Prediction 1: If China remains excluded from the West, they will continue to improve their parallel value chain in this era of state-sponsored capitalism.

As the United States and China engage in a high-stakes competition for dominance in cloud computing and AI, the global AI ecosystem is being reshaped by the intricate web of alliances, rivalries, and innovations that are emerging. 

Since 7 October 2022, only a select handful of US-allied countries have been able to import sensitive technologies, including high-performance computing (HPC) chips, without significant restrictions. China’s exclusion may not diminish tensions but could lead to a higher equilibrium in terms of AI infrastructure capacity in this multipolar world. At the extreme, China might adopt a more aggressive stance, targeting AI’s fault line, their neighbour Taiwan. 

US restrictions on chip exports to China create a dilemma, as they aim to curb China’s AI advancement while avoiding lost sales. However, loopholes and smuggling have allowed sensitive technologies to reach China, complicating the situation. This cat-and-mouse game has not slowed China’s AI ambitions, as evident in the emergence of DeepSeek, a Chinese company offering a robust model suite at a negligible cost to end users, which is now sparking interest and raising important questions about the future of AI development. 

The dynamics of AI’s complex value chain are integral to the global economy. In the following, we will highlight several trends worth noting.

Prediction 2: The evolution toward reasoning-based models marks a pivotal moment in AI; consequently, the demand for compute in the inference process is expected to increase moving forward.

Today, most compute remains allocated to training, from pretraining on vast datasets to capture patterns and post-training refinements. As Large Language Models (LLMs) become increasingly accessible and affordable, the focus shifts to optimising their underlying infrastructure to support efficient reasoning and inference.

With new technologies such as autonomous vehicles and smart homes on the rise, the demand for bespoke AI accelerators with high-performance inference will rise in parallel. Agentic AI is expected to emerge within the next 18 months, further increasing demand for inference compute as these systems operate more autonomously.

Prediction 3: Hyperscalers (large cloud providers) are shifting from the Ferraris to the Toyotas of chips.

The rapid expansion of AI-powered applications is fueling a sharp increase in demand for specialised computing hardware. Custom silicon solutions—such as application-specific integrated circuits (ASICs) and optimised graphics processing units (GPUs)—are becoming essential for handling AI workloads more efficiently.

To meet this demand, hyperscalers are increasingly designing and producing their own custom chips, often in partnership with top semiconductor manufacturers. By 2028, these tailor-made chips could make up more than 25% of the total chip market for hyperscalers, reflecting a major shift toward in-house innovation.

Prediction 4: The current level of computing power and quality of models will be the worst we will ever have.

As technology advances, we anticipate continued growth in data centre capacity across various segments, including generative AI, hyperscalers, cloud, and non-hyperscale facilities. The business model, which often involves real estate companies leasing data centres to tech firms, remains profitable due to high utilisation rates. 

Moreover, hyperscalers and LLM operators require substantial clusters to minimise latency and ensure reliable power distribution, ideal for training purposes. Although smaller data centres can be effective for inference tasks, the industry is poised for expansion, with the current capacity likely being the lowest in the next decade. To support this growth, hyperscalers are investing heavily, with $335 billion in capital expenditures projected for this year, nearly half of which will be dedicated to AI-related initiatives. 

Prediction 5: The future of AI holds promise, with breakthroughs in model architecture unfolding at a remarkable pace. This is leading to a commoditisation, and freemium offerings are becoming the norm.

In a highly complex value chain that spans from energy, data connectivity, and deep-sea cable systems to the silicon industry and semiconductor manufacturers, we see opportunities for investors amongst others in the following areas: strategic AI adopters, data centre services & infrastructure, cloud providers & helpers, computing power units as well as memory and storage.

While the long-term growth potential remains strong, short-term uncertainty, driven by macroeconomic and geopolitical factors, is currently clouding the outlook for the theme.