Intel foundry business to make custom chip for Amazon, chipmaker's shares jump

[hist78]

"Sept 16 (Reuters) - Intel's (INTC.O), opens new tab foundry, or contract manufacturing business, has signed up Amazon's (AMZN.O), opens new tab cloud services unit as a customer for making custom artificial intelligence chips, the companies said on Monday, a deal that gives the chip maker a vote of confidence.

Intel's shares rose roughly 8% in extended trading after CEO Pat Gelsinger released a memo to employees announcing Intel had secured the Amazon unit as be a multibillion-dollar customer, paying Santa Clara, California-based Intel for design services and manufacturing. The memo also outlined Intel's planned cost cuts."

https://www.reuters.com/technology/intel-aws-expand-strategic-collaboration-chip-manufacturing-2024-09-16/

 

[hist78]

"Today we announced that we will expand our strategic collaboration with Amazon Web Services (AWS). This includes a co-investment in custom chip designs, and we have announced a multi-year, multi-billion-dollar framework covering product and wafers from Intel.

Specifically, Intel Foundry will produce an AI fabric chip for AWS on Intel 18A. We will also produce a custom Xeon 6 chip on Intel 3 that builds on our existing partnership, under which Intel produces Xeon Scalable processors for AWS. More broadly, we expect to have deep engagement with AWS on additional designs spanning Intel 18A, Intel 18AP and Intel 14A.

This framework reflects the power of our “better together” strategy, anchored on our integrated portfolio across foundry services, infrastructure and x86 products. And with the 5N4Y finish line in sight, we are beginning to see a meaningful uptick in interest from foundry customers. This includes continued momentum in advanced packaging, which remains a meaningful differentiator for Intel Foundry as we have tripled our deal pipeline since the beginning of the year."

Pat Gelsinger on Foundry Momentum, Progress on Plan

A message from Intel CEO Pat Gelsinger to employees regarding the next phase of Intel's transformation.

www.intel.com

 

[hist78]

This is great news for Intel Foundry. However, the custom-made Amazon chip project will likely have a limited impact on Intel Foundry's overall utilization. Below is my analysis, though you can adjust the numbers to reflect your own assumptions.

When compared to the AMD-Sony PlayStation 6 deal, estimated at 10 million units per year for 10 years, this progress should be measured accordingly.

 

[Y.H]

Sorry. Can I ask how do you figure out that Amazon need 1million chips per year ?

 

[tooLongInEDA]

Sorry. Can I ask how do you figure out that Amazon need 1million chips per year ?

I guess it's simply an estimate of demand. If this goes into data centres (let's assume "AI fabric" isn't for consumer edge devices), it seems reasonable to assume a demand of 10% of games consoles.

The other variable is die size. 200 die off a 300m wafer is something like a 32mm square die (rough estimate) which is at/possibly over the reticle limit (max die size). I doubt the die is that large (yield reasons, it's not nVidia), so chips/wafer almost certainly greater than 200 and wafers needed lower.

This really does bring it home how enormous the capacity of these fabs now is.

Now, can someone add the estimated revenue $ for each of these table lines ? What might an 18A class wafer be selling for these days ? We should be able to get a handle on the annual revenue estimate here.

 

[hist78]

Sorry. Can I ask how do you figure out that Amazon need 1million chips per year ?

The following is my estimation or guessing process:

Estimating Amazon's Potential Procurement of Amazon-Intel Processors:

Given the difficulty of accurately estimating the number of processors Amazon might procure, we can use several assumptions and available data to make an educated guess. Here are the key factors influencing the estimate:

1. AWS Server Growth:
   • AWS had an estimated 1.4 million servers in 2014, but AWS does not publicly disclose its current server count or annual growth rate.
   • 
     
