That blistering pace, fuelled by what chief executives and analysts call “insatiable demand” for compute, is throwing up nasty problems. Vital components are scarce, while the industry is fretting about when AI companies will reliably make enough money to keep buying chips.
Hardware designers like Nvidia, which more than doubled revenue year on year, are selling the picks and shovels for this digital gold rush. Nvidia is staring down growing competition from Alphabet’s Google and Amazon.com as the battlefield continues to shift.
Last week, Nvidia signed a $20 billion licensing deal with chip startup Groq, which builds chips and software to accelerate AI inference, in which trained models dish out answers to prompts. The last leg of the race was training; now the giants are fighting to deliver the fastest and cheapest inference.
Analysts at Bernstein wrote that "Inference workloads are more diversified and may open up new areas for competition.”
Data centre operators, AI labs and business customers have been scrapping for Nvidia’s H200 and B200 graphics processing units. Google’s custom TPUs and Amazon’s Trainium and Inferentia chips are pulling in customers, while OpenAI is teaming up with Broadcom to design its own silicon.
AMD is launching a GPU in 2026 that it is pitching as its first serious shot at Nvidia’s AI processors. Microsoft said in October it would double its data centre footprint in the next two years, which has analysts pencilling in extra chip revenue in 2026.
Recent quarterly numbers and next-quarter guidance across Nvidia, Broadcom, Troubled Chipzilla, Qualcomm and AMD point to another big year, even if their fiscal calendars do not line up neatly. Nvidia’s fiscal year ends in January, Broadcom’s ends on the Sunday closest to 31 October, AMD’s and Chipzilla’s end in December, and Qualcomm’s ends in September.
Goldman Sachs reckons Nvidia will shift $383 billion of GPUs and other hardware in calendar 2026, up 78 per cent from the year before. Analysts polled by FactSet expect combined sales from Nvidia, Chipzilla, Broadcom, AMD and Qualcomm to top $538 billion, without counting Google’s TPU business or Amazon’s custom chips because they are not broken out.
Then come the real headaches. A shortage of electrical transformers and gas turbines is slowing data centre construction, and operators are scrambling to secure the power needed to run these compute clusters.
There is a global squeeze on parts used in AI data centre servers, including ultrathin silicon substrate layers and memory chips. As inference ramps up, demand for high-bandwidth memory has surged because inference is more likely to be “memory-bound” than training, which is usually capped by processor grunt.
Micron Technology, chief business officer Sumit Sadana, told the Wall Street Journal: “We’re significantly short of our customers’ needs, and it’s going to persist for a while, as memory suppliers enjoy the pricing power. Micron’s shares are up 229 per cent so far this year, and rivals like South Korea’s Samsung and SK Hynix have benefited too, even if new clean rooms and fabs take time."
The financing propping up the data centre build-out is another pressure point, especially if big buyers like OpenAI cannot raise cash fast enough to keep the chip orders flying. Investors have been conditioned to expect lumpy quarter-to-quarter leaps, then panic at the first hint of a slowdown.
This autumn, investors dumped AI stocks, including major chip designers, amid concerns that the funding behind AI infrastructure purchases was less solid than it appeared. OpenAI has multibillion-dollar agreements with Amazon, Microsoft, Oracle, and others for compute, and hyperscalers like Microsoft say they will continue to invest in construction in 2026. However, some analysts think the boom could cool in 2027.
DA Davidson analyst Gil Luria said: “There’s a chance that 2026 is a peak. If it’s the end of March and we don’t hear that OpenAI has raised a hundred billion dollars, then the market may start pumping the brakes.”
As more chip companies pile into AI products, margin pressure is another lurking threat. Broadcom’s stock sank even after it posted record quarterly revenue in December, partly because investors worried growth could slow in its higher-margin lines.
Circular Technologies, global head of research, Brad Gastwirth said: “I don’t think this will be the top. The race to artificial general intelligence is still powering a huge appetite for compute, across the spectrum of customers.”