Craig Barrett: Intel is back—stop talking about breaking it up

[hist78]

Fair point, but if you think Nvidia qualifies then I don't see why Amazon, Microsoft, and Google wouldn't also qualify, given that all three appear to gravitating to custom chips at both the CPU, XPU, and DPU level. Broadcom's massive 2027 guidance appears to make the case for hypergrowth in custom ASICs going forward.

I understand how Gelsinger could have overestimated external customer demand but I'm not sure it could have been by THAT much.

Google, Amazon, and Microsoft do not sell their in-house designed high-performance computing or AI chips to other companies, whereas Nvidia does. Consequently, their production volumes are on a different scale.

For example, Amazon's AI chips are exclusively used in its own data centers, while Nvidia's chips are sold to a much broader client base, including Microsoft, Amazon, Google, Facebook/Meta, Tesla, Oracle, OpenAI, xAI, and Lambda, as well as system vendors such as Dell, HP, Supermicro, Asustek, Foxconn, Gigabyte, Inventec, Quanta, Wistron, and Wiwynn.

 

[Daniel Nenni]

You can have all the customer relationships you want, but if MSFT is able to build a superior ASIC on 18A, don't you think Amazon and Google will quickly follow? Otherwise they get left behind. Amazon already dipped its toe in, likely because it sees this as a possible scenario. Understand there is debate about whether 18A is superior to 2nm -- I believe there is a high probability that 18A is, and is the node when Intel forges ahead. If that is correct (and Craig Barrett apparently agrees, I'd trust him over the other four board members who say in their op-ed Intel is "at least a generation behind") then there is no reason why Intel can't onboard a few high-volume HPC customers. AMZN and GOOG are behind NVidia in line anyway at TSMC for capacity, so all the more reason to move over. Then there's TSMC jacking up 2nm wafer prices to $30k. All plays in Intel's favor. It would be stupid to break the company up right before 18A shows what it can do.

That is a very big if. I do not see a huge difference between Intel 18A and TSMC N2, based on the PDKs, with the exception of BSPD and I know of no external customers using it, yet. We will hear more next month at the Intel Foundry event.

 

[Bdubs82]

That is a very big if. I do not see a huge difference between Intel 18A and TSMC N2, based on the PDKs, with the exception of BSPD and I know of no external customers using it, yet. We will hear more next month at the Intel Foundry event.

I believe MSFT and AMZN have already confirmed they are 18A customers, no? For exactly what chips and how many of them, we don't know for sure, but should find out soon/this year, hopefully at the IFS event. Also, hyperscaler ASICs are still on 4nm at TSMC, and even the upcoming Rubin will be on 3nm. the first TSMC 2nm should be Apple/mobile or PC CPUs I think. So an 18A AI ASIC would still have the lead this year, with a backside power advantage (10%?).

 

[Bdubs82]

Google, Amazon, and Microsoft do not sell their in-house designed high-performance computing or AI chips to other companies, whereas Nvidia does. Consequently, their production volumes are on a different scale.

For example, Amazon's AI chips are exclusively used in its own data centers, while Nvidia's chips are sold to a much broader client base, including Microsoft, Amazon, Google, Facebook/Meta, Tesla, Oracle, OpenAI, xAI, and Lambda, as well as system vendors such as Dell, HP, Supermicro, Asustek, Foxconn, Gigabyte, Inventec, Quanta, Wistron, and Wiwynn.

You are double counting some of those customers, for instance xAI uses SuperMicro and Dell servers, and cloud guys use the ODMs. And I believe the vast majority of AI training takes place in the cloud. Anthropic just announced late last year it's going to train all of its models on Trainium/Inferentia on AWS going forward. Nvidia has the CUDA advantage, but it also makes ~80% gross margins, so custom ASICs are 1/5 the cost. That's a strong incentive to go in the ASIC direction as long as the manufacturing process is on the same/similar nodes.

 

[Xebec]

That is a very big if. I do not see a huge difference between Intel 18A and TSMC N2, based on the PDKs, with the exception of BSPD and I know of no external customers using it, yet. We will hear more next month at the Intel Foundry event.

Not seeing a huge difference between Intel 18A and TSMC N2 sounds pretty good though?

5 years ago Intel was just starting to get "7nm" class products out while TSMC/Apple were shipping N5 chips. Now it looks like 18A products will be on store shelves quarters ahead of N2.

 

[Maximus]

Not seeing a huge difference between Intel 18A and TSMC N2 sounds pretty good though?

5 years ago Intel was just starting to get "7nm" class products out while TSMC/Apple were shipping N5 chips. Now it looks like 18A products will be on store shelves quarters ahead of N2.

