In the early days of the artificial intelligence (AI) infrastructure buildout, graphic processing units (GPUs) have been the go-to chips for companies to train AI models and run inference. GPUs are powerful chips that were originally designed to speed up graphics rendering in video games, but later became used for other applications after Nvidia created a software platform to allow developers to program it chips for other purposes.
However, a number of large tech companies have also started to use custom AI chips, or ASICs (Application Specific Integrated Circuits). ASICs are custom chips that are designed for very specific purposes. As such, they tend to perform better and be more efficient at these tasks compared to a GPU, but aren't as flexible in how they can be used.
Let's look a two companies that are helping lead the way with custom AI chips.
Broadcom
When it comes to custom AI chips, Broadcom (AVGO -1.93%) has been the leader in the space. Alphabet was its first AI chip customer, with Broadcom helping to design its tensor-processing unit (TPU) called Trillium. This chip was created to specifically work within Google Cloud's TensorFlow (a software library for AI and machine learning). Alphabet has said these custom chips have several features that make them better than GPUs for AI training and inference. These include having matrix multiply unit (MXUs), which are components used to efficiently perform large matrix multiplications, and SparseCores, which are "dataflow processors that accelerate models relying on embeddings found in recommendation models."
Last quarter, Alphabet credited using a combination of GPUs and TPUs as being a differentiator for its cloud computing segment. It said its TPUs helped lower costs while also reducing inference time.
Since then, Broadcom has added more custom AI customers, which are believed to be Meta Platforms, ByteDance, OpenAI, and most recently Apple. The company saw strong growth from AI chips in its fiscal 2024 ending early November, with its AI revenue topping $12 billion, well above its expectations of around $7.5 billion going into the year.
The company excited investors last quarter when it indicated each of its three initial AI customer could deploy up to 1 million AI chips in 2027, which it said would represent a $60 billion to $90 billion revenue opportunity in 2027 alone. It added that if its two new customers were able to get their chips into develpment by that time that this opportunity would be even greater.
While Broadcom isn't likely to win all this business, with some going towards GPUs, it shows the strong customer AI chip opportunity the company has in front of it. The stock, meanwhile, trades at a forward price-to-earnings (P/E) ratio of 35.5 based on 2025 analyst estimates, which is attractive if it can capture a good share of the 2027 opportunity it has talked about.
Image source: Getty Images
Marvel Technology
Broadcom is not the only company that is helping customers design custom chips. Marvel (MRVL -2.60%) is also in the custom silicon game as well. It's best known for helping Amazon with its Trainium chip that is used to help train large language models (LLMs).
While Broadcom is believed to have completely designed chips for Alphabet based on its needs, Marvel is primarily providing some intellectual property around high speed SerDes (Serializer-Deserializer) with Amazon doing most of the work, according to analysts at Morgan Stanley. On its fiscal Q3 earnings call in December, Marvel said that it recently signed a a five-year multi-generational deal with Amazon's cloud unit AWS for custom AI products as well as various other data center switches and components. It said the deal should lead to a significant ramp in volumes between the two companies.
Amazon is not Marvel's only AI customer, with the company saying it has a range of design wins for both AI accelerators and compute and that other programs would go into production in 2025. It also said it will have a third large customer soon. At its AI day, the company said it sees custom AI chips as a $40 billion market opportunity and that it thinks it can take a 20% market share, or $8 billion.
Marvel said it was on track to exceed $1.5 billion in AI revenue this fiscal year, with more growth expected in fiscal 2026. It said much of this well come from its custom silicon program.
Outside of custom AI chips, Marvel is very tied to the data center market. Its data center revenue soared 98% year over year last quarter to $1.1 billion. Data center revenue now accounts for 73% of its total revenue, up from it being 39% of its total revenue a year ago. However, overall revenue rose just 7% due to weakness in other end markets.
Overall, with data center spending continuing to ramp up, Marvel should see strong growth in this area. The stock currently trades at 42 time next year's earnings estimates, which is a bit on the high side.
