Outra semana se passou, e a incerteza continua em relação à exportação dos chips avançados de IA da Nvidia para a China. Proponents of maintaining the controls argue that these chips would help build Chinese military systems that threaten the US and its allies. Eles também afirmam que os controles são necessários para manter e ampliar a liderança americana no mercado de serviços de IA.

But this is wrong. These arguments assume that China would not be able to advance in AI without access to these advanced chips — which is not true.

Also read: China’s leap: new technological era grows even with record real estate crisis

Chips avançados de IA simplesmente reduzem o custo da inteligência artificial. Os modelos de ponta atuais exigem um grande número de chips para serem treinados e operados. An advanced chip has superior performance; por isso, são necessários menos chips para atingir o mesmo nível de performance.

China’s decision to make its AI models open source, in particular, allows it to take advantage of the best in software and algorithms to cut costs. Além disso, os chips de IA representam só uma parte do custo total.

AI-based systems involve numerous other expenses — engineering, data, software and licenses, regulation, energy and infrastructure — in which China has significant advantages.

Chips avançados também reduzem o custo de energia da IA. These chips are made using the latest technology from TSMC (and sometimes Samsung) — each new generation is more efficient than the last.

High energy consumption increases the cost and slows down the pace of deployment, as it is difficult to guarantee rapid access to large amounts of energy, especially in the US.

However, China is expanding its energy supply much faster than the US and, therefore, has a better chance of meeting the energy demands of its AI data centers, even though they consume more energy because they do not have access to advanced chips.

High consumption also increases the carbon footprint, but this should not limit Chinese ambitions in technologies it considers strategic.

Além disso, muitas aplicações de IA não precisam de chips avançados. Various applications in network security, facial recognition, medical image analysis, advanced driver assistance systems, logistics and robotics can be operated with much simpler models than the cutting-edge ones.

Esses modelos podem ser treinados e executados em chips que a própria China consegue produzir. The country intends to dominate these areas. Even in more complex applications, recent research suggests that cutting-edge models can be replaced with a much simpler set — and that set doesn’t require advanced chips to build or operate. Portanto, não está claro se a China vai ficar para trás também nessas áreas.

There is also the possibility that China will learn to produce its own advanced chips — after all, the country has already invested heavily in technologies with the potential to surpass the current state of the art.

No geral, a China consegue mitigar significativamente as desvantagens de não ter acesso aos chips avançados de IA. Furthermore, it will be willing to absorb higher upfront costs, especially in AI technologies for military and strategic purposes, because it knows it can reduce downstream costs thanks to the scale and strength of its manufacturing.

It is not surprising that China continues to produce competitive cutting-edge models and dominate AI-based applications such as robotics and autonomous vehicles, despite controls imposed in recent years.

The argument in favor of controls may even seem reasonable — why not take advantage of the chance to raise AI development costs in China, even if only a little, if it would have no impact on the US? But the costs are significant.

A China poderia ter sido um dos maiores mercados para as empresas americanas de chips avançados. The US lost that market. Furthermore, the controls turned the issue into a matter of national pride for the Chinese and triggered a wave of investment in a domestic AI chip ecosystem.

Diversificar, fortalecer e proteger as cadeias de suprimentos de IA. Trabalhar com aliados para liderar o desenvolvimento de padrões e práticas internacionais. Reduce the cost of AI (through selective open source initiatives or public-private partnerships, for example) to ensure that American AI — and its values ​​— are the most prevalent in the world.

And prioritize highly complex applications aimed at companies, where the US competitive advantage is greater compared to a fast competitor with a lot of talent, resources and cost and speed advantages.