AI is seldom out of the headlines in 2025, with defining developments coming one after another. We look at where AI is today and how its promise is matched against technological, economic, and geopolitical challenges.
October 3, 2025
Frédérique Carrier Managing Director, Head of Investment StrategyRBC Europe Limited
On Jan. 29, 2025, a little-known Chinese tech company, DeepSeek, released an AI model, R1, that shook the industry. R1’s cutting-edge capabilities make it seemingly as good a model as those created by U.S. leader OpenAI, the maker of ChatGPT. But what unsettled the industry most was its shoestring development cost – just $6 million, a fraction of what comparable U.S. models required.
Barred from importing state-of-the-art chips from the U.S. due to the export restrictions imposed by the former Biden administration in October 2022, DeepSeek had to rely on older hardware. To compensate, the company pushed efficiency to the limit, making careful tradeoffs between accuracy and computing power, replicating this method at scale, and fine-tuning every other aspect of performance. This approach kept development costs remarkably low while still producing a highly capable model.
Adding to the excitement, DeepSeek-R1 was released as an open-source model – i.e., publicly available at no cost. Users can download and run it on their own computers or servers, keeping data private, and they can retain, modify, or adapt it for their own needs. This means anyone, from individuals to large companies, can build tools or applications without seeking permission or paying for expensive access.
Open-source models stand in contrast to most of today’s leading AI models, such as OpenAI’s ChatGPT, Anthropic’s Claude, or Google’s Gemini, which are closed-source. It wasn’t always so. OpenAI initially produced open-source models, but the San Francisco-based lab shifted to a closed approach in 2018.
Closed-source model creators share only parts of the code, or describe some of the training process but keep the full details private, much like a secret recipe. Users must rely on those companies to maintain the entire AI system, and are thus vulnerable to changes in model creators’ prices or access rules.
For instance, in the case of ChatGPT, OpenAI is responsible for training the model, maintaining the code that runs the model, running the model on powerful servers to enable users to access ChatGPT instantly, and adding features. If OpenAI were to stop doing all that, users could not run it themselves.
DeepSeek’s decision to make its cutting-edge AI model both cheap and widely available gave China an unexpected competitive edge, as its model could spread and be adapted far more widely and cheaply than those of its Western rivals. This shook the U.S. tech ecosystem to its core. Until then, the U.S. had been confident in its position as the world’s leading AI force.
DeepSeek may have captured the attention of the West, but China’s AI achievements go far beyond a single company. That should not be a surprise as millions of engineers and scientists graduate from Chinese universities every year, the country has the spare grid capacity necessary to run the power-hungry AI models, and its permissive planning laws allow data centers to be built swiftly.
Still, China faces a critical constraint: it lacks the sufficient domestic supply of cutting-edge chips. Huawei, the country’s hardware champion, cannot yet produce top-end chips in sufficient quantities. Nevertheless, the combination of these favourable conditions, remarkable ingenuity, and a relentless effort at squeezing as much efficiency as possible from older-generation chips has enabled Chinese firms to release frontier AI models (i.e., highly capable general-purpose AI models that can perform many tasks matching or exceeding the capabilities of today’s most advanced models).
In July 2025 alone, Alibaba, one of China’s largest tech and e-commerce companies, released Qwen3, a model approximately one-quarter the size of the most prevalent AI models, making it significantly more energy-efficient while maintaining comparable performance. Meanwhile, Moonshot AI, a lesser-known AI company, unveiled Kimi K2 in July, one of the largest open-source models ever released at that time. Kimi K2 excelled in benchmarks like MATH-500, which tests mathematical reasoning, outperforming frontier class U.S. models from OpenAI (GPT-4) and Anthropic, according to Venturebeat, a publication that focuses on technology news.
Despite this progress, China still lags the U.S. in productization – turning AI models into agentic tools, or fully integrated tools that can autonomously assist users in complex workflows. Agentic tools take initiative, make decisions, and complete multistep tasks for the user, such as reading incoming customer queries, selecting the urgent ones, drafting replies, and escalating complex issues to a human.
The rivalry in technology between China and the U.S. has been ongoing for years. Both countries see winning the AI race as a strategic advantage – not only will the winner be able to extend their geopolitical influence through the supply of AI systems to other countries, but there are also important implications for military applications.
In 2025, the White House initiatives to secure an edge in the AI race included releasing its AI Action Plan in July and changing chip export restrictions to China.
