Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/threads/google-and-nvidia-consider-intel-as-backup-chip-manufacturer-the-information.25255/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/EmailDomainReplace] => 1000010
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2031070
            [XFI] => 1060170
        )

    [wordpress] => /var/www/html
)

Google and Nvidia Consider Intel as Backup Chip Manufacturer - The Information

Brady

Well-known member
The chipmaker has secured billions of US dollars of investments from the Trump administration, Nvidia and SoftBank

  • Google has been pushing to make its in-house AI chips a viable alternative to Nvidia’s dominant GPUs.

Google has been pushing to make its in-house AI chips a viable alternative to Nvidia’s dominant GPUs. PHOTO: BLOOMBERG


[BENGALURU] Alphabet’s Google has placed an order with Intel to manufacture more than three million tensor processing units in 2028, The Information reported on Monday (Jun 8), citing sources with direct knowledge of the discussions.

Nvidia is also evaluating whether Intel’s technology can be used to make a processor that combines four graphics chips into a single unit, although it has not placed an order with the company yet, the report said.

Intel’s shares rose more than 9 per cent in early trading, set to add to the nearly 169 per cent gain so far this year, on the back of signs of steady turnaround progress under CEO Tan Lip-Bu.

Intel declined to comment on the report, while Alphabet and Nvidia did not immediately respond to requests for comment.

Reuters could not independently verify the report.

The potential order for Google’s in-house AI chips would bolster Intel’s contract chip manufacturing business and comes as the company tries to claw back its chipmaking crown that it lost to Taiwan Semiconductor Manufacturing Company (TSMC) following years of management blunders.

Soaring chip demand from the AI boom has left TSMC struggling to bring in adequate supply. That capacity crunch has prompted several major AI chip design companies to turn to Intel, The Information said.

The news is “evidence that AI’s biggest players are racing to diversify a supply chain still heavily concentrated in TSMC”, said Jacob Bourne, technology analyst at eMarketer.

Since Tan took charge, Intel has secured billions of US dollars of investments from the Trump administration, Nvidia and SoftBank.


It also landed Tesla as the first major customer for its next-generation 14A manufacturing process to make chips for Elon Musk’s Terafab project, an advanced AI chip complex he has envisioned in Austin.

The Trump administration has also been trying to drum up business for Intel, an official said last month.

“Beyond the standard need to diversify, Google and Nvidia are even more motivated than usual to work with Intel. Supporting Intel supports US-based manufacturing, which is important for the relationship with the US administration,” said DA Davidson analyst Gil Luria.

Last month, The Wall Street Journal reported that Intel has reached a preliminary deal to make some chips for Apple devices following intensive talks for more than a year.

Google, meanwhile, has been pushing to make its in-house AI chips a viable alternative to Nvidia’s dominant GPUs, with sales of its tensor processing units becoming a growth driver for the company’s cloud revenue. REUTERS



From Reuters because paywall:
June 8 (Reuters) - Alphabet's (GOOGL.O)Google has placed an order with Intel (INTC.O) to manufacture more than three million tensor processing units in 2028, The Information reported on Monday, citing people with direct knowledge of the discussions.
Nvidia (NVDA.O) is also evaluating whether Intel's technology can ‌be used to make a processor that combines four graphics chips into a single unit, although it has not placed an order with the company yet, the report said.
https://www.reuters.com/business/go...-manufacturer-information-reports-2026-06-08/
 
Last edited by a moderator:
Is the Not TSMC market taking off?

3 million sounds like a lot but in reality isn’t many chips compared to, say, consumer PCs shipping 100M+ devices a year.
 
Is the Not TSMC market taking off?

3 million sounds like a lot but in reality isn’t many chips compared to, say, consumer PCs shipping 100M+ devices a year.
depends on chip size. there are 25M DC CPUs per year, 250M PC CPUs, 1.4B phone processors, IDK how many GPUs/TPUs

3M TPUS is a good chunk but if it was one chip it is about 1/10th of a Fab. 1-2B in Revenue including packaging.
 
