The initial boom in artificial intelligence (AI) data center investment centered around hardware for training AI models -- workloads that demand massive parallel-processing capabilities. Graphics processing units (GPUs) are powerful and flexible parallel processors, and that property propelled the rise of GPU specialist Nvidia (NVDA +0.59%) from an ordinary large cap worth around $350 billion at the start of 2023 to its current position as the world's most valuable company. Today, it's worth more than $5 trillion.
But data centers are growing in size, creating cost constraints and an AI energy bottleneck. What's more, hyperscalers' needs are evolving as AI inference becomes a growing part of the overall workload. While training is needed to build a model's intelligence, inference applies that intelligence in real-world applications, such as through AI chatbots, AI agents, robotics, and self-driving cars.
Application-specific integrated circuits (ASICs) aren't nearly as flexible as GPUs, but they can be highly cost-effective at scale. This is why investment bank Goldman Sachs forecasts that demand for ASICs will surpass GPU demand in the coming years.
Broadcom (AVGO +5.23%) is the leading designer of ASICs, so it stands to benefit from that boom. But is the semiconductor stock a buy before it reports earnings on June 3?
Image source: Getty Images.
ASICs at the hyperscale
Alphabet's (GOOGL 4.33%) (GOOG 3.70%) Tensor Processing Units (TPUs), Meta Platforms' (META 0.28%) Meta Training and Inference Accelerator (MTIA), and Amazon's (AMZN 2.14%) Trainium chips are examples of ASICs.
Meta and Alphabet's Google work with Broadcom's custom accelerator platform to design their chips, while Amazon Web Services (AWS) has an in-house semiconductor division, Annapurna Labs.
Google and AWS have achieved cost savings and efficiency improvements by deploying their custom chips at scale for cloud services. Google uses TPUs to power its Gemini large language model and other AI-powered applications like Google Search, Google Maps, and Google Photos. Meta designed MTIA for its internal infrastructure, including its search and content recommendation algorithms.
Broadcom hasn't been shy about quantifying the surging demand for ASICs. It's forecasting that it will book $100 billion in fiscal 2027 sales from its AI chips alone. Beyond that, the company also has massive non-AI semiconductor and infrastructure software businesses. However, it would be a mistake for investors to assume ASICs will overtake GPUs in data centers.
NASDAQ: AVGO
Key Data Points
ASICs won't replace GPUs
ASICs are cost-effective at scale, but rigid because they are hardwired for highly specific functions, such as machine learning workloads. That said, the software stack remains flexible, which is why Google and AWS have been discussing selling their custom chips to select third parties for similar AI training and inference tasks.
For Meta, using ASICs for a high-volume, repetitive inferencing task like its content recommendation algorithms for Instagram and Facebook makes perfect sense because it leverages what it has learned previously and applies it to new inputs without needing to be retrained.
Nvidia's GPUs, particularly when paired with its CUDA software platform, can be easily reprogrammed to handle changing needs. As such, GPUs are more suitable for customers where innovation and flexibility are required, such as those in high-performance computing, regulatory-heavy industries, cybersecurity, and healthcare.
In other words, GPUs will always be needed at the frontier of AI for evolving workloads, whereas ASICs are useful for maximizing efficiency for fixed workloads.
A top AI inferencing play
Even with the stock trading near its all-time high and carrying a premium valuation, Broadcom remains a generational buying opportunity for long-term investors because it has a diversified business model across non-AI semiconductors and infrastructure software, plus a clear runway for AI-driven growth.
As AI use cases and adoption become more widespread, many simple, straightforward, inference-heavy tasks that do not require retraining will benefit from ASICs. Investors who believe AI will be increasingly used to automate stable, repetitive workloads should take a closer look at buying Broadcom in June.
