Semiconductor R&D spending by company

[blueone]

In terms of percentage of revenue spent on R&D, Intel's $16.52 billion R&D spending is the second highest among the top 11 semiconductor companies by R&D spending. At such 31.25% high level, Intel doesn't really have any room to spend more on R&D unless it can increase its revenue much faster and bigger. Intel needs new killer products/services to increase revenue quickly but new products and services also require more R&D. Then it goes to the Chickens or the Eggs dilemma.

View attachment 1540
In US$ Billion
TTM = Trailing 12-month
Sorted by R&D Expense

How old are these statistics? For 2023 Nvidia's R&D/revenue ratio looks like ~27%.

NVDA | NVIDIA Corp. Annual Income Statement - WSJ

NVIDIA Corp. annual income statement. View NVDA financial statements in full, including balance sheets and ratios.

www.wsj.com

I think the more interesting question is: how close to TSMC's ~9% R&D expense IFS can get and when? In theory, if you have a lot of fab volume the geographic differentials in R&D costs should have a relatively minimal effect, because the R&D impacts CAPEX amounts that are much larger. Chip design companies have low CAPEX compared to companies with fabs (almost any non-fab company in any industry does).

 

[Mooredaddy]

How old are these statistics? For 2023 Nvidia's R&D/revenue ratio looks like ~27%.

NVDA | NVIDIA Corp. Annual Income Statement - WSJ

NVIDIA Corp. annual income statement. View NVDA financial statements in full, including balance sheets and ratios.

www.wsj.com

I think the more interesting question is: how close to TSMC's ~9% R&D expense IFS can get and when? In theory, if you have a lot of fab volume the geographic differentials in R&D costs should have a relatively minimal effect, because the R&D impacts CAPEX amounts that are much larger. Chip design companies have low CAPEX compared to companies with fabs (almost any non-fab company in any industry does).

We’ll never get the numbers but I would really love to see the ROI on R&D spend for each company. Intel spends a whole lot, but is much more spread out with the things it has to spend that money on.

 

[blueone]

We’ll never get the numbers but I would really love to see the ROI on R&D spend for each company. Intel spends a whole lot, but is much more spread out with the things it has to spend that money on.

Average ROI on R&D spending is just a matter of calculating the ratio of R&D to gross income for multiple years. In Intel's case, up to 2021, it was a ratio of about 3:1 to 4:1, which until recently was in the same ballpark as Nvidia. Obviously the ratios for these two companies have diverged in 2022 and 2023. More interestingly is Intel's high ratio of non-R&D SG&A expenses compared to R&D expenses versus Nvidia's results. Intel has been inefficient in sales and marketing expenses for years (decades, IMO), and they got away with it due to their historically dominant marketshare position which allowed premium pricing. One of Intel's challenges is, IMO, is that they have not reigned in their sales and marketing expenses as their market position has deteriorated.

 

[hist78]

How old are these statistics? For 2023 Nvidia's R&D/revenue ratio looks like ~27%.

NVDA | NVIDIA Corp. Annual Income Statement - WSJ

NVIDIA Corp. annual income statement. View NVDA financial statements in full, including balance sheets and ratios.

www.wsj.com

I think the more interesting question is: how close to TSMC's ~9% R&D expense IFS can get and when? In theory, if you have a lot of fab volume the geographic differentials in R&D costs should have a relatively minimal effect, because the R&D impacts CAPEX amounts that are much larger. Chip design companies have low CAPEX compared to companies with fabs (almost any non-fab company in any industry does).

I'm using most recent trailing 12-month (TTM) revenue and R&D Expense whenever it is possible or available.

I think you pointed out an important question that Pat Gelsinger and his team are probably working hard to find a solution. Otherwise IFS' long term prospect is very questionable. Like I said previously:

"Another way to evaluate the possible $5.62 billion R&D Intel spent on IFS is to compare it with IFS revenue. If IFS can reach $10 billion non-Intel revenue by 2026 and Intel will still maintain the same level of yearly R&D expenses on IFS, will $5 billion IFS R&D on a $10 billion IFS revenue make sense? To put it into perspective, Samsung Foundry 2022 revenue was US$5.39 billion."

 

[rjbrown747]

In terms of percentage of revenue spent on R&D, Intel's $16.52 billion R&D spending is the second highest among the top 11 semiconductor companies by R&D spending. At such 31.25% high level, Intel doesn't really have any room to spend more on R&D unless it can increase its revenue much faster and bigger. Intel needs new killer products/services to increase revenue quickly but new products and services also require more R&D. Then it goes to the Chickens or the Eggs dilemma.

