Orbital datacenter startup CEO admits launch economics don't fly, presses ahead regardless
Needs SpaceX et al to drop prices and give competitors a ride into space to make it work
A startup called Orbital has revealed a plan to build a 10,000-satellite neocloud in space – if Elon Musk delivers on his ambitious plans to increase launch capacity and reduce costs.
Speaking with The Register, Orbital CEO Euwyn Poon admitted orbital datacenters aren't economically feasible today. "The constellation approach to run AI inference makes sense, but as of today, the economics of launch don't quite work yet," he said.
Poon said the cost to launch one kilogram into space is currently around $7,000, assuming you can catch a ride on one of SpaceX’s Falcon 9 missions.
"Elon's stated goal is around $10 per kilogram, and we need to get there," Poon said. "I believe we'll get there. I think all the pieces are in place. I see that happening, but it's a big, couple orders of magnitude to get to where we need to have an economically viable solution."
Even if Elon’s engineers can hit that target – a valid question given the Musky One’s failed past predictions like putting humans on Mars by 2022 or fielding a million robo-taxis by 2020 – Orbital’s ambitions assume launch companies like SpaceX and Blue Origin, which both plan to create their own orbiting datacenters, will have spare capacity they’re happy to sell to a competitor.
While Orbital waits to see if launch capacity becomes available, the company is working on a satellite roughly the size of four refrigerators put together that houses 100kW worth of gear.
On Earth, cooling a 100kW rack is challenging. In orbit, dispersing heat requires large radiators.
But Poon, whose experience includes running the electric scooter company Spin before it was acquired by Ford, argues the power and heat rejection issues aren't as big a problem as they've been made out to be.
He tells us the satellites will employ a solar array "roughly the size of two halves of a tennis court and a radiator about half that."
Getting the heat from the chips to the radiators is another matter entirely. "Redesigning chip cooling for the space use case is a fun and interesting challenge,” Poon said.
In fact, he argues that 100kW per satellite is really the sweet spot. If they were "10 times the scale, the radiator and solar panels sizes would be unwieldy."
Speaking of the chips, Orbital is keeping things simple. At GTC last month Nvidia teased its Space-1 Vera Rubin module, and that's exactly the chip Orbital aims to launch as part of a proof-of-concept flight planned for next year.
The test flight will use a smaller system to evaluate Orbital’s radiation hardening technology, an important requirement for any computer in space because stray particles that Earth’s atmosphere deflects or absorbs can damage computers in orbit, or just corrupt data.
Fixing flaws in orbiting datacenters isn’t an option, so reliability is essential.
"Servicing is certainly impractical, even at scale we're thinking about it's sort of a non-starter. The real mitigation is avoiding the need for service by testing chips before we put them out there, adding radiation hardening," Poon said.
"I see that as a pretty good first proof-of-concept, but we're excited to just get something off the ground," Poon told us.
This won't be the first proof-of-concept datacenter GPU in space. That launched aboard Starcloud-1, launched by another startup called Starcloud, last December.
Perhaps the biggest challenge in Poon's mind is manufacturing at scale. "Most satellites have been designed for larger payloads with less power requirements. GPU-type server requirements are very different, power hungry, and very light," he said.
Orbital is tentatively targeting 2030 for its first full-scale satellite, but again that's contingent on SpaceX's Starship being ready and launch costs becoming affordable.
Once operational, Orbital plans to scale the constellation much like a terrestrial neocloud using a combination of equity and debt financing to grow the business and acquire the GPUs necessary to power the datacenters.
Poon hasn't ruled out the possibility of offering sovereign constellations.
However, Orbital still has a long way to go before it can realize that goal. On Tuesday, the startup announced it received funding from venture capital fund a16z's speedrun program to get its first flying servers off the ground.
VC funds, of course, assume the majority of companies they invest in won’t get off the ground. Orbital may struggle to do so literally and metaphorically, as may the whole idea of orbiting datacenters: Gartner analyst Bill Ray certainly isn't, describing the trend as peak insanity.
"Datacenters in space won't analyze data on Earth for Earth applications for decades, if ever," he wrote in a February report. ®