Who will win the coveted AI race? Will it be the US, which by all accounts is still way ahead, with its access to parallel processing chips like Blackwell Ultra designed by Nvidia or Advanced Micro Devices’ MI300 chip, as well as access to the 2nm chip-making prowess of Taiwan Semiconductor Manufacturing, backed by cutting-edge tools shipped by Dutch equipment giant ASML? Or, will it be China — the tortoise with access only to inferior 7nm chip equipment — that will plod along steadily and eventually pip past America?
Until a year ago, the US was considered five years ahead of China in AI with access to top-of-the-line chip design, chip equipment and manufacturing. Moreover, America’s large language AI model pioneers — such as ChatGPT creator OpenAI, Gemini maker Google, Claude producer Anthropic and Grok maker Elon Musk’s start-up xAI, which will spend over US$400 billion ($523 billion) on AI infrastructure this year and US$600 billion next year — are way ahead of Chinese tech firms Alibaba Group Holding, Baidu and others.
In January, Chinese AI start-up DeepSeek stunned the world when it released its open-sourced R1 large language model, tailored for reasoning and structured problem-solving, particularly for tasks like math and coding. What made DeepSeek so special was that, while its capabilities were close to OpenAI’s ChatGPT models, it was developed using far fewer resources — reportedly US$5.58 million compared with over half a billion spent on ChatGPT’s o1 model.
DeepSeek’s breakthrough notwithstanding, US large language models like ChatGPT, Gemini, Claude and Grok are still far superior to anything that Alibaba, Baidu, Tencent and others like Shanghai-based MiniMax AI are working on. Jefferies tech analyst Edison Lee believes that China’s AI model performance gap with its US peers continues to narrow. The performance of MiniMax M2, the best Chinese AI model, may currently only be 10% lower than that of the best US AI model, OpenAI’s GPT-5 according to tests done by Artificial Analysis, an independent AI research firm. Beijing-based AI start-up Moonshot AI released its Kimi 2 Thinking model with a performance score of 67, which even exceeded MiniMax. Only Google’s Gemini 3 Pro with a score of 73 and OpenAI’s GPT-5.1 with 71 are better.
See also: China warns of bubble risk in booming humanoid robotics industry
Moonshot AI, 36% owned by Alibaba, trained its model on Nvidia’s H800 chip, a downgraded version of the Nvidia H100 specially designed for China. The H800 chip has been replaced by a slightly downgraded Nvidia H20 chip specially designed for China. Beijing has urged its tech firms to shun the new chip and look for local alternatives. Lee notes that it is difficult to extrapolate the US-China AI model gap, given that US players continue to have access to the next generation of AI chips from Nvidia, Advanced Micro Devices and customised AI chip design firm Broadcom, thereby widening the gap with Chinese upstarts. That means a more significant focus for investors is whether these Chinese models could be monetised by developing applications that will be adopted by consumers and enterprises, Lee notes.
Alibaba, Baidu and start-ups like MiniMax and Moonshot AI want whatever AI chips they can get hold of from Nvidia and AMD, but China wants to make its own AI chips and drastically cut its dependence on the US. Ahead of his meeting with President Xi Jinping in Seoul late last month, US President Donald Trump had said he was willing to allow China to buy a slightly degraded version of Nvidia’s Blackwell chips, but the matter wasn’t even brought up at the summit. Trump believes that the export controls on chips introduced by the Biden administration in October 2022 have proved to be a massive “own goal” for the US. That is because they have triggered an overwhelming incentive for China to set up its own chip value chain while depriving US firms of a major customer.
Beijing and Washington have not yet agreed on AI chip sales to China. Xi reportedly didn’t want to be seen as weak by asking for degraded AI chips. The Chinese president believes that his country’s export control on rare earth elements gives it sufficient leverage with far-reaching, long-term impact on the chip supply chain, including defence, technology and automotive industries. He would rather have Chinese tech giants buy lower-end domestic chips from Huawei Technologies, which for the first time publicly outlined detailed long-term plans for its Ascend AI and Kunpeng server chips. Beijing is using Nvidia’s market access as leverage in trade negotiations to push Washington for wider access to advanced semiconductors. What China really wants is access to top-of-the-line chip-making tools from ASML, which will allow Huawei and others to make state-of-the-art chips in China.
See also: Tesla touts China ties as automakers scrutinise supply chain
While American proprietary or closed-source AI models like GPT, Gemini, Claude and Grok dominate in overall performance, China leads in performance for open-sourced models. US-based closed-source models have higher intelligence scores than Chinese open-source models. But that is a reflection of their technological edge, driven by access to superior data, larger compute resources, more advanced training and alignment pipelines, and proprietary inference techniques. Open-source models, on the other hand, offer flexibility, transparency, lower costs and ecosystem benefits. They can be fine-tuned and deployed locally without vendor lock-in while thriving communities and toolchains accelerate innovation and broaden adoption. Over the longer term, analysts expect Chinese AI models to slowly catch up with the US because some Chinese AI players will continue to conduct pre-training in data centres outside China, thereby avoiding chip restrictions. Already, US companies are using Chinese AI models. Airbnb’s CEO Brian Chesky recently revealed that his firm was relying on Alibaba’s Qwen model instead of OpenAI’s GPT because the connective tools were not “quite ready”. Did I mention that Chesky is a long-time close friend of OpenAI’s CEO, Sam Altman?