Good enough for Intel's own products but probably not attractive enough for serving external customers. Although Intel plans to begin mass production of 18A early this year, their model indicates that serving external customers will take at least a year from the 18A launch. If Intel starts producing for external customers in the second half of 2026, it will coincide with TSMC's A16 entering HVM. The strategic timing and positioning of TSMC's A16 seem carefully planned to significantly diminish the appeal of Intel's 18A.

 

[siliconbruh999]

That is a very big if. I do not see a huge difference between Intel 18A and TSMC N2, based on the PDKs, with the exception of BSPD and I know of no external customers using it, yet. We will hear more next month at the Intel Foundry event.

So in terms of PPA N2/18A are very close ?
If that's true than I wonder why they are outsourcing some Compute tiles to TSMC N2

 

[Xebec]

Good enough for Intel's own products but probably not attractive enough for serving external customers. Although Intel plans to begin mass production of 18A early this year, their model indicates that serving external customers will take at least a year from the 18A launch. If Intel starts producing for external customers in the second half of 2026, it will coincide with TSMC's A16 entering HVM. The strategic timing and positioning of TSMC's A16 seem carefully planned to significantly diminish the appeal of Intel's 18A.

A16 is only one year behind N2 then? (I think Apple N2 products appear for purchase in 2026)

 

[Maximus]

A16 is only one year behind N2 then? (I think Apple N2 products appear for purchase in 2026)

Yes, the A16 is just one year behind the N2. The N2 is set to launch in the second half of 2025, which doesn't align with the 2025 iPhone timeline. Therefore, the first wave of products will likely feature M series chips. This situation is similar to the N3B launch in the second half of 2022.

TSMC plans to launch the N2P and A16 in the second half of 2026. The N2P will not have a super-power rail, while the A16 will include it.

Most people compare the N2 with the 18A, but for foundry customers, the real competition will be between the N2P (likely for mobile) and the A16 (likely for HPC) versus the 18A in 2026.

 

[siliconbruh999]

Yes, the A16 is just one year behind the N2. The N2 is set to launch in the second half of 2025, which doesn't align with the 2025 iPhone timeline. Therefore, the first wave of products will likely feature M series chips. This situation is similar to the N3B launch in the second half of 2022.

TSMC plans to launch the N2P and A16 in the second half of 2026. The N2P will not have a super-power rail, while the A16 will include it.

Most people compare the N2 with the 18A, but for foundry customers, the real competition will be between the N2P (likely for mobile) and the A16 (likely for HPC) versus the 18A in 2026.

we would have 18AP with 10% more perf/watt vs 18A to compete with N2P/A16

 

[Maximus]

we would have 18AP with 10% more perf/watt to compete with N2P/A16

Again, according to Intel's model, the 18AP will initially be used internally. If it becomes available to external customers one year after its launch, the earliest it could serve external customers in HVM would be the second half of 2027. TSMC hasn't explicitly announced what will follow the A16 and N2P, but they should have something new in 2027H2.

 

[siliconbruh999]

Again, according to Intel's model, the 18AP will initially be used internally. If it becomes available to external customers one year after its launch, the earliest it could serve external customers in HVM would be the second half of 2027. TSMC hasn't explicitly announced what will follow the A16 and N2P, but they should have something new in 2027H2.

i don't know for sure but maybe it is a Intel 4 to Intel 3 case where not much work is needed but i agree with your comment that it will be Intel exclusive for at least 6 months.

 

[MKWVentures]

Simplified: Intel had 4 years. their external foundry revenue dropped. Intel has no external foundry committed production in 2025 or 2026. Based on this, the finances are horrible if Intel ramps and horrible if Intel does not ramp. There is no happy ending to "I invested billions in creating an external foundry and there are no significant sales until 2027". If you build it and they do not come, there is no recovery. This is why Intel had the strategy it did in 2020, and why this was predicted to happen.

Side note: they also sold of output rights to their two most advanced factories. Those investors will get their money from Intel one way or another, Intel is already paying penalties to them.

Its not about the tech (i hear 18A is doing well). Its about the financial spreadsheet and how will Intel pay for what it wants to do.

 

[XYang2023]

Simplified: Intel had 4 years. their external foundry revenue dropped. Intel has no external foundry committed production in 2025 or 2026. Based on this, the finances are horrible if Intel ramps and horrible if Intel does not ramp. There is no happy ending to "I invested billions in creating an external foundry and there are no significant sales until 2027". If you build it and they do not come, there is no recovery. This is why Intel had the strategy it did in 2020, and why this was predicted to happen.

Side note: they also sold of output rights to their two most advanced factories. Those investors will get their money from Intel one way or another, Intel is already paying penalties to them.

