it.

I see this will only boost the AI hardware market.

(comment deleted)
Nvidia is at a premium because it's so dominant. But the sentiment now is that China can build cheap, quality AI too and presumably they can make top dog or second dog GPUs too.

US's anti-China policies further forces China to develop their own GPUs, breaking Nvidia's dominance within 10 years or something.

Deepseek was trained on Nvidia GPUs H800. Nvidia is still selling GPUs to China and the reason even the reduced chip performance is easily negated by scaling.

US government is stupid because they asked for certain limits on chip base first but AI uses GPU clusters. In a GPU cluster you don't have full utilization anyway so slower GPUs don't matter as much as slower networking. China still gets pretty high bandwith Nvidia HW for building large clusters for training/inferencing.

Chips from Huawei still seem to be way too unstable for the job. In training/inference stability is even more important than performance. Imagine you have a fast chip but it can't run without errors for 2 months and you training never gets done. That's DoA.

It doesn't. Inference is still expensive, and demand for it is high, as evidenced by Anthropic's frequent "we're out of quota" messages and Deepseek's crap-out under load last night. On the training side right now only the top flight labs can conduct serious, ambitious research, and even they don't do as much research as they'd like. Witness Meta effectively train the exact same architecture on similar data mixtures for the past couple of years. More or less the same situation is happening across the board - compute bandwidth (and therefore the ability to experiment) is scarce. What this means is inference will remain quite expensive in the foreseeable future, especially multimodal and long-context inference. Believe it or not, even Google is compute constrained. When I was there some days I couldn't even get a handful of TPUs to do my job - everything was allocated to training Gemini. Even if it gets a lot cheaper to train models, you could just train larger, more capable models and do more architectural / efficiency research, and iterate faster, with tremendous payback in the long run. NVIDIA is the only viable seller of shovels for this gold rush for everyone but Google and Anthropic. Bypassing the gatekeepers, and making capable AI models available to more people makes their product more valuable.
> NVIDIA is the only viable seller of shovels for this gold rush for everyone but Google and Anthropic.

Why do you except Google and Anthropic?

Google makes its own hardware, they are vertical integrated .Dont know about Antrophic
Anthropic uses a ton of TPU in addition to GPU, so presumably has the expertise to use both, and shift workloads as needed. Note that large scale TPU pretty much means Jax and not just "platform independent" flavor of Jax but Jax with TPU-specific optimizations.
Anthropic are the only (?) heavy users of Amazon's chips. Or maybe they aren't heavy users. It's hard to say, they use NVIDIA too. Amazon is a big investor.
Amazon's chips at this point are marketing for Amazon. I've seen the benchmarks, they're not quite ready for serious use yet. I suspect Anthropic got a good discount on GPUs in return for using Amazon's own chips in any possible capacity (or maybe just for the press release claiming such use). The only real alternative to NVIDIA on the inference side that you can actually buy hardware for is Intel Gaudi which costs less and performs rather well, but everyone seems to have written it off, along with Intel itself, and it's not available in any cloud last I checked. On the training side there's really no alternative at all - PyTorch is the de-facto standard, and while there is PyTorch XLA, it's even less popular than Jax, which is already like 20x less popular than PyTorch. Bottom line: capable Jax engineers able to optimize distributed Jax programs on TPUs are unobtainable unicorns for anyone but the top labs and Google itself. Note that the training side has significantly different requirements than inference side. Inference side is much simpler to optimize and wring the performance out of.
Yes I've been expecting AMD to eventually get inference working because it's so much simpler. Supposedly Meta do use some AMD for inference. It's sad that you can implement llama inference on the CPU in a few thousand lines of Java yet somehow AMD isn't cleaning up there.
It's a bit like "the Cisco moment" (and lots of people have been observing this). The company was building hardware needed for building out networks. The web looked like it was going to be the next big thing, and people couldn't get enough of CSCO. The web didn't pan out the way people hoped (or as quickly), and CSCO fell quickly.

Cisco kept making and selling network hardware, and probably (citation needed) sold more from 2000-2006 than 1994-2000, but the stock trade was over. The web did become a serious thing, but only once people got broadband at home.

