Broadcom Rides Rocketing Trend For Custom AI Accelerators

retained text

Broadcom Rides Rocketing Trend For Custom AI Accelerators
Back in March, when looking out into the future of the custom XPU and networking ASIC chips as well as the other chippery that Broadcom sells that goes into AI clusters, I posited that Broadcom, more than AMD or Intel, might become the biggest counterbalance to Nvidia in the GenAI boom. And that it would be getting a little competition from and help from Marvell in this task.
This is turning out to be even more prophetic than I thought, based on the revenue forecast that Hock Tan, chief executive officer and legendary conglomerate builder at Broadcom, provided to Wall Street when going over the third quarter of fiscal 2026 financial results. We will cut to the chase scene here and just show you a table of Broadcom’s AI-related chip sales since F2022 including the blunt – and ebullient –forecast Tan provided:
Broadcom was already helping Google get its homegrown TPUs designed and through the Taiwan Semiconductor Manufacturing Co foundry and packaging for seven years when the GenAI boom hit. And basically, between F2022, when the GenAI boom started, and the F2028 forecast, that custom AI XPU business plus AI networking and that other chippery going into AI systems will grow by two orders of magnitude from $2.6 billion to $230 billion. The US military budget for the fiscal year just ended was $1 trillion by comparison. Then again, the worldwide consumption of pizza – the world’s most multinational food – was $230 billion in 2024, which I just learned through reading Ten Tomatoes That Changed The World by William Alexander.
Broadcom has secured the HBM memory, chip wafers, and packaging to fulfill this revenue stream for AI compute and networking capacity, and Tan said if the demand is higher, it can add more capacity.
Tan warned Wall Street to not expect quarterly updates to that forecast, but wanted to make it clear that some big money was going to rolling in during the coming years and basically to tell Wall Street to take a chill pill on the GenAI bubble bursting.
Across the six XPU customers that Broadcom has – Google, Anthropic, OpenAI, and Meta Platforms for sure, and possibly ByteDance (which owns TikTok) or Apple for the other two, but Amazon Web Services is using Broadcom for packaging and integration for Trainium 4 – there is about 30 gigawatts of total compute capacity, which suggests that you can install XPU systems for somewhere around $12 billion per gigawatt – not including the CPUs, memory, and racks. You are probably around $15 billion per gigawatt when you add those components in. This compares favorably to the $18 billion per gigawatt for Nvidia “Grace-Hopper” systems (which do include CPUs and memory as well as all of the GPUs and interconnects needed to make a rackscale machine) and $40 billion per gigawatt for Nvidia “Vera-Rubin” systems.
Tan said on the call with Wall Street that it would ship “multi-tens of billions of dollars of TPUs annually over the next several years” to Google. “We have consistently delivered the fastest time to market for TPUs from product definition to production without the need for resubmits. We believe these are very deep moats for any competitor to cross.” So forget all of the talk about Marvell, MediaTek, or Alchip stealing the TPU business. It is all about who gets the HBM allocations anyway. . . .
Anthropic is deploying 1 gigawatt of “Ironwood” TPU 7 accelerators this year, will be deploying 5 gigawatts of TPU 8i systems in 2027, and in 2028 Anthropic wants 10 gigawatts of the follow-on TPU 9. Tan said that Anthropic was on track to be its largest XPU customer in 2027 and would sustain that in 2028. (Presumably these are numbers for Broadcom’s fiscal years, not calendar years, but Tan did not say fiscal when referring to these numbers.) That means Anthropic is buying more TPU iron by the grace of Google than Google itself is. As one of the AI model builders with tons of cash flow and a strong balance sheet, you would think Google would be buying more TPUs than anyone else. But Google is a cloud, and cloud customers want Nvidia GPUs.
Crazy, right?
OpenAI has chosen Broadcom to do its chip shepherding for its “Jalapeno” XPU, and is on track to buy 1.3 gigawatts of stuff from Broadcom for its Jalapeno clusters in 2027. In 2028, Broadcom thinks OpenAI will deploy another 5 gigawatts of Jalapeno ASICs and its follow-on, which will make OpenAI its second biggest XPU customer, knocking Google down to number three on the list.
Crazy, right?
The follow-on OpenAI chip, which we do not know the code-name of, is approaching tapeout, according to Tan, and the two companies are working on the third generation XPU design now.
With Meta Platforms, Broadcom will deliver three generations of MTIA devices between now and the end of fiscal 2027, and across those three generations of MTIA accelerators – which we drilled down into back in April – from now and the end of fiscal 2028, Broadcom has a line of site on 3 gigawatts of capacity.
