Google's orbital AI bet forces Nvidia, SpaceX, and the power-grid crowd to reprice — at the same time

Alphabet unveiled Project Suncatcher on November 4, 2025 — a research moonshot to fly clusters of solar-powered satellites carrying the company's own Trillium (v6e) TPUs, networked by free-space optical links at tens of terabits per second. The announcement landed with a designed-in hardware partner (Planet Labs) and a concrete milestone: two prototype satellites in orbit by early 2027. That single move converts orbital compute from a sub-$200M venture-funded curiosity into a hyperscaler-validated engineering category, and the rest of the field has been scrambling to price it in ever since.

Google's pre-print lays out an aggressive but specific target: 81 satellites flying in formation within a1-km-radius cluster at ~650 km altitude in a dawn–dusk sun-synchronous orbit, with next-nearest-neighbor spacing oscillating between 100 m and 200 m. The bench-scale demonstrator already hit 1.6 Tbps (800 Gbps each way) on a single transceiver pair; the paper argues that multi-channel DWDM and spatial multiplexing can push aggregate per-link bandwidth to roughly 10 Tbps — enough to behave, in aggregate, like a data-center fabric.

Radiation testing is the load-bearing engineering claim. Google's Trillium chips were zapped with a 67 MeV proton beam to simulate a five-year mission: no hard failures attributable to total ionizing dose up to 15 krad(Si) on a single chip, while high-bandwidth memory started showing irregularities after 2 krad(Si) — inside the 750 rad(Si) shielded envelope Google assumes for a real spacecraft. That's a credible five-year survival envelope for a compute payload, not a lab curiosity.

Starcloud beat Google to orbit by two days. On November 2, 2025, Starcloud-1 launched aboard a SpaceX rocket carrying an NVIDIA H100 — the first data-center-grade GPU in space — and used it to train Google's open-weight Gemma model in orbit. By August 2026, Starcloud had closed a $250M Series A extension at a $2.3B post-money valuation led by Manhattan West, with NVIDIA joining existing backers Benchmark and EQT. The company's FCC filing asks for 88,000 orbital computing satellites at 20 GW of capacity. Total raised since 2024 founding: $450M.

SpaceX filed with the FCC on January 30, 2026 for up to one million solar-powered satellites at 500–2,000 km altitude — the same Shell FCC application that backs its reported IPO thesis. The orbital data-center tie-in was made explicit at Starbase Louisiana's August 25–26, 2026 unveiling: a $100B, 125,000-acre second spaceport whose first Starship flight is targeted for 2029, sized to support \"thousands of Starship flights a year.\" Orbital compute has become a documented IPO narrative.

NVIDIA hedged both ways. CEO Jensen Huang publicly called orbital data-center economics \"not compelling yet\" in February 2026, then launched the NVIDIA Space-1 Vera Rubin Module on March 16, 2026, claiming up to 25x more AI compute per watt than the H100 for space-grade inferencing. IGX Thor and Jetson Orin are available today; the Vera Rubin module ships later. NVIDIA is now selling silicon into the same category Google just legitimized — and funding the most credible startup in it.

Jensen Huang's repeated refrain has been that \"every data center of the future will be power-limited.\" That thesis is the bedrock of NVIDIA's $5.46T market cap: as long as the world's AI compute growth runs into a U.S. grid that Gartner expects to power-constrain 40% of AI data centers by 2027, every incremental training and inference dollar flows to GPU suppliers. Project Suncatcher threatens the demand side of that thesis — not by replacing GPUs tomorrow, but by offering Alphabet a non-grid-bound scaling lane with a five-year radiation envelope already validated.

Google Cloud revenue grew 82% year-over-year to $24.8B in Q2 2026, with a $514B contracted backlog — the demand that justifies the spend. Roughly 78% of AI servers shipped to Alphabet in 2026 carry TPUs, not NVIDIA GPUs, by analyst tally. Suncatcher extends that mix into orbit. For NVIDIA, the read-through is mixed: orbital compute is still decades from meaningful share, but the marginal hyperscaler is now publicly demonstrating a non-GPU scaling option with Google-branded silicon at the core.

The more immediate pressure is on the on-ground buildout. If Alphabet believes even a fraction of post-2027 AI workloads can run in space, the capex that funds U.S. utility buildouts, transformers, and gas turbines slows — directly. NVIDIA's power-bound thesis assumes that constraint binds for at least another decade; Suncatcher is now a public option to bypass it.

Three listed names have a designed-in seat at the orbital-compute table, and the evidence for each is concrete.

Days to quarters (through 2027): The two Suncatcher prototype satellites fly with Planet Labs by early 2027, validating — or refuting — the Trillium TPU's radiation envelope under real mission conditions. Alphabet's Q3 2026 earnings (October) and full-year 2027 capex guide will be the next read on whether Suncatcher draws material spend. Rocket Lab's Neutron first flight is the catalyst for non-SpaceX orbital-launch credibility. Starcloud's 88,000-satellite FCC application and any pre-IPO disclosure from SpaceX are the timing risks to watch.

One to three years (2027–2029): The path to a 1-km, 81-satellite cluster runs through Starship's $/kg to LEO, currently estimated in the low-thousands. If SpaceX hits the implied $200/kg target and Alphabet executes on the 1.6 Tbps inter-satellite link in volume, orbital compute could approach ground-cost parity on inference workloads that tolerate latency — a small share of total AI today, but the fastest-growing share. Watch the NVIDIA Space-1 Vera Rubin module ship date, Broadcom's TPU mix disclosures, and any utility-company (e.g., AES, Vistra) revision to data-center load forecasts as the second-order signals. The structural risk to the entire on-ground capex stack isn't that orbital compute wins — it's that hyperscalers now have a publicly disclosed option to bypass U.S. grid constraints, which changes how much on-ground capex they commit to first.