I agree. There will be a point where the LLM running with home hardware will be nearly as good as enterprise GPUs. Not for a while though. Maybe 10-15 years.
Qwen team has a roadmap that scales from 1M context to 100M and 1 Trillion param models to 10 trillion.
Last edited:
Yes, I'm doing the same. I've been 100% invested in semis since ChatGPT came out 3 years ago.
Not only that, in my profession (software engineering in a major Silicon Valley company), I've gone from using Stackoverflow via Google to asking ChatGPT via copy/pasting code, to now having agents write 95% of the code I ship. This all happened within 3 years.
Every single engineer on my team has gone through the same journey. And before anyone asks, no, our software quality did not drop. It increased significantly because LLMs can be leveraged for testing as well.
In my mind, there is no way the AI boom is done. We've only begun. There will be dips here and there but the market size in 2030 will be several times bigger than 2025. That's what I'm betting on. That's the majority wall street opinion as well.
Last edited:
Keep in mind that it is 27% of x86 client market share. AMD is winning at the expense of Intel only. In other words, AMD is getting a bigger piece of the shrinking x86 pie.
115 P/E is based on the useless (for AMD) GAAP results.
A good filter to separate those who know something about finances of the company in question and those just using Google.
People in the know (company executives, financial analysts who attend earnings conferences, investors following the stock) always talk in non-GAAP figures.
Maybe shrinking slightly in inflation adjusted total, but in fixed dollars, kind of hovering around $10 billion total (per quarter) between Intel and AMD and likely sowing some growth in both 2025 and 2026.
Nothing remotely close to growth of the AI GPU market, but for AMD, with steadily growing Client Revenue and Market Share, it is a good, growing business.
Some people project that in ~3-5 years, material percentage of AI will shift from Cloud to Client, which could result in some AI fueled growth in client as well.
GAAP figures aren’t “useless”. They’re the legally required accounting standard that ensures comparability and auditability across companies. Non GAAP metrics can highlight operational trends but they also exclude real expenses like stock based compensation and amortization.
Analysts and investors who understand a business look at both. Ignoring GAAP entirely isn’t sophistication. It’s cherrypicking the numbers that make a company look better.
x86 has been steadily losing marketshare to Arm laptops over the past 5 years. I expect this to accelerate drastically as Qualcomm and Nvidia enter the market in big ways and Windows on Arm gets better quietly.
Furthermore, Arm owns the highend laptop market which is where most of the profit is. It wouldn't surprise me if Apple owns the lion's share of total laptop profits already.
In case of AMD, it is not just some cherry picking, but a big elephant in the room, in form of Xilinx acquisition. IIRC, the biggest acquisition in semiconductor industry.
Cost of the acquisition has then been gradually written off as depreciation of Good Will, which truly made AMD GAAP results useless especially for any company to company comparisons.
I'm aware of Xilinx amortization.
Regardless, don't be so attached to AMD. Had you invested in Nvidia/TSMC/Broadcom instead of AMD, you would have done far better over the last few years.
AMD is the darling amongst tech enthusiast crowd, which happens to be the Anandtech crowd. I get the bias favoring AMD here. I totally do. 90% of posters here are probably AMD fans because they own Ryzen CPUs. They feel like AMD saved them from Intel 4-core 14nm years.
A rising boat (AI boom) lifts all boats. But AMD isn't the boat I'm riding on.
The discussion turned to bubbles, and what happens if it were to pop.
So, the same way as people like YouTube channel "Ticker Symbol You" were not too crazy about AMD (Vs. NVidia, Broadcom) on the way up, precisely because of AMD's limited exposure to AI, but the same token, on the way down, the companies with greatest exposure to AI would perform the worst and those with smallest exposure would perform the best.
That's summary of my point. The point of disagreement waw when you said that in case of AI collapse, AMD would do worse than NVidia, which is something I disagreed with.
AMD's valuation $380.21B hinges on its ability to grab AI market share. That's why Wallstreet doesn't care about anything other AI in AMD's earnings.
In any AI collapse, orders for AMD GPUs will dry up because Nvidia GPU supplies will free up. Further more, because Nvidia has technical and margin superiority, it will lower prices to take remaining AMD GPU orders in a down market
The size of non-AI business is irrelevant because the AI business is so big. No, AI isn't collapsing to $0 so it's a moot point to discuss what happens if Nvidia and AMD's AI revenue go to $0. A more realistic scenario is what happens if AI chip sales fall 30% for one year.
Mac has reached a multiyear plateau and is now has correlation of ~1 to x86 shipments.
The ChromeBooks, one area of Arm penetration - they are now kaput.
