Portfolio Manager Insights

Chip off the old block: AI goes custom

As AI giants race to secure powerful, proprietary chips, Fidelity’s Aidan Brandt sees a new wave of custom silicon emerging that could reshape the competitive landscape.

  • Artificial-intelligence hyperscalers are seeking a competitive edge by moving to in-house development of the proprietary chips that underpin their AI infrastructure, according to Fidelity Portfolio Manager Aidan Brandt. In turn, this shift is creating significant opportunities for chip-design partners that can help bring these specialized solutions to market.
  • “The first wave of AI was powered by off-the-shelf hardware,” says Brandt, who co-manages Fidelity Advisor® Semiconductors Fund with Sonu Kalra. “I think the next chapter could belong to the companies designing custom silicon tailored to the purpose-built workloads that matter most.”
  • In leading the industry-focused equity strategy, Brandt and Kalra believe the value of technology stocks is driven in large part by companies’ potential to generate earnings and cash-flow growth. Their investment framework focuses on identifying themes that influence the largest end markets, assessing potential winners and losers, and evaluating how technology disruptors could affect incumbent players.
  • Brandt has watched with interest as companies at the forefront of AI, including Google parent Alphabet, Amazon, Microsoft, Facebook parent Meta Platforms, and OpenAI, have increasingly shifted toward developing custom computing chips optimized for the memory- and compute-intensive workloads required to run AI applications.
  • The goal, he says, is to trade some flexibility for a lower total cost of ownership on workloads that businesses know they will be running at massive scale for years to come.
  • Brandt points to Google’s Tensor Processing Units, or TPUs, as an early example of what he prefers to call a “custom accelerator” rather than an ASIC. Originally built to support the company's search applications, as well as AI and machine-learning workloads, TPUs have since expanded into a broader range of use cases.
  • “I expect to see a proliferation of these custom accelerators, primarily among the handful of organizations globally with the resources, including enormous AI spending, technical expertise and purchasing scale, to economically justify developing these chips in-house,” Brandt explains.
  • Many of these firms are savvy enough to recognize the risks of going it alone as a chipmaker, according to Brandt, and are therefore enlisting the support of specialized semiconductor design and manufacturing partners, such as Broadcom and Marvell Technology – large fund holdings as of September 30, 2026.
  • These companies help customers, including Amazon, OpenAI and Anthropic, implement internally designed logic, manage supply chains, oversee packaging and integrate custom silicon into large-scale AI systems, according to Brandt. He notes that custom accelerators are optimized for specific workloads, particularly AI inference tasks such as chatbot applications.
  • From Brandt’s perspective, the trade-off is clear: customers are willing to sacrifice some versatility in exchange for improved economics and potentially lower long-term operating costs. 
  • “As AI spending ramps into the hundreds of billions, the race is no longer just about building bigger models,” he concludes. “It’s increasingly about controlling the silicon underneath them, potentially creating a new generation of leaders in AI infrastructure.”
Featured Fund

Fidelity Advisor Semiconductors Fund (FELIX)

Seeks to provide capital appreciation.