What is the biggest narrative change in AI in 2026?

2026 may become a turning point for AI inference workloads to surpass training. According to the Wind Trading Desk, a team of analysts led by Vivek Arya at Bank of America has released a new research report stating that despite concerns over funding, valuations, and interest rate fluctuations, the continued growth of AI capital expenditures is driving strong performance in the chip sector, with industry focus shifting from training to inference, which can truly deliver investment returns. Bank of America predicts that of the $1.2 trillion in AI capital expenditures expected by 2030, inference may ultimately account for the majority share, reaching 75%. Despite numerous concerns in the market regarding AI financing, valuations, and interest rate fluctuations, chip stocks have started the year strongly. The Philadelphia Semiconductor Index (SOX) has risen about 13% year-to-date, marking the second-best January performance in the past 20 years, far exceeding the S&P 500's mere 1% increase. Notably, this rally has not been driven by computing giants Nvidia and Broadcom, but rather led by memory chips, semiconductor equipment manufacturers, and analog chip producers. The AI capital expenditures from hyperscale cloud service providers have injected confidence into the market. These companies emphasize three points: AI investment is crucial for maintaining double-digit growth; sales growth could have been higher if not for supply constraints; and there is no evidence of a "bubble." Bank of America Securities expects computing stocks to regain momentum based on these positive signals. This narrative shift will create differentiated opportunities for various types of chip suppliers, with a range of silicon solutions from GPUs, CPUs to ASICs finding their place in the inference market, while continuously impacting memory and semiconductor equipment suppliers. Inference Workloads Welcome a Turning Year Inference is a key link in achieving AI investment returns, requiring a series of chip solutions optimized for cost and performance. According to Bank of America Securities, while training remains important, 2026 may mark the year when inference becomes a larger workload, ultimately accounting for the majority share of the projected $1.2 trillion in AI capital expenditures by 2030, reaching 75%. In this transition, Nvidia maintains its leading position with the most extensive product pipeline, covering training (Blackwell, Vera Rubin) and inference (Rubin CPX, Groq, Vera CPU), and has supply assurance advantages. Broadcom has good partnerships with Google and Anthropic and has gained new opportunities with OpenAI, Apple, and xAI. Bank of America Securities believes that market concerns over Google's internal self-developed (COT) risks have been exaggerated, leading to an excessive decline in Broadcom's stock price. AMD, as a reliable second supplier of Nvidia's general-purpose chips, saw its stock price drop last Friday due to market overconcerns about TSMC's 2-nanometer process, but Bank of America Securities believes that the progress of this process is still on track. Opportunities for memory and semiconductor equipment suppliers continue to exist, although volatility may increase after recent surges The Demand for Optical Connections is Real but the Price Increase is Excessive Optical transceiver and component suppliers have become the second strongest performing chip sector after storage chips. With the expansion of AI cluster scale and bandwidth demand, the necessity for optical connections is undeniable. NVIDIA's upcoming photonic switch (a webinar will be held on February 3rd during the LITE earnings report release, with COHR announcing results at the same time) is a potential catalyst, and Marvell's acquisition of Celestial AI further amplifies market interest in new optical architectures. However, it is important to note that, aside from Meta, there is limited evidence of interest in Co-Packaged Optics (CPO) from hyperscale cloud service providers, due to its higher operational complexity and the transfer of more bill of materials control to NVIDIA and Broadcom. According to industry-leading analysis firm Lightcounting, CPO sales are expected to account for only about 1% (approximately $500 million) of the $46 billion Ethernet transceiver market in 2026/27. Copper cables remain relevant, and it is recommended to buy the leading active cable company CRDO