News Intel foundry unit loses $7 billion in 2023, company outsources 30% of its production to TSMC, others

I guess now we know why Intel was begging for another $10,000,000,000 from the US government. They don't have a viable business model. No tax dollars should be given to a company with a nonviable business model.
 
  • Like
Reactions: dalek1234
I guess now we know why Intel was begging for another $10,000,000,000 from the US government. They don't have a viable business model. No tax dollars should be given to a company with a nonviable business model.
How do you know that? Were you expecting a massive new business investment in western geographies that don’t traditionally run foundaries to not come with a massive up front cost? Either way, the incentives are there to absorb this huge up front cost in exchange for gaining geopolitical strategically important redundancies to critical defense and industry supply chains. We won’t know if it’s a boondoggle until 2027 or thereabouts really. Do you have alternative solution to achieve this goal that’s less expensive? Maybe perhaps we start paying China tribute?
 
  • Like
Reactions: artk2219
Taking a loss because you are investing in equipment and facilities for the future isn't a bad thing.
Yes. Exactly. But wkm isn't totally off either. Intel wants this news to look as bad as possible to get as much money from the government as possible. It stands to make far more than the losses over the past few years once the foundry side of things is really moving, and if it can get the US govt to pay for it, so much the better.

The US meanwhile wants to get away from being so dependent on Taiwan, China, etc. for tech manufacturing. It needs Intel to succeed perhaps even more than Intel needs to succeed. $10 billion (or more) is a pittance compared to where other money goes.
 
Yes. Exactly. But wkm isn't totally off either. Intel wants this news to look as bad as possible to get as much money from the government as possible. It stands to make far more than the losses over the past few years once the foundry side of things is really moving, and if it can get the US govt to pay for it, so much the better.

The US meanwhile wants to get away from being so dependent on Taiwan, China, etc. for tech manufacturing. It needs Intel to succeed perhaps even more than Intel needs to succeed. $10 billion (or more) is a pittance compared to where other money goes.
Yeah, but I think we both know that the post wasn't meant that way, but just as another "INTEL IS DOOMED, DOOMED I SAY, YOU HEAR ME?!?"-post. The article literally says that every other division was profitable, so why claim there is no viable business model when the only losses are investments for the future?
 
I wonder what are the odds Intel just ends up keeping a 30% outsource rate for transitionary periods and for certain products that their own fabs just can't meet the performance/efficiency minimums for.

That said, such an expense loss does explain why they'd want to open up their fabs, even to competitors. They need to keep the fabs operating, and all that matters is keeping the fab books full, esp. to recoup upgrade costs. Making up that missing 20-30% would help them recoup costs much faster.
 
  • Like
Reactions: artk2219
I wonder what are the odds Intel just ends up keeping a 30% outsource rate for transitionary periods and for certain products that their own fabs just can't meet the performance/efficiency minimums for.

That said, such an expense loss does explain why they'd want to open up their fabs, even to competitors. They need to keep the fabs operating, and all that matters is keeping the fab books full, esp. to recoup upgrade costs. Making up that missing 20-30% would help them recoup costs much faster.
Honestly for some non essential or low margin parts, if you can get someone to make it quickly, and at around the same cost or lower than you could manage. Then it makes sense to free up your capacity for more sensitive and likely more profitable parts.
 
Last edited:
If you want GPU prices to come down, Intel having a competitive node would be a good thing. Even if they need a little bit of help right now.
It would free up TSMC from producing ARC, and Intel could sell their GPUs for profit, rather than at cost.
 
  • Like
Reactions: KyaraM and artk2219
Honestly for some non essential or low margin parts, if you can get someone to make it quickly, and at the same cost or lower than you could manage. Then it makes sense to free up your capacity for more sensitive and more likely profitable parts.
Even if it turns out more expensive than making it yourself, if it frees up capacity for something you can sell for a higher margin, or plain more money than the extra cost, then it's still worth it.
I mean even if iGPUs/GPUs that are made at tsmc are more expensive, if that allows them to make more 14900ks (or whatever, server parts) then it's still worth it.
 
  • Like
Reactions: KyaraM and artk2219
The 10nm node has been a disaster from Intel. The reason why, will never revealed, but it's interesting to ponder.

In my perspective, the foundry business has been an extremely risky cut-throat industry since entering the sub-100 nm era.

