You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
Intel really has no meaningful DC AI sales as of now and the market is not giving them credit for any. Intel product spends probably 2B per year on DC AI development and has minimal sales. Intel does not seem to have meaningful mobile (Phone/tablet) sales
So for a valuation, Look at the profitability of CPU products and market share. Essentially take Intel corporation and remove the losses and expense that is IFS.
Intel product has 50B in revenue and 13B in Operating Margin per year. Both are relatively healthy IMO and both are 2x AMD for 2024. It seems like Product are still worth 120+ vs Intel Corp at 85B today.
The Cash impact and the losses of IFS are what is killing the Intel stock price. Once people saw the reality of Intel Manufacturing, the stock price collapsed.
Intel really has no meaningful DC AI sales as of now and the market is not giving them credit for any. Intel product spends probably 2B per year on DC AI development and has minimal sales.
How do you figure Intel is spending $2B/year on data center AI product development? Habana Labs has only ~1000 employees. They might have spent $300-500M developing Gaudi 3, if you add in associated software work. The Falcon Shores team might be 1000 people too. Assuming a fully loaded cost per engineer in Intel is $500K/year (that's probably high, especially in Israel), $2B in R&D would be 4000 engineers. What am I missing or just plain wrong about?
How do you figure Intel is spending $2B/year on data center AI product development? Habana Labs has only ~1000 employees. They might have spent $300-500M developing Gaudi 3, if you add in software associated software work. The Falcon Shores team might be 1000 people too. Assuming a fully loaded cost per engineer in Intel is $500K/year (that's probably high, especially in Israel), $2B in R&D would be 4000 engineers. What am I missing or just plain wrong about?
Thats just labor. Labor is not the only cost in development. What do you think the expenses were for all DC AI development (Falcon shores and Gaudi are not the only projects.)?
Thats just labor. Labor is not the only cost in development. What do you think the expenses were for all DC AI development (Falcon shores and Gaudi are not the only projects.)?
Fully loaded costs per engineer includes things like office space, IT, software licenses (Cadence, etc.), whatever it takes to add an incremental engineer to the team. Direct payroll and benefits costs for a G7-G9 engineer will be a lot lower even in the US than $500K. Chip production costs (mask sets, shuttles/test chips, test equipment, etc) are high, especially for, say 5nm processes, but they're a lot smaller than employee costs because they're fewer in number, however expensive individually. (I've read a 5nm mask set is $10M-$20M.)
Nvidia spent just under $9B on R&D in 2024 for the entire company, so $2B for the Intel projects still feels high.
The only other DC AI projects in Intel I can think of are the Xeon accelerators, like AMX. Those can't be huge projects.
I don't believe there's any resolution for Intel's problem unless Intel splits up,
... and it should be. The design house and fab business models are so different that there's no way to manage both effectively.
I don't believe there's any resolution for Intel's problem unless Intel splits up,
... and it should be. The design house and fab business models are so different that there's no way to manage both effectively.
I don't believe there's any resolution for Intel's problem unless Intel splits up,
... and it should be. The design house and fab business models are so different that there's no way to manage both effectively.
I don't think that is the main reason. The main reason is that TSMC holds a monopoly position. As such, no other fab companies (such as UMC, GF, etc.) can compete with it.
Intel could do that because it has its own products.
The U.S. should introduce effective tariffs to push fabless companies to adopt domestic foundries, helping to break TSMC’s monopoly to some extent.
I don't think that is the main reason. The main reason is that TSMC holds a monopoly position. As such, no other fab companies (such as UMC, GF, etc.) cannot compete with it.
Intel could do that because it has its own products.
The U.S. should introduce effective tariffs to push fabless companies to adopt domestic foundries, helping to break TSMC’s monopoly to some extent.
This is deadly wrong.
Intel used to have the monopoly because it used its fab advantage to dominate the CPU market. Now, Intel is the result of a failed monopoly.
NVIDIA's market value is much larger than TSMC's.
This is the best indicator of the absence of monopoly.
This is deadly wrong.
Intel used to have the monopoly because it used its fab advantage to dominate the CPU market. Now, Intel is the result of a failed monopoly.
NVIDIA's market value is much larger than TSMC's.
This is the best indicator of the absence of monopoly.
I don't think it is right or wrong. It is just a practical way to achieve the objective. Without economic incentives, fabless companies would choose lower cost suppliers.
I noticed Taiwan media in the past week seemed to adopt the view that tariffs on semi is unavoidable.
