K2 Space, founded by former SpaceX engineers, will launch Gravitas—a 2-metric-ton satellite designed to test data center operations in orbit—by the end of March 2026. Announced yesterday, the mission carries 12 payload modules (including Department of Defense customers) and the most powerful electric thruster ever flown in space. If successful, Gravitas validates a radical bet: moving AI compute infrastructure from power-constrained terrestrial data centers to solar-powered orbital platforms.
As AI compute demand explodes and terrestrial data centers hit limits on power, cooling, and real estate, K2 is testing whether space solves the infrastructure bottleneck.
The Most Powerful Satellite in Its Class
Gravitas isn’t a typical satellite. The 2-metric-ton spacecraft spans 40 meters when its dual 10 kW solar arrays unfold, generating 20 kW of power—10 times more than comparable satellites. That power feeds 12 undisclosed payload modules from Department of Defense and commercial customers, plus a 20 kW Krypton-fed Hall-effect thruster.
That thruster is the mission’s centerpiece. It’s the most powerful electric propulsion system ever flown in space, capable of raising Gravitas from Low Earth Orbit to Medium Earth Orbit in under 90 days. That’s significant: faster orbit raising means K2 can deploy four times more satellites per launch, cutting deployment costs substantially.
The satellite is designed for multi-orbit operations—LEO, MEO, geosynchronous, and cislunar space—positioning it for different use cases depending on latency and coverage requirements. K2 raised $450 million at a $3 billion valuation and secured a $60 million STRATFI contract from the U.S. Space Force, SpaceWERX, and the Air Force Research Laboratory. This isn’t vaporware.
The Economic Debate: 3× Cost vs 15× Cheaper Energy
Here’s where it gets contentious. Skeptics, including Varda Space Industries, calculate orbital compute costs at roughly 3× more per watt than terrestrial equivalents. Current launch costs run $1,500-$3,000 per kilogram; viability requires dropping to $200-500 per kilogram. Quentin A. Parker, director of the Laboratory for Space Research at the University of Hong Kong, is blunt: “To do a cost effective, true, objective analysis of it, it doesn’t really stand up to scrutiny.” OpenAI co-founder Sam Altman ridiculed the concept, saying “we’re not there yet.”
Proponents counter with hard numbers. Space solar panels generate roughly 8 times more energy than terrestrial panels because they’re above the atmosphere with 95% sunlight exposure (versus 24% capacity factor for terrestrial solar farms). Vacuum cooling eliminates the need for water—terrestrial 40 MW data clusters consume 1 million tons of water annually for cooling. Eliminating that, plus accessing unfiltered solar power, could cut energy costs by 15 times, down to approximately $0.005 per kilowatt-hour compared to wholesale electricity prices.
And here’s what skeptics can’t dismiss: February 2026 marked the first month in history where multiple orbital data center operators ran production workloads in space simultaneously. This is already happening.
Where Orbital Compute Makes Sense (and Where It Doesn’t)
The debate isn’t whether orbital computing can work—it’s where it makes economic sense. K2’s 12 DOD payloads suggest defense applications: compute infrastructure physically inaccessible to adversaries. Earth observation is another obvious fit. Satellites with cameras can process imaging data on-board before downlinking results, avoiding the cost and bandwidth of transmitting raw data to terrestrial data centers.
Batch AI training could work too. Workloads that tolerate 20-40 millisecond latency (plus Doppler shifts and handover delays) can run in space. Real-time inference? No. The latency makes it impractical for applications requiring sub-10ms response times.
And let’s be clear about scale. Training and serving frontier AI models at scale requires hundreds of thousands of GPUs. Deploying that in space means launching millions of satellites. Radiation, hardware replacement every 5-6 years, and space debris mitigation remain unsolved at that scale.
So orbital computing isn’t replacing terrestrial data centers. It’s addressing niche use cases where the advantages—solar power, vacuum cooling, physical inaccessibility, regulatory arbitrage—justify the costs.
The Stakes: This Launch Matters
Gravitas is a test mission, not a full data center. K2 must hit clear success metrics: deploy the satellite and generate 20 kW power reliably, operate 12 payload modules, test the 20 kW thruster, and execute the LEO-to-MEO orbit raise in under 90 days. The company also needs to demonstrate actual operational costs versus projections.
If Gravitas succeeds, it validates orbital computing for niche applications and likely accelerates investment from SpaceX, Google, and others watching closely. Google CEO Sundar Pichai said we’re “a decade away” from orbital data centers, but that timeline collapses if K2 proves the economics work.
If Gravitas fails—deployment issues, thruster underperformance, payload failures—skeptics are vindicated, and the industry cools off. Orbital computing gets delayed by years, maybe longer.
What’s Next
The launch window is end of March 2026 aboard a SpaceX Falcon 9. The mission’s success or failure will provide real data to settle the economic debate. Developers should watch this: AI compute infrastructure is a bottleneck affecting model training costs, deployment economics, and sustainability. Orbital computing is one proposed solution among many (better chip efficiency, nuclear-powered data centers, distributed edge computing are others).
K2’s bet is that power, water, and real estate constraints on terrestrial data centers create an opening for orbital infrastructure in niche applications. By end of March, we’ll know if former SpaceX engineers building the most powerful satellite thruster ever flown can validate that bet—or if Sam Altman was right to dismiss it.
