Artificial intelligence is the new steel. The “DeepSeek Moment” is redefining global competition

The following contribution presents the author’s analysis in the context of 15 April 2026. References to upcoming models and market valuations should be read in light of that date. A final editorial note distinguishes subsequent developments and clarifications from the original content.

The Year of the Fire Horse is taking shape as a race between two profoundly different AI models. On one side of the Pacific, the big US technology companies are pouring huge amounts of capital into advanced chips and into the development of ever more powerful models, in pursuit of Artificial General Intelligence — sometimes provocatively described as “God-in-a-Box”, in other words an extremely autonomous and capable superintelligence. A pursuit that is driving up costs and putting corporate balance sheets under pressure. On the other side of the ocean, China is aiming to make AI inference as accessible and low-cost as possible, accelerating its adoption on a large scale.

These divergent strategies reflect both cultural factors and structural differences. Chinese businesses have always been reluctant to pay for software licences, preferring free, open-source or digitally integrated solutions. The SaaS (Software as a Service) industry in the country has long suffered from a monetization problem and, even today, the sector’s overall revenues are about one-tenth of those in the United States. As a result, direct monetization of Large Language Models (LLMs) through retail and corporate subscriptions is much more difficult in China, forcing developers to favour the open-source approach.

Even more decisive are the structural differences between the two economies. If the United States has the most advanced chips, China can rely on a much larger installed electricity-generation capacity. While Beijing is unable to produce cutting-edge semiconductors, because it is excluded from the main global suppliers of advanced chipmaking equipment, Washington has to contend with a significant energy bottleneck that makes AI expansion complex and costly.

Whoever wins the AI race will depend in part on who first overcomes their own structural limits: energy for the United States, chips for China. In January 2025 DeepSeek drew global attention with the launch of R1, a model designed for advanced reasoning tasks, able to come close to the performance of the best OpenAI system of the time, namely “o1”, but at drastically lower cost. R1 showed that open-weight, cost-efficient models can sit close to the technological frontier and be suitable for most commercial applications. This development challenged the idea that a handful of large US labs could dominate the sector.

A year later, China still lacks the most advanced US chips. Chinese models operate under strict regulatory constraints, avoiding politically sensitive content, and DeepSeek, like other developers, has been accused of using US models to train its own.

Yet external constraints often become a catalyst for innovation. The lesson of the “DeepSeek shock” is not that ambitious startups take shortcuts, but that algorithmic progress can significantly offset hardware limits. And what if the “DeepSeek moment” were not just a technical advance, but the signal that the ecosystem has reached critical mass? The growth in scientific output in artificial intelligence points precisely to the expansion of China’s research and innovation ecosystem in recent years.

China’s 15th Five-Year Plan (2026–2030) identifies “embedded intelligence” as the country’s strategic goal: the priority will be to integrate artificial intelligence as quickly as possible into the widest possible range of products and services. In this perspective, AI models are conceived as national infrastructure, a basic commodity destined to serve as an enabling platform for the entire economy, much like electricity or steel. In 2025 China added about half a terawatt of new power capacity, a measure of the rapid expansion of the country’s generation capability. Through the “East Data West Computing” initiative, digital workloads are being shifted from the more developed eastern coast to the energy-rich western regions, via high-capacity data and electricity transmission infrastructure.

To encourage adoption and ensure economic sustainability, local governments have introduced subsidies for the use of computational tokens and cloud storage, with the explicit aim of buying time while the domestic semiconductor industry closes the energy-efficiency gap with Nvidia’s most advanced GPUs. At the same time, the central government is sending clear signals that domestic companies should favour national hardware. Meanwhile, Chinese labs have established themselves as particularly effective fast followers, developing algorithmic solutions capable of reducing computational requirements while maintaining nearly comparable performance.

After the release of Opus 4.6 by Anthropic, Zhipu AI unveiled a model with performance comparable to top-tier systems, aiming to cut inference costs significantly. After Google’s announcement of Genie 3, ByteDance launched Seedance 2, capable of turning a storyboard into a full cinematic preview. Meanwhile, DeepSeek is preparing to release the V4 model, which should reduce inference costs further. The logic of open source and large-scale adoption is that broader diffusion improves hardware economics, enriches available data, multiplies use cases and stimulates further algorithmic innovation, in a virtuous cycle capable of producing ever more powerful models.

