Generative AI apps, like ChatGPT and Sora, took the world by storm. Due to these programs, computational power demand increased considerably, signaling good news for silicon wafer manufacturing companies. After all, AI and computing devices all depend heavily on the semiconductor industry.
From advanced logic chips to CPUs and storage units, this complex hardware has one or several semiconductors to function. Because of that, growing demands push wafer manufacturers to innovate and produce faster and in mass.
In this article, we’ll explain why AI depends on wafers and what that means for the semiconductor industry.
Last year, generative AI soaked the entire world. The capabilities of these new artificial intelligence models were like nothing we had ever seen, opening business opportunities and optimizing processes throughout every industry.
At least 5% of businesses worldwide state they used generative AI for work in 2023, with the number rising to 14% in technology and media-related enterprises. In this environment, the demand for computational power also skyrocketed.
According to McKinsey, AI and machine learning could potentially generate between $35 billion and $40 billion in value annually within the next two to three years. Over a longer time frame, this figure could rise to between $85 billion and $95 billion per year.
Gen AI applications require substantial computing resources for many tasks, from training to inference. An increase in computing demand translates to a proportional increase in semiconductor demand, a basic component in today’s technology.
To understand AI’s demand for silicon wafers, we must explain their hardware requirements.
Gen AI applications typically run on dedicated servers and in data centers. These servers use high-performance graphics processing units (GPUs) or specialized AI chips, known as application-specific integrated circuits (ASICs), to manage the heavy workloads associated with Gen AI.
There are two basic processes AI conducts: training and inference. For training, the server architecture consists of two central processing units (CPUs) and eight GPUs. The infrastructure for inference workloads is similar.
In these processes, there are different types of chips and other components involved:
Logic chip demand depends on the type of Gen AI and the type of server for training and inference workloads. In 2030, McKinsey estimates that the demand for logic wafers from AI applications will be approximately 15 million wafers.
This will imply a supply gap of one million to about four million wafers compared to actual production numbers. Three to nine new fabs will be needed to meet that quota.
As generative models grow, companies need to expand their memory capacity. By 2030, the demand for wafers for memory chips will be between 13 and 1 million wafers, which translates to the need for between 12 and 18 new dedicated fabs.
NAND memory stores input and output, user data, and operating system data. Data servers dedicated to multimodal and video generation will drive the need for NAND.
Two to eight million wafers, or one to five fabs, are anticipated to be needed for NAND overall.
Gen AI needs high-bandwidth, low-latency connectivity for its servers and other components. To do that, a huge quantity of switches and network interfaces—mostly composed of copper—are needed.
However, optical connectivity is anticipated to become more common. The benefits of optical technology in communication include cost-effectiveness, high bandwidth, low power consumption, and fast speed.
Additionally, optical connectivity relies on silicon photonics. Which, as you can guess, works with silicon wafers.
AI servers consume a lot of electricity. They’re expected to use more than 10% of global electricity in 2030. For that, power semiconductor manufacturing will have to grow.
Power semiconductors are specialized transistors that play a key role in the world's electrical infrastructure.
There has been a severe global shortage of semiconductors since 2020, and there does not appear to be an end in sight. The COVID-19 pandemic exacerbated and highlighted all the vulnerabilities and weak points in the semiconductor supply chain.
That, coupled with a widespread need for semiconductors in modern electronics, led to an uncontrollable scarcity of semiconductors. Many industries are finding it difficult to keep up with the high demand from consumers.
AI is no exception. In particular, AI workloads are impacted by the scarcity of expensive chips like GPUs, which are needed for model training.
In response, industry leaders are allocating large sums of capital to grow semiconductor fabrication facilities and data centers. A good example is the CHIPS Act of the USA. Meanwhile, advancements in chip design, materials, and architecture paint a positive picture.
However, there are still many obstacles to overcome.
The answer is yes.
Generative AI made a big impact on society, with both advocates and critics emerging as a response.
However, mass AI adoption cannot take place without the proper physical hardware needed for these technologies to function. As of today, the AI revolution will be forced to wait until semiconductor and chip manufacturers can meet their needs.
This is something that will be very hard to achieve in a world where every tech-related industry is in dire need of semiconductors. Moreover, geopolitical conflicts put another stick in the wheel for AI.
With China being the biggest wafer producer worldwide and Taiwan an important (if not, the number one) chip manufacturer, growing tensions in the region pose a real severe risk for AI applications and their development.
The effects of this technology were felt throughout all areas and industries worldwide, and immense growth is accurately predicted. Nevertheless, wafer shortage directly affects the semiconductor industry’s ability to meet the growing demand for computer power driven by AI and Gen AI applications.
If you’re interested in learning more, give us a call!