Given that CC Wei is saying that "Supply Won’t Meet AI-Fueled Demand for Years" the NOT TSMC Market has never been better. Lip-Bu is doing a great job but his timing could not have been better with the Intel Foundry push.

TSMC N2 won the first wave of 2nm business but Intel 18P and Samsung SF2 are now viable alternatives. 1.4nm is another story. Intel has a good shot at some first wave business. Samsung SF1.4 probably not.

As it stands today TSMC A14 and Intel 14A will start HVM in 2028 and Samsung is a year or more behind that if you match yield. Intel Foundry is working closely with PDF Solutions on 14A yield so I have very high hopes.
 
Is the Not TSMC market taking off?

3 million sounds like a lot but in reality isn’t many chips compared to, say, consumer PCs shipping 100M+ devices a year.
And guess what, that 60% of these PC market i.e. laptop and desktop is now move away from TSMC to Intel Fab, guess what the name is Panther Lake, Wildcat Lake from Lunar Lake
 
Given that CC Wei is saying that "Supply Won’t Meet AI-Fueled Demand for Years" the NOT TSMC Market has never been better. Lip-Bu is doing a great job but his timing could not have been better with the Intel Foundry push.

TSMC N2 won the first wave of 2nm business but Intel 18P and Samsung SF2 are now viable alternatives. 1.4nm is another story. Intel has a good shot at some first wave business. Samsung SF1.4 probably not.

As it stands today TSMC A14 and Intel 14A will start HVM in 2028 and Samsung is a year or more behind that if you match yield. Intel Foundry is working closely with PDF Solutions on 14A yield so I have very high hopes.
If your A14 is referred to a Renamed A16 then A14 will be available in 2028, while 14A will be in 2028 (what 14A is 18AP will PowerVIA 2 i.e. direct BSPD and EUV High NA), but for A16/A14 it is Quad Patterning and that is good luck to you.

Look at the timeline (forecast). 2026 N2 2027 N2P, Then 2028 is A16 that just make sense.

Because TSMC and its Fanboy and its PR Robot, is so into BSPD is for HPC, they just don't have the desktop / laptop / mobile phone market to test run BSPD to get their yield up and going. A16 is going to HPC directly ??? Wrong move.
 
As it stands today TSMC A14 and Intel 14A will start HVM in 2028 and Samsung is a year or more behind that if you match yield. Intel Foundry is working closely with PDF Solutions on 14A yield so I have very high hopes.
Intel has been working with PDFS for a while .. back to the Pat G days. Pat is a financial supporter and drives their Advisory Board.
 
And guess what, that 60% of these PC market i.e. laptop and desktop is now move away from TSMC to Intel Fab, guess what the name is Panther Lake, Wildcat Lake from Lunar Lake
where is Nova Lake made? The only factory Intel wants to ramp down is Intel 7 and 10, correct? TSMC has loadings for some time.... please correct me if I am wrong.... seems like you have some good info.
 
Intel has been working with PDFS for a while .. back to the Pat G days. Pat is a financial supporter and drives their Advisory Board.
I don't know about PG being their financial supporter and being on advisory board but they have been working together for a long time.


1781136031679.png
1781136091283.png
1781136545193.png


 

Attachments

  • 1781136139422.png
    1781136139422.png
    255.7 KB · Views: 18
  • 1781136517813.png
    1781136517813.png
    255.7 KB · Views: 15
3M TPUS is a good chunk but if it was one chip it is about 1/10th of a Fab. 1-2B in Revenue including packaging.
I believe this is only an advanced packaging design win for Intel Foundry not wafer order.

But lets do rough math exercise on what if it is actually a wafer + adv.packaging order.
TPU 7 had two compute die with die area of 720sq-mm (got it online, so not a credible source). If we assume similar design for TPU v9 (I am pretty sure I am underestimating this, could be 4 dies by the way things are progressing in AI accelerator land) die is 25.8 mm x 27.9 mm dies size, with a defect density 0.1/Sq-cm (we all know 18A yield sucks, LBT says so, so this is me being generous), that is 40 good die per 300mm wafer. Lets assume parametric yield is 80% (again we all know Intel Foundry sucks), so that is 32 Known Good Die. So (3e6/16) @ 2 die per TPU = 187,500 of 18A-P wafers (equal to 15,625 wspm = ~78% utilization for a fab the size of Fab 62), At $22,500 per wafer pricing, this amounts to $4.22B of foundry revenue. Assuming another 50% for packaging costs & other services. i.e. = $6.328B of total foundry revenue.