Accelerating adoption does not simply mean saturating existing capacity, but also speeding up technological progress. If the new generation of AI-based systems succeeds in automating production lines and robotic systems or in carrying out pharmaceutical testing, productivity gains would more than offset the initial investments, making Chinese products and services even more competitive in global markets. After all, a generation ago the world economy became dependent on low-cost Chinese manufacturing. What if tomorrow it became dependent on Chinese data and artificial intelligence?

It may be tempting to think that the current geopolitical environment makes the Chinese approach unsustainable, limiting the potential for these technologies to expand internationally. The property crisis continues to act as a significant drag on economic growth and the challenges facing China’s AI strategy are real: decision-making is far from perfect and the limited transparency regarding the level of state support for training and inference activities makes any linear economic assessment extremely difficult. Yet, from the investor’s point of view, even a scenario of only partial success appears to be far from fully priced in. According to Bloomberg data cited in the April analysis, the MSCI China index was trading at around 12 times expected 2026 earnings, versus 22 times for the S&P 500, against broadly comparable expected growth rates in the mid-to-high double digits.

If China’s AI bet were to work, the country’s productivity could reaccelerate significantly over the next decade. The beneficiaries will not necessarily be the AI infrastructure providers or model developers. The government’s aim is for the benefits to spread across the economy as a whole and to extend to a much wider range of sectors, such as manufacturing and healthcare. In the 2000s, the state subsidized the production of critical resources such as steel, enabling national champions to dominate sectors including shipbuilding, logistics and advanced manufacturing. Today, the real value of Chinese AI could be captured downstream, by companies that use low-cost inputs to further strengthen their competitive advantages.

Editorial note of 14 October 2026

The reference to DeepSeek V4 as a model still in preparation belongs to the April framework: DeepSeek announced V4 Preview on 24 April 2026 and V4.1-Flash on 10 September. These releases show that the development path has continued; however, the performance and efficiency indications released by the producer do not on their own demonstrate an increase in Chinese productivity or investment returns. The author’s thesis remains a scenario to be assessed through actual adoption, total costs and the results of user companies.

On the energy front, China’s National Energy Administration reports a total installed generation capacity of 3.89 terawatts at the end of 2025, up 16.1%. This is installed capacity, not electricity actually produced or constantly available to data centres. The original comparison with “50% of the entire US grid” has been replaced by a more cautious formulation: without a homogeneous statistical perimeter, it is not a directly usable comparison. Grid connections, supply continuity and the location of loads remain separate issues from the size of national capacity.

The 12 and 22 times expected earnings multiples are those cited by the author in the April framework and not October updated quotes. The comparison between models also requires caution: benchmark performance, API prices and total development cost do not measure the same thing. Genie and Seedance, mentioned in the piece, also belong to products with different purposes, and the sequence of their announcements does not constitute a direct technical comparison. These clarifications do not alter the central hypothesis: some of the value may emerge in companies that adopt AI, but economic success does not automatically follow the release of new models.

DeepSeek — V4 Preview announcement of 24 April 2026

DeepSeek — V4.1-Flash announcement of 10 September 2026

National Energy Administration — 2025 electricity statistics

Original disclaimer of the contribution

This analysis relates to Plenisfer Investments SGR S.p.A. (“Plenisfer Investments”) and is not a marketing communication relating to a Fund, an investment product or investment services in your country. This document does not constitute an offer or invitation to sell or purchase securities or any asset or business described here and does not form the basis of any contract. Any opinions or forecasts provided are current as of the specified date, may change without notice, do not predict future results and do not constitute a recommendation or an offer of any investment product or service. Past performance does not predict future returns. No guarantee can be given that an investment objective will be achieved or that there will be a return on capital. This analysis is addressed exclusively to professional investors in Italy pursuant to Directive 2014/65/EU on markets in financial instruments (MiFID). It is not intended for retail investors or US Persons, as defined in Regulation S of the United States Securities Act of 1933, as amended. The information is provided by Plenisfer Investments, authorised as an OICVM management company in Italy, regulated by the Bank of Italy - Via Niccolò Machiavelli 4, Trieste, 34132, Italy - CM: 15404 - LEI: 984500E9CB9BBCE3E272. All data used in this analysis, unless otherwise indicated, are provided by Plenisfer Investments. This material and its contents may not be reproduced or distributed, in whole or in part, without the express written consent of Plenisfer Investments.