Theoretically this should be good for OpenAI - in that they can reduce their costs by ~27x and pass that along to end users to get more adoption and more profit.
I wish more people had understood that spending a lot of money processing publicly available commodities with techniques available in the published literature is the business model of a steel mill.
It’s the business of commodities. The magic is in tiny incremental improvements and distribution. DeepSeek forces us to question if AI—possibly intelligence—is a commodity.
No, but it's good enough to replace some office jobs. Which forces us to ask, to what degree is intelligence--unique intelligence--required for useful production? (We can ask the same about physical strength.)
I find it interesting that so much discussion about “LLM’s can do some of our work” is centred around “are they intelligence” and not what I see as the precursor question of “are we doing a lot of bullshit work?”
My partner is in law, along with several friends and the amount of completely _useless_ work and ceremony they’re forced to do is insane. It’s a literal waste of their talent and time. We could probably net most of the claimed AI gains by taking a serious look at pointless workloads and come out ahead due to not needing the energy and capital expenditure.
Surely that would be amazing for NVDA? If the only 'hard' part of making AI is making/buying/smuggling the hardware then nvidia should expect to capture most of the value.
No. Before Deepseek R1, Nvidia was charging $100 for a $20 shovel in the gold rush. Now, every Fortune 100 can build an O1-level model with currently existing (and soon to be online) infra. Healthy demand for H100 and Blackwell will remain, but paying $100 for a $20 shovel is unlikely.
Nvidia will definitely stay profitable for now though, as long as Deepseek’s breakthroughs are not further improved upon. But if others find additional compression gains, Nvidia won’t recapture its old premium. Its stock hinged on 80% margins and 75% annual growth, Deepseek broke that premise.
There still isn't a serious alternative for chips for AI training. Until competition catches up or models become so efficient they can be trained on gaming cards Nvidia will still be able to command the same margins.
Growth might take a short-term dip, but may well be picked up by induced demand. Being able to train your own models "cheaply" will cause a lot more companies and departments want to train their own models on their own data, and cause them to retrain more frequently.
The time of being able to sell H100 clusters for inference might be coming to an end though.
Maybe they even suppressed algorithmic improvements in their company to preserve moat. Something akin to Kodak suppressing internal research on digital cameras because they were world leading company that produced photo film.
You're fooling yourself if you think OpenAI is going to pass up implementing the same strategies to get a ~27x cheaper model.
> Unlike a social network, network effects won't help them - their users don't care how many other users they have, only about the AI output quality.
Google Search doesn't have a network effect. Everyone on HN has been saying Google Search is complete garbage for a decade. It still has the same market share (roughly) as it did a decade ago.
Not directly. The 27x is about costs. What it means is some order of magnitude of more competition. That reduces natural market share, price leverage and thus future profits.
Valuations are based on future profits. Not future revenues.
You can theoretically lower your costs by 27x and end up with 2x more future profits - if you're actually 45x cheaper (which DeepSeek's method claims to be).
You mean charge a 27x lower price, but have 45x lower costs, so your profit margin has doubled?
Your relative margin may have doubled, but your absolute profit-per-item hasn't. Say you had a 10% margin before, at a $100 price and $90 cost, for a $10 profit-per-item. Reduce price 27x and cost 45x, so $3.7 price, $2 cost, and $1.7 profit-per-item. 6x less profit - not as bad as 27x, but not good if you're OpenAI.
> Your relative margin may have doubled, but your absolute profit-per-item hasn't.
ChatGPT doesn't have any profits right now.
We have no idea what investors are expecting future profits to be.
> Say you had a 10% margin before, at a $100 price and $90 cost, for a $10 profit-per-item. Reduce price 27x and cost 45x, so $3.7 price, $2 cost, and $1.7 profit-per-item. 6x less profit - not as bad as 27x, but not good if you're OpenAI.
Now do the same thing but assume you have 10x more subscribers because the prices are ~27x lower.
You end up with almost 2x more total profit.
Just take ChatGPT's ~$200 subscription. Hardly anyone is going to pay ~$200 a month. Reduce that by 27x - and you're at $7.5 per month. Maybe 10% of people on the planet will pay that.
> Now do the same thing but assume you have 10x more subscribers because the prices are ~27x lower.
You're in various spots of this thread pushing the idea that their 1B MAUs make them unassailable. How are they gonna get to 10B in a world with less than that total people?
> Just take ChatGPT's ~$200 subscription. Hardly anyone is going to pay ~$200 a month. Reduce that by 27x - and you're at $7.5 per month. Maybe 10% of people on the planet will pay that.
if ChatGPT starts selling ads on chat results that will probably improve revenue. I've seen social media ads recently for things I've only typed into ChatGPT so that leads me to believe they're already monetizing it to advertising platforms.
Google spends immense amounts of resources every year to ensure that their search is almost always the default option. Defaults are extremely powerful in consumer tech.
> Google Search doesn't have a network effect. Everyone on HN has been saying Google Search is complete garbage for a decade. It still has the same market share (roughly) as it did a decade ago.
