Broadcom ASIC – As hyperscalers shift from training AI models to inference at scale, ASICs are pulling fresh attention from investors. Broadcom, positioned as a leading ASIC designer, has forecast $100 billion in fiscal 2027 AI-chip sales. But the picture isn’t as simple as “
On a market day when so much AI money is still chasing GPUs. another kind of chip is quietly moving up the stack. ASICs—application-specific integrated circuits—don’t look as flexible on paper. but they can be far more cost-effective when the work is repetitive and the hardware never needs to change.
That’s the tension investors are weighing as they look at Broadcom (NASDAQ: AVGO) heading into a major moment: it reports earnings on June 3.
The broader backdrop is familiar to anyone watching the AI boom. The early surge in artificial intelligence data center investment centered on hardware for training AI models—workloads that demand massive parallel-processing capabilities. Graphics processing units (GPUs) are powerful and flexible parallel processors. and that flexibility helped Nvidia (NASDAQ: NVDA) rise from an ordinary large cap worth around $350 billion at the start of 2023 to a company valued at more than $5 trillion. making it the world’s most valuable company.
But data centers are getting bigger, which tightens cost constraints and feeds an AI energy bottleneck. At the same time, hyperscalers’ needs are shifting as AI inference becomes a larger part of the overall workload. Training builds the model’s intelligence. Inference applies that intelligence in real-world applications such as AI chatbots, AI agents, robotics, and self-driving cars.
The bet being made now is that ASICs can outperform for a lot of that inference work—if you can scale the right designs.
Application-specific chips are already showing up in hyperscaler roadmaps. Alphabet (NASDAQ: GOOGL) (NASDAQ: GOOG) has Tensor Processing Units (TPUs). Meta Platforms (NASDAQ: META) uses its Meta Training and Inference Accelerator (MTIA). Amazon (NASDAQ: AMZN) relies on Trainium.
Meta and Alphabet work with Broadcom’s custom accelerator platform to design their chips. Amazon Web Services (AWS) runs an in-house semiconductor division, Annapurna Labs.
The case for custom silicon isn’t vague. Google and AWS have achieved cost savings and efficiency improvements by deploying their custom chips at scale for cloud services. Google’s TPUs power its Gemini large language model and other AI-powered applications including Google Search. Google Maps. and Google Photos. Meta designed MTIA for its internal infrastructure, including its search and content recommendation algorithms.
Broadcom has been explicit about how far it thinks the demand could go. It is forecasting that it will book $100 billion in fiscal 2027 sales from its AI chips alone.
Even with that kind of runway, a crucial question hangs over the June 3 earnings event: is this the start of ASICs overtaking GPUs inside data centers—or something more nuanced?
ASICs aren’t nearly as flexible as GPUs. They can be highly cost-effective at scale, but they are hardwired for highly specific functions, such as machine learning workloads. 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, the logic for inference-heavy tasks is straightforward: using ASICs for a high-volume, repetitive inferencing job like content recommendation algorithms for Instagram and Facebook leverages what it has already learned and applies it to new inputs without needing to be retrained.
GPUs, especially Nvidia’s, take a different path. Nvidia’s GPUs can be easily reprogrammed to handle changing needs—an approach that fits customers where innovation and flexibility matter. Those include high-performance computing, regulatory-heavy industries, cybersecurity, and healthcare.
The end result is a split role that investors can’t ignore: GPUs are positioned for the frontier of AI as workloads evolve, while ASICs are better suited for maximizing efficiency when the workload is fixed.
That’s why even if Goldman Sachs forecasts that demand for ASICs will surpass GPU demand in the coming years. the “no-brainer” framing can be misleading. The point isn’t whether ASICs will matter. It’s whether they will matter in a way that shows up cleanly in Broadcom’s numbers before the market rewrites expectations around the training-versus-inference mix.
With Broadcom trading near its all-time high and carrying a premium valuation. the core pitch for long-term investors still leans on diversification and the AI runway. Broadcom has a diversified business model across non-AI semiconductors and infrastructure software, alongside its forecast for AI-chip growth.
If AI use cases and adoption keep widening, many inference-heavy tasks that do not require retraining could indeed benefit from ASICs—making June 3 a real test for the market’s confidence.
And that test is arriving fast: Broadcom’s earnings report is scheduled for June 3.
Broadcom ASICs GPUs Nvidia AI inference AI training hyperscalers TPUs MTIA Trainium June 3 earnings