I'm having trouble understanding Broadcom's claimed AI revenue results and projections

[blueone]

Broadcom announced results for for 1Q26:

Broadcom is claiming Q1 "AI Revenue" of $8.4B, based mostly on "robust demand for custom AI accelerators and AI networking". Looking at these reports, Broadcom doesn't break down revenue contributions from Ethernet Products, which they just ascribe everything to "AI Networking", and "custom AI accelerators", like the Google TPUs, which Broadcom calls XPUs.

The S&P Global revenue chart for 2025 states that XPU revenue is much bigger than the Ethernet products revenue.

The XPU revenue looks to be about $11B-12B in 2025. My question is, how could Broadcom achieve such high unit revenue for chips unless they were actually selling the chips made and packaged by TSMC (or others) to the hyperscalers? This implies that Broadcom is managing the entire chip engineering process and owning the relationship with TSMC, and selling completed units to the hyperscalers, just like Intel or AMD would.

This isn't what I thought was going on. I thought the hyperscalers owned the business relationship with TSMC, especially Google. Volume-wise, the only two hyperscale customers that matter are AWS and Google. AWS chip design, AKA Annapurna Labs, does its own backend engineering. So where is all of Broadcom's revenue and unit volume for AI XPUs coming from?

Hence my confusion.

https://investors.broadcom.com/news-releases/news-release-details/broadcom-inc-announces-first-quarter-fiscal-year-2026-financial

https://www.spglobal.com/market-intelligence/en/news-insights/research/2026/02/broadcom-q1-seen-up-29percent-yoy-as-ai-revenue-nears-usd8-billion-margins-set-to-ease

 

[Markwrob]

It looks to me like a fairly textbook case of building customer specific SoCs. Granted, these designs may also include custom packaging also handled by the vendor. Hock is just demanding a better rate of pay now.

Many years back I worked for the storage BU for Agere/LSI Logic* and that model was the same: customer has some secret sauce IP along with a digital "back end", but contracted us to provide the Phys and other analog "front end" IP and all the connecting logic like link layers and etc. We would handle all the integration, layout, test and sometimes even multi chip modules**, even back then. Chip fab was usually at TSMC (occasionally with a second source) at that time, and we worked closely with the fab to handle any issues. We got paid pretty decently to take on those roles and risks.

I think Hock and Co. are running that same model, on steroids. One thing Hock is great at, is finding a price that the customer will swallow that also makes great margins for him.

*: now part of AVGO

**: grudgingly, for there was great risk involved with handling the Known Good Dies model at that time, and the volumes for these was rather low

 

[blueone]

It looks to me like a fairly textbook case of building customer specific SoCs. Granted, these designs may also include custom packaging also handled by the vendor. Hock is just demanding a better rate of pay now.

That's what the revenue would appear to say, but in previous conversations on this forum it was postulated by some (including Dan) that Google owned the relationship with TSMC, not Broadcom. This made financial sense to me.

Many years back I worked for the storage BU for Agere/LSI Logic* and that model was the same: customer has some secret sauce IP along with a digital "back end", but contracted us to provide the Phys and other analog "front end" IP and all the connecting logic like link layers and etc. We would handle all the integration, layout, test and sometimes even multi chip modules**, even back then. Chip fab was usually at TSMC (occasionally with a second source) at that time, and we worked closely with the fab to handle any issues. We got paid pretty decently to take on those roles and risks.

I was a customer of LSI Logic when they were still fabricating chips.

I understand getting high prices from second-tier companies, but from Google and Microsoft? Both of these customers could simply do their own end-to-end chip design, and I would suspect for less money.

I think Hock and Co. are running that same model, on steroids. One thing Hock is great at, is finding a price that the customer will swallow that also makes great margins for him.

*: now part of AVGO

**: grudgingly, for there was great risk involved with handling the Known Good Dies model at that time, and the volumes for these was rather low

 

[Daniel Nenni]

I understand getting high prices from second-tier companies, but from Google and Microsoft? Both of these customers could simply do their own end-to-end chip design, and I would suspect for less money.