At this stage, the size of investments no longer matters, and (with a more controversial note) I'd also argue that talent/expertise do not even matter. Things can easily go wrong; one wrong decision, say betting on the wrong material or not implementing a technology sooner, can completely destroy a fab's effort to chase the leading-edge. For a company like Intel, this could be fatal to their very existence.

A bit of history:

Intel had been quite confident early on about their 10 nm process. They showed a working wafer and process in IEDM (Dec 2017), only to meaningfully release it in (very late) 2020 in the form of Tiger Lake which went up to only 45 W (except the special ed. -HK). Sapphire Rapids was released very late at 2021, but was known to be a failure both in performance and economics. Only its improved derivative (later rebranded as Intel 7) was decent, though immense damage was inevitable as Intel lost its long-held advantage.

Currently:

Intel 4 (or 3) worries me as the only meaningful product fabbed with this specification to date is the compute/CPU "chiplet" of Meteor Lake. It's both not much (in quantity), and not impressive (in performance). Granite Rapids, slated for late this year, might be the make-or-break situation, because it's going to be very hard for Intel to cover its fab tech deficit with ambiguous message when GNR replicates Sapphire Rapids.

Let's not begin about 20A yet. This thing is, chronologically, an order of magnitude more ambitious than Intel's 10 nm, and not even close to TSMC's N3. They have to go through that almost impossible window of success to successfully mass produce this as they promised (2H 2024). It's also very hard to see them coming back to lead when 20A only ends up in "real products" in 2H 2026 or later.

It's hard to predict the success, because it's more like shooting from the halfway line. No matter how good you are, you have to have a lot of luck. I'd say that the last ditch effort for Intel is to increase this "outsourcing percentage" and become fabless once and for all. I'd even argue that this might be the most likely decision that they've to take.

Unless of course, the US govt steps in and gave tens of billions to them. That kind of "bail-out" hasn't happened to a fab before, like ever, so honestly couldn't comment on that.
 
  • Like
Reactions: artk2219
Things can easily go wrong; one wrong decision, say betting on the wrong material or not implementing a technology sooner, can completely destroy a fab's effort to chase the leading-edge. For a company like Intel, this could be fatal to their very existence.
For intel leading edge doesn't matter, what are people going to do?! NOT buy CPUs because they are not on the leading edge node wise, even though their performance is fine?!
At worst some people might wait a few years more but they will still upgrade.

For their IFS it might play some role but even there, older more stable nodes are always still lucrative. If they can match the bulk of what tsmc offers, even without the top leading edge one, at similar prices then they will do fine.

You are talking as if intel is only a FAB.
 
  • Like
Reactions: KyaraM and artk2219
It turns out when you're building capacity for customers you don't have yet in a capital intensive business you take losses? Who could possibly have guessed this was the case?!
A bit of history:

Intel had been quite confident early on about their 10 nm process. They showed a working wafer and process in IEDM (Dec 2017), only to meaningfully release it in (very late) 2020 in the form of Tiger Lake which went up to only 45 W (except the special ed. -HK). Sapphire Rapids was released very late at 2021, but was known to be a failure both in performance and economics. Only its improved derivative (later rebranded as Intel 7) was decent, though immense damage was inevitable as Intel lost its long-held advantage.
This is... not particularly accurate. Intel released Cannon Lake (so they could say they didn't lie about 10nm being production ready even though it wasn't) on the original 10nm node in 2018. This was followed by Ice Lake on a revamped 10nm which they just called 10nm pretending Cannon Lake never happened in 2019. Tiger Lake was built on 10SF and this was the first major node improvement to 10nm which allowed them to get respectable clock speeds in 2020. Alder Lake was built on 10ESF (renamed to Intel 7 when Intel did the node name normalization) bringing us to the current version of the 10nm manufacturing process in 2021. Sapphire Rapids is built on the same process node and its problems were largely twofold: bugs found in silicon which required restarting mass production and the complex quad tile nature of the design.
Intel 4 (or 3) worries me as the only meaningful product fabbed with this specification to date is the compute/CPU "chiplet" of Meteor Lake. It's both not much (in quantity), and not impressive (in performance). Granite Rapids, slated for late this year, might be the make-or-break situation, because it's going to be very hard for Intel to cover its fab tech deficit with ambiguous message when GNR replicates Sapphire Rapids.
Intel 4 is for all intents a one and done node for the Intel primary markets. It's their first EUV node and despite that it's still cheaper to manufacture than Intel 7 is. MTL doesn't feature a significant core redesign, but rather more of what would have been a "tock" in Intel's old tick tock parlance. Intel 3 is slated to be the long term node which we'll first see GNR and SF this year.
Let's not begin about 20A yet. This thing is, chronologically, an order of magnitude more ambitious than Intel's 10 nm, and not even close to TSMC's N3. They have to go through that almost impossible window of success to successfully mass produce this as they promised (2H 2024). It's also very hard to see them coming back to lead when 20A only ends up in "real products" in 2H 2026 or later.
20A seems to be another one and done with regards to Intel high volume parts. Chances are it's still the node for Arrow Lake CPU tiles which should be out this year. We still don't have enough information from Intel to say any of this is concrete though. 18A brings Intel to approximate node parity with their competition according to the slides released with the presentation mentioned in the article. It begins shipping next year and is the first leading node Intel is selling to external customers.
 
