The AI hype train was built on the premise that AI will progress linearly and eventually end up replacing a lot of well paid white collar work, but it failed to deliver on that promise by now, and progress has flatlined or sometimes even gone backwards (see GPT-5 vs 4o).
FAANG companies can only absorb these losses for so long before shareholders pull out.
Also never forget that in technology moreso than any other industry showing a loss while actually secretly making a profit is a high art form. There is a lot of land grabbing happening right now, but even so it would be a bit silly to take the profit/loss public figures at face value.
Numbers prove we aren't. Sales figures show very few customers are willing to pay $200 per month for the top AI chatbots, and even at $200/month, OpenAI is still taking a loss on that plan so they're still loosing money even with top dollar customers.
I think you're unaware just how unprofitable the big AI products are. This can only go on for so long. We're not in the ZIRP era anymore where SV VC funded unicorns can be unprofitable indefinitely and endlessly burn cash on the idea that when they'll eventually beat all competitors in the race to the bottom and become monopolies they can finally turn a profit by squeezing users with higher real-world price. That ship has sailed.
[1] https://www.axios.com/2025/08/15/sam-altman-gpt5-launch-chat...
Ermmm dude they are competing with Google. They have to keep reinvesting otherwise Google captures the users OAI currently has.
Free cash flows matter. Not accounting earnings. On a FCFF basis they largely in the red. Which means they have to keep raising money, at some point somebody will turn around and ask the difficult questions. This cannot go on forever.
And before someone mentions Amazon... Amazon raised enough money to sustain their reinvestment before they eventually got to the place where their EBIT(1-t) was greater than reinvestment.
This is not at all whats going on with OAI.
If you're gonna buy at face value whatever Scam Altman claims, then I have some Theranos shares you might be interested in.
Anyway, I'd just point out that users don't even need to depend on the bots for increase productivity, they just need to BELIEVE it increases their productivity. Exhibit A being the recent study which found that experienced programmers were actually less productive when they used an LLM, even though they self-reported productivity gains.
This may not be the first time the tech industry has tricked us into thinking it makes us more productive, when in reality it's just figuring out ways to consume more of our attention. In Deep Work, Cal Newport made the argument that interruptive "network tools" in general decrease focus and therefore productivity, while making you think that you're doing something valuable by staying constantly connected. There was a study on this one too. They looked at consultants who felt that replying as quickly as possible to their clients, even outside of work hours, was important to their job performance. But then when they took the interruptive technologies away, spent more time focusing on their real jobs, and replied to the clients less often, they started producing better work and client feedback scores actually went up.
Now personally I haven't stopped using an LLM when I code but I'm certainly thinking twice about how I use it these days. I actually have cut out most interruptive technology when I work, i.e. email notifications disabled, not keeping Slack open, phone on silent in a drawer, etc. and it has improved my focus and probably my work quality.
To save others a search, here is a blog post and the paper:
https://metr.org/blog/2025-07-10-early-2025-ai-experienced-o...
https://arxiv.org/abs/2507.09089
Thanks for mentioning it.
Correct, but said technology needs to be self sustaining commercially. The cost the white collar worker pays needs to be enough to cover the cost of running the AI + profit
It seems like we are a long way off that yet but maybe we expect an AI to solve that problem ala Kurzweil
This is comically premature.
When you follow the progress in the last 12 months, it really isn't. Big AI companies spent "hella' stacks" of cash, but delivered next to no progress.
Progress has flatlined. The "rocket to the moon" phase has already passed us by now.
They don’t make AI chips really, they make the best high-throughput, high-latency chips. When the AI bubble pops, there’ll be a next thing (unless we’re really screwed). They’ve got as good chance of owning that next thing as anybody else does. Even better odds if there are a bunch of unemployed CUDA programmers to work on it.
There will undoubtedly still be a market for Nvidia chips but it won’t be enough to keep things going as they are.
A new market opening up with the same demand as AI just at the point that AI pops would be a miracle. Something like being an unsecured bond holder in 2010.
And what is that post-AI bubble "next big thing" exactly?
If there were, you'd already see people putting their money towards it.
A shocking surprise needs to be a surprise for it to work. Call it strategic naivety if you want.
Donald anounces tariffs and the markets react. He postpones tariffs and the markets react again. Only Donald and his friends know what he will announce next.
This feels like a misreading of what I wrote. The discovery that he is using tariffs to make a personal profit should be surprising.
> Donald anounces tariffs and the markets react. He postpones tariffs and the markets react again. Only Donald and his friends know what he will announce next.
That wouldn’t surprise me at all, I just don’t think a hypothesis about how he could abuse his power will be very compelling to anybody who doesn’t already think he’s prone to corruption. If anything, I think it starts inoculating people to the idea.
