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

[ChrisGar]

I've been thinking about the question you're raising for several days now. To me, the answer is non-obvious. Google Cloud has an operating margin of 30.1% (4Q25), but they are hardly an AI pure play. The majority of Google Cloud's revenue, CAPEX, and operating expenses are probably not AI-related. Also, everything about TPUs is expensive. The interconnects (scale-up and scale-out) are proprietary, and I suspect they also use high-margin partners like Broadcom (but I'm not sure who for optical interconnects), so I suspect TPUs and their chassis, racks, and networks are higher cost than AMD's or Nvidia's (even for NVLink and InfiniBand). No one else has enough volume yet to count.
.....

So, to get back to your question, I think it is not clear that Google Cloud TPUs are more profitable than AMD GPUs.

TPUs are used to mainly drive their internal workloads -- more than the Cloud. TPUs are used to drive the ~$300B in advertising for example.
I'm not sure how an outsider could calculate that ROI or margins.

 

[blueone]

I assume you meant Nvidia GPUs in the last line?

No, I specifically meant AMD GPUs. Nvidia's gross and net margins are higher than anyone else's, including all three major cloud computing companies'.

Nvidia margins are enormous (more than 70%?) because they can charge what the market will stand (silicon + IP + CUDA) and there's little merchant competition. I would expect that any custom chip (like a TPU) will be cheaper than this, even allowing for Broadcom's margins -- which are also high, but unlikely to be *as* high as Nvidia (especially in volume) because they add less value to the raw silicon than Nvidia do. Cheapest of all is an all-in-house custom chip, but few companies are capable of doing this for such humongous devices.

You're forgetting about the cost of networking. Nvidia's networking chips and other hardware certainly have lower unit costs than Google's proprietary designs do. Nvidia's networking business is #1 in the world now, and they do all of the design in-house, which has got be cheaper than what Broadcom charges.

 

[blueone]

TPUs are used to mainly drive their internal workloads -- more than the Cloud. TPUs are used to drive the ~$300B in advertising for example.
I'm not sure how an outsider could calculate that ROI or margins.

I agree that TPUs are probably mostly for in-house use, but their contribution to Google's advertising business must also be modest, because TPU unit volume is only about 2.5 million units for 2025. I also agree, and stated, that I'm guessing.

 

[blueone]

I just saw TP Morgan's article on Broadcom's results on his site, THENEXTPLATFORM. While the article contains a lot of his usual editorializing, more than reporting, it does have one paragraph that is especially interesting.

Broadcom May Become The Biggest Counterbalance To Nvidia

Here’s a funny thought. If Broadcom had not bought legacy software maker Computer Associates in Nove ...

www.nextplatform.com

Broadcom has built relationships with five big AI customers for hone-designed XPUs, and now has added a sixth who will be ramping this year. Tan said that demand from Google for its “Ironwood” TPU v7 accelerators would be strong in 2026 and in 2027 and beyond demand for TPUs from Google would be even stronger. Separate from Google’s own TPU demand, Broadcom is also benefitting from Anthropic’s decision to buy $10 billion in TPU racks, with Google’s permission. The plan is for Anthropic to install 1 gigawatt of TPU v7 capacity in 2026 (we presume Tan is speaking in fiscal years) and that this will more than triple to in excess of 3 gigawatts in 2027 (again, a fiscal year).

Basically, what Morgan and Broadcom are implying is that Broadcom benefits financially from Google's TPU deployment volume. I take this to mean that Google is actually buying TPUs from Broadcom, not TSMC. This only makes sense if Broadcom really is a private label merchant chip vendor, and, if true, IMO it changes the entire view of Broadcom in the AI industry, and suddenly Broadcom's AI financials make more sense.

The rest of the paragraph regarding the Anthropic TPU deployment continues to make no sense to me, unless Broadcom is actually already Google's merchant rack provider. The latest information I've read is that Foxconn and Celestica produces the rack systems and integration for Google, according to Google's specifications. So it's more mysterious. I'm still not convinced TPUs are going anywhere but Google Cloud.

SemiAnalysis has an excellent article on TPUv7 and its proprietary scale-up and scale-out interconnects (which unfortunately is only available behind their paywall), but the complexity it details so well only reinforces the notion that the Anthropic deal to do their own TPU systems looks very difficult to believe.

I've never run into a computer industry company of Broadcom's size where their one of their key businesses and how their financials are actually earned is such a mystery.