Intel really has no meaningful DC AI sales as of now and the market is not giving them credit for any. Intel product spends probably 2B per year on DC AI development and has minimal sales. Intel does not seem to have meaningful mobile (Phone/tablet) sales
So for a valuation, Look at the profitability of CPU products and market share. Essentially take Intel corporation and remove the losses and expense that is IFS.
Intel product has 50B in revenue and 13B in Operating Margin per year. Both are relatively healthy IMO and both are 2x AMD for 2024. It seems like Product are still worth 120+ vs Intel Corp at 85B today.
The Cash impact and the losses of IFS are what is killing the Intel stock price. Once people saw the reality of Intel Manufacturing, the stock price collapsed.
All I did was change one line of code in PyTorch, so on the software side, it is catching up.
If Intel had not canceled Rialto Bridge, the conversation about Intel’s role in AI could be very different.
I think Intel is using Falcon Shores to bootstrap a competitive product line against Nvidia’s data center product offerings. For the meantime, Gaudi should be sufficient to run Deepseek R1 models on enterprises' ends.
This is deadly wrong.
Intel used to have the monopoly because it used its fab advantage to dominate the CPU market. Now, Intel is the result of a failed monopoly.
NVIDIA's market value is much larger than TSMC's.
This is the best indicator of the absence of monopoly.
No, not wrong, complicated. TSMC is an effective monopoly for high volume leading edge fab processes. Who is the competition?
TSMC's market capitalization is depressed by geopolitical risk. Nvidia's PE Ratio is more than twice TSMC's. (46 versus 23). Nvidia has competition, AMD. Unfortunately, AMD is currently hobbled by leading in the "Not CUDA" market, and "Not" markets are usually second-class profit makers by definition. Intel is not an effective competitor in GPUs because of a CPU-centric mentality for decades, and then they re-hired one of the high priests of "every compute chip should be a general-purpose CPU", Pat Gelsinger, as CEO, who even mounted a silly defense of the old x86-based Larrabee project, and didn't make competing with Nvidia a priority.
If only Intel didn’t kill his Larrabee project, laments Pat Gelsinger.
www.tomshardware.com
In a wide-ranging interview with Gelsinger, hosted by the Massachusetts Institute of Technology (MIT), the Intel boss told attendees that Nvidia CEO Jensen Huang “got extraordinarily lucky.” He went on to lament Intel giving up the Larrabee project, which could have made Intel just as ‘lucky,’ in his view.
Sorry to say this, but I think Intel are boned. The fact that they are using TSMC N2 for Nova Lake high end NVL-SK / NVL-HX is an indication that Intel 18A has inferior performance. They have to use N2 if they want to compete with AMD.
Remember they said "Intel is betting the house on 18A"? Well, it looks like they lost the house!
Sorry to say this, but I think Intel are boned. The fact that they are using TSMC N2 for Nova Lake high end NVL-SK / NVL-HX is an indication that Intel 18A has inferior performance. They have to use N2 if they want to compete with AMD.
Remember they said "Intel is betting the house on 18A"? Well, it looks like they lost the house!
Is one of the advantages of using chiplets the ability to multi-source? I don't think that really matters. What matters is that they make their foundry break even.
Is one of the advantages of using chiplets the ability to multi-source? I don't think that really matters. What matters is that they make their foundry break even.
That could be one of the reasons the 5090 is very expensive. I think for Intel, if they can get the foundry business to break even, it would reflect positively on its valuation. Perceptions could change after that.
Is one of the advantages of using chiplets the ability to multi-source? I don't think that really matters. What matters is that they make their foundry break even.
It does matter. If you don't have the fastest node it's going to get hard to get foundry customers. And if your own multi-billion fabs are under utilised because you have to use competitors node you are going to lose money.
It does matter. If you don't have the fastest node it's going to get hard to get foundry customers. And if your own multi-billion fabs are under utilised because you have to use competitors node you are going to lose money.
It does matter. If you don't have the fastest node it's going to get hard to get foundry customers. And if your own multi-billion fabs are under utilised because you have to use competitors node you are going to lose money.
They are using TSMC cause they don't have that much capacity on 18A we will have Panther Lake ,Nova Lake,Clearwater Forest Diamond Rappids and wildcat lake plus external tape out in H1 25 so EOY 26 Production?
They are using TSMC cause they don't have that much capacity on 18A we will have Panther Lake ,Nova Lake,Clearwater Forest Diamond Rappids and wildcat lake plus external tape out in H1 25 so EOY 26 Production?
Sorry, but that's simple not true. If it was capacity then they wouldn't be doing their high end SKU on N2. High end SKUs are lower volume than the mid range SKUs and they need the extra performance from the N2 node