The AI boom inside of Broadcom is not just AI XPUs, and in fact is dominated by networking. The 100 Gb/sec and 200 Gb/sec SerDes variants of the Tomahawk 6 Ethernet switch ASIC are selling like hot cakes, and all of the hyperscalers and clouds that are using Broadcom to shepherd their homegrown XPUs are using Tomahawk 6 for scale out networking on their AI clusters. The Tomahawk Ultra variant of the 6 series chips, which drives down latency in the Ethernet stack so it can be used for scale up coherent memory interconnects for XPUs inside of rackscale machines, is having surprising adoption and is being deployed now at some customers with the ramp progressing through fiscal 2027 as new XPUs are coming to market.
And by the way, the 204.8 Tb/sec Tomahawk 7 Ethernet switch ASIC, which has not been announced yet but which uses 400 Gb/sec SerDes and which gives you 3.2 Tb/sec per port with eight lanes per port, has just taped out and will presumably be seen in switches starting in early calendar 2028 and maybe in late calendar 2027.
One last thing and then I will dig into the Q3 F2026 numbers with you. I fully expect for Broadcom to start shepherding CPUs, and very likely focusing on packaging with a little design work as it is doing with Fujitsu for its forthcoming “Monaka” Arm Server CPU. Fujitsu as well as the six XPU customers are using Broadcom’s 3.5D Extreme Dimension System in Package (XDSiP) chiplet stacking, which puts stacked DRAM as well as stacked CPU chiplets into a package. With the CPU total addressable market growing from $26 billion in 2025 to $220 billion by 2030 and with every hyperscaler and cloud builder making its own Arm server CPUs, I think Broadcom will branch into CPU shepherding. I do not think Broadcom has any interest in designing its own Arm server chip and trying to sell it to these tech titans. This is why it sold off the “Vulcan” processor to Cavium Networks all those years ago.
I do think it is possible that Broadcom will help customers integrate CPU tiles onto their XPU tiles if they decide to hardwire their CPU to XPU ratios. But I also think that having a big memory address space on the CPU (albeit a slower one) is beneficial to XPUs that have to skimp on HBM or other stacked memories because of the scarcity of capacity for those memories. So an external CPU might be the safest bet.
We shall see.
In the third quarter ended in early August, Broadcom raked in $29.59 billion in overall sales, up 85.5 percent year on year and up 33.4 percent sequentially. Operating income was $15.96 billion, up by 2.71X compared to Q3 F2025. Net income grew even faster, rising by 3.17X to $13.09 billion, representing a very nice 44.2 percent of revenues.
Say what you will, but Tan knows how to squeeze a lot of money out of the companies he buys and runs. It is so much better to be a chip shepherd indulging the egos of the tech titans then to try to sell a finished XPU or CPU to them. None of them want to use whatever any of their rivals are using, except X86 CPUs and Nvidia and sometimes AMD GPUs, and they only tolerate that because they are clouds and their customers demand it and because of the shortage of CPU and GPU capacity.
Broadcom ended the quarter with $23.98 billion in cash, but it still has $59.42 billion in debt thanks to the VMware acquisition.
In the quarter, Broadcom’s Semiconductor Solutions group had $20.84 billion in sales, up 127.4 percent, with an operating income of $12.71 billion, up 143.3 percent.
The Infrastructure Software group, which includes VMware and a bunch of other legacy enterprise systems software, had sales of $8.75 billion in Q3, up 29 percent, and operating income was an IBM mainframe-class 84 percent of revenue at $7.35 billion, up 40.7 percent. My model says that $5.52 billion of that was for VMware licenses and support, with the remaining $3.23 billion for other legacy software. But take those numbers with a quarter of a shaker of salt, given that I am extrapolating into the void because Broadcom doesn’t give out much in the way of details about its software business.
Above is my guesstimate on how the old Broadcom divisions are doing. Use another quarter of the salt shaker on that chart, again because Broadcom doesn’t really talk much about these businesses anymore. I probably will stop after this quarter, too.
And finally, that brings us to Ai versus non-AI sales in the chip business. Here is a pretty picture on that:
And there is at table outlining the AI versus non-AI breakdown:
AI-related revenues added up to $16.74 billion, up 3.22X from a year ago and up 54 percent sequentially. Of that, $12.22 billion was for AI compute and $4.52 billion was for AI networking.

document fetched 2026-09-11 02:58 · canonical https://nextplatform.com/connect/2026/09/10/broadcom-rides-rocketing-trend-for-custom-ai-accelerators/5295681