Inside Taiwanese Chip Giant, a U.S. Expansion Stokes Tensions

[hist78]

How so? I don't see two small fabs changing that part of IFS's value proposition. R&D is still done in the ROC, and presumably mask data of some type will also live in the ROC. TSMC AZ will not have the capacity to run anything more than all DOD and a tiny fraction of the total merchant/systems company orders. If you are serious about a diversified fab that is "out of the line of fire" or want the assurance that R&D will never be disrupted and that your mask data could never fall into the hands of the PLA, then your only foundry choices would be GF, Skywater, and IFS. The way I see it TSMC AZ is a nice option for current TSMC customers that offers some piece of mind, but it is not a replacement for a foundry that does all of their manufacturing in low risk areas and has all of their R&D in the west.

Intel and Pat Gelsinger have repeatedly said majority of Chips Act grant money should go to American companies, like Intel. But it looks like US government has a different thought.

We have to understand a weapon system development process is long, expensive, and high risk. By now in 2023 DoD, DoE, Lockheed, Ratheon, Northrop, and Boeing probably already decided and/or ordered the chips for a particular weapon system planned to be coming out in the next 3, 5 or 7 years. Those chips can be used for a new F35 fighter jet, a B21 bomber, a sixth generation fighter jet, or a hypersonic missile.

Why it needs to be done so early? That's because if they haven't decided or ordered the chips by now, those weapon system contractors won't have enough time to design, build supply chain, test, and assemble the system at all. And the military itself needs a lot of time to test, verify, train the users, and build the whole logistics/maintenance organization.

If many chips needed by military are made by TSMC in Taiwan today, why not subside TSMC to build a fab in US to secure those chip supply for the next 3, 5 or 7 years? It's a practical, responsible, and cost efficient way to satisfy the need for national security and ensure the on time execution of the weapon project. Similar type of practices have been used in the past several decades.

Can Intel Foundry Service (IFS) or Globalfoundries (GF) do the same thing as TSMC does for US military and weapon system contractors? IFS or GF may or may not meet the expectation but that's nothing to do with those chips that have already decided/ordered. IFS and Globalfoundries are welcomed to compete in the new projects and Chips Act will subside their manufacturing and R&D anyway. But in no way they are allowed to delay or interrupt those existing ones that have already passed the project check points.

DoD, DoE, and US policy makers have little interest to build a new supply chain to replace an existing and efficient one. They know they need real products to complete a weapon system. They understand US can't build strong national security based on the good intention and good roadmap promised by IFS or Globalfoundries.

Intel and Globalfoundries (with the old IBM fabs) have been providing US national security related chips for a long time. But they are facing several serious challenges such as bad execution, bad leadership, bad financial management, bad product designs, and bad strategic decisions. All those bad things are coming from company internal that US government can't help at all. Bringing more TSMC and Samsung fab capability and capacity to US is a way for US to hedge the risk and strengthen domestic production for national security.

Intel believes they are the only one and the chosen one. But the US government doesn't think exactly that way. How can Intel be happy?

 

[nghanayem]

Intel and Pat Gelsinger have repeatedly said majority of Chips Act grant money should go to American companies, like Intel. But it looks like US government has a different thought.

We have to understand a weapon system development process is long, expensive, and high risk. By now in 2023 DoD, DoE, Lockheed, Ratheon, Northrop, and Boeing probably already decided and/or ordered the chips for a particular weapon system planned to be coming out in the next 3, 5 or 7 years. Those chips can be used for a new F35 fighter jet, a B21 bomber, a sixth generation fighter jet, or a hypersonic missile.

Why it needs to be done so early? That's because if they haven't decided or ordered the chips by now, those weapon system contractors won't have enough time to design, build supply chain, test, and assemble the system at all. And the military itself needs a lot of time to test, verify, train the users, and build the whole logistics/maintenance organization.

If many chips needed by military are made by TSMC in Taiwan today, why not subside TSMC to build a fab in US to secure those chip supply for the next 3, 5 or 7 years? It's a practical, responsible, and cost efficient way to satisfy the need for national security and ensure the on time execution of the weapon project. Similar type of practices have been used in the past several decades.