Plus all the added benefits of foundry utilization, yield learning etc etc...

Except for the 50% added cost for adv.packaging, other numbers are reasonable imo. I have no idea how much advanced packaging like EMIB cost to model that out reasonably. But we do know Intel CFO commentary saying the projected revenue opportunity of an advanced packaging orders could be in the $billions of dollars.
 
Last edited:
I believe this is only an advanced packaging design win for Intel Foundry not wafer order.

But lets do rough math exercise on what if it is actually a wafer + adv.packaging order.
TPU 7 had two compute die with die area of 720sq-mm (got it online, so not a credible source). If we assume similar design for TPU v9 (I am pretty sure I am underestimating this, could be 4 dies by the way things are progressing in AI accelerator land) die is 25.8 mm x 27.9 mm dies size, with a defect density 0.1/Sq-cm (we all know 18A yield sucks, LBT says so, so this is me being generous), that is 40 good die per 300mm wafer. Lets assume parametric yield is 80% (again we all know Intel Foundry sucks), so that is 32 Known Good Die. So (3e6/16) @ 2 die per TPU = 187,500 of 18A-P wafers (equal to 15,625 wspm = ~78% utilization for a fab the size of Fab 62), At $22,500 per wafer pricing, this amounts to $4.22B of foundry revenue. Assuming another 50% for packaging costs & other services. i.e. = $6.328B of total foundry revenue.

Plus all the added benefits of foundry utilization, yield learning etc etc...

Except for the 50% added cost for adv.packaging, other numbers are reasonable imo. I have no idea how much advanced packaging like EMIB cost to model that out reasonably. But we do know Intel CFO commentary saying the projected revenue opportunity of an advanced packaging orders could be in the $billions of dollars.

TPU v7 is already in operation, and TPU 8t and TPU 8i are expected to go online by the end of 2026. For this potential TPU deal with Intel to materialize in 2028, there isn't much time left. I tend to think it would involve a special version of Google's TPU with a much smaller die size than the current TPUs for two reasons:

1. Both Google and Intel want to reduce and manage risk.

With less than two years remaining until 2028 and Google's AI hosting business booming, Google cannot afford much uncertainty. A smaller die size would help both Google and Intel improve yields and reduce manufacturing risk.

2. Based on your wafer volume calculations, Google's 2028 demand for this Intel Foundry order is not large enough that TSMC could not accommodate it.

I don't think Google would bet a major TPU generation on its first attempt with Intel Foundry. While Google wants to improve supply-chain diversity, it would be unwise to risk creating a hole in its TPU roadmap if something happens.
 
With less than two years remaining until 2028 and Google's AI hosting business booming, Google cannot afford much uncertainty. A smaller die size would help both Google and Intel improve yields and reduce manufacturing risk.
Like I said, we already know only the advanced packaging is being done at Intel Foundry for this TPU. My math exercise is to show if it wasn't just a packaging order, it would have been a bigger oppurtunity for Intel. Something to note, Intel has been making similar or larger die size DC CPUs for a long time and packaging them with EMIB including HBM.

1781144863225.png

2. Based on your wafer volume calculations, Google's 2028 demand for this Intel Foundry order is not large enough that TSMC could not accommodate it.
TSMC is really capacity constrained, current estimate by a leading wall st investment bank is 600k wafer supply-demand deficit on N3 in 26/27. TSMC being too conservative in the last couple of years, turning out to be a gift to Samsung/Intel Foundry provided they can leverage this. Looks like "build it, they will come" does work.
 
I believe this is only an advanced packaging design win for Intel Foundry not wafer order.