Chinese AI firms will reportedly spend US$80 billion or less than one-fifth of what US firms are spending on AI infrastructure in 2025. That’s not just because Beijing has told its AI players not to buy any more Nvidia chips and use their existing inventory of American AI chips or Chinese-made chips. China’s low AI capex is also due to higher model efficiency, rather than limited access to expensive advanced AI chips from the US. While US players like OpenAI and Meta are focused on model performance and artificial general intelligence (AGI) or AI that can perform any intellectual task a human can, Chinese open-source models are more focused on efficiency, which helps drive down token cost and, in turn, will help improve application development and user adoption.
American AI giants are still raking in huge profits. On Nov 19, Nvidia, whose stock had been under pressure for the past two weeks over valuation concerns, reported blowout earnings. The chip giant reported 62% higher revenue of US$57.01 billion for its fiscal third quarter ending October, compared with analysts’ estimates of US$54.92 billion for the quarter. Nvidia’s net income in the quarter rose 65% to US$31.91 billion, from US$19.31 billion. Nvidia’s growth should be seen in the context of its exports of advanced AI chips to China being completely halted. Excluding China, Nvidia’s sales to the rest of the world nearly doubled over the same quarter last year. CEO Jensen Huang said in late October that Nvidia has US$500 billion in orders, for 2025 and 2026 combined, for its AI chips.
“China is going to win the AI race,” Huang told the Financial Times’ Future of AI Summit on Nov 5. Hours later, he attempted to walk back a bit. “As I have long said, China is nanoseconds behind America in AI,” he said in a statement posted on X. “It’s vital that America wins by racing ahead and winning developers worldwide.” That was a subtle hint that the US should allow Nvidia to sell whatever chip it wants to China.
Energy is the ‘secret sauce’
For AI Compute, the US firms need power-hungry chips in the massive data centres that are being built at home and abroad. Data centres consume 4.4% of US energy generation today. That is expected to grow to over 12% by 2028 and 18% by 2040. The International Energy Agency believes the US should be adding 80 GW of new power generation capacity a year to meet demand. China, on the other hand, is investing twice as much as the US in power plants, storage, as well as the grid. Beijing’s energy subsidies make power more affordable for Chinese tech firms, allowing them to run local alternatives to Nvidia’s AI chips cheaply. Compared to rising energy costs in the US, power is ridiculously cheap in China, and that gives China a leg up in its competition with the US on AI supremacy. Energy costs are low in China due to its superior battery storage technology. Falling costs of battery storage and plummeting prices of photovoltaic cells have made solar more competitive than coal in China. Solar and wind power have long been thought of as intermittent energy, but when combined with cheaper storage, they are now seen as a regular source of energy like coal, LNG or indeed nuclear power.
Increasingly, US firms are pressing the Trump administration to address the country’s growing energy deficit. OpenAI last month urged the White House to close the “electron gap” by setting up an ambitious national target of building 100 GW a year of new energy capacity to “fuel American AI dominance”. Just to put it in perspective, 1 GW can power a major US city of several million people. William Thompson, an analyst at Barclays in New York, notes “whoever wins the energy race likely will win the AI race.” That is “increasing emphasis on energy needs and risks to AI spending, economic growth and American exceptionalism,” he notes.
Sink your teeth into in-depth insights from our contributors, and dive into financial and economic trends
China added 429 GW of new power capacity, or one-third of the entire US grid and over half of all global electricity growth. Indeed, China added 240 GW of solar generation capacity in the first nine months of 2025 — more than the entire installed solar capacity in the US of about 178 GW. While China’s total power demand increased by 170 TWh (or terawatt hours) in the first half of this year, power generation from solar and wind sources increased by 250 TWh.
Here is why cheap energy costs in China matter: Electricity required to generate the same amount of “tokens” from Chinese AI chips is between 30% and 50% higher than Nvidia’s H20 chips. Little wonder, then, that Beijing has begun encouraging local governments to incentivise Chinese tech giants with data centres by offering subsidies that slash their electricity bills by as much as 50% provided they are powered by domestic chips like those made by Huawei. Like the tortoise in Aesop’s Fables, China is catching up to the American hare in the AI race. Not moving on energy needs now would be akin to the hare napping as the tortoise plods on.
Assif Shameen is a technology and business writer based in North America