Right now if the US wants to go to war with China, or anyone China really really likes, they can expect with high probability to very quickly encounter major problems getting the best chips. AIUI the world has other fab capacity that isn't in Taiwan, and some of it is even in the US, but they're all on much older processes. Some things it's not a problem that maybe you end up with an older 500MHz processor, but some things it's just a non-starter, like high-end AI.
Sibling commenters discussing profits are on the wrong track. Intel's 2024 revenue, not profits, was $53.1 billion. The Federal Government in 2024 spent $6,800 billion. No entity doing $1.8 trillion in 2024 in deficit spending gives a rat's ass about "profits". The US Federal government just spends what it wants to spend, it doesn't have any need to generate any sort of "profits" first. Thinking the Federal government cares about profits is being nowhere near cynical enough.
The US government always ought to have the interest of US companies in mind, their job is to work in the interest of the voters and a lot of us work for US companies.
The US is desperate to not have that war, because they spent so long in denial about how sophisticated China has become that it would be a total humiliation. What you see as the US wanting war is them simply playing catch up.
Destroy the other country?
Take it over?
Be in a 1984 style „fake“ war forever?
The countries wouldn't fire nukes against each other's mainlands but maybe against each other's fleets. Pretty likely
I don’t think geographically restricting a war is even possible, really. The US’s typical game plan involves hitting the enemy’s decision-making capabilities faster than they can react. That goes out the window if we can’t hit each other’s mainlands. A war where we don’t get to use our strongest trick and China keeps their massive industrial base is an absurd losing one that the US would be totally nuts to sign up for.
Anyway, we and China can be perfectly good peaceful competitors.
Don't mistake talking about a thing as advocating for that thing. It leaves you completely unable to process international politics, and frankly, a lot of other news and discussion as well. If you can only think about things you approve of, your model of the world is worse than useless.
Never imagined politics so obviously manipulating the talking heads with nary a care about perception.
There is no such thing.
But now, are they really going to undermine this partnership for that? Their GPUs probably aren't going to become a cash cow anytime soon, but this thing probably will. The mindset among American business leaders of the past two decades has been to prioritize short-term profits above all else.
I think the assumption there is that the strategic partnership that is part of the deal would in effect preclude Intel from aggressively competing with NVIDIA in that market, perhaps with the belief that the US governments financial stake in Intel would also lead to reduced anti-trust scrutiny of such an agreement not to compete.
I feel bad for gamers - I’ve been considering buying a B580 - but honestly the consumer welfare of that market is a complete sidenote.
...and yet Nvidia is not gambling with the odds. Intel could have challenged Nvidia on performance-per-dollar or per watt, even if they failed to match performance in absolute terms (see AMD's Zen 1 vs Intel)
I don't think so:
> The chip giant hasn’t disclosed whether it will use Intel Foundry to produce any of these products yet.
It seems pretty likely this is an x86 licensing strategy for nvidia. I doubt they're going to be manufacturing anything on intel fabs. I even wonder if this is a play to get an in with Trump by "supporting" his nationalizing intel strategy.
Any down the road repercussions be damned from their perspective.
Intel can no longer fund new process nodes by itself, and no customers want to take the business risk to build their product on a (very difficult) new node when tsmc exists. They're in a chicken and egg situation. (see also https://stratechery.com/2025/u-s-intel/ )
I'm sorry that's just not correct. Intel is literally just getting started in the GPU market, and their last several releases have been nearly exactly what people are asking for. Saying "they've lost" when the newest cards have been on the market for less than a month is ridiculous.
If they are even mediocre at marketing, the Arc Pro B50 has a chance to be an absolute game changer for devs who don't have a large budget:
https://www.servethehome.com/intel-arc-pro-b50-review-a-16gb...
I have absolutely no doubt Nvidia sees that list of "coming features" and will do everything they can to kill that roadmap.
224 GB/s
128 bit
I love GPU differentiation, but this is one of those areas where Nvidia is justified shipping less VRAM. With less VRAM, you can use fewer memory controllers to push higher speeds on the same memory!
For instance, both the B50 and the RTX 2060 use GDDR6 memory. But the 2060 has a 192-bit memory bus, and enjoys ~336 GB/s bandwidth because of it.
To that point, they've been "just getting started" in practically every chip market other than x86/x64 CPUs for over 20 years now, and have failed miserably every time.
If you think Nvidia is doing this because they're afraid of losing market share, you're way off base.
They've been making discrete GPUs on and off since the 80s, and this is at least their 3rd major attempt at it as a company, depending on how you define "major".
They haven't even just started on this iteration, as the Arc line has been out since 2022.
The main thing I learned from this submission is how much people hate Nvidia.