Can Intel Foundry Service (IFS) or Globalfoundries (GF) do the same thing as TSMC does for US military and weapon system contractors? IFS or GF may or may not meet the expectation but that's nothing to do with those chips that have already decided/ordered. IFS and Globalfoundries are welcomed to compete in the new projects and Chips Act will subside their manufacturing and R&D anyway. But in no way they are allowed to delay or interrupt those existing ones that have already passed the project check points.

DoD, DoE, and US policy makers have little interest to build a new supply chain to replace an existing and efficient one. They know they need real products to complete a weapon system. They understand US can't build strong national security based on the good intention and good roadmap promised by IFS or Globalfoundries.

Intel and Globalfoundries (with the old IBM fabs) have been providing US national security related chips for a long time. But they are facing several serious challenges such as bad execution, bad leadership, bad financial management, bad product designs, and bad strategic decisions. All those bad things are coming from company internal that US government can't help at all. Bringing more TSMC and Samsung fab capability and capacity to US is a way for US to hedge the risk and strengthen domestic production for national security.

Intel believes they are the only one and the chosen one. But the US government doesn't think exactly that way. How can Intel be happy?

I mostly agree, however like I said, for existing customers TSMC AZ is a good option to partially derisk your supply chain. However just like how intel roadmap is risk for DoD contractors, TSMC's roadmap is not exactly guaranteed either. Mind you, I'm not one of the PLA invasion doomsayers, but when it comes to defense you have to plan for the worst case scenario. A takeover of the ROC would shatter TSMC's roadmap into nothing, and potentially allow the PLA to reverse engineer or develop countermeasures for US technology. Hence my statement that TSMC AZ is not a full stop replacement for a viable US or EU based foundry. The DoD uses TSMC for parts that are impractical to be made on TI, GF, or Skywater fabs, but if TSMC had a viable western competitor on the leading edge you can bet your biscuits they would use that competitor for at least some future contracts. It is not all that different from CIA shell companies buying Ti from the USSR for USAF aircraft production. Did the US want to do this? Hell no. Did they have any other choice at the time? Also no.

Intel and Pat Gelsinger have repeatedly said majority of Chips Act grant money should go to American companies, like Intel. But it looks like US government has a different thought.

Everybody will be getting a cut. TSMC, Samsung, Ti, Micron, and Globalwafers are getting cuts as established players that will for sure strengthen the American industry and provide a "safe" production base to the DoD. Intel, GF, and Skywater will also get cuts given they already contribute to the US economy/service the DoD, but also on the hope that they can grow into viable competitors. After all to spend $0 would be to admit failure and just accept that the dream would never happen. Going all in can also blow up in the US Gov's face. The best bet is to hedge your bets, hope for the best, and plan for the worst.

As for Pat's comments what is he supposed to say "Don't invest in us we should just shrivel up and die without trying"? Of course he will say intel should receive the lion's share he is the CEO of intel.

How can Intel be happy?

Getting a fair cut of CHIPs money so they can have the capital to build the capacity necessary to become the number two foundry by the end of the decade. They will still be the only member of the big three with all of their fabs (or even just major share of capacity) in the west. If they execute to their roadmap they will have PPW leadership by a wide margin in 2024/25. They already have a large stack of compelling IP for customers to licence. If they play their cards right, they might even have better foundry services and PDKs than Samsung who seems to be lacking in these departments (as well as trust). To think that three Samsung and TSMC fabs will at the stroke of a pen kill IFS seems nihilistic to me.

 

[cliff]

Can somebody break the cost and cost differences (Taiwan vs Arizona) based on NRE (masks), wafer setup, cost per wafer, testing, packaging, and whatever else I missed. Let's assume the 16'ish processes.

Aren't these steps mostly automated?

I saw the discussion regarding the 4x cost number and I suspect that was pure BS. This forum seems to be dominated by foundry people, so you guys may know.

