Investing.com – The trajectory of the AI semiconductor ecosystem has an evolving landscape, driven by growing demand for the computational power needed to fuel AI developments.
According to analysts at Barclays, the sector is at a critical juncture as the global appetite for AI-enabled solutions, especially large language models, continues to outpace current chip supply and performance.
The selling of AI chip names, such as NVIDIA (NASDAQ:), following earnings reports, has raised concerns about whether the market has peaked.
However, Barclays stresses that the future of the industry is still full of growth, driven by the ever-increasing computational needs of AI models.
Barclays points out that the AI semiconductor ecosystem is in the early stages of ramping up, and this period is characterized by significant supply constraints.
Expectations indicate that the computational resources required to train the next generation of master’s degree holders, some of which amount to 50 trillion teachers, are enormous.
The brokerage estimates that by 2027, approximately 20 million chips will be needed just to train these models. This number highlights the stark reality that demand for AI computing is growing much faster than current chip technology can keep up, even as the performance of AI accelerators improves.
The gap between AI computing demand and chip supply becomes even more apparent when considering the training requirements for models like GPT-5, which is expected to require a 46-fold increase in computing power compared to GPT-4.
However, over this same period, the improvement in performance of high-end chips, such as NVIDIA’s next-generation Blackwell, is expected to be only seven-fold.
This problem is further exacerbated by limited chip production capacity, with Taiwan Semiconductor Manufacturing (NYSE:), for example, limited to producing about 11.5 million Blackwell chips by 2025.
Adding to the complexity is the expected demand for inference chips. Inference, the stage in which AI models generate outputs after being trained, is set to consume a large portion of the AI computational ecosystem.
Barclays notes that heuristics could account for up to about 40% of the AI chip market, as evidenced by NVIDIA’s claims that a significant portion of its chips are used for this purpose. Total demand for chips in both training and inference could exceed 30 million units by 2027.
As the industry grapples with these challenges, Barclays suggests a dual-track approach to the AI accelerator market, where commercial and custom silicon solutions can flourish.
On the one hand, companies like NVIDIA and AMD (NASDAQ:) are well positioned to provide chips for training and inference of frontier AI models at scale. On the other hand, hyperscalers — companies that run massive data centers — will likely continue to develop custom silicon for more specialized AI workloads.
This split approach will allow flexibility in the market and support different use cases outside of the large LLM.
Inference is expected to play an increasingly important role, not only as a driver of demand but also as a potential revenue generator.
New approaches to improving inference, such as reinforcement learning implemented in OpenAI’s latest “o1” model, suggest breakthroughs in AI performance are possible.
With better resource allocation and cost-effective inference strategies, the return on investment in AI models could improve dramatically, providing incentives to continue investing in both training and inference infrastructure.