I think there's a lot of frustration with Nvidia as of late. Their monopoly was mostly won on the merits of their technology but now that they are a monopoly they have shifted focus from building the best technology to building the most lucrative technology.
They've demonstrated that they no longer have interested in producing the best gaming GPUs because those might cannibalize their server technology. Instead they seem to focus on crypto and AI while shipping over priced knee capped cards at outrageous prices.
People are upset because they fear this deal will somehow influence Intel's GPU ambitions. Unfortunately I'm not sure these folks want to buy Intel GPUs, they just want Nvidia to be scared into competing again so they can buy a good Nvidia card.
People just need to draw a line in the sand and stop supporting Nvidia.
My RTX 5090 is about 10x faster (measured by FP32 TFLOPS) and I still don't find it to be fast enough. I can't imagine using something so slow for AI/ML. Only 2.2 tokens/sec on an 8B parameter Llama model? That's slower than someone typing.
I get that it's a budget card, but budget cards are supposed to at least win on a pure price/performance ratio, even with a lower baseline performance. The 5090 is 10x faster but only 6-8x the price, depending on where in the $2-3,000 price range you can find one at.
the intel card is great for 1080p gaming. especially if you're just playing counterstrike, indie games, etc, you don't need a beast.
very few people are trying to play 4k tombraider on ultra with high refresh rate.
I've been using Mistral 7B, and I can get 45 tokens/sec, which is PLENTY fast, but to save VRAM so I can game while doing inference (I run an IRC bot that allows people to talk to Mistral), I quantize to 8 bits, which then brings my inference speed down to ~8 tokens/sec.
For gaming, I absolutely love this card. I can play Cyberpunk 2077 with all the graphics settings set to the maximum and get 120+ fps. Though when playing a much more graphically intense game like that, I certainly need to kill the bot to free up the VRAM. But I can play something simpler like League of Legends and have inference happening while I play with zero impact on game performance.
I also have 128 GB of system RAM. I've thought about loading the model in both 8-bit and 16-bit into system RAM and just swap which one is in VRAM based on if I'm playing a game so that if I'm not playing something, the bot runs significantly faster.
I don't know exactly how the scaling works here but considering how LLM inference is memory bandwidth limited you should go beyond 100 tokens/sec with the same model and a 8 bit quantization.
Clearly I'm doing something wrong if it's a net loss in performance for me. I might have to look more into this.
If you're using llama.cpp run the benchmark in the link I posted earlier and see what you get; I think there's something like it for vllm as well.
Its also orders of magnitudr slower than what I normally see cited by people using 5090s; heck, its even much slower than I see on my own 3080Ti laptop card for 8B models, though usually won’t use more than an 8bpw quant for that size model.
> The 5090 is 10x faster but only 6-8x the price
I don't buy into this argument. A B580 can be bought at MSRP for 250$. A RTX 5090 from my local Microcenter is around 3250$. That puts it at around 1/13th the price.
Power costs can also be a significant factor if you choose to self-host, and I wouldn't want to risk system integrity for 3x the power draw, 13x the price, a melting connector, and Nvidia's terrible driver support.
EDIT: You can get an RTX 5090 for around 2500$. I doubt it will ever reach MSRP though.
Preempting a (potential) future competitor from entering a market is also an antitrust issue.
Yes.
> Other than the market segmentation over RAM amounts, I don't see very much difference.
The difference between CDNA and RDNA is pretty much how fast it can crunch FP64 and SR-IOV. Prior to RDNA, AMD GPUs were jacks of all trades with compute bias. Which made them bad for gaming unless the game is specifically written around async compute. Vega64 has more FP64 compute than the 4080 for context.
I think if AMD was able to get a solid market share of datacenter GPUs, they wouldn't have unified. This feels like CDNA team couldn't justify its existence.
Why would it matter if not? This is a nice partnership. Each gets something the other lacks.
And it strengthens domestic manufacturing. Taiwan is going to be subumed soon, and we need more domestic production now.
Besides, who would actually use them if they don’t support CUDA?
Everyone designs better GPUs than Intel - even Apple’s ARM GPUs have been outpacing Intel for a decade even before the M series.
But that's exactly what they started doing with Battlemage? It's competitive in its price range and was showing generational strides.
> Besides, who would actually use them if they don’t support CUDA?
ML is starting to trend away from CUDA towards Vulkan, even on Nvidia hardware, for practical reasons (e.g. performance overhead).
Though fair and free markets is not at all what the current regime in the US believes in, instead it will be consolidation, leading waste, and little innovation and progress.
https://www.asiafinancial.com/taiwan-says-tsmc-not-allowed-t...
That may have changed since then. But do you really want to depend on a foreign government for chip manufacturing?
For what it's worth, its TSMC's expertise in semiconductor manufacturing that has been loaned to the US, not bought, settled, and forgotten.