It always seemed to me that the goal (for ASICs) is IP protection, functionality, turnaround times, NRE costs, and availability is far more important than the recurring cost of silicon. If the recurring costs are 20% more for a more guaranteed supply, it's worth it. I don't think I am the only one nervous about dictators making good on their promises.

 

[nghanayem]

Can somebody break the cost and cost differences (Taiwan vs Arizona) based on NRE (masks), wafer setup, cost per wafer, testing, packaging, and whatever else I missed. Let's assume the 16'ish processes.

My educated guess is that the variable costs are not a dime over 10%, more likely in the single digits. Tool equipment/depreciation (biggest single cost in a wafer) 0% more expensive. Cost of building a fab shell, that is probably significantly more expensive in the US, but given that concrete is at the end of the day just concrete my guess as a non civil engineer might be around 20-50% more expensive.

Aren't these steps mostly automated?

Unless a tool errors out and needs manual intervention, they are all automated for 300mm fabs. My understanding is 200mm fabs had to actually have operators load the wafer cassettes onto the tools with a PGV .

It always seemed to me that the goal (for ASICs) is IP protection, functionality, turnaround times, NRE costs, and availability is far more important than the recurring cost of silicon. If the recurring costs are 20% more for a more guaranteed supply, it's worth it. I don't think I am the only one nervous about dictators making good on their promises.

Yeah I think the people who would be most concerned would be merchant chip vendors where their margins are far lower

 

[cliff]

I believe you (as always). You know what I'm thinkin... and I ain't sayin it.

 

[hist78]

I mostly agree, however like I said, for existing customers TSMC AZ is a good option to partially derisk your supply chain. However just like how intel roadmap is risk for DoD contractors, TSMC's roadmap is not exactly guaranteed either. Mind you, I'm not one of the PLA invasion doomsayers, but when it comes to defense you have to plan for the worst case scenario. A takeover of the ROC would shatter TSMC's roadmap into nothing, and potentially allow the PLA to reverse engineer or develop countermeasures for US technology. Hence my statement that TSMC AZ is not a full stop replacement for a viable US or EU based foundry. The DoD uses TSMC for parts that are impractical to be made on TI, GF, or Skywater fabs, but if TSMC had a viable western competitor on the leading edge you can bet your biscuits they would use that competitor for at least some future contracts. It is not all that different from CIA shell companies buying Ti from the USSR for USAF aircraft production. Did the US want to do this? Hell no. Did they have any other choice at the time? Also no.


Everybody will be getting a cut. TSMC, Samsung, Ti, Micron, and Globalwafers are getting cuts as established players that will for sure strengthen the American industry and provide a "safe" production base to the DoD. Intel, GF, and Skywater will also get cuts given they already contribute to the US economy/service the DoD, but also on the hope that they can grow into viable competitors. After all to spend $0 would be to admit failure and just accept that the dream would never happen. Going all in can also blow up in the US Gov's face. The best bet is to hedge your bets, hope for the best, and plan for the worst.

As for Pat's comments what is he supposed to say "Don't invest in us we should just shrivel up and die without trying"? Of course he will say intel should receive the lion's share he is the CEO of intel.


Getting a fair cut of CHIPs money so they can have the capital to build the capacity necessary to become the number two foundry by the end of the decade. They will still be the only member of the big three with all of their fabs (or even just major share of capacity) in the west. If they execute to their roadmap they will have PPW leadership by a wide margin in 2024/25. They already have a large stack of compelling IP for customers to licence. If they play their cards right, they might even have better foundry services and PDKs than Samsung who seems to be lacking in these departments (as well as trust). To think that three Samsung and TSMC fabs will at the stroke of a pen kill IFS seems nihilistic to me.

Intel, Globalfoundries, TI, Micron, Samsung, and TSMC will all get US subsides. The problem is that Intel keeps promoting that they are the one should get the lion share of the subsidy and foreign companies, like TSMC or Samsung, should get little or nothing.

Obviously US policy makers have a very different approach on this. After many years of tolerance and assistance from US government, Intel's words, actions, and results frequently are not consistent. The long overdue Aurora Supercomputer project is one of several examples that have forced US government to think what Intel can do.