  • Like
Reactions: JarredWaltonGPU
I dont know exactly hoe these things work, but i dont agree on a company that has billions of dollars in the bank (lots of it thanks to illegal actions) to receive 10 billions more in the form of a gift, because its not even a loan.

And to the poster that mentioned gpu prices, abundant capacity wont make dear leader jensen cut prices, especially when media and influencers keep hyping his hardware without criticizing of their actions and consequences.

Only thing that will make them cut prices is if people stop buying.
 
It turns out when you're building capacity for customers you don't have yet in a capital intensive business you take losses? Who could possibly have guessed this was the case?!
I think this is the gist of the topic. I'm not sure where this discussion will lead to, but let me first say that investors need assurance and must be able to gauge the risks. This is the basis of my argument. I don't really want to debate whether the loss was really "a literal loss", but rather I want to imagine the risks involved in that quote-on-quote loss (or high capex, if preferable) and its implications for the next few years.

For their IFS it might play some role but even there, older more stable nodes are always still lucrative. If they can match the bulk of what tsmc offers, even without the top leading edge one, at similar prices then they will do fine.

You are talking as if intel is only a FAB.
I have to reiterate. I am not specifically referring to the loss of lead, but I specifically refer to the FAILURE of attempting to reach to the top. GloFo is doing fine, TSMC is still taking a lot of 28 nm orders too. But let's not confuse intentional loss and attempted failure. Two of which are in the opposite direction in terms of annual capex. Intel is struggling to keep its cash flow and revenue healthy during the last two years, and I'd argue the main driving factor (other than the competition) is still the overreliance on their fab.

Of course, I agree that my mention of "fabless" is a completely unnecessary metaphor, but let's not get carried away with that. They have the option to abandon or slow down their pursuit for the leading-edge and increase that outsourcing on the leading-edge nodes, something that Intel should strongly consider at this point.

For intel leading edge doesn't matter, what are people going to do?! NOT buy CPUs because they are not on the leading edge node wise, even though their performance is fine?!
At worst some people might wait a few years more but they will still upgrade.
Intel has to prepare a cut on gross margins, in that case. It's already below 40% last year, down from the 60%+ on their pre-pandemic era.

This is... not particularly accurate. Intel released Cannon Lake (so they could say they didn't lie about 10nm being production ready even though it wasn't) on the original 10nm node in 2018. This was followed by Ice Lake on a revamped 10nm which they just called 10nm pretending Cannon Lake never happened in 2019. Tiger Lake was built on 10SF and this was the first major node improvement to 10nm which allowed them to get respectable clock speeds in 2020. Alder Lake was built on 10ESF (renamed to Intel 7 when Intel did the node name normalization) bringing us to the current version of the 10nm manufacturing process in 2021. Sapphire Rapids is built on the same process node and its problems were largely twofold: bugs found in silicon which required restarting mass production and the complex quad tile nature of the design.
Sorry, I did mixed up in the ADL and SPR part, you did correct me there, I have to say thanks.

But I wanted to mention that 10 nm actually only turned out to be "a considerable thing" effectively in 2021, and "a proper node" effectively in 2022, more than 4 years after Intel demoed it for the first time. The recent average is around 16 to 20 months, for reference.