   • Microsoft Azure had 4 million servers in 2021. Using Azure's growth as a rough proxy for AWS could provide some insight, but the comparison is limited due to differences in client base and server utilization and configuration.
2. Server to Processor Ratio:
   • Servers come in various configurations: single-CPU, dual-CPU, quad-CPU, or even 8-CPU setups. This makes it challenging to translate server count directly into the number of processors needed.
3. Revenue Comparison Between AWS and Azure:
   • In 2023, AWS had $54.648 billion in revenue compared to Azure's $32.197 billion. However, this revenue difference doesn't necessarily correspond directly to the number of servers, given variations in clientele and server efficiency (e.g., AWS hosts many video streaming services like Netflix, which uses just 100,000 server instances but serves 260 million members).
4. Virtual vs. Physical Servers:
   • AWS and other cloud providers use virtual machines, which further complicates estimating processor demand.
5. Competition Between Processors:
   • AWS uses processors from Intel, AMD, Nvidia, Amazon (Graviton), and now the Amazon-Intel project. The allocation among these providers is unknown and will depend on cost, function, and timing.
6. Server Lifecycle:
   • AWS has extended its server lifecycle from 3 years to 6 years, meaning that fewer servers may be replaced annually, but we also don’t know how many new servers will be deployed each year.
7. Timeframe for Amazon-Intel Project:
   • Assuming the Amazon-Intel processor will be available starting from 2026 because it will likely take 2–3 years to design, test, and reach high-volume manufacturing (HVM).

My Wild Guess Estimate:​

Assumption: The new Amazon-Intel processor will start delivery from 2026 through 2028.

1. Server Count Estimate:
   • Start with Azure’s 2021 server count of 4 million.
   • Apply 10% annual growth over 7 years (from 2021 to 2028):
   • 
     
   • 4 million servers (2021 Azure server number) * 1.10^7 ( 7 years of 10% annual server growth through 2028) = Total 7.795 million servers in 2028 (I brutally assigned this Azure 2028 number to AWS 2028).
2. AWS Server Share Powered by Amazon-Intel Processors:
   • Assume equal distribution among 5 processor providers for AWS (Intel, AMD, Amazon Graviton, Nvidia, and Amazon-Intel):
   • 
     
   • 7.795 million servers×20%=1.559 million servers using Amazon-Intel processors by 2028.
3. 
   
4. Server Rollout Over 3 Years (2026–2028):
   1.559 million servers / 3 years=0.52 million Amazon-Intel based servers deployment per year
5. 
   
6. Processor Count Estimate:
   • Assuming an average of 2 processors per server:
   • 
     
   • 0.52 million servers per year×2 processors per server=1.04 million processors per year 2026 ~ 2028

Conclusion:​

Based on these assumptions, Amazon could procure around 1.04 million processors per year from the Amazon-Intel project during the rollout period from 2026 to 2028. However, this is a rough estimate, subject to many unknown factors, including growth rates, processor allocation, server configurations and market situation.

 

[hist78]

I guess it's simply an estimate of demand. If this goes into data centres (let's assume "AI fabric" isn't for consumer edge devices), it seems reasonable to assume a demand of 10% of games consoles.

The other variable is die size. 200 die off a 300m wafer is something like a 32mm square die (rough estimate) which is at/possibly over the reticle limit (max die size). I doubt the die is that large (yield reasons, it's not nVidia), so chips/wafer almost certainly greater than 200 and wafers needed lower.

This really does bring it home how enormous the capacity of these fabs now is.

Now, can someone add the estimated revenue $ for each of these table lines ? What might an 18A class wafer be selling for these days ? We should be able to get a handle on the annual revenue estimate here.

If we assume $50 per processor and 300 die cuts from a 300mm wafer for this Amazon-Intel processor, this would translate to about $15,000 in revenue per wafer for Intel. Of course, @Daniel Nenni is more qualified to assess these numbers accurately.