Its not about the tech (i hear 18A is doing well). Its about the financial spreadsheet and how will Intel pay for what it wants to do.

https://twitter.com/x/status/1896676447672303978

Let TSMC manage geo-risks. It should not be Intel's responsibility anyways. Intel just needs to narrow the scope.

 

[MKWVentures]

https://twitter.com/x/status/1896676447672303978

Let TSMC manage geo-risks. It should not be Intel's responsibility anyways. Intel just needs to narrow the scope.

Intel doesnt solve or want to solve Geo stability. They just want to avoid financial disaster at this point.

You are correct that Pats strategy was too optimistic. But thats why Intel planned to outsource in 2020. You cannot afford the fabs and TD work with Intels internal volume. the scale is too small.... even if all was done internal. Now we know there is no meaningful external foundry volume in 2025 or 2026. How do you pay for the 18A ramp given that? its not possible. So.......

Question: Will the 18A Intel wafer cost be higher or lower than TSMC N2 wafer Price in 2026? Intel knows both numbers very well (Intel is a major TSMC customer)

 

[siliconbruh999]

Intel doesnt solve or want to solve Geo stability. They just want to avoid financial disaster at this point.

You are correct that Pats strategy was too optimistic. But thats why Intel planned to outsource in 2020. You cannot afford the fabs and TD work with Intels internal volume. the scale is too small.... even if all was done internal. Now we know there is no meaningful external foundry volume in 2025 or 2026. How do you pay for the 18A ramp given that? its not possible. So.......

Question: Will the 18A Intel wafer cost be higher or lower than TSMC N2 wafer Price in 2026? Intel knows both numbers very well (Intel is a major TSMC customer)

For Intel 18A is cheaper and better for margin and profit for external I don't know but Intel sure knows that as well.

I hope they use 18A nicely like they did with 14nm cause after 18A the performance improvements are good but density improvement is not much it's not like Intel 14nm vs TSMC N7 gap or Intel 10nm vs N5 gap so they should really use it in volume for every product going forward.

 

[XYang2023]

For Intel 18A is cheaper and better for margin and profit for external I don't know but Intel sure knows that as well.

I hope they use 18A nicely like they did with 14nm cause after 18A the performance improvements are good but density improvement is not much it's not like Intel 14nm vs TSMC N7 gap or Intel 10nm vs N5 gap so they should really use it in volume for every product going forward.

I think for margin, it should be viewed relatively. Given TSM US investments, their margin numbers are also depressed. For Intel, assuming they remain IDM, they can stack their margins.

 

[siliconbruh999]

I think for margin, it should be viewed relatively. Given TSM US investments, their margin numbers are also depressed. For Intel, assuming they remain IDM, they can stack their margins.

Yes that's what I mean but TSMC US Fabs will charge more how much more I don't know but I expect a minimum of 10-15%

 

[MKWVentures]

I think for margin, it should be viewed relatively. Given TSM US investments, their margin numbers are also depressed. For Intel, assuming they remain IDM, they can stack their margins.

@siliconbruh999

What do you think the operating margins will be for 18A for IFS in 2025 and 2026.

 

[Xebec]

Simplified: Intel had 4 years. their external foundry revenue dropped. Intel has no external foundry committed production in 2025 or 2026. Based on this, the finances are horrible if Intel ramps and horrible if Intel does not ramp. There is no happy ending to "I invested billions in creating an external foundry and there are no significant sales until 2027". If you build it and they do not come, there is no recovery. This is why Intel had the strategy it did in 2020, and why this was predicted to happen.

Side note: they also sold of output rights to their two most advanced factories. Those investors will get their money from Intel one way or another, Intel is already paying penalties to them.

Its not about the tech (i hear 18A is doing well). Its about the financial spreadsheet and how will Intel pay for what it wants to do.

Intel doesnt solve or want to solve Geo stability. They just want to avoid financial disaster at this point.

You are correct that Pats strategy was too optimistic. But thats why Intel planned to outsource in 2020. You cannot afford the fabs and TD work with Intels internal volume. the scale is too small.... even if all was done internal. Now we know there is no meaningful external foundry volume in 2025 or 2026. How do you pay for the 18A ramp given that? its not possible. So.......

Question: Will the 18A Intel wafer cost be higher or lower than TSMC N2 wafer Price in 2026? Intel knows both numbers very well (Intel is a major TSMC customer)

Just curious - what node do you think was Intel's last "good node" for only internal volume? (Assuming Intel also migrated GPUs such as Battlemage to Intel fabs). Were they financially broken already at "10nm/Intel 7"? earlier? later?

The board switching tactics twice in 18 months (during Bob Swan's tenure) definitely hurt the success rate of going all in internal fabrication. The Tower acquisition failure certainly also put a nail in the coffin too.