Aurora Supercomputer: https://semiwiki.com/forum/index.ph...ling-out-39-billion-chip-aid.17486/post-58702

The Chips Act manufacturing subsidy in certain way acts like a venture capital. The potential reward is enhanced national security, stronger economy, reliable supply chain for manufacturing, and technology advancement. As a venture capital, are we going to bet the whole house only on Intel or TSMC or Samsung? Definitely not.

And Intel is not too happy about this.

 

[blueone]

Obviously US policy makers have a very different approach on this. After many years of tolerance and assistance from US government, Intel's words, actions, and results frequently are not consistent. The long overdue Aurora Supercomputer project is one of several examples that have forced US government to think what Intel can do.

Aurora Supercomputer: https://semiwiki.com/forum/index.ph...ling-out-39-billion-chip-aid.17486/post-58702

I doubt any member of Congress, or the President, has any knowledge of what Aurora is or what the issues with it are. It is a nit.

The Chips Act manufacturing subsidy in certain way acts like a venture capital. The potential reward is enhanced national security, stronger economy, reliable supply chain for manufacturing, and technology advancement. As a venture capital, are we going to bet the whole house only on Intel or TSMC or Samsung? Definitely not.

And Intel is not too happy about this.

I doubt anyone is happy about the CHIPS Act. I'm certainly not. I think it's a waste of taxpayer money, and creates additional useless bureaucracy.

 

[hist78]

I doubt any member of Congress, or the President, has any knowledge of what Aurora is or what the issues with it are. It is a nit.

I doubt anyone is happy about the CHIPS Act. I'm certainly not. I think it's a waste of taxpayer money, and creates additional useless bureaucracy.

"I doubt any member of Congress, or the President, has any knowledge of what Aurora is or what the issues with it are. It is a nit."

The President or Congressman might not know the Aurora Supercomputer but many people who help them to formulate the policy do know. There are many experienced and knowledgeable Under Secretaries, Deputies, Directors, and staffers across DoD, DoE, NSA, NSC, Commerce Department, Whitehouse, Congress and Senate. They have significant impact in the policy and rule making.

After Intel paying ~$300 million to the Federal government in Q4 2021 for not completing the Aurora Supercomputer and still hasn't deliver it at this moment, important people inside the US government would know what the Aurora Supercomputer is about, for the sake of national security and their own career prospect.

 

[blueone]

"I doubt any member of Congress, or the President, has any knowledge of what Aurora is or what the issues with it are. It is a nit."

The President or Congressman might not know the Aurora Supercomputer but many people who help them to formulate the policy do know. There are many experienced and knowledgeable Under Secretaries, Deputies, Directors, and staffers across DoD, DoE, NSA, NSC, Commerce Department, Whitehouse, Congress and Senate. They have significant impact in the policy and rule making.

After Intel paying ~$300 million to the Federal government in Q4 2021 for not completing the Aurora Supercomputer and still hasn't deliver it at this moment, important people inside the US government would know what the Aurora Supercomputer is about, for the sake of national security and their own career prospect.

The Secretaries and Directors report to the executive branch, not to Congress. There are congressional reviews of various topics, but if you've ever watched any of these on CSPAN you wouldn't have much faith in Congress knowing anything at this detail level. The only people who would have deep knowledge of Aurora are in the Department of Energy, and Aurora is targeted for what appears to be unclassified projects at Argonne National Laboratory.

As for that $300M charge, Intel has not specified what the charge is for, other than it's due to a Federal supercomputing contract. And Intel is not the primary contractor for Aurora, HPE is, and the penalty may have been paid to HPE for loss of revenue, and not the USG. The only reference Google could point me to on this $300M is on Reddit, and Reddit discussions often have a weak foundation in facts, and lots of unsubstantiated conjecture.

 

[hist78]

The Secretaries and Directors report to the executive branch, not to Congress. There are congressional reviews of various topics, but if you've ever watched any of these on CSPAN you wouldn't have much faith in Congress knowing anything at this detail level. The only people who would have deep knowledge of Aurora are in the Department of Energy, and Aurora is targeted for what appears to be unclassified projects at Argonne National Laboratory.

