Digitalisation, led by the proliferation of computing power and greater internet connectivity, affects every corner of our lives. Importantly, digitalisation is transforming how we consume data, work and interact with others.
There is a universal acknowledgement that recent advancements in generative AI have been achieved thanks to the extraordinary innovations in the semiconductor industry.
Indeed, while neural networks and other AI models have been around for many decades, at least in concept, they only became functional recently because the hardware finally has the right amount of computational, memory and connecting power to train a neural network.
Interestingly, training these models is so intense that data centres, which are rooms filled with thousands of interconnected servers and racks, have become the main computers that can handle AI workflows.
For example, GPT 4.0 (the evolution of the model behind ChatGPT) has 1 trillion parameters, which require four terabytes of memory to store the model weights.
According to Semi-Analysis research, the parameters are trained on many petabytes of data, which move in and out of the model and require over 16,000 graphic processing units (GPUs) to be trained.
See also: US equities, IG, fixed income strategies, gold and copper among top investment picks: UBS
It is no surprise then that companies in the semiconductor industry are considered key enablers of AI, as they build extremely complex solutions to address the seemingly infinite demand of compute horsepower and memory space.
In the future, as AI products and solutions become more pervasive, complex and tailored to specific use cases, there will be a greater need for more powerful and efficient chips.
This means companies in the semiconductor value chain will need to innovate and deliver even better solutions. Failing to do so might constrain the sustainability of the AI industry.
See also: With Trump win boosting stocks, investors hunt for next winners
The development of more efficient chips, a shift towards a new computing architecture, parallel processing units and the need to upgrade the data centre infrastructure are key drivers of growth in the semiconductor industry.
The Moore’s Law debate
There has been a debate on the future of Moore’s Law, which theorised that the number of transistors on a chip would double every 18 to 24 months as the transistors become smaller and the production and energy consumption per component (like transistors) become cheaper.
Despite scepticism, major chip manufacturers still have plans to innovate and shrink the size of transistors, with some key players planning to start mass production of smaller 2-nanometre technology in 2025.
Innovations keeping Moore’s Law alive include advanced lithography techniques, multi-patterning and new structures and materials.
However, the law is slowing down and becoming economically more challenging due to the increasing complexity and cost of transitioning from one process node to another.
As Moore’s Law slows, the industry is adopting new growth drivers to improve chip performance. These include increasing chip die size, adopting advanced packaging solutions and shifting towards parallel processing units.
For more stories about where money flows, click here for Capital Section
The demand for computing power is increasing, especially for complex AI systems. This has led to the emergence of Huang’s Law, which states that the performance gains of GPUs more than double every two years. Synergies between hardware and software drive this law.
On top of leading-edge hardware, software is equally important in delivering better computing power. Indeed, parallel processing would not be possible without the appropriate software, which divides complex tasks into blocks, distributes them among different cores and finally reassembles the output in a precise order.
Servers must be equipped with cutting-edge chips and hardware to handle AI workloads. The shift towards parallel processing units, which can simultaneously handle separate parts of an overall task, is expected to drive a strong demand for cutting-edge chips such as GPUs.
The acceleration of GPUs has outgrown that of central processing units (CPUs) due to more AI training and penetration into inference with an additional boost from higher selling prices as they leverage more advanced semiconductor process nodes.
Overall, we are just at the beginning of an infrastructure upgrade cycle that has the potential to last for many years, creating a strong demand for leading-edge chips such as GPUs.
Who is buying?
Hyperscalers, which are large cloud service providers, have built huge data centres to provide their services and are among the main buyers of semiconductors and AI chips.
Based on industry GPU data, we estimate that around 50% of AI chips have been purchased by hyperscalers, with the remaining half purchased by consumer internet companies, enterprises and, to a lesser extent, government agencies and supercomputing labs.
Impact on semiconductor demand
Demand for semiconductors saw a slowdown in 2022 and a downturn in 2023 but is expected to reaccelerate in 2024. One of the key drivers of accelerating demand is expected to be the rise of generative AI, which should drive an investment cycle by cloud providers and enterprises, with GPUs expected to benefit the most, followed by networking solutions and memory.
However, AI is not the only positive driver of future demand, as the major semiconductor end markets are also improving. Indeed, we should see a rebound in the personal computer and smartphone markets as shipping rates normalise post-Covid-19 and the subsequent air pocket. Semiconductor demand from these markets and strong demand for AI chips should support a recovery in the semiconductor industry.
As outlined above, growth in AI is generating tremendous demand for semiconductors, especially for leading-edge chips. According to Gartner, chips dedicated to AI workloads were worth around US$44 billion in 2022, corresponding to 7.4% of total semiconductor sales. Revenues from AI chips are expected to improve to US$53 billion ($72 billion) in 2023, US$67 billion this year and US$119 billion in 2027 when they will reach almost 16% of the total semiconductor markets. Sales of AI chips are expected to grow at around 22% per annum for the next five years, thus outpacing the sales of other semiconductors, which should grow at around 5% per annum.
Semiconductor innovations
The ability of semiconductor companies to continue to innovate in line with, if not beyond, what has been predicted by Moore’s Law has been paramount for developing the AI industry.
Among these recent innovations, the advent of parallel processing units and the application of GPUs for the training and inference of AI models are the most important. This new computing architecture, which brings innovation to the whole hardware, software and networking solutions ecosystem, speeds up the computational process so that training a model with trillions of parameters is no longer an impossible challenge.
While we remain confident about the ability of semiconductor companies to innovate, we expect that the strong demand from AI applications will continue to outpace what semiconductors can supply to the market, at least in the short-to-medium term.
According to Nvidia CEO Jensen Huang, around US$1 trillion worth of infrastructure needs to be upgraded with AI-related chips to host AI workloads. Unfortunately, nobody knows the real size of this market nor how long it might last. That said, we remain confident that AI chips will play a crucial role in recovering semiconductor sales as AI becomes a key end market for the industry.
To conclude, we reiterate our constructive view on the Cloud Computing theme. We have observed earnings and price momentum in the driving seat with no signs of a deceleration.
We continue to monitor earning results and bottom-up observations for any signs of fatigue or deceleration in demand, which might call for a change in our rating. So far, this is not the case, and we will stay invested and ride the momentum.
Luca Menozzi is next generation research analyst at Julius Baer
