2026 may mark a pivotal turning point when AI inference workloads surpass training.
According to the Chasing the Wind Trading Desk, an analyst team led by Bank of America's Vivek Arya released a new research report stating that despite concerns over funding, valuations, and interest rate volatility, the continued growth in AI capital expenditure is driving strong performance in the chip sector, with the industry's focus shifting from training to the inference segment, which can truly deliver a return on investment. Bank of America indicates that by 2030, inference could ultimately command the majority share, or up to 75%, of the projected $1.2 trillion in AI capital expenditure.
Despite numerous market concerns regarding AI financing, valuations, and interest rate fluctuations, chip stocks have demonstrated a steady start to the year. The Philadelphia Semiconductor Index (SOX) has risen approximately 13% year-to-date, marking its second-best January performance in the past 20 years and significantly outperforming the S&P 500's mere 1% gain. Notably, this rally has not been led by computing giants NVIDIA and Broadcom, but rather by memory chip makers, semiconductor equipment suppliers, and analog chip manufacturers.
Positive commentary on AI capital expenditure from hyperscale cloud service providers has injected confidence into the market. These companies emphasize three key points: AI investment is critical for sustaining double-digit growth; sales growth could have been even higher if not for supply constraints; and there is no evidence of a "bubble." Bank of America Securities anticipates that computing stocks will regain momentum based on these positive signals.
This narrative shift is set to create differentiated opportunities for various types of chip suppliers. A range of silicon solutions, from GPUs and CPUs to ASICs, will find their place in the inference market, while also generating a sustained impact on memory and semiconductor equipment suppliers.
The inflection point for inference workloads is approaching. Inference is the crucial phase for realizing a return on AI investment, requiring a suite of chip solutions optimized for both cost and performance. According to Bank of America Securities analysis, while training remains important, 2026 may signify the year inference becomes the larger workload, ultimately claiming the majority share, or 75%, of the projected $1.2 trillion AI capital expenditure by 2030.
Amid this shift, NVIDIA maintains a leading position with the most extensive product pipeline, spanning training (Blackwell, Vera Rubin) and inference (Rubin CPX, Groq, Vera CPU), coupled with advantages in supply security. Broadcom sustains strong partnerships with Google and Anthropic and is securing new opportunities with OpenAI, Apple, and xAI. Bank of America Securities believes market concerns over the risk of Google's internal custom chip (COT) efforts are overstated, leading to an excessive decline in Broadcom's stock price.
AMD serves as a reliable second source for general-purpose chips against NVIDIA. Its stock price declined last Friday due to market overreaction to concerns about TSMC's 2-nanometer process, though Bank of America Securities views the process development as remaining on track. Opportunities for memory and semiconductor equipment suppliers persist, although volatility may increase following recent significant gains.
The demand for optical connectivity is real, but recent gains appear overextended. Optical transceiver and component suppliers have emerged as the second-strongest performing chip subsector, trailing only memory chips.
As AI cluster sizes and bandwidth requirements expand, the necessity for increased optical connectivity is undeniable. NVIDIA's upcoming photonic switch (a potential catalyst, with a webinar scheduled for February 3rd coinciding with LITE's earnings this week, and COHR reporting results around the same time) and Marvell's acquisition of Celestial AI have further amplified market interest in new optical architectures.
However, it is important to note that evidence of interest in co-packaged optics (CPO) from hyperscale cloud providers, aside from Meta, remains limited. This is due to higher operational complexity and a greater transfer of Bill of Materials control to NVIDIA and Broadcom. According to leading industry analysis firm Lightcounting, CPO sales are projected to account for only about 1% (approximately $500 million) of the estimated $46 billion Ethernet transceiver market in 2026/27. Copper cables remain relevant, with a recommendation to buy into active cable leader CRDO.
Comments