Early in his second term, U.S. President Donald Trump directed his administration to develop an AI strategy for the country. The AI Action Plan is built on three core pillars: accelerating innovation, expanding data center infrastructure, and promoting American technology abroad.
According to the Brookings Institution, an American think tank, the plan can be lauded for focusing on advancing and democratizing basic and applied AI research, and addressing the need to develop an AI-ready workforce. However, it has a few concerns, including that adequate safety measures may be compromised by the strong emphasis on accelerating U.S. AI innovation and global competitiveness. Insufficient oversight – particularly in financial AI applications – could pose significant systemic risk.
Despite these misgivings, Brookings scholars think that overall these factors could strengthen the regional innovation ecosystems in the U.S. so long as the federal government provides adequate support and funding.
Source – Executive Office of the President, July 2025
The Trump administration is also using export controls, a strategy used in Trump’s first term and followed up by former President Joe Biden to respond to the Chinese threat. This time, however, the controls have sent confusing signals.
In April 2025, the Trump administration banned exports of NVIDIA’s H20 chips to China over concerns the technology could strengthen Beijing’s defense industry. After strong lobbying from the semiconductor industry, the ban was lifted in July. A few weeks later, Washington announced a new framework: NVIDIA and Advanced Micro Devices would be granted export licenses to sell specific chips to China so long as they shared 15 percent of their revenue from these chip sales with the U.S. government.
April 2018: Initial restrictions
Trump administration blocks sales of advanced chips to Chinese telecom firm ZTE. (R)
June 2020: Expanded restrictions
Trump administration extends prohibition of advanced chips sales to Chinese conglomerate Huawei. (R)
2020–2021: Pressure on allies
Trump administration pressures allies to impose similar restrictions (e.g., the Dutch government restricts ASML from selling its most advanced semiconductor equipment to China). (R)
October 2022: Continued export ban
Biden administration imposes broad restrictions on export of high-end chips (e.g., H100) to all Chinese entities based on national security concerns. (D)
November 2023
NVIDIA unveils the H20 chip, which complies with U.S. export restrictions.
April 2025: Further expansion of restrictions
Trump administration bans export of H20 chips, believed to be powerful enough to support Chinese defense industry. (R)
May 2025: Further expansion of restrictions
Export ban expanded to include newer models (e.g., cutting-edge H200 chip) to curb China’s technological advancement. (R)
July 2025
U.S. government reverses ban on H20 chips; H100 and H200 export bans remain in effect. (R)
August 2025: Revenue sharing agreement
NVIDIA and AMD agree to pay 15% of their China chip revenue to the U.S. government in exchange for export licenses to resume sales to China. (R)
R = Republican government initiative; D = Democratic government initiative
Source – RBC Wealth Management
This policy U-turn points to the difficulty of calibrating the security and economic interests of the U.S., and exposes tensions between China hawks in the U.S. government pushing for tighter export controls and businesses eager to access the world’s second-largest economy.
The Trump administration seems to have adopted NVIDIA CEO Jensen Huang’s view that providing China access to NVIDIA’s AI chips could serve both the company’s interests and U.S. strategic goals by creating Chinese dependence on American technology. NVIDIA supplies not just the chips, but also the hardware and infrastructure that support entire data centers. If major Chinese AI firms such as Alibaba, ByteDance, and Tencent build their data centers around U.S. hardware, it could give Washington a geopolitical advantage and greater leverage in any future negotiations with China. The logic is that if U.S. technology were completely banned from China, Beijing would likely accelerate efforts to develop its own AI infrastructure.
The H20 chip had been developed two years earlier by NVIDIA specifically for the Chinese market and to comply with the Biden administration’s 2022 export controls. It was designed mainly for AI inference – the process by which trained models generate insights and suggest decisions – but lacks the power needed to train new models. By offering a chip that China could use but one that is markedly weaker than the next-generation H100 and H200, Washington sought to maintain China’s dependency on American hardware, while at the same time limiting its ability to advance in frontier AI.
Many observers worry that engineering a financial payout for the U.S. government has now taken precedence over national security. Reversing the H20 export ban, they argue, may be a strategic mistake, effectively providing China with the hardware it needs to surge ahead in AI.
