The grip of artificial intelligence (AI) has continued to hold, with strong competition to develop more powerful AI models, active innovation in the semiconductor industry, and continued investment in the high-performance computing (HPC) data centre market.
AI has been a hot topic, particularly since generative AI’s (GenAI’s) introduction with large language models (LLMs), writes Jameel Rogers, technology analyst at IDTechEx.
In today’s landscape, there is an ever-growing plethora of generative AI applications, not limited to generating images, videos, code, and data. Training the AI models for these applications hinges on expanding HPC infrastructure, cutting-edge HPC compute innovations, and strong investment to support this.
Subsequently, this broadens the opportunity for hardware supply chain players, putting them in a position to capture some of the $581-billion market share that IDTechEx has forecasted the HPC market to reach by 2035, in its latest report, “Hardware for HPC and AI 2025-2035: Technologies, Markets, Forecasts”.
$500bn Stargate project
Bold investment plans in the US have dominated headlines since 21 January 2025, with President Trump announcing plans for a $500-billion project backed by OpenAI, SoftBank, Oracle, and MGX to build out the US’s AI infrastructure.
Under the newly found company “Stargate”, the joint venture is committing an initial $100-billion to immediately build out data centres in Texas, with a further $400-billion promised in the upcoming four years.
Some critics have questioned how solid the grounds of these investments are; however, this bullish attitude toward scaling AI infrastructure strikes confidence in technology providers.
AI chip giants revenue growth driven by HPC, AI
AI chip giant Nvidia will be a large facilitator for this data centre build-out.
Nvidia has been the supplier of choice for graphics processing units (GPUs), for HPC and AI workloads, as well as supplying leading interconnect, networking, CPU, and other data centre infrastructure.
Of Nvidia’s quarterly revenue (Q3 FY2025), 87,7% came from data centre sales, at $30,8-billion. This is an impressive 112% year-on-year growth.
AMD has also entered the room in recent years, with growing sales of its EPYC CPU line and various GPU lines. Making up 52% of its quarterly revenue in the same period, it has amassed $3,5-billion in revenue in its data centre segment.
More specifically, both Nvidia and AMD deliver hardware specifically for HPC and AI workloads, such as Nvidia’s Hopper GPUs (H100) and upcoming Blackwell GPUs (B200), and AMD’s Radeon Instinct GPUs, such as the latest MI300 series.
IDTechEx forecasts a six-fold increase in yearly unit sales for HPC and AI GPUs between 2025 and 2035. The main markets for these GPUs are supercomputers, enterprises, and cloud providers.
HPC GPUs, custom chips makes up hyperscalar data centres
Cloud providers are a large market for HPC and AI hardware. Of Nvidia’s recent $30,8-billion quarterly data centre revenue, half of this came from cloud providers.
Cloud providers have built out distributed HPC and AI infrastructure in the form of hyperscaled data centers, the largest being Amazon Web Services (AWS), Microsoft Azure, Google Cloud Platform (GCP), and Alibaba Cloud.
These services are delivered through GPU instances, where hyperscalars give customers access to Nvidia Hopper, Ampere, and Volta GPUs, as well as AMD Radeon Instinct GPUs – all examples of HPC and AI hardware.
However, in the interest of optimisation, cloud providers have also begun building custom in-house application-specific integrated circuits (ASICs). These are domain-specific, making the chips efficient for specific workloads.
Google set the trend with its tensor processing units (TPUs), and others have followed, like AWS with its Trainium and Inferentia chips.
But, perhaps more importantly, this gives companies the power to have full-stack control by developing their software, hardware, and frameworks in-house. This allows these companies the power to tailor AI infrastructure to their desired needs and takes reliance away from third-party providers.
However, extensive resources and market shares are required, which are available to a few companies. Perhaps this has been a warning that general-purpose compute, such as GPUs, might not be the optimal solution.
DeepSeek shocks investors
Following the development of the largest AI models, a massive wake-up call has been delivered to the US in the form of Chinese start-up DeepSeek’s new R1 AI model.
R1 is a reasoning model that is competitive with OpenAI’s o1 model but has allegedly been delivered at a fraction of the cost and uses less advanced hardware due to US-imposed AI chip trade restrictions.
Training costs of $5,58-million are a stark difference compared to the more than $40-million required for GPT-4 training, especially when using Nvidia H800 GPUs, which are a version of the Nvidia H100 with half the bandwidth.
DeepSeek has reportedly done this through various software optimisations, including some that have curbed limitations to memory bandwidth.
This has raised massive skepticism in investors as to why US firms have been pledging billions of dollars into massive data center build-outs when a leading AI model can be built on much less HPC and AI infrastructure.
Suddenly, hardware markets are in a position where buying the latest GPUs with the best interchip bandwidth is no longer the only route to realizing AI demands, as weaker hardware can also produce remarkable results.
With artificial general intelligence (AGI) potentially on the horizon, and governments putting their hands into the mix, will investing in efficiency or mass scale-out be prevalent?
The answer is likely a combination of both. On the flip side, the longer-term picture looks promising – what could these efficient AI models achieve using the latest hardware, such as the GB200 chips from Nvidia? In comparison to ASICs, like Google’s TPUs, general-purpose GPUs may have been fed a lifeline, as suddenly, with potential efficiency gains in GPUs, ASICs could potentially lose their edge.
Either way, Jevon’s paradox will win in the long term – with efficiency improvements comes higher usage – if AI can be built for cheaper, and more intelligence can be reached with less, this is a good thing.
This competition will only ignite the demand for more compute, and ultimately cutting-edge AI hardware will remain a pillar for the future growth of AI.