Everyone wants this pricey chip for their AI. But what is it?

This article has been updated.

It’s been hailed as the microchip that’s powering the artificial intelligence (AI) boom. At $40,000, it’s not cheap - but even this powerhouse chip is already being outpaced by more advanced successors. The Nvidia H100, introduced in 2022, is the graphics processing unit (GPU) that set the standard and paved the way for the Blackwell B200 and the upcoming Rubin architecture.

Creating generative AI applications is a costly business, both in cash terms and in the amount of computing power needed to train AI. In fact, it’s so demanding that only a handful of chips can cope - and as the focus shifts towards 'Physical AI' and robotics, that demand is only accelerating.

Of course, microchips are all around us. The laptop on which I’m writing this and the device you’re reading it on couldn’t function without them. And, as our expectations of what our devices can do expands, they’ve been getting more powerful.

Way back in 1965, Gordon Moore, one of the founders of tech giant Intel, predicted that the number of transistors in a microchip – and hence its processing power – would double every year. It became known as Moore’s Law and experts say it still holds true today.

You don’t need to be a computer expert to appreciate the change in processing power – the numbers speak for themselves. The computer that navigated the Apollo space missions had just 4 kilobytes (KB) of RAM and 72KB of ROM - a tiny fraction of modern devices, yet enough to guide astronauts to the Moon. The H100 has 80 gigabytes (GB) – that’s 80 billion bytes. Meanwhile, the newer Blackwell B200 features 192GB. This extra space is what allows AI to move from simple chatbots to 'Agentic AI' that can reason through multi-step problems.

Remarkable though that is, H100 is also part of another revolution in computer power. H100 is a GPU – a graphics processing unit – which, as the name suggests, is a type of microprocessor originally developed to deliver on-screen graphics for games and videos.

In fact, Nvidia created the first GPU back in 1999. Since then, their superior processing power has made GPUs the default choice for the massive amounts of data handling needed for AI.

The company controls up to 95% of the accelerator market, but the nature of this market has changed. Data centre operators like Amazon (AWS), Google Cloud and Microsoft (Azure) are no longer just customers – they are now competitors, developing custom AI ASICs (application-specific integrated circuits) such as Amazon's Trainium and Microsoft's Maia series to reduce reliance on third-party silicon.

While the H100 was the industry standard, it was the GH200 Grace Hopper Superchip – named for US programming pioneer Grace Murray Hopper – which fundamentally changed the game by merging the CPU and GPU into a single, unified 'brain' with a massive 624GB of shared memory. This architecture has paved the way for the current Grace Blackwell B200, which offers 30 times the performance compared to the H100 for inference tasks.

“To meet surging demand for generative AI, data centres require accelerated computing platforms with specialized needs,” Jensen Huang, founder and CEO of Nvidia, said at the launch of the GH200.

Today, that legacy lives on in the company's Rubin platform, which is being rolled out later in 2026. Combining six new chips, this system aims to slash AI costs by another factor of ten, turning data centres into full-scale 'AI factories' capable of Physical AI - the science of teaching machines to interact with the real world.

It will also improve efficiency five-fold, crucial when, as AMD's CEO Lisa Su recently pointed out, AI is currently using 100 zettaflops (sextillion floating point operations) up from just 1 zettaflop in 2022. "We don't have nearly enough compute for everything that we can possibly do," she pointed out.

Future generations of chips and platforms are being designed to handle these ever-growing datasets, with enormous memory and processing power that make advanced AI workloads feasible at scale.

But the rapid pace of development has left many worried that regulation is not keeping up. The World Economic Forum has launched the AI Global Alliance, bringing together industry leaders, governments, academic institutions and civil society organizations, to champion responsible global design and release of transparent and inclusive AI systems.

In April 2023, the Forum convened “Responsible AI Leadership: A Global Summit on Generative AI” in San Francisco, which was attended by over 100 technical experts and policymakers. Together they agreed on 30 recommendations for ensuring AI is developed responsibly, openly and in the service of social progress.

Writing on Forum Stories, Cathy Li, Head of AI, Data and Metaverse at the Forum and Jeremy Jurgens, a Forum Managing Director, said: “Generative AI … is proving to be a transformative force on our economies and societies.

“Its rapid development in recent months underscores AI's vital role as the foundational infrastructure upon which the digital landscape will evolve – and reinforces the need for ensuring its responsible, ethical use.”

A 2025 Forum and Accenture report found that just 1% of organizations have fully operationalized responsible AI, as evidenced in the graphic below.

In January 2026, the Forum's focus evolved further to include "Sovereign AI". Global investment in AI is accelerating, with annual investment expected to increase to $400 billion for AI infrastructure by 2030. But countries risk falling behind as AI sovereignty becomes increasingly conflated with infrastructure ownership, the white paper said.

Nvidia's Huang agreed during a featured session at Davos 2026 that AI has "started the largest infrastructure build-out in human history." He argued it should be on par with other infrastructure, like electricity and roads.

Huang urged countries to create their own models to process their own data, rather than exporting it, framing a nation's data as its most valuable asset: "With your local expertise, you should be able to create models that are helpful to your own country. And so I really believe that every country should get involved, develop AI infrastructure, build your own AI, take advantage of your fundamental natural resource, which is your language and culture, develop your AI, continue to refine it, and have your national intelligence be part of your ecosystem."

This move towards sovereign, efficient models is now seen as essential to ensure the economic benefits of the AI boom are shared globally, rather than being concentrated in a few tech hubs.

Watch the full session with Huang below: