TPU challenges GPU, but Bank of America suggests: buy NVIDIA, Broadcom, and AMD

Despite the challenge posed by custom chips like Google's TPU to GPUs, intensifying competition in the AI accelerator market, Bank of America still recommends investors buy Nvidia, Broadcom, and AMD amid the rapid expansion of the overall market.

On November 27, according to Hard AI news, Bank of America stated in its latest research report that although Google is trying to expand its ecosystem of custom chips by renting TPUs to Meta, which theoretically poses a challenge to Nvidia and AMD, the bank maintains a bullish stance on Nvidia, Broadcom, and AMD.

Analyst Vivek Arya emphasized that custom chips are primarily suited for companies like Google and Meta that have large-scale internal workloads, while GPUs still hold irreplaceable advantages in the public cloud and enterprise markets.

Bank of America expects that by 2030, the total addressable market (TAM) for AI data centers will grow fivefold, exceeding $1.2 trillion. Even if Nvidia's market share normalizes from the current 85% to 75%, its absolute revenue will still experience explosive growth.

Bank of America recommends a full buy on the three AI chip giants—Nvidia, Broadcom, and AMD—believing that current valuations do not fully reflect their long-term profitability.

Model War Escalates: Is TPU Starting to Encroach on GPU Territory?

The AI hardware sector is currently undergoing a critical architectural battle.

The report points out that Google has just released the Gemini 3 model, and shortly after, Anthropic quietly launched Claude Opus 4.5 yesterday, claiming to surpass Google's new model in coding and reasoning tasks.

Bank of America analysts noted that this indicates LLM development is a long-term marathon, and the current snapshot may not reflect the long-term evolution of market share.

Bank of America specifically pointed out that competition at the hardware level is more critical. Google relies on its TPU, which has been refined over a decade, for model training (both Gemini 2 and 3 are 100% trained using TPU). Crucial market rumors suggest that Google may rent TPUs to Meta next year and could achieve local deployment cooperation with Meta by 2027.

If this news is true, it would directly impact Meta's current GPU suppliers—Nvidia and AMD. This marks a shift where custom chips (ASICs) are no longer just tools for cloud vendors' internal use but begin to show potential for external output.

Trillion-Dollar Track Explosion: Nvidia's Share Normalization Does Not Change Growth Logic

Despite the aggressive advance of TPUs, Bank of America believes this does not alter the core logic of "growing the pie." The bank provided three reasons: the market size is exploding, share logic, and the moat remains strong, as detailed below:

Bank of America predicts that the overall addressable market (TAM) for AI data centers will grow from $242 billion in 2025 to approximately five times that, exceeding $1.2 trillion by 2030. Against this backdrop of rapid growth, even if Nvidia's market share declines, absolute revenue will still increase significantly.

Bank of America's model assumes that as customers develop custom chips to handle diverse workloads, Nvidia's market share will gradually normalize from the current 85% to 75%

Bank of America believes that while custom chips have cost advantages in specific internal workloads (such as Google's and Meta's internal demands), they lack flexibility in public cloud environments.

Microsoft Azure, AWS, and over 100 emerging cloud service providers require high flexibility, making general-purpose GPUs still the irreplaceable choice, even Google's own public cloud (GCP) is still using NVIDIA's GPUs.

Additionally, Bank of America analysts point out that general-purpose GPU chips have advantages such as spot supply, multi-cloud portability, a complete software stack, and a larger developer ecosystem. Furthermore, the tight supply chain situation and NVIDIA's scale advantages make it unlikely for market share to experience drastic changes in the short term.

Buy Rating on AI Chip Giants

The research report states that Bank of America maintains a buy rating on NVIDIA, Broadcom, and AMD.

Although NVIDIA is expected to achieve over 40% sales and earnings per share growth, its trading multiple is only about 25 times, which Bank of America believes undervalues the company, maintaining a buy rating with a target price of $275.

Based on data and share forecasts, NVIDIA's earnings per share capability is expected to exceed $10 by 2027 and break $20 by 2030.

Broadcom is seen as the biggest beneficiary of the custom chip trend. Bank of America gives it a buy rating with a target price of $400, corresponding to a 37 times price-to-earnings ratio for 2026.

Bank of America expects the company's AI business revenue to achieve over 100% year-on-year growth in 2026 due to additional TPU and Anthropic projects.

However, analysts warn that if Google directly licenses more TPUs, it could cut Broadcom's direct market opportunities for developing ASICs for other customers.

The report states that AMD, as NVIDIA's main competitor, is still worth holding. Bank of America gives it a buy rating with a target price of $300, based on a 32 times price-to-earnings ratio for 2027.

This also reflects the company's broad growth drivers in areas such as CPU, GPU, embedded systems, and gaming.

Despite facing risks of cyclical slowdowns in the embedded market, there is still significant upside potential in gaining market share in AI computing and CPUs.

As of Wednesday's close, the current stock prices of the three companies are $180.33, $397.57, and $214.24, respectively.