Semiconductor businesses are notoriously cyclical – a key reason why we are wary of them. While the good times can be very good indeed, there can be prolonged slumps, which is an unfortunate characteristic for investors seeking stocks to hold over the very long term.
Broadcom is a rare exception.
Often compared to Nvidia, Broadcom is similarly capitalising on surging demand for compute capacity for AI. But it is different in fundamental ways and today it is a much more diversified and less cyclical business than it was in the past.
Custom chips
Broadcom is a leading global semiconductor innovator with a rapidly growing and strategically important AI chips business. It specialises in custom AI chips, rather than off-the-shelf Graphics Processing Units (GPUs), which Nvidia designs and are renowned for strength in training AI models.
Custom chips are particularly well-suited to AI inference, which is where AI assistants make decisions and generate results, and demand for AI inference is surging.
ChatGPT is now estimated to have 900 million weekly active users, whilst in late 2025 Google said it was processing around 1.3 quadrillion tokens (the basic units of text that AI models process) across its AI platforms every month – up 20-fold year over year.
The rise of agentic AI models will only supercharge demand. According to Nvidia CEO Jensen Huang, the amount of computation required for agentic AI could be 1,000 times greater than ‘one-shot’ prediction chatbots.
All this means that custom chips’ share of total compute capacity is likely to rise materially from its current level of around 12%, not least since the large-cap tech companies will increasingly seek to develop their own custom chips.
Broadcom, as the dominant player in custom chips with around a 75% market share, is expected to be a main beneficiary of this accelerating demand.
A growing customer base
A large part of Broadcom’s AI semiconductor revenue has historically come from Alphabet, which has long partnered with Broadcom for its custom Tensor Processing Unit (TPU) chips, on which Google’s state-of-the-art Gemini 3 model was trained.
Google and Broadcom recently announced a long-term supply agreement for TPUs and networking chips that runs until 2031.
However, Broadcom has also been signing other major league clients including OpenAI, Meta and Anthropic, which has just signed an extended agreement with Google and Broadcom for multiple gigawatts of next-generation TPU capacity.
These agreements collectively add strong credibility to Broadcom CEO Hock Tan’s recent comment that he has “line of sight” to more than $100bn of AI chips revenue for 2027, which, incredibly, is a roughly five-fold increase on 2025.
IT software
Importantly, Broadcom is not just a leader in custom chips: it is also one of the dominant players in IT infrastructure software.
In 2023, Broadcom completed a very successful acquisition of VMware, a critical enabler of the virtual networks that typically run in today’s corporate offices (i.e. multiple machines run on an individual server). This high-margin software is extremely hard to switch out and is therefore ‘sticky’.
Moreover, the deal brought with it around 375,000 enterprise customers paying predictable subscription fees, which smooth the cyclicality in the semiconductor business and provide cashflow visibility.
Today, nearly 50% of Broadcom’s profits are driven by infrastructure software. While the market has recently focused on the potential for AI to disrupt the software space, Broadcom is confident that generative AI and agentic AI will “create the need for more VMware, not less”.
Capital light
One concern currently weighing on investors’ minds is the sustainability of the rise in AI capex, particularly among the hyperscalers. Whilst Broadcom investors will watch any change closely, custom chips are typically used by the hyperscalers for mission-critical internal workloads, insulating them from an attenuation in AI infrastructure spend.
Given the fierce competition in the generative AI space, a fall in capex seems unlikely in any case: companies simply cannot afford to halt spending while their rivals speed ahead.
For its part, Broadcom outsources manufacturing of chips to the likes of TSMC, limiting its own capex to only circa 1% of sales. This means that AI-related demand should drop directly to its free cashflow growth, comfortably supporting both double-digit dividend growth and buybacks.
Hock Tan, in place since 2006, has hugely improved the quality of the business over his tenure, partly through strong capital allocation and a focus on operational efficiency.
Tan has noted that Broadcom’s hyperscaler clients are “doubling down on inference in order to monetise their platforms”.
His compensation is well aligned with shareholders to drive strong growth in the custom chips business. He receives an equity payout currently worth $200m if custom chip sales reach $90bn by 2030 or triple this amount if sales reach $120bn.
While Broadcom still trades at a premium valuation to Nvidia (its current forward price-to-earnings ratio is around 27x versus the latter’s 22x), we believe it is well deserved given the accelerated growth profile of Broadcom’s AI semiconductor business and rise in contribution of software to its profit mix.
Overall, Broadcom has delivered strong shareholder returns over a number of years. We believe its diversified growth profile, backed by strong fundamentals and ability to capture accelerating structural AI tailwinds, remains compelling for long-term investors in quality companies.
Carolyn Bell is lead portfolio manager of the Stonehage Fleming Global Best Ideas Equity fund. The views expressed above should not be taken as investment advice.