However, at $50 per processor and 1 million processors per year, Intel Foundry would generate only $50 million in revenue. Even if we double the price to $100 per processor, that’s still just $100 million in revenue. If we also double the yearly volume from 1 million to 2 million units and increase the per-processor price to $200, Intel Foundry’s revenue would still be only $400 million—hardly a significant figure for a company of Intel’s size.

The key questions are:

1. What’s a reasonable wafer price for Intel 18A?
   Understanding the cost structure for Intel’s 18A process is crucial, as it will significantly impact pricing and profitability for these custom Amazon processors.
2. How big will the processor be?
   Is 300 processors per 300mm wafer a reasonable assumption? The die size could vary greatly depending on the complexity and the target performance of the chip, influencing both cost and yield.
3. What will be the volume of processors Amazon requires from this Amazon-Intel project?
   Estimating Amazon’s total processor demand is challenging, especially with Intel, AMD, Nvidia, and Amazon’s own Graviton processors also in the mix. The scale and purpose of this custom chip will dictate how many units are ultimately needed.
4. Who gets the most?
5. Intel stated that this Amazon-Intel multi-year collaboration will generate multi-billion dollars in revenue with the combination of Intel's design services, Intel Foundry, and Xeon processor sales. Given that Amazon already has processors from Intel, AMD, Nvidia, and its own Graviton lineup, it’s likely that the Intel custom-made chips will be on the lower end of the cost spectrum and Intel Foundry will get the least of the share.

 

[fansink]

Since this AWS order has countless contingency clauses, I wonder how long after TSMC customers are flush with 2nm parts, AWS will wait on Pat’s excuses before triggering one of those clauses, and simultaneously triggering their TSMC 2nm contingency plan.

 

[tedc]

While I can't imagine this is a hugely profitable deal for Intel, it is positive that Intel is focused on competing for business (vs. getting stuck on profit margins for the PS6 deal). If they're serious about getting IFS off the ground, they need the reps / volume, especially at 18A.

Btw, does anyone know what "AI fabric" for AWS means?

 

[blueone]

Btw, does anyone know what "AI fabric" for AWS means?

Just a guess... fabric is a term for a switched network coined by the T11 INCITS group while defining Fibre Channel in the 1990s. Fibers... fabrics... get it? Cute at the time, but IMO it didn't age well. FC switched networks were centrally managed with specific engineered topologies, while Ethernet networks allow arbitrary physical network topologies with algorithmic routing (Spanning Tree Protocol) to find a path and avoid loops. When InfiniBand came along in about 2000, it looked more like FC than Ethernet (IB uses centralized management and programmed switching tables), and the IB guys hijacked the term fabric to describe their switched networks, in an apparent attempt to convey a positive performance differentiation from Ethernet. After all the IB marketing that went on, and the fact that Mellanox IB was the preferred HPC interconnect in the TOP 500 supercomputer list, the term fabric has come to be associated with scalable high performance interconnects. Amazon could be referring to a NIC for either Ethernet or a proprietary network they're designing.

It is interesting to note that the emerging IB competitor specification, Ultra Ethernet, describes itself as a network rather than a fabric, I think because fabric has become an IB-associated word. Ultra Ethernet advertises itself as being optimized for AI, though it is not clear what that means. But it is interesting to note that Amazon is not an Ultra Ethernet Consortium member, and while Meta and Microsoft are. Google isn't a member either. Google has its own proprietary network called Jupiter. However, Intel is a founding member of Ultra Ethernet.

One possibility is that Amazon wants Intel to build an Ultra Ethernet NIC for it. I think this is unlikely, because Amazon isn't a member of the Consortium. The second possibility is that the Annapurna Labs group that Amazon acquired, and is the design team for Amazon's CPUs and AI chips, wants to build a proprietary AWS network akin to Google's Jupiter. This is not so far-fetched as it sounds, since Annapurna Labs was founded to design Ethernet SuperNICs. They know a lot about networking. Amazon is not a member of the InfiniBand Trade Association either, so... an AWS proprietary network? Maybe, but I wouldn't bet on it. Not yet. The AI fabric chip in the announcement remains a mystery.