As for that $300M charge, Intel has not specified what the charge is for, other than it's due to a Federal supercomputing contract. And Intel is not the primary contractor for Aurora, HPE is, and the penalty may have been paid to HPE for loss of revenue, and not the USG. The only reference Google could point me to on this $300M is on Reddit, and Reddit discussions often have a weak foundation in facts, and lots of unsubstantiated conjecture.

The US Federal government policy and rule making process is much sophisticate than what we saw on the CSPAN. For the legislation and budget as big as the Chips and Science Act went through many internal and interagency meeting and negotiation with the Congress and Senate. There are many agencies who are staffed with people who know semiconductor and technology.

You mentioned Aurora Supercomputer is for unclassified projects. That's not relevant to our discussion here because Intel is 4 to 5 years late on delivery is a fact, no matter what Aurora is intended to use for. It looks even worst is that the Frontier Supercomputer at Oak Ridge National Lab started after Aurora yet completed in 2022 with AMD CPU and GPU.

In terms of the ~$300 million one time charge Intel had in 2021, here is one of such reports that Intel didn't disputed it back then.

"And while we don’t normally track the operating margins of the individual groups, it’s noteworthy that in explaining a drop for DCG’s operating margins, Intel has cited a “one-time Intel Federal charge.” This would appear to be an allusion to the Aurora supercomputer, where Intel apparently owes the US Government penalty payments (in essence, a partial refund) for not completing the supercomputer on time. "

Source: https://www.anandtech.com/show/1723...d-fy-2021-earnings-ending-2021-on-a-high-note

 

[blueone]

The US Federal government policy and rule making process is much sophisticate than what we saw on the CSPAN. For the legislation and budget as big as the Chips and Science Act went through many internal and interagency meeting and negotiation with the Congress and Senate. There are many agencies who are staffed with people who know semiconductor and technology.

The US Congress includes the House of Representatives and the Senate. It is incorrect to discuss the Congress and the Senate as separate entities.

The CHIPS Act is about allocating funding, defining eligibility criteria, and creating a list of new agencies. An example new agency is the National Semiconductor Technology Center, which is funded by the bill and defines a strategy, but there's no technical detail in the legislation. There are several new agencies specified. This is a good description I've been using for awhile:

The CHIPS and Science Act: Here’s what’s in it

We look at what the CHIPS and Science Act is and how it's helped to boost US semiconductor capacity, catalyze R&D, and create new regional tech hubs.

www.mckinsey.com

What an expensive mess of bureaucracy! I'm sure a bunch of lobbyists were involved in defining the agencies.

You mentioned Aurora Supercomputer is for unclassified projects. That's not relevant to our discussion here because Intel is 4 to 5 years late on delivery is a fact, no matter what Aurora is intended to use for. It looks even worst is that the Frontier Supercomputer at Oak Ridge National Lab started after Aurora yet completed in 2022 with AMD CPU and GPU.

The fact that Aurora is not classified is why we know about its problems, so it is relevant. Many (most?) federal supercomputing projects are classified.

Most of the computing technology expertise in the US government is in the Department of Energy and the Department of Defense. The DoD includes the NSA, which is a significant driver of supercomputing technology, but all the projects are classified. So are many projects in the DoE labs like Sandia and Livermore.

In terms of the ~$300 million one time charge Intel had in 2021, here is one of such reports that Intel didn't disputed it back then.

"And while we don’t normally track the operating margins of the individual groups, it’s noteworthy that in explaining a drop for DCG’s operating margins, Intel has cited a “one-time Intel Federal charge.” This would appear to be an allusion to the Aurora supercomputer, where Intel apparently owes the US Government penalty payments (in essence, a partial refund) for not completing the supercomputer on time. "

Source: https://www.anandtech.com/show/1723...d-fy-2021-earnings-ending-2021-on-a-high-note

That is speculation on Anandtech's part. I don't look to the Wall Street Journal for technical information, and I wouldn't look to Anandtech for business details.