But Chris Miller, acclaimed author of Chip War and professor of international history at Tufts University, offers a more nuanced perspective, drawing on his expertise in the global semiconductor industry and geopolitics. He believes that despite the 15 percent financial arrangement, national security is still at the heart of Trump’s policies. He argues that despite criticizing Biden’s CHIPS Act, the Trump administration continues to disburse grants promised under the 2022 law to semiconductor companies and research institutions. Moreover, the U.S. government has taken a 9.9 percent equity stake in Intel, the only U.S. semiconductor firm that both designs and manufactures leading-edge logic chips. Miller believes the U.S. administration sees Intel as relevant to the broad future of U.S. technological leadership.
Whether the strategy of making China “addicted” to U.S. tech, in Commerce Secretary Howard Lutnick’s terms, will prove successful remains to be seen. In August, China urged local companies to avoid using NVIDIA’s H20 processors particularly for government-related purposes, according to Bloomberg. Then in September, it banned its largest tech firms from buying NVIDIA’s AI chips in an effort to foster domestic production.
Slowing China’s technological progress has long been a preoccupation of U.S. administrations, and it is instructive to look at how U.S. export control policy evolved over time.
The first Trump administration realized that:
As a result, its strategy was two-pronged:
Under the Biden administration the strategy evolved as follows:
Another key AI development in 2025 has been the shift in funding sources for the substantial investment needed to build the infrastructure to support AI models.
According to McKinsey, a consulting company, global data centers will need between $3.7 trillion and $5.2 trillion by 2030 to meet demand for AI computing power, including hardware, processors, memory, storage, and energy.
Much of this will be shouldered by Big Tech. Companies like Alphabet (Google), Amazon, Microsoft, and Meta Platforms (Facebook) – also referred to as “hyperscalers” – are building large data centers to support their cloud services and AI initiatives. Traditionally, they often preferred to self-fund these investments, and were able to while maintaining robust balance sheets with minimal debt. But this is changing due to the scale of financing needs.
Other tech companies are boosting the demand for data center financing as well. OpenAI has formed a joint venture with Oracle and SoftBank to invest up to $500 billion in AI infrastructure across the U.S. in the next four years. Property developers are also increasingly building data centers, further fueling the financing demand.
Financing requirements are so large that companies are turning to different sources of funding. Debt is gaining in popularity: investment-grade borrowing by U.S. tech firms was up 70 percent year over year in the first half of 2025, according to Bloomberg, with Alphabet issuing bonds for the first time in five years in April 2025. Smaller or fast-growing firms, such as CoreWeave, are even turning to borrowing arrangements that use graphics processing units (GPUs) – specialized chips that accelerate AI computations – as collateral.
Debt securitization is growing, whereby data center-related borrowing is pooled and sold to investors in tranches, much like mortgages are. The data center-related debt securitization market, virtually negligible five years ago, has grown rapidly and is now valued at around $30 billion, according to AInvest.
Private capital is playing an important role too, with large private equity firms increasingly acting as direct lenders to businesses and infrastructure projects, alongside their traditional equity investments. In August 2025, Meta finalized a $29 billion deal for its Hyperion data center project including a debt portion of $26 billion led by PIMCO, the investment management firm.
Data center lending and investing carry additional risks beyond cost overruns. Overcapacity from rapid capital investment can leave assets underutilized. For example, in the late 1990s, U.S. telecom companies laid more than 80 million miles of fiber optic cables across the country after overestimating future demand. Prices plummeted and many companies entered bankruptcy proceedings.
Technology risk is also substantial. Much of the current spending goes to data centers built to train powerful AI models, but as demand shifts toward running those models, the need for computing power could drop, lowering the value of these assets. Newer, higher-performance chips could make older facilities less useful, and some may even require innovative cooling systems, leaving existing data centers obsolete.
Hyperscalers are diversified enough to weather these challenges, in our view, though they now carry substantial infrastructure and capital commitments – they are no longer asset light. We believe smaller investors and lenders will need to be particularly vigilant.
Some observers are optimistic that progress in AI will be swift, raising hopes that Artificial General Intelligence (an AI model with human-like cognitive abilities) and even Artificial Superintellingence (an AI model with an intellectual scope beyond human intelligence) could be achieved within their lifetimes. On July 30, 2025, Mark Zuckerberg, CEO of Meta Platforms, perhaps driven by ambition and competitive positioning, stated that “developing superintelligence is now in sight.”
