Graphic Processing Units, or GPUs. You might have come across them, and maybe you think they're a bit of a buzzword. Yet they’ve been around for nearly thirty years, thanks to gaming company-turned-AI gamechanger Nvidia, which first developed them in 1999, putting them in graphics cards to render complex 3D visuals in real-time. 

Fast-forward three decades, and gaming seems to be the last use for GPUs. They are both the key and kryptonite of the artificial intelligence boom that has consumed global imagination (and economies) for the last six years. But with card prices potentially reaching US$5,000 later this year, and Nvidia reportedly considering cutting production of mid-range cards by 30 to 40 per cent, are we facing a run on GPUs?

What exactly are GPUs?

It's easier to understand why GPUs are so important by diving into what they actually are. In simple terms, GPUs are electrical circuits, or chips, that can perform millions of simple mathematical equations at the same time. This process, known as parallel processing, allows them to render complex graphics, ideal for making realistic video games. What gamers buy is the physical, assembled hardware component that hosts the GPU chip, known as the graphics card.

They've also got far more cores per chip, which are the primary computing units that do all this math work. CPUs on the other hand are able to do only a few bits of more complex maths (sequentially), and have far fewer cores.

In the late 2000s, researchers realised that this combination of parallel processing and core power had applications far beyond video games. Able to perform technical calculations at rapid speeds and with high efficiency, these chips transformed from something you bought when you wanted to play games on your computer, to the foundation of the AI revolution. 

The AI makers and AI takers

The ability of GPUs to process thousands of calculations simultaneously means they are massively in demand for AI and machine learning purposes - so much so that the GPU market size is expected to increase from US$128.17bn in 2025 to US$144.83bn in 2026, and reach a whopping US$296.34bn by 2031. 

Nvidia has the lion's share of the market, but companies like Advanced Micro Devices (AMD), Intel and Google all develop their own GPUs or specialised AI processors to power models. China has also become a major player, with companies like Lisuan Technology and Moore Threads shaking up the industry.

The primary customers of these companies are hyperscale cloud providers and Big Tech – so that's entities like Amazon, Microsoft, and Alphabet (the parent company of Google). Another GPU mega-customer is Mark Zuckerberg's Meta, having committed to not one but two mammoth chip deals earlier this year, with Nvidia and AMD respectively. 

Sam Altman's OpenAI is another, operating over one million GPUs to train their massive GPT models. They are reportedly nearing a deal for a massive Ohio data center project with Nvidia, which could become one of the largest artificial intelligence infrastructure developments ever undertaken, with a price tag of over $500bn.

Countries also put in orders for these chips, with European governments increasingly pivoting towards a sovereign AI strategy that involves purchasing chips to power their own data centers. The recent announcement from the UK's prime minister Keir Starmer that a billion pounds would be allocated to strategic purchases of semiconductor equipment and AI infrastructure underscores the government's goal to prevent domestic tech companies fleeing to Silicon Valley.

Is the supply meeting the demand?

In the early 2020s, GPU prices surged when pandemic-hit supply chains struggled to meet massive demand spikes, as home gaming and crypto-mining surged. Gamers today however may look back to those times with wishful thinking: a recent report by TechSpot revealed that the price of the average gaming graphics card has increased by 15 per cent over the course of just one year. 

In the past month however, the squeeze has gotten tighter still. Global shortages of other components required to make GPUs is pushing up production costs further, with some premium cards recording price hikes of over 90 percent. Midrange cards are expected to follow as stocks deplete. For the gamers and PC builders looking to upgrade their setup, these shocks will be anything but welcome.

What are the solutions?

For consumers and gamers, one solution is going back in time. As the demand for GPUs is not evenly spread, instead concentrating around AI and enterprise-focused devices such as those focused on model training, older graphics cards and refurbished GPUs are now being sold on second-hand markets as a viable budget alternative for anyone looking to upgrade their compute. Some manufacturers, such as Manli Technology Group, are also taking this line, re-releasing old-generation cards. 

For tech firms on the other hand, it's more of a waiting game. While the big players are able to take advantage of their sheer spending power and snap up the GPUs at inflated prices, they can't speed up production. Memory fabrication plants take years to build, and existing capacity is already stretched thin trying to satisfy the AI data centre boom. 

Lead times across the GPU market are extending, and availability is becoming more inconsistent from week to week. Growing demand in older chip lines shows that some buyers are even turning to previously overlooked devices just to secure enough compute.

The emergence of GPU-as-a-service (GPaas) also nods to a growing prerogative towards flexibility over ownership. Rather than sinking enormous capital into buying and housing their own hardware (which may become outdated within a couple of years, or simply unobtainable at a reasonable price), companies can rent GPU capacity by the hour or month from specialist cloud providers. This spreads the cost, reduces the risk of being locked into a single chip generation, and gives smaller AI startups a route to the compute power they need without competing directly with hyperscalers for limited supply.

Are we missing the bigger picture?

The chips are not the only problem facing AI-driven economies these days. There is growing concern about being able to actually use chips once they have been accessed, as they will all be fitted into massive data centers that require almost unfathomable amounts of power to run them. 

In the UK, a 256MW data centre, one of the largest in the country, has recently been approved in West London If built, it will swallow enough energy to power every household in Birmingham, the country's second-largest city.

Climate campaigners have been raising questions over the sustainability of such projects, with some data center developers in London reportedly looking to power their servers using gas rather than plugging into the main grid. 

While data centers for now remain an integral part of tech infrastructure and the AI boom, the pressure on country's grids and the environmental concerns of using oil and gas mean we could be seeing a move towards using solar panels, turbines and nuclear energy to power our growing AI needs. 

This is the perspective for Nvidia's founder and CEO Jen-Hsun Huang, who said to the BBC last year that he was hopeful that in the UK, more gas turbines could be used "off the grid so we don't burden people on the grid". 

His optimistic outlook that AI itself would be used to produce more cost effective sustainable energy further underscores how AI may not just be informing our energy needs, but at the heart of designing them too.