View attachment 1540
In US$ Billion
TTM = Trailing 12-month
Sorted by R&D Expense

I estimated Samsung Semiconductor 2022 R&D Expense by using the following calculations:

View attachment 1541

Good stuff... how does the appx 15% compare historically for Samsung?... I'm especially interested in the inflection of EUV.

 

[hist78]

Capital expenditures (building and outfitting fabs) are not included in R&D.

It's not that straight forward anymore. The rule had been changed since January 2022.

"Why is this relevant? Prior to 2022, companies could deduct full R&D costs—also known as research and experimental expenditures (R&E) costs in the tax code—on their tax return. The 2017 Tax Cuts and Jobs Act put a stop to this practice.

As of January 1, 2022, companies are required to capitalize and amortize the cost of research and development—including software development cost.

So rather than immediately deducting R&D costs in the same fiscal year, you have to spread those costs out across five or 15 years—five years for R&D expenditures incurred within the U.S. and 15 years for R&D expenditures incurred abroad.

It’s worth noting that there’s no change to the generally accepted accounting principles (GAAP). R&D will continue to be expensed in companies’ financial reporting statements."

What is R&D capitalization?

R&D capitalization is a process companies use to categorize research and development costs as investments rather than expenses.

www.pluralsight.com

 

[blueone]

It's not that straight forward anymore. The rule had been changed since January 2022.

"Why is this relevant? Prior to 2022, companies could deduct full R&D costs—also known as research and experimental expenditures (R&E) costs in the tax code—on their tax return. The 2017 Tax Cuts and Jobs Act put a stop to this practice.

As of January 1, 2022, companies are required to capitalize and amortize the cost of research and development—including software development cost.

So rather than immediately deducting R&D costs in the same fiscal year, you have to spread those costs out across five or 15 years—five years for R&D expenditures incurred within the U.S. and 15 years for R&D expenditures incurred abroad.

It’s worth noting that there’s no change to the generally accepted accounting principles (GAAP). R&D will continue to be expensed in companies’ financial reporting statements."

What is R&D capitalization?

R&D capitalization is a process companies use to categorize research and development costs as investments rather than expenses.

www.pluralsight.com

These points are not really relevant to my point that R&D expenses (whether capitalized or not), are small compared to the cost of building and outfitting fabs. Everything is financially small compared to the cost of building and outfitting leading edge fabs. The only comparable thing I can think of quickly is a US nuclear aircraft carrier (the most recently built one cost about $20B). Even the Freedom Tower in New York City only cost $4B.

 

[Daniel Nenni]

Intel’s global push – investing billions in new production​

Intel’s global push – investing billions in new production

Since 2021, Intel has announced several massive projects – all in the multi-dollar scale – across the globe. From a USD 20 billion investment to build two new wafer fabs in Ohio, to a new USD 4.6 billion cutting-edge semiconductor assembly and test facility in Poland.

evertiq.com

 

[hkwint]

These points are not really relevant to my point that R&D expenses (whether capitalized or not), are small compared to the cost of building and outfitting fabs. Everything is financially small compared to the cost of building and outfitting leading edge fabs. The only comparable thing I can think of quickly is a US nuclear aircraft carrier (the most recently built one cost about $20B). Even the Freedom Tower in New York City only cost $4B.

So the names of ITER (between €20B-€65B) or Hinkley C (>£33B and growing) don't ring a bell?

The LHC in comparison was cheap, 'only' €5B.

 

[hkwint]

Efficiency of capital and research expenditures is key. Bang for the buck as they say and there are radical differences in application, use and effectiveness of expenditures. Strategy is also key, business can be like a game of chess or go.

Indeed!

ASML developed the physically/ mechanically almost impossible litho EUV for <€2B R&D a year, and for >€10B / yr Intel didn't seem to be able to just _use_ it in that timeframe.

Just listened to an hour long podcast with Martijn van den Brink.

Deel 2 van het interview met ASML's technische topman Martin van den Brink

omny.fm

ASML's strategy is that you are no supplier to them if more than 40% of your revenue depends on them, and that way, R&D is always shared with other companies, of which many not even semi-conductor related. And where Apple has a strategy "there's an app for that", ASML has the strategy "there's a supplier for that".

This only wasn't possible for Cymer & Zeiss, their tech could only be used for EUV. so instead ASML had to buy / invest in those companies.

At a time in the podcast, Martijn tells how solving hard problems became 'just normal' within ASML.