Apple used ASIC service companies when they first started doing the iProducts but transitioned fully to internal development. Apple has a lot of software to drag along with chip development so it made complete sense for them to do end-to-end SoC development. Apple also has a big secrecy shroud around them. I remember meetings about tools and IP where Apple asked questions and you answered, that was it.

Google is a different animal. They write some very BIG ecosystem checks. Seriously, money is not an issue nor is compute power for design and simulations. Google is looking for the best silicon. Can Google do it better internally than Broadcom? I do not believe so. Broadcom has the experience and IP to ensure the best of the best chips which is what Google wants. There is a lot of ego involved in semiconductor design and all of this chip investment has to be justified to the stake holders. Google has to have the best silicon and arguably they do. And remember, writing big ecosystem checks is what Google has always done.

 

[IanD]

Apple used ASIC service companies when they first started doing the iProducts but transitioned fully to internal development. Apple has a lot of software to drag along with chip development so it made complete sense for them to do end-to-end SoC development. Apple also has a big secrecy shroud around them. I remember meetings about tools and IP where Apple asked questions and you answered, that was it.

Google is a different animal. They write some very BIG ecosystem checks. Seriously, money is not an issue nor is compute power for design and simulations. Google is looking for the best silicon. Can Google do it better internally than Broadcom? I do not believe so. Broadcom has the experience and IP to ensure the best of the best chips which is what Google wants. There is a lot of ego involved in semiconductor design and all of this chip investment has to be justified to the stake holders. Google has to have the best silicon and arguably they do. And remember, writing big ecosystem checks is what Google has always done.

We used Broadcom as a custom ASIC supplier in the past, they did all the design and dealt with packaging and testing and shipped packaged chips with the customer name on, which is their normal business model.

The downside was the cost, a four-digit number for something we knew the foundry cost (for packaged tested chips) was a fraction of this. OK for small volumes, not so good for larger ones -- at some point it becomes cheaper to do your own design, *if* you have the expertise and capability -- which many companies don't, especially for massive chips like AI accelerators...

 

[blueone]

Apple used ASIC service companies when they first started doing the iProducts but transitioned fully to internal development. Apple has a lot of software to drag along with chip development so it made complete sense for them to do end-to-end SoC development. Apple also has a big secrecy shroud around them. I remember meetings about tools and IP where Apple asked questions and you answered, that was it.

Apple is a different animal altogether. Total A-series + M-series SoC volume is about... 300 million units per year? And each SoC has DRAM dies in it. Apple would be foolish to out-source direct relationships with suppliers and TSMC over SoC production, and Apple execs are as far from fools as exist in the high-tech industry (I would argue).

Google is a different animal. They write some very BIG ecosystem checks. Seriously, money is not an issue nor is compute power for design and simulations. Google is looking for the best silicon. Can Google do it better internally than Broadcom? I do not believe so. Broadcom has the experience and IP to ensure the best of the best chips which is what Google wants. There is a lot of ego involved in semiconductor design and all of this chip investment has to be justified to the stake holders. Google has to have the best silicon and arguably they do. And remember, writing big ecosystem checks is what Google has always done.

Google is a lot different from Apple. Google TPU volume was about 2.5 million units for 2025. That's a rounding error on Apple's volume. Even if Google's Arm CPUs (Axion) are included, I suspect that's only a few million more units per year. And Axion's key advantages for Google are lower power than x86 products, and lower price. Broadcom's claimed AI revenue numbers are huge. Something smells fishy here.

 

[IanD]

Apple is a different animal altogether. Total A-series + M-series SoC volume is about... 300 million units per year? And each SoC has DRAM dies in it. Apple would be foolish to out-source direct relationships with suppliers and TSMC over SoC production, and Apple execs are as far from fools as exist in the high-tech industry (I would argue).

Google is a lot different from Apple. Google TPU volume was about 2.5 million units for 2025. That's a rounding error on Apple's volume. Even if Google's Arm CPUs (Axion) are included, I suspect that's only a few million more units per year. And Axion's key advantages for Google are lower power than x86 products, and lower price. Broadcom's claimed AI revenue numbers are huge. Something smells fishy here.

