Broadcom (AVGO -0.66%) grew its artificial intelligence semiconductor revenue by 221% year over year in the third quarter, and that growth could accelerate in the fourth quarter. Management's outlook calls for $21.7 billion worth of AI-related revenue, up 236% from a year ago.
There's a clear driver behind that revenue -- custom AI accelerators, or what management calls XPUs. These chips are designed in partnership with hyperscalers, including Alphabet (GOOG -0.52%) (GOOGL -0.61%), Meta Platforms (META -1.82%), Anthropic, and OpenAI. And while Nvidia (NVDA -0.80%) and its GPUs have been the workhorse of AI training and inference, custom semiconductor chips could end up being a much bigger business.
Here's why and what it means for both Broadcom and its customers.
Image source: Getty Images.
The biggest bottleneck in the AI build-out
There are a handful of bottlenecks in AI training. Much focus has been placed on how quickly hyperscalers can stand up new data centers and the glut of demand for memory chips against a short supply. But another bottleneck investors need to consider is how much capital the hyperscalers can actually access.
Alphabet reported negative free cash flow last quarter, the first time it's ever done so as a publicly traded company. Meta is on track to do the same. Meanwhile, Anthropic and OpenAI are extremely cash-flow-negative. As a result, they've had to raise capital via debt and equity.
NASDAQ: GOOG
Key Data Points
With limited capital, hyperscalers must allocate it to assets with the highest potential return on investment. To be sure, there's a baseline demand for GPUs, which are often better for training or mixed-workload tasks. However, custom silicon offers better power-to-performance ratios and faster inference speeds while requiring less upfront capital. As a result, custom silicon can exhibit higher returns on capital.
That was crystalized in a recent interview given by Google Cloud CEO Thomas Kurian. He said the payback period on its AI servers is less than two years, but on Google's own silicon, it's less than half that. If the useful life of an AI server is six years, that increases the return on investment from about 200% to about 500%. It also reduces the risk of a slowdown in compute demand through smaller capital outlays and shorter payback periods.
Keeping inference costs low
There's a reason Anthropic and OpenAI are designing their own chips as well. They see the value in Alphabet's full-stack business model, where it designs the custom silicon that its large language models run on. Google Gemini offers a very low price per token for frontier models, and Kurian says that's entirely because it runs on Google's TPU chips on Google Cloud.
As more businesses use AI for more tasks and agentic AI becomes increasingly useful, token costs will play a significant role in determining which model enterprises choose. Being able to offer competitive pricing with Google will require OpenAI, Anthropic, and other LLM developers to use the most efficient chips available. More often than not, that will mean running most of their compute on custom silicon.
NASDAQ: AVGO
Key Data Points
That further tilts the scales in favor of XPUs, as not only can they provide a higher return on invested capital for hyperscalers selling access, but they can also reduce the biggest ongoing cost for frontier labs (and, by extension, other enterprises).
Will XPUs overtake GPUs?
Nvidia's data center business exited the second quarter at a $356 billion run rate. That's absolutely massive, more than 4 times the size of Broadcom's XPU business.
But XPU adoption is growing faster, and the trend favors continued growth.
Earlier this year, Amazon CEO Andy Jassy said his company would add more custom silicon servers to its data centers than Nvidia GPUs. Comments from Kurian, as well as Microsoft executives, suggest they're also leaning heavily toward adding more custom silicon than GPUs.
By volume, XPUs already seem to be winning the race among Nvidia's biggest customers, and that lead should continue to expand over the next few years. However, considering XPUs garner lower prices than GPUs, it will be a long time before revenue catches up. That shouldn't matter to investors, though, as prices should reflect the business's future growth. While Broadcom trades at a premium to Nvidia based on future earnings, that premium may be worth paying, considering the growth outlook for custom silicon.
