The US, China, Europe are racing to achieve self-reliance in AI chips while the tech world witnesses a series of events amidst the GPU drought.
The AI generative explosion sweeps the tech sector in 2023. But behind the market excitement is an undercurrent focused on AI chips, especially GPUs. In AI training, graphics GPU chips have an advantage over CPUs thanks to their ability to perform massive parallel computations simultaneously. From the US-China tech war to the mystery behind the sudden firing of OpenAI CEO Sam Altman, the shadow of AI chips looms everywhere.
Thirst for AI chips
The explosion of AI generatives has made Nvidia the brightest star in hardware. Its H100 is currently the most powerful graphics GPU chip on the market, used to operate super AI machines. Jensen Huang, Nvidia’s CEO, described it as “the world’s first system designed for supersized AI.” During the fever, on the black market, each chip costs up to $46,000.
According to the research firm Omdia, in just one quarter, Nvidia sold half a million H100 chips. The chip thirst has also made Nvidia the world’s largest chip designer. The company’s market capitalization exceeded $1 trillion, higher than Netflix, Nike and Novo Nordisk combined.
The chip thirst has also prevented semiconductor manufacturers from sitting still. Despite facing a US ban, Huawei is believed to have successfully produced basic AI chips to fill the gap left by Nvidia. According to Reuters, by the end of October, Huawei delivered over 60% of the 1,600 Ascend 910B chip orders to replace Nvidia’s A100 on 200 servers for Baidu. Experts said 1,600 chips are not a large order, but a sign that the US no longer holds a monopoly and Huawei is being given a chance to conquer the $7 billion domestic market.
US actions and China’s response
After passing the CHIPS Act in August 2022, since early this year, the US has continuously strengthened the embargo to prevent China from benefiting from the $53 billion aid package. AI chips are a product of particular concern. Nvidia, the company currently dominating the GPU market share, is banned from selling advanced A100 and H100 chips to Chinese companies. Subsequently, lower-tier chips like the H800 and A800 were also tightened. The US move is seen as having dealt a heavy blow to China’s AI ambitions, where tech giants like Tencent, ByteDance, Baidu, Alibaba are accelerating the development of large language models (LLMs) and AI similar to ChatGPT.
Immediately after the ban, Chinese companies frantically hunted for and tried to stockpile AI chips. According to analysts, the US blocking Nvidia from selling chips could set China back in the AI race. In early August, SCMP cited supply chain sources that a number of major Chinese tech companies had quickly placed orders for about 100,000 A800 units worth over $1 billion from Nvidia, to be delivered this year. In addition, other orders totaling $4 billion will be delivered in 2024.
In the opposite direction, Chinese chip manufacturers are optimistic about the embargo and see it as an incentive to easily access the $41 billion fund. According to Reuters, the Big Fund is being quietly prepared by the Chinese government to achieve chip self-reliance, balancing the US’s $53 billion fund.
Europe also did not sit still when it approved a $47 billion semiconductor support package in April, with the goal of doubling the continent’s chip production market share to 20% by 2030. SCMP cited Thierry Breton, Commissioner for the Internal Market at the European Commission, that this is the EU’s latest effort to catch up with the US and Asia in terms of manufacturing capabilities.
A new race
Chips are not only at the center of the tech war between superpowers or between semiconductor manufacturers, they are also related to many undisclosed secrets of the tech world.
On November 17, OpenAI CEO Sam Altman was suddenly fired but reinstated after 5 days. In documents accessed by Wired, in 2019 OpenAI signed a $51 million deal to buy chips from Rain AI. The company is headquartered in San Francisco, less than a mile from OpenAI, and researches a type of neuromorphic processing unit (NPU) that mimics the human brain. Rain AI was expected to launch its first product in October the following year. Notably, Altman, in his personal capacity, also invested $1 million in the company.
For many years, Altman has often complained about the exorbitant costs of AI chips and predicted that artificial intelligence could create a “brutal crisis” around chips, while affirming that the pace of AI development will depend on new chip designs and supply chains.
2023 ends with a series of “battle” announcements from the world’s leading semiconductor manufacturers. AMD says its MI300X processor, part of the Instinct MI300 series dedicated to generative AI models, has specifications comparable to Nvidia’s most powerful H100. Last month, Microsoft unveiled the Azure Maia 100 chip to compete with Nvidia. On December 14, Intel unveiled the Gaudi3 chip dedicated to generative AI. Meanwhile, on November 13, Nvidia also announced the H200 model, expected to hit the market in Q2/2024, and is expected to “make a huge leap forward in performance, especially in the reasoning ability of large AI data models”.
Experts predict that AI chips will remain one of the focal points of the tech industry next year as generative AI continues to explode, while general AI (AGI) may soon become a reality as specialized GPUs are upgraded.