Such enthusiasm is understandable. The pace of AI progress has been remarkable. OpenAI’s GPT-2, released in 2019, could write coherent paragraphs but often lapsed into meaningless output, while by early 2023, the company’s GPT-4 model had advanced enough to pass the U.S. bar exam, scoring in the top 10 percent of test takers, as reported by Reuters. Remarkably, this leap in performance was achieved without changing the science behind AI, but rather just by feeding the model more data and using more powerful GPUs.
Yet others guard against too much enthusiasm. Rodney Brooks, robotics pioneer and former director of the MIT Computer Science and Artificial Intelligence Laboratory, and best known as the founder of the company that developed and supplied the search-and-rescue robots used at Ground Zero after the 9/11 attacks, offers a clear-eyed take on AI’s current hype.
In a February 2025 interview with Newsweek, he emphasized that while AI models can use language fluidly, they are essentially pattern recognizers, really good at spotting and repeating patterns in data. That, in his view, is not the same as truly understanding or thinking for themselves.
He believes change will come more slowly than is generally expected because rolling out new technology almost always runs into practical hurdles like cost, integration with other systems, and regulation. And because today’s AI models are pattern-matchers, they still need a great deal of careful oversight – they are far from being plug-and-play as the hype often suggests. Finally, he emphasized that corporate adoption of the new technology will be based on return on investment, not “glitziness.”
On that front, a July 2025 report from the Massachusetts Institute of Technology revealed that 95 percent of generative AI (GenAI) pilot programs in enterprises yielded no measurable return on investment, despite the $30 billion–$40 billion enterprise investment in GenAI so far. The authors conceded that over 80 percent of organizations surveyed have explored or piloted tools such as ChatGPT and Copilot, with 40 percent reporting to have purchased an official large language model subscription. Interestingly, they found that workers at more than 90 percent of companies surveyed say they use personal AI tools like chatbots for work, but the report highlighted that these tools seem to primarily enhance individual productivity, not profitability.
2025 has been a pivotal year for the AI industry, and 2026 will likely be as eventful. Some of the significant developments to watch in the next year include:
Beyond these developments, which will likely keep enthusiasm high, we believe investors should also keep on eye on whether such investments and the application of AI in business are generating adequate returns. As has almost always been the case in the past, the risk remains that investors may overestimate what the new technology can deliver in the short to medium term.
The material herein is for informational purposes only and is not directed at, nor intended for distribution to or use by, any person or entity in any country where such distribution or use would be contrary to law or regulation or which would subject Royal Bank of Canada or its subsidiaries or constituent business units (including RBC Wealth Management) to any licensing or registration requirement within such country.
This is not intended to be either a specific offer by any Royal Bank of Canada entity to sell or provide, or a specific invitation to apply for, any particular financial account, product or service. Royal Bank of Canada does not offer accounts, products or services in jurisdictions where it is not permitted to do so, and therefore the RBC Wealth Management business is not available in all countries or markets.
The information contained herein is general in nature and is not intended, and should not be construed, as professional advice or opinion provided to the user, nor as a recommendation of any particular approach. Nothing in this material constitutes legal, accounting or tax advice and you are advised to seek independent legal, tax and accounting advice prior to acting upon anything contained in this material. Interest rates, market conditions, tax and legal rules and other important factors which will be pertinent to your circumstances are subject to change. This material does not purport to be a complete statement of the approaches or steps that may be appropriate for the user, does not take into account the user’s specific investment objectives or risk tolerance and is not intended to be an invitation to effect a securities transaction or to otherwise participate in any investment service.
To the full extent permitted by law neither RBC Wealth Management nor any of its affiliates, nor any other person, accepts any liability whatsoever for any direct or consequential loss arising from any use of this document or the information contained herein. No matter contained in this material may be reproduced or copied by any means without the prior consent of RBC Wealth Management. RBC Wealth Management is the global brand name to describe the wealth management business of the Royal Bank of Canada and its affiliates and branches, including, RBC Investment Services (Asia) Limited, Royal Bank of Canada, Hong Kong Branch, and the Royal Bank of Canada, Singapore Branch. Additional information available upon request.
Royal Bank of Canada is duly established under the Bank Act (Canada), which provides limited liability for shareholders.
® Registered trademark of Royal Bank of Canada. Used under license. RBC Wealth Management is a registered trademark of Royal Bank of Canada. Used under license. Copyright © Royal Bank of Canada 2025. All rights reserved.
We want to talk about your financial future.