China’s recipe for tech leadership has long missed a main ingredient: chips. It isn’t for lack of trying. Since the launch of the “531 Development Plan” in 1986 to the formation of the $50 billion integrated circuits fund in 2014, considerable state resources have been poured into China’s semiconductor aspirations.
But there has been little to show for it. China today still buys 90% of its chips each year (~$200 billion) from foreign companies. To a large degree, China has lost the battle on existing commodity chips, such as memory chips and mobile processors, as American and South Korean firms hold a strong grip across virtually the entire production chain. As Chair of the China Computer Federation has put it, “In the realm of traditional chips, the game is over.” Indeed, Chinese companies are nowhere to be seen among top global chipmakers (see Figure 1).
Figure 1. Market Share of Top Semiconductor Suppliers (2008 – 2018)Source: Gartner.
But the Fourth Industrial Revolution—driven in large part by artificial intelligence (AI) technologies—has given a second life to China’s fledgling semiconductor industry. If AI’s application across industries will define the future, then chips with AI capabilities will be in high demand.
China’s bet on AI chips makes both technological and market sense, as it is viewed as a clean slate where Chinese companies have a shot at “overtaking competitors around the bend” (弯道超车). Traditional chips have powered everything from smartphones and laptops to televisions, but are starting to see their growth plateau as the consumer electronics market becomes saturated. AI chips, while making up just 1% of global semiconductor sales, are poised to experience robust growth over the next decade (see Figure 2).
Figure 2. Growth of AI Chips Market, 2017-2027 (in $ billion)*projected values of AI chip sales.
Source: The Insight Partners.
AI chips are built specifically to handle complex machine learning (ML) algorithms, particularly for neural networks. Just like a human brain, a neural network has countless nodes (neurons) across several layers that process an inordinate amount of data. The end result is essentially pattern recognition: a distinctive human trait. But ML requires a level of computational power beyond what current central processing units (CPUs) can handle. Take Google Brain. In 2010, Google was training AI systems to recognize photos of cats. Yet even that simple task required running 16,000 CPUs, which couldn’t be realistically deployed.
So a new type of chip was designed to specifically handle ML tasks. These specialty chips, often called AI accelerators, can divide up the processing and perform parallel computations, making them much faster and efficient than normal CPUs. The AI chips are then integrated with the CPU on a single set, which is typically called system-on-a-chip (SoC). One example of a SoC is Apple’s A13 Bionic in the latest model of the iPhone, which features an AI “neural engine” that processes facial recognition.
Betting on ASICs
To be sure, AI chips come in different forms, including graphics processing unit (GPU), field-programmable gate array (FPGA), and application-specific integrated circuit (ASIC), among others. But it is the ASIC that offers China a shot to end its woeful record on manufacturing semiconductors.
When NVIDIA launched the first GPU chip GeForce 256 in 1999, it was originally meant to be used for processing graphics for increasingly complex computer games. But that all changed in 2010, when Nvidia made a splash by achieving roughly the same performance as 16,000 CPUs, but with only 48 of its GPUs, as part of the aforementioned Google cat photo challenge.
Today, the global GPU market is practically an oligopoly dominated by American firms like Nvidia, Intel, and AMD (see Figure 3). When it comes to FPGA chips, which were originally used in satellites, US firms Xilinx and Intel became major players holding 70% of China’s FPGA market and more than 10,000 patents.
Figure 3. Market Share of GPU Chips (2013 – 2019) Source: Jon Peddie Research.
Since the entry barrier for those chips is so high, ASICs offer a relative “white space” in which China hopes to compete, for several reasons:
- Efficiency gains
ASIC chips are custom designed and optimized to handle a specific set of functions, so they tend to be significantly superior to general purpose chips in terms of efficiency. Some estimate that ASIC chips could be as much as 10 times more efficient than GPUs. Therefore, ASIC is viewed as the future and is projected to account for 70% of market demand, according to McKinsey.
- Low production cost
Custom designs require significant upfront costs, and those designs often require tweaking. But since the chip is designed for narrow functions rather diverse use cases, once the design is complete, its manufacturing cost can be a few cents per chip, much lower than that of FPGA or GPU.
Since all ASIC chips are custom-made, one company making ASIC chips does not preclude another from making their own. This also means there will likely be more push for vertical integration so that software and hardware makers can have their own chips made for their specific devices and functions.
In fact, Google, as well as more than a dozen Chinese startups, have started designing their own ASIC chips to improve the performance of their data centers and products, respectively. As production volume goes up, ASICs can achieve leverage the China market’s advantage in economies of scale.
- Market potential
AI-enabled chips will likely be found in a host of edge devices, such as smartphones, laptops, barcode scanners, cameras, sensors. This means the total addressable market for ASIC chips is significant. Although GPUs still make up the bulk of the AI chips market today, by 2025 more than half of these are projected to be ASIC.
AI technologies are poised to deliver additional global economic output of around $13 trillion by 2030. To ride the wave of the fourth industrial revolution, China seems to believe that the ASIC chip affords a unique opportunity to rid the chip on its shoulder.
Joy Dantong Ma is Associate Director of MacroPolo. You can find her work on AI, trade, and investment here.
Get Our Stuff
Get on our mailing list to keep up with our analysis and new products.Subscribe