So the corrected paragraph would be:

They showed a working wafer and process in IEDM (Dec 2017), only to *meaningfully* release it [for clarity: refers to ANY 10 nm process] in (very late) 2020 in the form of Tiger Lake which went up to only 45 W (except the special ed. -HK). Sapphire Rapids [EDIT: NOT SPR, but ADL] was released very late at 2021, as one of its derivatives (later rebranded as Intel 7) that did end up decent [for clarity: went above traditional DT and HEDT TDP limits]. Though, immense damage was inevitable as Intel lost its long-held advantage [reason: 2022 was a very bad year for them, followed by an even worse 2023].

20A seems to be another one and done with regards to Intel high volume parts. Chances are it's still the node for Arrow Lake CPU tiles which should be out this year. We still don't have enough information from Intel to say any of this is concrete though. 18A brings Intel to approximate node parity with their competition according to the slides released with the presentation mentioned in the article. It begins shipping next year and is the first leading node Intel is selling to external customers.
Yes, correct, at least that's what Intel said.

But also, let's consider the facts first. 20A will be Intel's first mass produced GAAFET/MGFET (dubbed ribbonFET) AND backside interconnect (dubbed powerVIA) process. Samsung is also currently doing MGFET for the 3 nm class, but performance improvements are still minimal vs. its 5/4 nm class process. Samsung had no mentions of backside interconnect on their 3 nm demo.

20A/18A is hence, tough on the technical and financial level, so to mass produce this at late 2024, the margins for the targets to meet is razor thin. Considering the odds, I think a delay to 2025, or even a 1H 2026 is perfectly reasonable. But to cover that, you either need the 4/5 nm to "work" (in order to keep profit margins and cash flow in check) or increase the outsourcing rate. Profit on the fab will go down, but added value would be imposed to other segments stemming from higher performance and volume. I of course, hope that Intel will be able to successfully pick either.

If either is not fulfilled, I've to say that there's no other choice here but to rely on huge bailouts by the US govt. I don't know how that will pan out, but I'd say that thanks to the USA's bureaucracy, the GOP, and Dem hardliners like Bernie, only 38bn USD is allocated towards subsidizing the Si-manufacturing sector.

Interesting times ahead for me. I hope you are as interested.
 
Last edited:
  • Like
Reactions: thestryker
Sorry, I did mixed up in the ADL and SPR part, you did correct me there, I have to say thanks.

But I wanted to mention that 10 nm actually only turned out to be "a considerable thing" effectively in 2021, and "a proper node" effectively in 2022, more than 4 years after Intel demoed it for the first time. The recent average is around 16 to 20 months, for reference.
TGL was absolutely considerable seeing as it was the entire mobile stack which moves significantly more parts than the desktop side of things. It was a proper node at that point and that was the most significant change made to the 10nm process to date. If you're not familiar with why I'm saying that feel free to read about the specific changes disclosed:
https://www.anandtech.com/show/1597...-soc-detailed-superfin-willow-cove-and-xelp/2

10nm being a full node was still around 4-5 years late by the time TGL launched.
But also, let's consider the facts first. 20A will be Intel's first mass produced GAAFET/MGFET (dubbed ribbonFET) AND backside interconnect (dubbed powerVIA) process. Samsung is also currently doing MGFET for the 3 nm class, but performance improvements are still minimal vs. its 5/4 nm class process. Samsung had no mentions of backside interconnect on their 3 nm demo.
BSP is actually optional for the 20A/18A process (Intel will absolutely be using it for their own volume products), but I cannot recall whether or not GAA is. This was likely a benefit of developing BSP on Intel 4 to ensure they weren't tying too many technologies to a single node in case something didn't pan out. It's a very shrewd way of developing process technologies and one that in the past Intel never bothered to do (hence the 14nm delays and 10nm debacle).
20A/18A is hence, tough on the technical and financial level, so to mass produce this at late 2024, the margins for the targets to meet is razor thin. Considering the odds, I think a delay to 2025, or even a 1H 2026 is perfectly reasonable. But to cover that, you either need the 4/5 nm to "work" (in order to keep profit margins and cash flow in check) or increase the outsourcing rate. Profit on the fab will go down, but added value would be imposed to other segments stemming from higher performance and volume. I of course, hope that Intel will be able to successfully pick either.
We'll find out, but I think they already made that choice with regards to their own products. It's pretty clear that Lunar Lake is going to be TSMC which makes sense given the integrated and low power nature of the part. ARL is probably going to be 20A, but it seems like it won't be covering the entire stack based on the most recent rumors, but again nothing directly clarifying from Intel yet.
 