 

[MKWVentures]

If we assume $50 per processor and 300 die cuts from a 300mm wafer for this Amazon-Intel processor, this would translate to about $15,000 in revenue per wafer for Intel. Of course, @Daniel Nenni is more qualified to assess these numbers accurately.

However, at $50 per processor and 1 million processors per year, Intel Foundry would generate only $50 million in revenue. Even if we double the price to $100 per processor, that’s still just $100 million in revenue. If we also double the yearly volume from 1 million to 2 million units and increase the per-processor price to $200, Intel Foundry’s revenue would still be only $400 million—hardly a significant figure for a company of Intel’s size.

The key questions are:

1. What’s a reasonable wafer price for Intel 18A?
   Understanding the cost structure for Intel’s 18A process is crucial, as it will significantly impact pricing and profitability for these custom Amazon processors.
2. How big will the processor be?
   Is 300 processors per 300mm wafer a reasonable assumption? The die size could vary greatly depending on the complexity and the target performance of the chip, influencing both cost and yield.
3. What will be the volume of processors Amazon requires from this Amazon-Intel project?
   Estimating Amazon’s total processor demand is challenging, especially with Intel, AMD, Nvidia, and Amazon’s own Graviton processors also in the mix. The scale and purpose of this custom chip will dictate how many units are ultimately needed.
4. Who gets the most?
5. Intel stated that this Amazon-Intel multi-year collaboration will generate multi-billion dollars in revenue with the combination of Intel's design services, Intel Foundry, and Xeon processor sales. Given that Amazon already has processors from Intel, AMD, Nvidia, and its own Graviton lineup, it’s likely that the Intel custom-made chips will be on the lower end of the cost spectrum and Intel Foundry will get the least of the share.

Excellent posts. Very much in line with my quantitative analysis.
Two Items: Intel has always done custom chips for amazon, this is not new. the Foundry chip is low volume and I am sure they have backup plans.

So what will foundry revenue be in 2026? what will Intel revenue be in 2026.

 

[blueone]

Intel has always done custom chips for amazon, this is not new.

I'm curious... what custom chips are you referring to?

 

[hskuo]

"Today we announced that we will expand our strategic collaboration with Amazon Web Services (AWS). This includes a co-investment in custom chip designs, and we have announced a multi-year, multi-billion-dollar framework covering product and wafers from Intel.

Specifically, Intel Foundry will produce an AI fabric chip for AWS on Intel 18A. We will also produce a custom Xeon 6 chip on Intel 3 that builds on our existing partnership, under which Intel produces Xeon Scalable processors for AWS. More broadly, we expect to have deep engagement with AWS on additional designs spanning Intel 18A, Intel 18AP and Intel 14A.

This framework reflects the power of our “better together” strategy, anchored on our integrated portfolio across foundry services, infrastructure and x86 products. And with the 5N4Y finish line in sight, we are beginning to see a meaningful uptick in interest from foundry customers. This includes continued momentum in advanced packaging, which remains a meaningful differentiator for Intel Foundry as we have tripled our deal pipeline since the beginning of the year."

Pat Gelsinger on Foundry Momentum, Progress on Plan

A message from Intel CEO Pat Gelsinger to employees regarding the next phase of Intel's transformation.

www.intel.com

IMO, if intel stands as a follower in foundry leading edge nodes, there is always a "switch-cost" which is 10~20% lower than the leading competitor (die cost), unless there is a strong short supply or political reason, which seems not be currently. To do this, the followers need to have a deep pocket for at least 3-5 years (to catch up and become leading next tech node provider then). Just wondering what the strategy for financial balance will be? or just bet for better future not considering and mitigate the possibility and cost of failure.

 

[MKWVentures]

I'm curious... what custom chips are you referring to?