Intel does not build and deliver supercomputer systems. HPE/Cray does, in the case of Aurora. For example, Aurora uses the proprietary HPE/Cray Slingshot as a system interconnect, and that will pretty much define the system architecture. Intel used to build, deliver, and support supercomputer systems, and had a division for it, but that was disbanded a long time ago (like before 2000). Their system product was the Intel Paragon, which was actually a pretty nice piece of work for its day.

Intel does have a subsidiary for federal government contracts call Intel Federal, but by Intel's description they mostly do joint technology R&D with federal agencies. They don't deliver supercomputers.

 

[hist78]

The US Congress includes the House of Representatives and the Senate. It is incorrect to discuss the Congress and the Senate as separate entities.

The CHIPS Act is about allocating funding, defining eligibility criteria, and creating a list of new agencies. An example new agency is the National Semiconductor Technology Center, which is funded by the bill and defines a strategy, but there's no technical detail in the legislation. There are several new agencies specified. This is a good description I've been using for awhile:

The CHIPS and Science Act: Here’s what’s in it

We look at what the CHIPS and Science Act is and how it's helped to boost US semiconductor capacity, catalyze R&D, and create new regional tech hubs.

www.mckinsey.com

What an expensive mess of bureaucracy! I'm sure a bunch of lobbyists were involved in defining the agencies.


The fact that Aurora is not classified is why we know about its problems, so it is relevant. Many (most?) federal supercomputing projects are classified.

Most of the computing technology expertise in the US government is in the Department of Energy and the Department of Defense. The DoD includes the NSA, which is a significant driver of supercomputing technology, but all the projects are classified. So are many projects in the DoE labs like Sandia and Livermore.

That is speculation on Anandtech's part. I don't look to the Wall Street Journal for technical information, and I wouldn't look to Anandtech for business details.

Intel does not build and deliver supercomputer systems. HPE/Cray does, in the case of Aurora. For example, Aurora uses the proprietary HPE/Cray Slingshot as a system interconnect, and that will pretty much define the system architecture. Intel used to build, deliver, and support supercomputer systems, and had a division for it, but that was disbanded a long time ago (like before 2000). Their system product was the Intel Paragon, which was actually a pretty nice piece of work for its day.

Intel does have a subsidiary for federal government contracts call Intel Federal, but by Intel's description they mostly do joint technology R&D with federal agencies. They don't deliver supercomputers.

I think I will stop here for further discussion on this because I understand your main point is the Aurora Supercomputer project's 4-5 years delay in no way has damaged Intel's reputation and credibility. And among many US government agencies, few of them even noticed it

On the other hand there are many articles discussing the cause of the Aurora project delay and Intel's failure to deliver the needed CPUs and GPUs on time. If you are interested, you can find them through a quick Google search.

 

[blueone]

I think I will stop here for further discussion on this because I understand your main point is the Aurora Supercomputer project's 4-5 years delay in no way has damaged Intel's reputation and credibility. And among many US government agencies, few of them even noticed it

Completely incorrect. My point was that the specific delay of Aurora and and exactly why it occurred is unlikely to be known by members of Congress. Intel's credibility with DoE is very damaged, which is one reason why (among several, IMO) AMD and Nvidia chips were chosen for the next system.

On the other hand there are many articles discussing the cause of the Aurora project delay and Intel's failure to deliver the needed CPUs and GPUs on time. If you are interested, you can find them through a quick Google search.

You're very amusing.

 

[benb]

Getting back to the NYT article, the guys who wrote it are probably all the NYT has for this sort of thing, and they dug beyond the surface. They spoke with contractors, which is where you should look for knowledge, so that was good.

Articles like this deal in stereotypes, such as:
-US is too expensive
-US workers are lazy

There is some examination of the stereotypes but not much.

It doesn’t help that the fab isn’t operating, so who knows how US TSMC engineers will react to working weekends, it hasn’t started yet. Whether the US is too expensive or not is tough to answer unless TSMC releases figures, which they haven’t, so it’s speculative. They also have incentive to play a tiny violin very loudly, in the process of begging for Chips Act money.