But lets do rough math exercise on what if it is actually a wafer + adv.packaging order.
TPU 7 had two compute die with die area of 720sq-mm (got it online, so not a credible source). If we assume similar design for TPU v9 (I am pretty sure I am underestimating this, could be 4 dies by the way things are progressing in AI accelerator land) die is 25.8 mm x 27.9 mm dies size, with a defect density 0.1/Sq-cm (we all know 18A yield sucks, LBT says so, so this is me being generous), that is 40 good die per 300mm wafer. Lets assume parametric yield is 80% (again we all know Intel Foundry sucks), so that is 32 Known Good Die. So (3e6/16) @ 2 die per TPU = 187,500 of 18A-P wafers (equal to 15,625 wspm = ~78% utilization for a fab the size of Fab 62), At $22,500 per wafer pricing, this amounts to $4.22B of foundry revenue. Assuming another 50% for packaging costs & other services. i.e. = $6.328B of total foundry revenue.

Plus all the added benefits of foundry utilization, yield learning etc etc...

Except for the 50% added cost for adv.packaging, other numbers are reasonable imo. I have no idea how much advanced packaging like EMIB cost to model that out reasonably. But we do know Intel CFO commentary saying the projected revenue opportunity of an advanced packaging orders could be in the $billions of dollars.
I have very different numbers. Fab size, yield. wafer price, packaging price, even the fab name Would be gald to review them with you and you can decide
 
where is Nova Lake made? The only factory Intel wants to ramp down is Intel 7 and 10, correct? TSMC has loadings for some time.... please correct me if I am wrong.... seems like you have some good info.
And guess what the whole Desktop CPU market is just around 1/4 of laptop, so Nova Lake is not that important to start off with and the more you move onto the mini PC it they used laptop CPU, the volume for desktop vs laptop CPU favour the laptop, there isn't that much workstation laptop out there vs mini pc with laptop chip
 
And guess what the whole Desktop CPU market is just around 1/4 of laptop, so Nova Lake is not that important to start off with and the more you move onto the mini PC it they used laptop CPU, the volume for desktop vs laptop CPU favour the laptop, there isn't that much workstation laptop out there vs mini pc with laptop chip
I believe Intel will give some color on internal vs external use going forward. I believe the intention is to use TSMC at about the same level and ramp 3/18A to replace 10/7 (10/7 was 80% of revenue in 2025).

The other issue is price. Intel 7 was much lower price and still had higher margins. The Price and Cost should both change given the "shortage"... not sure how this will pan out (Does intel want more share or higher margins)

Lets see if Intel provides an update here in the next two quarters report out . I am told Q2/3 is inflection point. Not sure I buy that, but lets see
 
If your A14 is referred to a Renamed A16 then A14 will be available in 2028, while 14A will be in 2028 (what 14A is 18AP will PowerVIA 2 i.e. direct BSPD and EUV High NA), but for A16/A14 it is Quad Patterning and that is good luck to you.

Look at the timeline (forecast). 2026 N2 2027 N2P, Then 2028 is A16 that just make sense.

Because TSMC and its Fanboy and its PR Robot, is so into BSPD is for HPC, they just don't have the desktop / laptop / mobile phone market to test run BSPD to get their yield up and going. A16 is going to HPC directly ??? Wrong move.
how much capacity you think Intel avaiable for 14A in 2028? and how many HiNA they need for that capacity? is that match to ASML's HiNA shipment?
 
And guess what, that 60% of these PC market i.e. laptop and desktop is now move away from TSMC to Intel Fab, guess what the name is Panther Lake, Wildcat Lake from Lunar Lake
If by "Not TSMC market" you mean semiconductor manufacturing outside of Taiwan Semiconductor Manufacturing Company, then yes, it is growing.
 
how much capacity you think Intel avaiable for 14A in 2028? and how many HiNA they need for that capacity? is that match to ASML's HiNA shipment?
All depends on Elon Musk, what Intel needs is not a real cash from customer, it needed a real commitment to 14A, let say if Elon Musk sign a commitment of 30 million chip then the capacity will be 30 million chip i.e around 300K waffle.

Another one that is not on the surface is how many Apple needs.

Intel Internally will budget for 480K waffle p.a., anything extra need commitment from Customer.