Don't forget that there will be lots of chips out there with big names on them (like Google, or Meta, or many others...) which nobody knows were actually built by Broadcom, their name isn't anywhere on them... ;-)

 

[Daniel Nenni]

Don't forget that there will be lots of chips out there with big names on them (like Google, or Meta, or many others...) which nobody knows were actually built by Broadcom, their name isn't anywhere on them... ;-)

Right, and what part of the chip did Broadcom actually do? Google does design, Broadcom does implementation. I remember working with eSilicon a while back. They did not make a lot of money on the operations side of the business (package, test, etc...), not even close to the margins they make on design. The problem always was that the ASIC companies are literally training their customers to do it themselves. Not an easy business to be in but now that systems companies are all trying to get control of their own silicon business is booming. Exciting times in the semiconductor industry!

 

[blueone]

Don't forget that there will be lots of chips out there with big names on them (like Google, or Meta, or many others...) which nobody knows were actually built by Broadcom, their name isn't anywhere on them... ;-)

I know that, but the volume numbers imply very expensive chips. Thousands of dollars. Perhaps I'm just being naive, but I have a difficult time believing Google is paying thousands of dollars for a TPU.

 

[IanD]

Right, and what part of the chip did Broadcom actually do? Google does design, Broadcom does implementation. I remember working with eSilicon a while back. They did not make a lot of money on the operations side of the business (package, test, etc...), not even close to the margins they make on design. The problem always was that the ASIC companies are literally training their customers to do it themselves. Not an easy business to be in but now that systems companies are all trying to get control of their own silicon business is booming. Exciting times in the semiconductor industry!

Drawing the line between "design" (front-end) and "implementation" (back-end) is not so easy nowadays, especially if you want a power-optimized design. In many ways coming up with the architectures and algorithms is quicker and easier than actually turning the result into robust fully-functional silicon especially for massive chips, which is what Broadcom do. And some other big companies, of course... ;-)

 

[Daniel Nenni]

I know that, but the volume numbers imply very expensive chips. Thousands of dollars. Perhaps I'm just being naive, but I have a difficult time believing Google is paying thousands of dollars for a TPU.

The cost equation was much easier for fabless companies that sell the chips versus Google that uses them internally. How much money does Google make off a single TPU? Certainly more than fabless companies who sell the chips. Except of course maybe Nvidia.

 

[Daniel Nenni]

Broadcom rises as $100 billion AI forecast signals gains in Nvidia-dominated market

Broadcom shares rose nearly 3% on Thursday after the company predicted over $100 billion in AI chip sales next year, signaling rapid share gains in ‌the market dominated by Nvidia.

www.reuters.com

 

[IanD]

I know that, but the volume numbers imply very expensive chips. Thousands of dollars. Perhaps I'm just being naive, but I have a difficult time believing Google is paying thousands of dollars for a TPU.

I don't now what the TPU die size is but I expect it'll be pretty big, implying a KGD cost of several hundred dollars. In which case given Broadcom's typical margins, Google would be paying a four-digit sum. It's really not easy doing the implementation of devices like this, otherwise everyone would be doing it -- and it's a bit like using TSMC, the best way to ensure success is to use someone who's done it all before, lots of times... ;-)

Paying big margins to someone like Broadcom for a chip that is on time and works is *far* cheaper than trying to do it yourself and ending up with something flaky or with problems or that needs a respin, which nowadays means at least a 9-month delay in getting to market -- which can really kill your product... :-(

To avoid this and do the same in-house needs a *huge* investment in time and money and people, and this still doesn't help until it has succeeded, preferably more than once. That's a big cliff to climb...

