Charlie
Edwards
Senior
Adviser for Strategy and National Security
About
Charlie Edwards
Online
Analysis 29th January
2025
With
Stargate, will the US win the AI race?
Building AI
infrastructure is only one part of the story, as sovereign AI capabilities may
prove more decisive.
Donald Trump
launched his second presidential term with the announcement of a US$500 billion
initiative to build the physical and virtual infrastructure needed to support
developments in artificial intelligence (AI). The Stargate initiative has
attracted much attention, including questions over its financing, but its goal
is more prosaic. The initiative is Donald Trump’s attempt to unify the private
and public sectors in a common mission and reduce the political hurdles and
infrastructure bottlenecks that stymied progress under his predecessor.
The AI
ambitions of the United States, China and the European Union are constrained by
insufficient underlying infrastructure, including old national energy grids;
inadequate data centres; and unreliable and unsustainable energy supplies. But
the competition for AI dominance may increasingly come to be defined by states’
sovereign AI capabilities and their ability to export them.
Gridlock
A key
determinant of becoming a global leader in AI is the ability to build an
efficient, sustainable and resilient infrastructure that ensures energy is
available, reliable and constant. The state of national power grids in China,
the EU, and the US remains a significant barrier. China’s creaking grid
represents a major constraint to progress and the government is planning to
invest more than US$800bn over the next six years. The investment will support
Beijing’s Eastern Data, Western Computing initiative, which aims to tap into
China’s energy resources in the west and transfer computing power to economic
hubs along the coast.
The European
power grid is one of the oldest in the world. Moreover, around 40% of the grid
is around ten years off its expected lifespan, while over half of the physical
grid needs to be repaired or replaced. It remains uncertain whether the
estimated US$584bn in European grid investments needed this decade will
materialise. In 2024, the EU’s Modernisation Fund handed out almost US$3bn to
modernise member states’ energy systems, amongst other activities.
The ageing
and fragmented US grid comprises three main regions (Western, Eastern and
Texas), which remain inefficient, especially for interconnections between
regions. The US Department of Energy (DoE) estimates that power outages cost
the US economy US$150bn annually. Modernising the US grid will cost trillions
over the coming decades.
The
difficulty of modernising the US grid was one of the reasons why, as one of his
final acts as president, Joe Biden signed an executive order (EO) to build the
next generation of infrastructure that will underpin the growth in technologies
such as AI and quantum computing. Both technologies will require large-scale,
energy-intensive computing infrastructure. The EO directs US federal agencies,
including the Department of Defense and DoE, to lease federal sites for
private-sector construction. The Trump administration is unlikely to keep the
EO in its current form, but there is broad bipartisan agreement on the
direction of travel, not least in modernising the US grid and boosting its
capacity.
At the
centre of this new AI ecosystem are the power-hungry data centres that house
the large and complex computers that are used to train AI models. They are also
critical to advancing technological innovation. The global data-centre market
is projected to grow from US$242.72bn in 2024 to US$584.86bn by 2032. But this
carries significant national-security implications, not least for critical
national infrastructure, and drives up energy consumption.
Power
supply: the nuclear option
The
expanding power demands of AI represent a significant challenge to states’
ambitions. The International Energy Agency estimates the combined electricity
use by Amazon, Microsoft, Google and Meta more than doubled between 2017 and
2021, rising to around 72 terawatts in 2021. And while in the medium term data-centre
energy use is expected to grow moderately, in the long term growth rates are
highly uncertain given the expected build-out of infrastructure.
According to
the US DoE, hyperscale facilities (data centres that provide extreme
scalability capabilities and are engineered for large-scale workloads) are
‘stretching the capacity of local grids to deliver and supply power at that
pace’. And the challenge of expanding power demands is only likely to grow. In
September 2024, ECL, a data-centre startup, announced it was planning to build
a one gigawatt AI data centre in Texas (roughly the equivalent to powering 2.65
million homes in the UK).
To realise
the ambitions of frontier AI, governments and the private sector will have to
find new sources of energy. In 2024, nuclear energy gained increased attention
as a potential means to meet the growing demands of the AI sector. In September
2024, Microsoft signed a deal to help resurrect a unit of the Three Mile Island
nuclear plant in Pennsylvania. Nuclear energy – nearly carbon free and more
reliable than solar or wind – is favoured by tech companies for its ability to
generate consistent and sustainable power.
Small
modular reactors (SMRs), which are a fraction of the size of a conventional
nuclear power reactor and can be factory assembled before being installed on
location, will likely make nuclear-powered data centres possible within the
next decade. Some campuses plan to co-locate with existing nuclear reactors,
but SMRs potentially offer scalable, zero-carbon technology for broader
applications. Amazon, Google and Oracle have all committed to buying SMRs,
which President Trump has previously suggested may provide a ‘potential answer
to long-running cost concerns surrounding [nuclear] … [that] could avoid the
complexities associated with large reactors’.
Digital
sovereignty
Assuming
greater domestic control over the development of world-class computing and data
infrastructure will lead to increased digital sovereignty. For example, in
October 2024, Denmark revealed a new AI supercomputer powered by Nvidia. The
supercomputer, named Gefion, is an example of what Nvidia calls ‘sovereign AI’,
which the company defines as ‘a nation’s capabilities to produce artificial
intelligence using its own infrastructure, data, workforce and business
networks’.
Other
governments are following suit. In the UK, the newly published AI Opportunities
Action Plan sets out an approach to build ‘sovereign AI compute, owned and/or
allocated by the public sector’ which will ‘enable the UK to quickly and
independently allocate compute to national priorities’. The challenge, however,
for the UK and other governments that have championed AI safety, will be how
they manage the inherent tension between building sovereign AI capability and,
in parallel, promoting AI governance to establish international norms and
standards for the development and deployment of government-backed AI
systems.
For all the
fanfare that came with the Stargate initiative, a better illustration of how
the US might win the AI race came earlier this year when Microsoft announced it
would invest around US$80bn in 2025 to build out AI-enabled data centres in the
US and overseas. Brad Smith, the company’s CEO, set out the challenge facing
the US and its allies by recalling China’s dominance of global telecoms
infrastructure: ‘if a country standardizes on China’s AI platform, it likely
will continue to rely on that platform in the future’. The race is on.
Author
Charlie
Edwards
Senior
Adviser for Strategy and National Security
Sem comentários:
Enviar um comentário