The next is 18A capacity allocation, therefore how many waffle and its allocation of Intel 3 in 2028:
Question and Answer that the CEO needed to make:
How he think the market of iGPU, can the new "Lakes" desktop CPU can be within 5% of AMD will be offer.
A - if that is the case, then they can use a "weak" GPU and start arrange iGPU on top desktop GPU to have "weak" iGPU, since those machine will have 6080 anyway.
B - If the desktop still can't be within 5% of AMD will be offer then pair it with a strong iGPU to go for mid-range market. Then Intel 3 is out of the picture goes for Arc 3 on 18A.

Then the next best thing is how Steam going to react:
Since in pure performance the current panther lake is ahead of what AMD is offering, can Intel get the Steam 2 or another offering like PSP xbox ... . If that is the case then they needed 18A capacity to be doing iGPU.

Then of course the 14A capacity is going to be upward of 500K waffle p.a. for Intel internal alone. Because if the desktop design is great again, that will impact the server market and that 500K is just to signal the Intel server market recovery. Although that is 3 years away, but they can get Dell for commitment since the server market is running so hot. Ask Dell to be committed helps Intel to get its cash flow forecasted, allocated better availability for those who will be committing now

If Apple and SpaceX is all fully committed, then Intel can easily go to the bank and get some cash flow, but if not then 500K, the most likely situation is 750K p.a.

For ASML EXE 5200 is listed as 175 per hour, but remember that GAA is 4 layers for 18A (assume that is 4 in 14A), I am not an insider but I think they needed 2-3 runs for each layer, that is already 8 runs, plus the top 3 line layer will be High NA EUV, that make it 10-12 run, then the maths, but I don't have that much time and as I don't have perfect data it is so difficult to estimate, so in short I think Intel needs 10 - 16 High NA EUV to make any sense to reach a reasonable level.
 
All depends on Elon Musk, what Intel needs is not a real cash from customer, it needed a real commitment to 14A, let say if Elon Musk sign a commitment of 30 million chip then the capacity will be 30 million chip i.e around 300K waffle.

Another one that is not on the surface is how many Apple needs.

Intel Internally will budget for 480K waffle p.a., anything extra need commitment from Customer.

The next is 18A capacity allocation, therefore how many waffle and its allocation of Intel 3 in 2028:
Question and Answer that the CEO needed to make:
How he think the market of iGPU, can the new "Lakes" desktop CPU can be within 5% of AMD will be offer.
A - if that is the case, then they can use a "weak" GPU and start arrange iGPU on top desktop GPU to have "weak" iGPU, since those machine will have 6080 anyway.
B - If the desktop still can't be within 5% of AMD will be offer then pair it with a strong iGPU to go for mid-range market. Then Intel 3 is out of the picture goes for Arc 3 on 18A.

Then the next best thing is how Steam going to react:
Since in pure performance the current panther lake is ahead of what AMD is offering, can Intel get the Steam 2 or another offering like PSP xbox ... . If that is the case then they needed 18A capacity to be doing iGPU.

Then of course the 14A capacity is going to be upward of 500K waffle p.a. for Intel internal alone. Because if the desktop design is great again, that will impact the server market and that 500K is just to signal the Intel server market recovery. Although that is 3 years away, but they can get Dell for commitment since the server market is running so hot. Ask Dell to be committed helps Intel to get its cash flow forecasted, allocated better availability for those who will be committing now

If Apple and SpaceX is all fully committed, then Intel can easily go to the bank and get some cash flow, but if not then 500K, the most likely situation is 750K p.a.

For ASML EXE 5200 is listed as 175 per hour, but remember that GAA is 4 layers for 18A (assume that is 4 in 14A), I am not an insider but I think they needed 2-3 runs for each layer, that is already 8 runs, plus the top 3 line layer will be High NA EUV, that make it 10-12 run, then the maths, but I don't have that much time and as I don't have perfect data it is so difficult to estimate, so in short I think Intel needs 10 - 16 High NA EUV to make any sense to reach a reasonable level.

I thought Elon Musk wanted to license Intel's 14A process technology and produce chips at SpaceX–Tesla's own Terafab. It would have no effect on Intel Foundry's capacity and utilization. Has something changed recently?
 
Back
Top