Intel has been doing "semicustom" skus for large data center customers for years. for example Intel would do a custom Xeon (say sapphire rapids) chip optimized for Google, amazon, microsoft. Pat and amazon alluded to this in the announcement. In my experience, it is a custom mask set or in some cases a configuration. this is nothing new.... I think AMD was actually a leader in this activity and Intel caught up.

 

[DougDD]

One thing to note is that Amazon went out of their way to point out that this
is a "significant expansion" of the relationship that Amazon already has with
Intel. Some are alluding to that this will just be a little more of the same.
This is what Amazon and Intel are telling us in the joint press release:

September 16, 2024

<<This is a significant expansion of the two companies’ longstanding strategic collaboration
to help customers power virtually any workload and accelerate the performance of artificial
intelligence (AI) applications.>>

https://press.aboutamazon.com/aws/2...-helping-advance-u-s-based-chip-manufacturing

 

[MKWVentures]

One thing to note is that Amazon went out of their way to point out that this
is a "significant expansion" of the relationship that Amazon already has with
Intel. Some are alluding to that this will just be a little more of the same.
This is what Amazon and Intel are telling us in the joint press release:

September 16, 2024

<<This is a significant expansion of the two companies’ longstanding strategic collaboration
to help customers power virtually any workload and accelerate the performance of artificial
intelligence (AI) applications.>>

https://press.aboutamazon.com/aws/2...-helping-advance-u-s-based-chip-manufacturing

Its a good announcement. Perhaps we should wait and see what the details are. It is using Intel as a foundry which is a major expansion.

 

[blueone]

Intel has been doing "semicustom" skus for large data center customers for years. for example Intel would do a custom Xeon (say sapphire rapids) chip optimized for Google, amazon, microsoft. Pat and amazon alluded to this in the announcement. In my experience, it is a custom mask set or in some cases a configuration. this is nothing new.... I think AMD was actually a leader in this activity and Intel caught up.

That's what I thought you would say, and I have to admit I'm incredulous. Changing a CPU design in any way is very expensive; there's a lot more to it than just a mask set (though a CPU mask set must be very expensive). So I'm still wondering... what changes? More or less cores? I can't believe it's practical to alter caches. Clock speeds if you add or delete cores - that's easy. I suppose changing microcode would be the easiest, and I've heard of Intel allowing OEMs to do that, but alter the silicon? Seriously? I'm incredulous.

And then there's the money factor. Intel's entire server CPU unit volume per year is probably something like 20 million units or less. So say Amazon was 4 million CPUs over the life of a CPU generation, and I doubt it's that much since Graviton was released, there just aren't that many units to divide the development and production costs over. This smells like a market share defense loss leader.

I'm assuming you know what you're talking about, so you've blown my mind. I knew CPU changes were the most likely scenario, but seeing your post was a shock.

 

[Y.H]

The following is my estimation or guessing process:

Estimating Amazon's Potential Procurement of Amazon-Intel Processors:

Given the difficulty of accurately estimating the number of processors Amazon might procure, we can use several assumptions and available data to make an educated guess. Here are the key factors influencing the estimate:

1. AWS Server Growth:
   • AWS had an estimated 1.4 million servers in 2014, but AWS does not publicly disclose its current server count or annual growth rate.
   • 
     
