News Nvidia to spend hundreds of billions on U.S.-made chips, confirms Blackwell GPU production at TSMC Arizona

[Admin]

Nvidia is already using TSMC's Fab 21 in Arizona, plans to use it even more extensively in the coming years.

Nvidia to spend hundreds of billions on U.S.-made chips, confirms Blackwell GPU production at TSMC Arizona : Read more

 

[JTWrenn]

Well America chips might be a better way to put it. Still made by a Taiwan company.

Anyway, still a good thing just not as good as having an actual competitor for TSMC.

 

[mac_angel]

is that why there's so many problems with the 50 series?

 

[phead128]

Intel is like....my foundry needs some business.

 

[EzzyB]

It should be noted that while unit sales of discrete GPUs for client PCs are generally decreasing,

That's a loaded statement. One wonders if that's true in a competitive market where these things are readily available from various manufacturers.

 

[JTWrenn]

That's a loaded statement. One wonders if that's true in a competitive market where these things are readily available from various manufacturers.

I concur. No idea what data set they are talking about considering pc gaming has been doing quite well. I guess steam decks could have dinged it a bit but I haven't seen anything saying GPU sales have gone down except in slumps before new releases.

Without some data that statement...well it just seems false.

 

[lmcnabney]

This is an outright LIE and it is very easy to prove. Just look at their financials. Their total costs of revenue for the last four quarters is about $32B. That cost is what they pay TSMC for the chips, their cost for RAM and other board components for their own branded gaming and AI GPUs. Oh yeah, and all of their development and support costs. They probably paid TSMC at most $15B and 4x that is $60B not $500B.

 

[hotaru251]

What chips made in states does NVIDIA even use???

As we all know TSMC has stated they will not have the "most advanced" stuff outside of Taiwan...so that'll never be made in states & we all know NVIDIA will need to use said cutting edge chips for its dominance in market (else it falls behind and we know they'd never allow that)

 

[EzzyB]

I concur. No idea what data set they are talking about considering pc gaming has been doing quite well. I guess steam decks could have dinged it a bit but I haven't seen anything saying GPU sales have gone down except in slumps before new releases.

Without some data that statement...well it just seems false.

My point is that sales may very well be down. But is it because the market demand is decreasing or because the supply isn't there to satisfy the market? This is coming from someone who has, reluctantly, decided to spend $1000 on a video card, but can't find one on the market worth near that much money.

It shouldn't be hard to spend a thousand dollars on a video card, but it is.

 

[JTWrenn]

My point is that sales may very well be down. But is it because the market demand is decreasing or because the supply isn't there to satisfy the market? This is coming from someone who has, reluctantly, decided to spend $1000 on a video card, but can't find one on the market worth near that much money.

It shouldn't be hard to spend a thousand dollars on a video card, but it is.

I don't see anything out there that says it is down. it's just people waiting for a decent mid range upgrade. It's the lull between upgrades for the mid range, but demand for them has gone up because gamer numbers are increasing. When the 5060 hits the market that will be where numbers will jump.

 

[thestryker]

That's a loaded statement. One wonders if that's true in a competitive market where these things are readily available from various manufacturers.

The graph on this page starts in 2014 and the general decline is obvious: https://www.jonpeddie.com/news/q424-pc-gpu-shipments-increased-by-4-4-from-last-quarter/

Of course I think a reasonable argument can be made for people holding onto video cards for longer as they have been all computer hardware in general (time between client generations is also longer). I wish we had figures for average price because I bet that has gone up higher than inflation.

What chips made in states does NVIDIA even use???

As we all know TSMC has stated they will not have the "most advanced" stuff outside of Taiwan...so that'll never be made in states & we all know NVIDIA will need to use said cutting edge chips for its dominance in market (else it falls behind and we know they'd never allow that)

All of nvidia's current chips are made on derivatives of N5 which is the node family being produced at the TSMC fab which recently started volume production in Arizona.

 

[edzieba]

This is an outright LIE and it is very easy to prove. Just look at their financials. Their total costs of revenue for the last four quarters is about $32B. That cost is what they pay TSMC for the chips, their cost for RAM and other board components for their own branded gaming and AI GPUs. Oh yeah, and all of their development and support costs. They probably paid TSMC at most $15B and 4x that is $60B not $500B.

Not so much a lie as Toms being confused or not understanding financials.

When Huang said:

“Overall, we will procure, over the course of the next four years, probably half a trillion dollars’ worth of electronics in total,” Jensen Huang told the Financial Times. “And I think we can easily see ourselves manufacturing several hundred billion of it here in the U.S.”

The "[Goods worth] probably half a trillion" means goods that, when sold by Nvidia, will net Nvidia ~$500bn in revenue. If we look at Nvidia 2025 revenue, it's $130bn. times 4 is $520bn. which matches the "half a trillion" revenue projection by Huang.

Toms instead assumed that it means Nvidia would be paying TSMC ~$500bn to produce chips, which is just incorrect.

The "I think we can easily see ourselves manufacturing several hundred billion of it here in the U.S." means >$200bn of revenue from products with chips fabbed in the US, which means much lower than $200bn in outlay to fabs in the US to produce those chips. If we naively assume similar materiel margins to 2025 ($32bn COGS for $130bn revenue, or 75% margin) then for that $200bn US-sourced revenue we'd expect an actual spend of ~$50bn.

 

[watzupken]

Whatever the cost of production, that cost will anyway be passed on to US companies who are "dumping" money at them. Jensen already knew this and so making this bold investment. That is as long as the AI bubble continues to be pumped up.