The Nvidia valuation was getting pretty weak. Lots of FAANGs with deep pockets started to invest in their own hardware, and it got good enough to start beating Nvidia. Intel and AMD are still out there and under pressure to capture at least some of the market. Then this came along and potentially upended the game, bringing costs down by orders of magnitude. It might not be true, and it might even drive up sales long-term, but for now, but the NVDA trade was always a short-term thing.

NVDA trade was always a short-term thing

NVDA has been going up for the last 10 years (with 2022 being the only exception).

AI today is better than anyone could hope for, and I don’t see any reasons to not expect further advances.

> AI today is better than anyone could hope for,

I hope for an AI that can actually reason and doesn't bullshit its users though

Without Nvidia, you would still wait for ChatGPT moment and not even think about AI at all.

I have been invested in Nvidia for 9 years and I have not only witnessed what Nvidia has done in ML/AI but also how the entire field evolved.

To say Nvidia is toast is like saying US is toast after Sovjets send a rocket into space.

What AI are you using? I use o1 and it hasn’t tried to bullshit me so far. Would love to hear some examples of o1 bullshitting.
I also use o1 and sometimes it just gets stuff confidently wrong. For example it often makes up command line arguments or optional parameters to methods or configuration keys. Anything that is even remotely niche, anything that hasn't been posted 1000 times on the internet, GPT will get wrong without ever telling you that it does not know.

Anything else, you can just google it because 9/10 there's a SO thread about it, and when there isn't, the documentation is usually good enough. GPT is not going to help you there.

But you can see it in the small things. For example if you ask an intelligent person "how do I install X on docker?" it will search google and then find a tutorial or some resource to reference. Then they will break it down and adapt it to your scenario.

GPT instead will give you... whatever it feels like giving you in that moment. It will not look on the f*n internet to make sure that what it says is correct and up to date. It does not update an internal knowledge base with factual information that it can then reference to produce a coherent plan. It has no concept of truth so it can't use logic.

You can throw as much compute and chain of thought as you want to the problem, it's the architecture itself that is disappointing.

That's fair, even though I haven't encountered this myself. I think having it search internet for relevant info and cross-reference it is a good idea, and is being implemented (SearchGPT). I guess we'll just have to wait a couple more months till o3.
While I agree with your points, I think a larger factor is that NVidia's valuation has been driven higher by the seemingly insatiable demand for data center AI GPUs from large companies investing far in advance (and excess) of near-term revenue. In recent months signs have emerged that leading edge models like O3 require significantly higher GPU cycles for each additional increment of quality. This would tend to push GPU demand growth rates even higher.

* DeepSeek appears to be credible evidence there may be clever optimizations to achieve higher model quality with less GPU cycles than previously thought. Basically, if you're making scarce oil derricks in a gasoline shortage and your stock price has been bid way up on the expectation of insatiable future gas demand, a more gas-efficient engine design is going to be adverse to your valuation. Especially if it's free and easy to implement.

* DeepSeek's weights are open source under a permissive license. Much of OpenAI (and similar company's) current revenue is from AI startups and other companies buying usage hours of proprietary leading edge models (eg O3) as cloud services through an API and reselling the output in their own applications targeting various verticals. If some of those companies start using a free open source model like DeepSeek (or it's future descendants/competitors) for some of their offerings - that'll reduce the income and war chest of some of today's biggest GPU buyers. Lower current revenue lowers valuations meaning the equity OpenAI et al use to buy GPUs will be devalued.

>Lots of FAANGs with deep pockets started to invest in their own hardware, and it got good enough to start beating Nvidia.

It's not just hardware though: you can't run CUDA on non-Nvidia hardware, which in my understanding is a major moat for Nvidia. I'd love to hear rebuttals on this though, because GPU programming is something I've only dabbled with.

Agreed on CUDA being a big moat; Nvidia has spent a big chunk of the last 2 decades building this ecosystem and is now reaping dividends from it.

From what I've read, most of the investments by FAANGs/startups in building specialised hardware has been in the inference space.

