Feature Port Hedland is a town of just 16,000 people in a hot and dusty corner of northwestern Australia, 1,600 kilometers by road from the nearest substantial city. As such, it's seemingly an odd place for a datacenter run by Australian operator NEXTDC. Yet such locations are increasingly in demand, thanks to their proximity to energy sources - given the voracious appetite for compute capacity to train artificial intelligence models.

Nvidia CEO Jensen Huang has advanced the idea of remote datacenters, because he thinks they're where the heavy lifting for AI can be done with low energy costs and environmental impact.

As he explained during June's Computex exhibition, the world has enough energy to run the infrastructure that builds AI models - it's just not available in densely populated areas.

He therefore suggested building datacenters where people are scarce, but energy is plentiful.

Huang's idea is not entirely new. Early in the hyperscale age, the likes of Meta and Google discovered the Columbia River valley in the Pacific Northwest of the US - a location where hydro-electric electricity plants were being under-utilized, as the aluminum smelters they were built to power shut down. In addition to cheap power from those hydro plants, the region also offers a cool climate and enviable connectivity. It quickly became a prominent datacenter province.

It helps that the Pacific Northwest isn't far from population centers, as that means latency won't be high for users on the US west coast - home to many who consume services running in in the region's datacenters.

While latency matters for some users in Port Hedland - the datacenter was largely built to serve the local mining industry - training AI models is not a latency-sensitive workload. On the other hand, inferencing - putting AI models to work - needs low latency because users want chatbots and text-to-image systems to work fast.

Ed Anderson, a distinguished VP analyst at Gartner, told The Register "Architecturally, inference is better placed at the point of consumption. And if we need big compute farms, let's collocate them with energy sources."

Which explains why a datacenter in far-off Port Hedland can be viable - either to serve local clients or to take advantage of energy assets that aren't available closer to cities.

It also explains China's policy of "Eastern Data and Western Computing." That's an effort to migrate five million racks full of equipment to the country's west, where land and renewable energy - much of it from hydro-electricity - are plentiful, leaving datacenters in the densely populated east to store data and run less energy-intensive workloads.

China's plan addresses an issue called out by energy consultant Matthew Warren, who told The Register that powering datacenters with solar energy is not feasible.

For starters, the sun doesn't shine at night, but datacenters must run 24x7. Warren noted that while it's possible to build solar farms of sufficient size to power a datacenter by day and generate enough energy to be stored on-site to keep it going at night, cloudy days still present a problem. Extending other energy sources to a remote datacenter is not viable because building enormously long wires is expensive, and transmission loss means the amount of energy falls over long distances.

Hydroelectricity is a possible alternative, but Warren observed "Everybody who wants to secure zero emission baseload has already put their facility next to a hydro plant."

Operators who can find a well-priced energy resource in remote areas will at least find remote ops aren't overly painful.

Gartner's Anderson pointed out that Microsoft has run an underwater datacenter without incident.

"Datacenters take a lot of resources to build but don't take a lot of resources to operate," he explained.

But in remote locations, those building resources can be harder to secure.

Ben Crowe, associate director for colocation and cloud at datacenter design and construction outfit Vertiv, told The Register his firm's work on the Port Hedland datacenter highlighted the shallowness of the labor pool in remote spots.

"We got access to the right talent," he recalled, but when weather events slowed work it was hard to get skilled people back on site - because they were booked on other jobs. Delays can therefore compound. But he also noted that supply chains aren't a major issue: remote operators know they can't assume same-day delivery of hardware, and plan accordingly.

Weather also creates datacenter design challenges. Port Hedland is prone to tropical storms, and the climate is hot and humid. Crowe said cooling systems are therefore often bespoke on such sites and require extra testing. Knowledge like the tropical datacenter operations guidance created by Singapore's government can help, but datacenter operators' need for extreme reliability means off-the-shelf kit and approaches seldom satisfy.

Remote datacenters dedicated to AI training alone also have challenging economics.

David Hirst, group executive for Australian datacenter operator Macquarie Technology Group, told The Register that his company, and its peers, typically seek diverse tenants to balance their risks.

An AI training datacenter, by contrast, might have a single tenant, for a few months or longer-term. Hirst's industry peers have told him some clients are willing to commit to five- or ten-year tenancy deals for remote datacenters dedicated to training, at which point the numbers stack up. But he's uncertain if today's datacenter operators are comfortable with the risks and operational requirements of a remote center.

Which may explain why specialist AI datacenter operators like CyrusOne are emerging. Of course that class of datacenter operator also provides computing infrastructure - typically many thousands of GPUs, which increase their capital costs. Operators of conventional datacenters nearly always require tenants to bring their own kit.

Gartner's Anderson highlighted another risk created by remote datacenters: they could end up in jurisdictions with little regard for environmental concerns.

"When you look at most Western economies - the US, Canada and Europe - environmental regulations mean it is not possible to build new power plants that don't produce clean energy."

"But there are other places in the world that have less strict regulation - and it's an option for datacenter operators." ®

https://www.theregister.com//2024/07/15/remote_datacenters_on_the_rise/