 

[ChrisGar]

I don't now what the TPU die size is but I expect it'll be pretty big, implying a KGD cost of several hundred dollars. In which case given Broadcom's typical margins, Google would be paying a four-digit sum. It's really not easy doing the implementation of devices like this, otherwise everyone would be doing it -- and it's a bit like using TSMC, the best way to ensure success is to use someone who's done it all before, lots of times... ;-)

Earlier comments on SemiWiki suggested that Broadcom was "just doing some back-end support" for the TPUs / AI processors. According to analysts like Beth Kindig, Broadcom manages the design & manufacturing process .. and sells a finished product to their customers. (Nvidia CEO has said their main competitors are Nvidia's customers doing their own processors.) Clearly $100B of revenue for AI semiconductors suggests that Broadcom's business model is different from what is often described on this site. Broadcom CEO has talked about revenue from Anthropic ($21B) where Broadcom will sell TPUs directly to Anthropic. Products like Google TPUs are large processors with chiplets and many HBMs on the same package.

 

[blueone]

Earlier comments on SemiWiki suggested that Broadcom was "just doing some back-end support" for the TPUs / AI processors.

That's not really correct. Most of us were saying that Broadcom was supplying some IP (though for TPUs, which use proprietary Google networks, it's not clear what that IP would be, since Broadcom mostly talks about I/O and networking IP) and pretty much all of the circuit design, not just "some backend support". In a post not all that long ago, Dan said Google had a direct business relationship with TSMC. Even assuming Broadcom does the entire backend development, thousands of dollars per TPU still seems expensive for what Broadcom provides.

In other words, I think either we are underestimating Broadcom's role in TPU development, or Broadcom is including a bunch of revenue in "AI" that really isn't AI. The SP Global chart shows that Broadcom is indeed pumping up "AI" revenue with their Ethernet business, but the fraction of Broadcom's AI business that isn't Ethernet is still significantly larger than I can explain without Broadcom being a private label merchant chip provider, to coin a term.

 

[blueone]

Drawing the line between "design" (front-end) and "implementation" (back-end) is not so easy nowadays, especially if you want a power-optimized design. In many ways coming up with the architectures and algorithms is quicker and easier than actually turning the result into robust fully-functional silicon especially for massive chips, which is what Broadcom do. And some other big companies, of course... ;-)

Says the backend guy.

But seriously, I believe it is more difficult to hire great backend people than it is chip design people. The physical design concepts are more difficult. They involve actual physics. Digital design is more conceptual, and computer science and engineering concepts are easier to master than the mathematics and physics necessary to make electrons and photons do what you what them to do, not to mention power and thermal considerations. Also, from years working in the chip development management chain, it is difficult to employ your own backend team unless you have a chip development pipeline of a certain depth, otherwise you can't keep your backend people sufficiently busy over the long term. I'll admit, I did expect Google to have a deeper chip product pipeline for Google Cloud by now than they appear to have, and they must be relatively happy with Broadcom's results to pay that high pricing (assuming my guess is correct).

 

[Markwrob]

Perhaps Google look at these questions like they consider office space: renting the expertise vs owning it. There are pros and cons, which can and do shift depending on the observer.

 

[IanD]

Says the backend guy.

But seriously, I believe it is more difficult to hire great backend people than it is chip design people. The physical design concepts are more difficult. They involve actual physics. Digital design is more conceptual, and computer science and engineering concepts are easier to master than the mathematics and physics necessary to make electrons and photons do what you what them to do, not to mention power and thermal considerations. Also, from years working in the chip development management chain, it is difficult to employ your own backend team unless you have a chip development pipeline of a certain depth, otherwise you can't keep your backend people sufficiently busy over the long term. I'll admit, I did expect Google to have a deeper chip product pipeline for Google Cloud by now than they appear to have, and they must be relatively happy with Broadcom's results to pay that high pricing (assuming my guess is correct).

Hey, I resent that -- I'm not a backend guy, I'm a high-speed custom mixed-signal / technology guy... ;-)

(but I talk to both front-end and back-end teams on both digital and analog sides, and have been involved in technology and library benchmarking for both -- so I know where the *really* nasty pain points are...)

Also seriously, the backend is getting harder and harder for both analog and digital with each process node, the architecture/algorithm side isn't that much different apart from scaling up complexity, it's just dealing with more data.

If you want to find out just how many different swear words English has, talk to someone doing full-custom layout at transistor level in N2... ;-)