It's not a CISCO moment, more like a "Wavelength-division multiplexing" [1] moment of 2000 where the fiber optic craze was over capacity and crashed, causing a lot of "Dark Fibre" [2] left around. Deepseek found a way to squeeze more out of the same hardware, heck a key point is how to do more with the bandwidth bottleneck also with BI-DIRECTIONAL MULTIPLEXING [3] :)

Most of the biggest Nvidia clients are valued on speculation of future revenue from their closed models (secret sauce). Deepseek is fully open source so those revenue expectations crashed and investors are having second thoughts on throwing more money at companies like OpenAI. And this hits the expected sales growth of Nvidia for the next few years.

Dark fibre eventually was used but it took many years. And it was bought for cheap by companies like Google and CloudFlare.

[1] https://en.wikipedia.org/wiki/Wavelength-division_multiplexi...

[2] https://en.wikipedia.org/wiki/Dark_fibre

[3] https://arxiv.org/html/2412.19437v1 (3.2.1DualPipe and Computation-Communication Overlap)

> more like a "Wavelength-division multiplexing"

no joke, the hype around DWDM is why I got into networking -- waves are the future, man!

Yeah, I also have made the "right move" industry or tech-wise but for the wrong reason (I'm assuming being in advanced networking worked out pretty well anyway for you).

It's always funny when someone asks me "So how did you know so far ahead of everyone else [that (my startup's product) would launch (category)]", and I tell them it's because we were inexperienced and didn't fully understand the market, so we guessed based on assumptions which turned out to be wrong. I try to always fess up to the fact that success in tech startups is a combination of insanely hard work, being (mostly) smart, luck, and being stubborn enough to keep trying again after failing until your luck turns.

The key thing is that it's actually both, and the market is still in waiting to see mode even after a 20% dump. AI end applications are not growing fast enough, and compute costs may be crashing down.
There are people who are willing to say that deepseek has such a great team -- so great, in fact -- that they could still always crush competition absolutely all the time, despite having open source their models to some extent, if not all.

I am actually interested to see those who hold this view explain their logic here further. It's quite an interesting, if unorthodox take because we would think that algorithmic brilliance is not a moat, even when the code is not open source.

The market doesn't care about rational, it needed a reason ( for a very long time ) for a correction. The more fear it can drive, the lower it can drive those valuation down.
(comment deleted)
I'm not stock expert (but bought 80 shares of Nvidia yesterday). My rationale is, that if DeepSeek lowered barier for entry into AI model development, actually more companies can afford to join. Yes, big dogs may require less cards with efficient algorithms, but to me, more companies = more cards needed.
> DeepSeek won’t be top dog forever.

Which is... worse for Nvidia? If someone else disrupts DeepSeek, do they train a similarly performing model for $600k?

it doesn't hurt Nvidia like a cheaper light bulb wouldn't hurt electricity investing in 1900. On the contrary what it means is that you get more intelligence for your buck. Consequently it means we will get to much higher levels of intelligence faster. Also that inference (using the intelligence) will be a more common good.
Deepseek is not hardware; how could it hurt NVidia? It's found a way to use train data efficiently, but it seems DeepSeeks brings more potential customers to them. It's open a way for small enterprises and individuals. I have lost faith that AMD will fight back against Nvidia. If Apple or another RISC-V-based LLM accelerator hardware manufacturer doesn't respond to NVIDIA, NVIDIA will drive hardware prices skyrocketing, as any monopoly does. It looks like Apple is the one with the most potential to give the answer that will hurt Nvidia.
Nobody knows and investors are guessing. Their confidence has been rocked and uncertainty has been introduced.
Analysing the market and competitive situation:

Deepseek's cheaper LLM services + providing open models for other hosts to provide

=> overall prices for using LLM services will fall due to competition (lower prices + more hosts entering the market); AI users won't pay so much for LLM services

=> LLM hosts/providers won't be able to project such high revenues or even purchase as many GPUs (and will receive less capital investment to buy GPUs since revenues per dollar invested are lower)

=> demand for and prices of Nvidia cards will fall

On the basis of this possible logic, portfolio managers and algorithms project lower growth/revenue for Nvidia and sell off its stock, setting off the usual chain reaction as other managers notice the downward price action and follow suit in order to stop further losses.

it doesn't.

Market is upset because monopolies are basically busted with opensource AI