It absolutely does. People use Google for search -> Websites optimise for Google -> People get “better” results when searching with Google.
The fact that it’s market share is sticky and not responding quickly to change in quality is sort of indicative of the network effect.
It probably counts pretty much anyone on a newer iPhone/Mac (https://support.apple.com/en-au/guide/iphone/iph00fd3c8c2/io...) and Windows/Bing. Plus all the smaller integrations out there. All of which can be migrated to a new LLM vendor... pretty quickly.
« The thing I noticed right away when Claude came out is how little lock-in ChatGPT had established. This was very different to my experience when I first ran a search on Google, sometime in the year 2000. After the first time I used Google, I literally never used another search engine again; it was just light years ahead of its competitors in terms of the quality of its results, and the clarity of its presentation.
This week I added a third chatbot to the mix: DeepSeek »
My guess is that OS vendors are the real winners in the long run. If Siri/Goolge can access my stuff and core of LLMs is this replicable then I don't see anyone downloading any apps for their typical AI usage. Specially that users have to go out of their way to allow a 3rd party to access all their data.
This is why OpenAI is so deep in the product development phase right now. They have to become the OS to be successful but I don't see that happening
MySpace and Friendster both claimed ~115M peak users.
> It's literally orders of magnitude.
Sure, and the speed at which ChatGPT went from zero to a billion is precisely why they need a moat... because otherwise the next one can do it to them.
Your argument is like a railroad company in 1905 scoffing at the idea that airliners will be a thing.
Peek users is not the amount of users they had when Facebook started.
Facebook probably would've never became a thing if MySpace already had ~115M users when it started.
MySpace had ~1M.
That's why DeepSeek (or anyone else) is going to have an incredibly difficult time convincing ~1B to switch from ChatGPT to their tool instead.
Can it happen? For sure.
Will it happen? Less likely.
If anyone unseats ChatGPT - it's much more likely to be a usual suspect like Google, Apple, or Microsoft - then some obscure company no one has ever heard of.
There is no network effect (amazon, instagram, etc.) not an enterprise vendor lock-in (Microsoft Office/AD, Apple Appstore, etc.) In fact, it's quite the opposite, the way these companies deliver ouput is damn near identical. Switching between them is pretty painless.
> You don't need a moat when you're in first place
There are different moats [1]. You’re describing incumbency, an intangible moat. It’s nice, but it’s fickle. Particularly with something with low switching costs.
OpenAI could argue, before, that it had a natural monopoly. More people use OpenAI so it gets more revenue and more data which lets it raise more capital to train these expensive models. That may not be true, which means it only has that first, shallow moat. It’s Nike. Not Google.
> There are different moats [1]. You’re describing incumbency, an intangible moat. It’s nice, but it’s fickle. Particularly with something with low switching costs.
Google has a low switching cost, and hardly anyone switches.
> Google has a low switching cost, and hardly anyone switches
Google has massive network effects on its ad business and a natural monopoly on its search index. Crawling the web is expensive. It’s why Kagi has to pay Google (versus being able to pay them once and then stop).
just for the chatbot, it's trivial to switch, create a new account and start asking questions from deepseek instead. There is nothing holding the users in chatgpt.
Businesses don’t even want to maintain servers locally. They definitely aren’t going to start managing servers beefy enough to run LLMs and try to run then with the reliability, availability, etc of cloud services.
This will make the cloud providers - especially AWS, GCP and to a lesser extent the also ran clouds more valuable. The other models hosted by AWS on Bedrock are already “good enough” for most business use cases.
And then consumers are definitely not going to be running LLMs locally on their computers to replicate ChatGPT (the product) anymore than they are going to get an FTP account, mount it locally with curlftpfs, and then using SVN or CVS on the mounted filesystem and then from Windows or Mac, accessed the FTP account through built-in software instead of using cloud storage like Dropbox. [1]
Whether someone comes up with a better product than ChatGPT and overcome the brand awareness is yet to be seen.
[1] Also the iPod had no wireless, less space than the Nomad and was lame.
There is a reason I kept emphasizing the ChatGPT product. The (paid) ChatGPT product is not just a text based LLM. It can interpret images, has a built in Python runtime to offload queries that LLMs aren’t good at like math, web search, image generation, and a couple of other integrations.
The local LLM on iPhones are literally 1% as powerful as the server based models like 4o.
That’s not even considering battery considerations
> The local LLM on iPhones are literally 1% as powerful as the server based models like 4o.
Currently, yes. That's why this is a compelling advance - it makes local LLMs much more feasible, especially if this is just the first of many breakthroughs.
A lot of the hype around OpenAI has been due to the fact that buying enough capacity to run these things wasn't all that feasible for competitors. Now, it is, potentially even at the local level.
Training costs are not the same as inference costs. DeepSeek (or anyone hosting DS largest model) will still need a lot of money and a bunch of GPU clusters to serve the customers.