 

[count]

So the names of ITER (between €20B-€65B) or Hinkley C (>£25B and growing) don't ring a bell?

The LHC in comparison was cheap, 'only' €5B.

There are also things like major oil pipelines and refineries, large scale mines, satellites that are comparable order of magnitude, but I think the point still stands that building semiconductor fabs has become one of the most capex intensive activities on earth. In some ways fabs cost more, because the life of an oil pipeline, mine, aircraft carrier could be something like 20, 30, sometimes even 50 years, while a fab is basically fully depreciated and most of the technology inside is often obsolete in under 10 years.

 

[count]

Indeed!

ASML developed the physically/ mechanically almost impossible litho EUV for <€2B R&D a year, and for >€10B / yr Intel didn't seem to be able to just _use_ it in that timeframe.

Just listened to an hour long podcast with Martijn van den Brink.

Deel 2 van het interview met ASML's technische topman Martin van den Brink

omny.fm

ASML's strategy is that you are no supplier to them if more than 40% of your revenue depends on them, and that way, R&D is always shared with other companies, of which many not even semi-conductor related. And where Apple has a strategy "there's an app for that", ASML has the strategy "there's a supplier for that".

This only wasn't possible for Cymer & Zeiss, their tech could only be used for EUV. so instead ASML had to buy / invest in those companies.

At a time in the podcast, Martijn tells how solving hard problems became 'just normal' within ASML.

Until EUV was commercialized a Intel kind of looked down on them like "this is basically just a complex machine, nowhere near as sophisticated and high tech as what we are doing". Now Intel has get in line and beg them for these machines.

 

[blueone]

So the names of ITER (between €20B-€65B) or Hinkley C (>£33B and growing) don't ring a bell?

The LHC in comparison was cheap, 'only' €5B.

Good point about ITER. I remember reading once about Hinkley Point, but being an American means it wasn't top of mind for me.

As long as you brought up a fission power fiasco, a friend of mine who lurks here and lives in the US state of Georgia wanted to remind me about Vogtle Plant Units 3 & 4, that have ballooned to cost nearly $US35B. So there are examples of things that cost more than 3nm fabs, but they are few, and, as @count wisely pointed out, these other things have service lives measured in decades.

 

[hkwint]

Good point about ITER. I remember reading once about Hinkley Point, but being an American means it wasn't top of mind for me.

As long as you brought up a fission power fiasco, a friend of mine who lurks here and lives in the US state of Georgia wanted to remind me about Vogtle Plant Units 3 & 4, that have ballooned to cost nearly $US35B. So there are examples of things that cost more than 3nm fabs, but they are few, and, as @count wisely pointed out, these other things have service lives measured in decades.

Yeah agree to that....

However funny thing is that ITER will be depreciated really quick, it's basically a research facility. Other technologies for fusion in both the US (with lasers) and Germany (Wendelstein) look more promising already.

Voigtle C btw doesn't have a bad ROI compared to Virgil C Summer.....

I checked TSMC financial data, and machinery is depreciated within 5 years.

However, I remember somewhere they still have a joint venture fab with Philips in Singapore from 2000 still running (SSMC). So the accounting life is definitely different than technical life. SSMC has a service life measured in decades too! And almost all old TSMC fabs are still running.

I wonder what a High NA EUV machine costs, I think the first one is shipping soon. I believe to Intel. It seems the astronomically high prices for fabs only go further up. And now ASML is researching Hyper NA (NA.0.7)....

 

[count]

ITER won’t be depreciated that fast, it’s a research project the primary purpose is to generate research not profits. Regardless of new technology they will be running experiments there for decades.

 

[Barnsley]

Yeah agree to that....

However funny thing is that ITER will be depreciated really quick, it's basically a research facility. Other technologies for fusion in both the US (with lasers) and Germany (Wendelstein) look more promising already.

Voigtle C btw doesn't have a bad ROI compared to Virgil C Summer.....

I checked TSMC financial data, and machinery is depreciated within 5 years.

However, I remember somewhere they still have a joint venture fab with Philips in Singapore from 2000 still running (SSMC). So the accounting life is definitely different than technical life. SSMC has a service life measured in decades too! And almost all old TSMC fabs are still running.

I wonder what a High NA EUV machine costs, I think the first one is shipping soon. I believe to Intel. It seems the astronomically high prices for fabs only go further up. And now ASML is researching Hyper NA (NA.0.7)....

There is expansion going on at SSMC as well.

I assume its NXP or TSMC footing that bill as havent heard of any Govt inputs