Snapdragon had a good potential to make inroads, but has fizzled out even during the height of the hype cycle. May have shipped 720k units (but it is unknown from where to where it was shipped). Not necessarily to consumers.
Qualcomm projection of reaching 5% market share by the end of 2025 is now a very sad joke, may not reach 1%.
Apple in comparison ships ~ 6 million units per quarter
NVidia could make some inroads, we will see.
Apple does, not Arm. As Apple did when Apple used x86.
Snapdragon is being dumped at cheap prices, big discounts.
That's some of wildest copium I've seen yet here.
Apple has dominated high end laptops where most of the profits are. Since Apple Silicon, they're growing and it seems to be accelerating (minus Covid boom). Now that Apple is readying an inexpensive Macbook, growth should be robust from now on. Soon, a $599-$699 Macbook will have 15% faster ST speed than AMD's very best laptop chip.
M6 MBP redesign and low cost Macbook will drive a ton of growth for Macs over the next few years.
Qualcomm's latest phone chip has faster ST speed than AMD's best laptop chip. It's not going to look pretty in the Windows on ARM world for x86 very soon.
Last edited:
adroc_thurston
Diamond Member
- Jul 2, 2023
- 8,191
- 10,930
- 106
adroc_thurston
Diamond Member
- Jul 2, 2023
- 8,191
- 10,930
- 106
Ah cool
So you're taking the GAAP profitability which is depressed because of the way the xilinx deal is taken to the P&L to save tax. It's a much lower number of you take that out.
Apple with x86: 24, 25, 23, 26, 25, 26 looks like a multi year (6 year) plateau
Subsequent M series laptops have a multiyear (6 year) plateau
So, all that changed in 12 years was that M series Macs added some incremental revenue
At this pace, Apple / Arm will dominate the laptop market, clinching 50% market share, in about 80 years.
Let me guess: ST Geekbench 6 running Linux.
Last edited:
I understand the feeling and I also use extensively LLMs to code (bioinfo research/ML), but all of this doesn't disprove in any way that current AI economics is a bubble.
In fact I'm pretty sure that most programmers that could and would use LLMs already use them, and the rest of users (probably 90% or so of the total) could already stop paying anyone and use small open weight models without noticing the difference.
What's the point of the trillion dollar investments? A 5% better accuracy on benchmarks? A 10M longer context length for the 1% use cases that need it? Faster inference in order to save users 3 seconds per prompt? All good things, but none of them require the kind of money these companies are talking about and promising investors.
I'm pretty sure that the goal of OpenAI and their friends is simply to hoard resources and hope to get to "superintelligent agents" before anyone else, but they will fail, they are too arrogant to make feasible and sustainable long term plans.
The only hope for Nvidia quite frankly is for robotics to take off the way they want, and that would also possibly give them an edge against AMD and others, but I would bet quite a lot that the Chinese will get there faster and with better final products.
There were dozens of railroad bubbles yet we still have railroads.
The utility of an investment category does not preclude or limit the possibility of investor delirium. Often it seems to invite the most idiotic build outs that have no possibility of profit.
Ultimately I find it very hard to believe the current level of spending (causing shortages in many supply sectors, even as fundamental as turbines and electricity) cannot be the result of a bubble. It seems the most likely explanation. It does not seem like a new baseline of activity. It is eagerness to be the one who captures the market.
The utility of an investment category does not preclude or limit the possibility of investor delirium. Often it seems to invite the most idiotic build outs that have no possibility of profit.
Ultimately I find it very hard to believe the current level of spending (causing shortages in many supply sectors, even as fundamental as turbines and electricity) cannot be the result of a bubble. It seems the most likely explanation. It does not seem like a new baseline of activity. It is eagerness to be the one who captures the market.
sandorski
No Lifer
- Oct 10, 1999
- 70,835
- 6,380
- 126
I don't think this will happen any time soon. The only way that could be true is if we hit a cap on the effective size of a model, but since they are still so much smaller than the human brain, I think that any way a model gets cheaper will not just be used to fit what we already have into consumer hardware but also to make frontier models bigger.
And just remember that centralized inference can benefit from batching, which local inference can't really do. A centralized system that can run 20 tokens per second for a 1000 people at the same time costs a fraction of what it would cost to ship those 1000 people their own hardware that can each do 20 tokens per second.
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)- Started by DisEnchantment
- Replies: 25K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes + WCL Discussion Threads
- Started by Tigerick
- Replies: 24K
-
Discussion Intel current and future Lakes & Rapids thread- Started by TheF34RChannel
- Replies: 23K
-
-
Facebook
Twitter