 

[phead128]

Anton Shilov is not most accurate writer at Toms. Tons of factually mistakes.

 

[TJ Hooker]

Not so much a lie as Toms being confused or not understanding financials.

When Huang said:

The "[Goods worth] probably half a trillion" means goods that, when sold by Nvidia, will net Nvidia ~$500bn in revenue. If we look at Nvidia 2025 revenue, it's $130bn. times 4 is $520bn. which matches the "half a trillion" revenue projection by Huang.

Toms instead assumed that it means Nvidia would be paying TSMC ~$500bn to produce chips, which is just incorrect.

The "I think we can easily see ourselves manufacturing several hundred billion of it here in the U.S." means >$200bn of revenue from products with chips fabbed in the US, which means much lower than $200bn in outlay to fabs in the US to produce those chips. If we naively assume similar materiel margins to 2025 ($32bn COGS for $130bn revenue, or 75% margin) then for that $200bn US-sourced revenue we'd expect an actual spend of ~$50bn.

I agree that this seems the most likely explanation. But that certainly isn't what Huang said. 'We will procure $0.5T worth of electronics' is very different than 'we will sell $0.5T worth of electronics'

Maybe he just misspoke. If he had said something like "we will procure parts for $0.5T worth of electronics", that would have made sense.

 

[TerryLaze]

'We will procure $0.5T worth of electronics' is very different than 'we will sell $0.5T worth of electronics'

I don't see it as different.
They will get electronics that will be worth 0.5T... for them, If they sell them at current prices.

They don't make them for themselves, so it wouldn't make any sense for him to mean the price of manufacturing, they always make them to sell them so if they don't specify it will always mean sell price.

 

[edzieba]

I agree that this seems the most likely explanation. But that certainly isn't what Huang said. 'We will procure $0.5T worth of electronics' is very different than 'we will sell $0.5T worth of electronics'

Maybe he just misspoke. If he had said something like "we will procure parts for $0.5T worth of electronics", that would have made sense.

It's the "X WORTH" that's important there.

Say I'm a baker. I sell $1000 worth of cupcakes a day. I buy $100 of flour a day to make those cupcakes.
To the flour supplier, that $100 of flour is worth $100. To me, that flour is worth $1000 because I do work to transform it into $1000 of sales. That sounds silly, until suddenly my supplier can only ship me half of that flour one day and I can only sell $500 of cupcakes - that $50 of flour was worth $500 of my income.
This is a weird way of accounting if you're used to personal finances where cash is cash, but makes perfect sense for business accounting where ephemeral things like "time" and "assets" and "risk exposure" have pseudocash values (that may even change over time).

 

[JTWrenn]

The graph on this page starts in 2014 and the general decline is obvious: https://www.jonpeddie.com/news/q424-pc-gpu-shipments-increased-by-4-4-from-last-quarter/

Of course I think a reasonable argument can be made for people holding onto video cards for longer as they have been all computer hardware in general (time between client generations is also longer). I wish we had figures for average price because I bet that has gone up higher than inflation.

All of nvidia's current chips are made on derivatives of N5 which is the node family being produced at the TSMC fab which recently started volume production in Arizona.

Considering the extra big pile before it and Covid timing...then the uptick afterwards I'm not sure any of that is about the state of gpus or the market. Seems like a pretty steady through line except for the covid hang over there when we had major inflation.

That's a lull.

 

[thestryker]

That's a lull.

No it's not a lull and here it's more obvious when you go back further, and look at just discrete (both graphs, first page):
https://www.anandtech.com/show/10864/discrete-desktop-gpu-market-trends-q3-2016

 

[EzzyB]

The graph on this page starts in 2014 and the general decline is obvious: https://www.jonpeddie.com/news/q424-pc-gpu-shipments-increased-by-4-4-from-last-quarter/

Of course I think a reasonable argument can be made for people holding onto video cards for longer as they have been all computer hardware in general (time between client generations is also longer). I wish we had figures for average price because I bet that has gone up higher than inflation.

All of nvidia's current chips are made on derivatives of N5 which is the node family being produced at the TSMC fab which recently started volume production in Arizona.

There is some truth to all of that. I'm a really old dude for forums like this, I've been working both personally and professionally with PC's since there have been PCs. Certainly for the first two-and-a-half decades a two-year-old PC was getting old and a three-years-old rig was a dinosaur (like me) 😉

I think of it as the time that personal computers finally caught up to the constant software bloat right about the time Intel released quad-core processors around '07. All of a sudden you're computer could last 5 or 6 years without a lot of pain.

Still what's missing here in the charts is capacity. And I mean manufacturing capacity. Before AI it was COVID, and before COVID it was crypto. How much of that capacity went somewhere other than to our specific market which is gaming? Before that all of it, essentially, went to gaming (I'm sure some was for scientific applications, super-computing, etc, but it was pretty small.)

I mean we've literally seen AMD refuse to play in the high-end market. That's not because they don't want to sell those sweet-sweet, high profit GPU's to us, they can just get an even better deal from Mega-Corp Inc. Nvidia just waves at us and puts 100 units in a market that wants to buy 1000 to try and satisfy us peasants.

So we really have no real idea what a competitive gaming GPU market would look like today. It doesn't exist. Supply, particularly in TSMC's manufacturing capacity, is an artificial bottleneck which is why I'm so bullish to see Intel enter the fab-for-hire market if they can pull it off. We should all be cheering them on.