Starcloud, the US company building satellites that run AI inference in orbit, has added 250 million USD (about 40 billion yen) to its Series A. The round values the company at 2.3 billion USD (about 366 billion yen), with NVIDIA and Cisco among the backers. The main reason the company is stockpiling capital, however, is not the satellites themselves but the scramble for launch slots.

※1 USD = 159 JPY

The Series A now totals 420 million USD

This financing is an extension of the 170 million USD (about 27 billion yen) Series A that Starcloud closed in March 2026. With the new 250 million USD on top, the round alone comes to roughly 420 million USD.

Manhattan West Ventures led the extension, with NVIDIA and Cisco participating. A person familiar with the deal says NVIDIA put in 25 million USD (about 4 billion yen). Benchmark, EQT, Soma, NFX, 776, Cedar Capital, Goanna Capital and Standard Capital also joined.

The money is headed to two places. One is a larger manufacturing facility. The other is Starcloud-3, the company's biggest orbital data center spacecraft to date, which is slated to fly on SpaceX's Starship.

The capital is really about launch capacity

Chief Executive Officer Philip Johnston's motivation for raising more is a little unexpected: he expects seats on rockets to run short.

Securing launch capacity has become one of the largest cost items, he says. Behind that is the scheduled end of SpaceX's workhorse Falcon 9 program in 2028. SpaceX plans to retire the proven Falcon 9 in stages and shift to Starship, a far larger vehicle that remains unproven. For satellite operators, that makes planning several years out difficult.

The alternatives are not dependable yet either. Blue Origin's New Glenn and ULA's Vulcan are not flying regularly, and Rocket Lab's Neutron is not yet on the pad. Starcloud therefore wants to get under contract with vehicles like Starship as soon as it can. The company has already asked the Federal Communications Commission for permission to operate 88,000 spacecraft, which implies an enormous number of launches.

The near-term plan is more modest. Two 8 kW compute satellites from the new Starcloud-2 generation are set to fly on rideshare missions in 2027, handling orbital inference for customers that include US government agencies. To add more spacecraft, the company is also weighing a dedicated Falcon 9 launch and contracts with other providers.

Starcloud's entire business case rests on Starship driving launch costs down. An orbital inference layer cannot compete with terrestrial data centers unless transport gets cheaper. Johnston remains confident that SpaceX will show the world's most powerful rocket can be reused quickly and often. That said, Elon Musk said this week that the attempt to catch a returning Starship will slip by a few months, and that the first re-flight will come at the end of this year or in early 2027. If Starcloud cannot book SpaceX capacity in 2029, Johnston admits, things get difficult.

An H100 in orbit is what brought NVIDIA in

Asked why NVIDIA invested now, Johnston points to the data gathered in orbit.

Starcloud is, as far as anyone knows, the only company currently operating an NVIDIA H100 in orbit, a GPU designed for terrestrial data centers. It is also the first to train a model using one. Most other GPUs sent to space are built for edge processing, so putting a data center class part into orbit is a different proposition.

The data from that first spacecraft, Starcloud One, is being shared with NVIDIA. The chipmaker is developing Vera Rubin Space-1, its first GPU purpose-built for space, and Starcloud is feeding findings into that design. According to Johnston, NVIDIA did far deeper technical diligence than any other investor.

The space-ready chip does not exist yet. Starcloud hopes to fly it in late 2028. Its engineers are tracking three design questions: the relationship between the chip's operating temperature and the size of the radiators that shed that heat, where to place radiation shielding, and how much ruggedizing the parts need to survive launch. None of these show up in conventional server design.

25 employees and a 100,000 square foot production line

The company itself is still small, with 25 employees and hiring under way.

Its manufacturing plans run ahead of that headcount. Production lines are being built out in a 100,000 square foot (about 9,300 square meter) facility in Woodinville, Washington, near where SpaceX and Amazon build satellites for their communications networks. Settling into a region rich in satellite manufacturing talent and supply suggests Starcloud intends to build Starcloud-3 and later spacecraft in house.

Summary

Starcloud's 250 million USD looks less like money for orbital AI technology and more like money to guarantee a ride to orbit. Running an H100 in space is what drew NVIDIA in, and that thread now leads to a dedicated chip called Vera Rubin Space-1. The technology is coming into focus, but the end of Falcon 9 and the timing of Starship's ramp remain the largest variables, and they sit outside the company's control. Starcloud-2 in 2027 and the space GPU in late 2028 will be the checkpoints for how real this plan is.