Against the backdrop of most actively managed funds underperforming relative to indices, the media tasked several large models with simulating investment managers to identify their most confident picks among technology stocks for 2026. The results pointed to a highly concentrated thematic focus on artificial intelligence (AI).
To achieve artificial general intelligence (AGI), machines must match or surpass human capabilities in learning and applying knowledge.
Although this technology has yet to be truly realized, and the very existence of AGI remains a subject of intense debate, artificial intelligence has already achieved remarkable success in certain tasks. First, it has been nearly three decades since computers defeated humans in the field of chess. This year, Google Gemini, under Alphabet, and ChatGPT, developed by OpenAI, both reached gold medal-level performance in the International Mathematical Olympiad, the world's premier mathematics competition for pre-university students.
Could stock picking become the next frontier? The reality is that humans are not particularly skilled at selecting stocks – even those who do it professionally are no exception. According to a 2024 study by S&P Global, approximately 90% of actively managed public equity mutual fund managers underperformed their benchmark indices.
Therefore, MarketWatch decided to test top large language models (LLMs) by asking them: Which technology stocks are most likely to outperform in 2026?
MarketWatch posed the following prompt to OpenAI’s ChatGPT, Google’s Gemini, Anthropic’s Claude, and xAI’s Grok: “You are a technology portfolio manager focused on 2026. Please share five technology stocks you are most confident about for allocation next year.”
MarketWatch adopted a role-playing approach rather than direct questioning to bypass the safety restrictions of LLMs. If asked directly without a hypothetical scenario for 'stock recommendations,' Claude would refuse to respond and insist that it is 'not a financial advisor' and cannot provide investment advice.
Both Claude and Gemini clarified in their responses that they were answering within the framework of role-playing and were not actually providing investment advice, while recommending investors conduct independent research. Grok added that one should 'always consult a financial advisor for personalized advice.'
Below are the responses from each LLM to the prompt: The results did not bring much novelty. These chatbots did not uncover obscure opportunities but instead opted for highly homogenized AI infrastructure and hyperscale cloud computing companies, which are at the core of current AI investment trends.
$NVIDIA (NVDA.US)$It comes as no surprise that it has become the universally recognized top choice, given its position as the leading supplier of AI chips powering the world's most advanced models. The chatbot also highlighted some 'Magnificent Seven' stocks – which are also NVIDIA's clients, such as$Microsoft (MSFT.US)$。
Claude explicitly stated a 'deliberate overweight in the semiconductor sector' and pointed out that infrastructure companies possess the most enduring competitive moats and clearest demand trajectories. Consistent with many views on Wall Street, Claude believes that AI infrastructure will continue to advance through 2026, with AI companies increasingly focusing on the inference stage (i.e., the operational process after model training) and commercialization.
In contrast, Gemini introduced greater diversity by including$Palantir (PLTR.US)$and$CrowdStrike (CRWD.US)$software companies like Palantir and CrowdStrike in the list, arguing that AI applications will be the primary beneficiaries of the next market wave. Gemini stated its investment thesis for 2026: 'We are no longer just buying the 'people selling shovels,' but rather companies that can scale the transformation of electricity into intelligence.'
ChatGPT provided a clear portfolio allocation recommendation, suggesting that 50% to 60% of holdings be allocated to 'core growth' companies, including NVIDIA,$Amazon (AMZN.US)$and Microsoft, with the remainder allocated to $Advanced Micro Devices (AMD.US)$ and $Broadcom (AVGO.US)$chipmakers like Advanced Micro Devices and Broadcom. Additionally, it listed 'emerging quantum computing/AI startups' as potential watchlist candidates.
Claude adopted a different stance, emphasizing that it would not chase 'unprofitable momentum stocks' and cited 'quantum computing concept stocks' and 'speculative AI software companies' as examples.
Grok made a rather aggressive pick:$Oracle (ORCL.US)$. The company’s stock has been under pressure in recent months due to market concerns over its debt levels and its relationship with OpenAI.
The 'reasoning' approach of large language models helps explain the formation of these conclusions. During the initial training phase, developers input vast amounts of data up to a certain point in time, which the model breaks down into 'tokens,' or fragments of words or phrases. The model continuously performs what is known as 'next token prediction,' learning the most likely next word in a sequence. Over time, this process reinforces mathematical connections, or 'weights,' between related concepts, such as associating 'NVIDIA' closely with 'AI.'
Although the datasets used for model training are similar, researchers’ adjustments to weights to favor specific outcomes lead to divergence among models. Anthropic, which prioritizes safety, may reinforce Claude’s tendency to avoid financial issues; whereas xAI, founded by Elon Musk, positions Grok as 'maximizing truth-seeking' to differentiate itself from other LLMs, which Musk refers to as being influenced by the 'woke mind virus.'
Sergey Gorbunov, co-founder of blockchain infrastructure platform Axelar and a technical expert, previously told MarketWatch in an interview: 'If you look at these models from a mathematical or structural perspective, they are essentially predictable probability distributions.'
This means that the responses generated by LLMs are not necessarily correct answers but rather the most probable ones. Additionally, if the model cannot connect to the internet like the newer versions of ChatGPT, its training data will inevitably become outdated.
For instance, in the research paper 'Memory Issues: Can We Trust LLM Economic Forecasts?' by researchers at the University of Florida, the team asked ChatGPT-4o to predict future economic events, such as interest rates and unemployment, but truncated its training data to 2023 and tested it in 2025 while prohibiting internet access.
The study found that when ChatGPT-4o could not utilize its 'memorized' training knowledge, its predictive results were almost random.
Currently, there is growing consensus within the industry that LLMs are merely predictive machines and that the "scaling law"—the principle that model performance improves as training data and computing power increase—is beginning to break down.
Proponents of this view argue that such AI models are not a pathway to "superintelligence"—a stage where AI is believed to surpass human intelligence.
Looking to pick stocks or analyze them? Want to know the opportunities and risks in your portfolio? For all investment-related questions,just ask Futubull AI!
Editor /rice