   • Microsoft Azure had 4 million servers in 2021. Using Azure's growth as a rough proxy for AWS could provide some insight, but the comparison is limited due to differences in client base and server utilization and configuration.
2. Server to Processor Ratio:
   • Servers come in various configurations: single-CPU, dual-CPU, quad-CPU, or even 8-CPU setups. This makes it challenging to translate server count directly into the number of processors needed.
3. Revenue Comparison Between AWS and Azure:
   • In 2023, AWS had $54.648 billion in revenue compared to Azure's $32.197 billion. However, this revenue difference doesn't necessarily correspond directly to the number of servers, given variations in clientele and server efficiency (e.g., AWS hosts many video streaming services like Netflix, which uses just 100,000 server instances but serves 260 million members).
4. Virtual vs. Physical Servers:
   • AWS and other cloud providers use virtual machines, which further complicates estimating processor demand.
5. Competition Between Processors:
   • AWS uses processors from Intel, AMD, Nvidia, Amazon (Graviton), and now the Amazon-Intel project. The allocation among these providers is unknown and will depend on cost, function, and timing.
6. Server Lifecycle:
   • AWS has extended its server lifecycle from 3 years to 6 years, meaning that fewer servers may be replaced annually, but we also don’t know how many new servers will be deployed each year.
7. Timeframe for Amazon-Intel Project:
   • Assuming the Amazon-Intel processor will be available starting from 2026 because it will likely take 2–3 years to design, test, and reach high-volume manufacturing (HVM).

My Wild Guess Estimate:​

Assumption: The new Amazon-Intel processor will start delivery from 2026 through 2028.

1. Server Count Estimate:
   • Start with Azure’s 2021 server count of 4 million.
   • Apply 10% annual growth over 7 years (from 2021 to 2028):
   • 
     
   • 4 million servers (2021 Azure server number) * 1.10^7 ( 7 years of 10% annual server growth through 2028) = Total 7.795 million servers in 2028 (I brutally assigned this Azure 2028 number to AWS 2028).
2. AWS Server Share Powered by Amazon-Intel Processors:
   • Assume equal distribution among 5 processor providers for AWS (Intel, AMD, Amazon Graviton, Nvidia, and Amazon-Intel):
   • 
     
   • 7.795 million servers×20%=1.559 million servers using Amazon-Intel processors by 2028.
3. 
   
4. Server Rollout Over 3 Years (2026–2028):
   1.559 million servers / 3 years=0.52 million Amazon-Intel based servers deployment per year
5. 
   
6. Processor Count Estimate:
   • Assuming an average of 2 processors per server:
   • 
     
   • 0.52 million servers per year×2 processors per server=1.04 million processors per year 2026 ~ 2028

Conclusion:​

Based on these assumptions, Amazon could procure around 1.04 million processors per year from the Amazon-Intel project during the rollout period from 2026 to 2028. However, this is a rough estimate, subject to many unknown factors, including growth rates, processor allocation, server configurations and market situation.

Hi. Pardon me if I got it wrong. The above estimates are for the custom Xeon chips only and exclude the demand of ‘AI fabric chips’?

 

[MKWVentures]

That's what I thought you would say, and I have to admit I'm incredulous. Changing a CPU design in any way is very expensive; there's a lot more to it than just a mask set (though a CPU mask set must be very expensive). So I'm still wondering... what changes? More or less cores? I can't believe it's practical to alter caches. Clock speeds if you add or delete cores - that's easy. I suppose changing microcode would be the easiest, and I've heard of Intel allowing OEMs to do that, but alter the silicon? Seriously? I'm incredulous.

And then there's the money factor. Intel's entire server CPU unit volume per year is probably something like 20 million units or less. So say Amazon was 4 million CPUs over the life of a CPU generation, and I doubt it's that much since Graviton was released, there just aren't that many units to divide the development and production costs over. This smells like a market share defense loss leader.

I'm assuming you know what you're talking about, so you've blown my mind. I knew CPU changes were the most likely scenario, but seeing your post was a shock.

I can't say what is done today as I dont know and if I did know it would be proprietary, But I can say that some of the configurations are not as hard as one might think. I had a discussion with AMD a couple years ago and they showed me how they could do amazing configuration changes with their designs without a mask change.

the whole FPGA acquisition plan was part of this strategy years ago. I am not sure how it is used now.

Hyperscalers LOVE semicustom parts LOL.

 

[hist78]

Hi. Pardon me if I got it wrong. The above estimates are for the custom Xeon chips only and exclude the demand of ‘AI fabric chips’?

My estimate is for the Intel custom designed chips for Amazon, not including Xeon (custom or not).