Sapphire Rapids was released very late at 2021
No... that's when Intel wishes it launched, but it didn't reach general availability until January 2023:

Intel 4 (or 3) worries me as the only meaningful product fabbed with this specification to date is the compute/CPU "chiplet" of Meteor Lake. It's both not much (in quantity), and not impressive (in performance).
I read somewhere (probably in these forums) that Intel 4 is not a complete node. Its library consists only of the cells needed for compute chiplets, but not everything you'd need for other sorts of tiles. That explains why Intel 3 is the node being marketed to some foundry customers, as well as possibly why it's the one being used for the server products.
 
TGL was absolutely considerable seeing as it was the entire mobile stack which moves significantly more parts than the desktop side of things. It was a proper node at that point and that was the most significant change made to the 10nm process to date. If you're not familiar with why I'm saying that feel free to read about the specific changes disclosed:
https://www.anandtech.com/show/1597...-soc-detailed-superfin-willow-cove-and-xelp/2

10nm being a full node was still around 4-5 years late by the time TGL launched.
My diction choice might not be completely correct, English isn't my native language. But I think we have no differences here. TGL was the first 10 nm that can be "considered". ADL (Intel 7/10 ESF) was designated as the first "proper 10 nm" as that had the ability to be binned reliably for traditional, high-margin TDPs. That was 6 years off from Intel's plan, 4 years off from Intel's technical demo.

The accuracy of my statements might not be pinpoint, but this is to support the narrative that:
In my perspective, the foundry business has been an extremely risky cut-throat industry since entering the sub-100 nm era.

At this stage, the size of investments no longer matters, and (with a more controversial note) I'd also argue that talent/expertise do not even matter. Things can easily go wrong; one wrong decision, say betting on the wrong material or not implementing a technology sooner, can completely destroy a fab's effort to chase the leading-edge. For a company like Intel, this could be fatal to their very existence.
Hence, not even a technical demo can imply anything. Let alone roadmaps and "promises". Risks are too high. I'd like to see Intel preparing better.

BSP is actually optional for the 20A/18A process (Intel will absolutely be using it for their own volume products), but I cannot recall whether or not GAA is. This was likely a benefit of developing BSP on Intel 4 to ensure they weren't tying too many technologies to a single node in case something didn't pan out. It's a very shrewd way of developing process technologies and one that in the past Intel never bothered to do (hence the 14nm delays and 10nm debacle).
Thanks for reminding that part, that is a slight detail that I didn't consider. Intel's mention of BSPDN being scrapped if it isn't panning the way it was planned is, of course, a welcomed strategy to reduce risks. However, it could also be said that MGFET/ribbon is non negotiable on 20A. No 2 nm class FET with that kind of source-drain pitch is possible without GAA, and that itself is still a huge risk too, which I believe remains to be riskier than Intel's ambitious 10 nm plan. Regardless, they need to get GNR/Intel 4(3) work well or increase the outsourcing rate.

I believe, even without BSPDN, a 2026 mass production is still the most realistic goal, and if they had to postpone it beyond that, Intel will still not have the leading edge. Not necessarily a negative thing if they prepared for this, though.

We'll find out, but I think they already made that choice with regards to their own products. It's pretty clear that Lunar Lake is going to be TSMC which makes sense given the integrated and low power nature of the part. ARL is probably going to be 20A, but it seems like it won't be covering the entire stack based on the most recent rumors, but again nothing directly clarifying from Intel yet.
It's evident since last year when they decided to outsource ARC and some MTL tiles to TSMC, which were clearly decisions based on necessity. No one should deny that the fab business, in this decade, is going to be VERY tough and there's unfortunately a lot of people who still thinks that Intel is exaggerating. Not only that Intel knows that it's extremely risky and expensive, but it's also almost capital-independent. We should leave the narrative that it's "sinful" for Intel to no longer chase the leading edge or even go entirely fabless.

I of course, hope that Intel will be able to successfully pick either.
I've to correct myself here, the sentence is structurally ambiguous. I wanted to say "I hope that Intel will be successful after picking either".
 
Last edited: