Sophia Space's Orbital Data Center Patent Trades Compute Density for Radiator Area

The US Patent and Trademark Office issued US Patent No. 12,679,564, "Space-Based Data Centers," on July 14, 2026. The document lists Caltech and Sophia Space as assignees. Sophia Space announced the grant on July 30. Its core layout is planar: solar cells on one face, radiator panels on the other, and computing electronics between them.

This is a different form of evidence from the orbital data center constellation filings SCN examined in July. The grant establishes a particular claimed packaging architecture. It does not establish flight readiness, technical viability, freedom to operate, or priority over every earlier orbital-compute patent. What the disclosure does provide is enough geometry and power allocation to test the architectural bargain: lower compute density and much more area in exchange for simpler local heat rejection and potentially lower cooling mass.

That test needs three denominators. Watts per square meter show how tightly power generation, compute, and heat rejection can be packed. Watts per kilogram show whether the integrated surface saves more mass than it adds. Sustained useful compute, rather than nameplate generation, shows what the payload can deliver after thermal derating and radiation losses. No available source in this comparison provides an apples-to-apples figure for end-of-life sustained compute watts per delivered launch kilogram.

What claim 1 recites

The issued B2 contains 20 claims. Claim 1 is the sole independent claim; claims 2 through 20 refer back to it directly or indirectly. On its face, claim 1 recites multiple orbital server modules, communications and command subsystems, and planar tiles with solar cells over a first external surface, radiator panels over a second surface, and electronics between them. It separates compute tiles containing processors and memory from support tiles containing network switches and energy storage.

A published continuation application was filed on January 22 and published on June 4, 2026. It expressly identifies itself as a continuation of the application that produced the issued B2. The publication alone does not establish its current Patent Center status.

That wording does not make the spacecraft bus disappear. Claim 1 does not recite a central bus, but the illustrated 60 by 60 meter baseline retains a 2.5 by 2.5 meter bus, an 8 by 8 meter solar array for that bus, and a 10 by 10 meter aggregation-switch area. The tile network then feeds an aggregation layer and core routers. Distributed tiles reduce centralized payload power routing and long-distance heat transport; the baseline remains a bus-and-aggregation architecture. The specification lays out those dimensions and network tiers.

The local thermal path is the larger departure. Each tile can use a heat spreader and a rear radiator rather than sending all payload heat to separate radiator booms through a central pumped loop. The disclosure also permits heat pipes and microfluidic channels, leaving working fluid available as a local option. The same specification describes the thermal break, heat spreader, and optional transport methods.

What the 259 W tile example implies

One disclosed embodiment is a 60 by 60 meter module that harvests and uses about 1 MW in sun-synchronous orbit. Dividing those figures gives 278 W/m² as a nominal module-footprint average. Whether the radiator surface sees that full load depends on assumptions the patent leaves open. Elsewhere, a worked tile example assumes 259 W of solar output, allocates 64 W to each of four regions, and applies 80 percent conversion efficiency before the processor budget. It uses a 25 W Qualcomm Cloud AI 100 board as an illustrative component. The patent application publication supplies the dimensions and allocation.

The 259 W figure is an electrical budget for a nominal one-square-meter tile; no claim recites it as a limit. SCN's thermal interpretation begins with a separate assumption: in steady state, most electricity consumed by the tile ultimately becomes heat. If that makes 259 W/m² a uniform, single-sided radiator load, emissivity is 0.85, and the sink is ideal deep space, the Stefan-Boltzmann relation gives 270.7 K, or about -2.4°C. Applying the same assumptions to the 278 W/m² nominal module-footprint average gives 275.6 K, or about 2.4°C. These estimates exclude Earth infrared, albedo, imperfect view factors, front-to-back leakage, local hot spots, and end-of-life surface degradation. NASA's thermal-control guidance explains the external radiative balance and environmental inputs.

At this areal power level, the required solar-collection and radiator areas are of the same order, allowing them to be paired tile by tile. A separate University of Austin concept study models solar cells at 20 to 27 percent efficiency and up to roughly 370 W/m² in sun-synchronous orbit. Sophia's 259 W example sits in the same areal-power regime without defining a universal photovoltaic ceiling. The preprint reports those modeled solar cases, while SCN's May analysis covers the wider orbital-compute power and thermal problem.

Terrestrial AI infrastructure pushes the other way, concentrating heat into facility liquid loops. SCN's coverage of NVIDIA's 45°C warm-water MGX specification examines that density-first approach, and NVIDIA describes facility-water temperatures up to 45°C in its own technical overview. The orbital tile spends surface area to shorten the power and heat paths.

AI1's draft design uses hotter, more compact rejection

SpaceX's first-party Starmind page presents AI1 with a company-stated compute payload of 150 kW peak and 120 kW average, a 110 m² deployable liquid radiator, and redundant pumping loops. In an official June 8 video, SpaceX described AI1 as a draft design and stated an assumption of about 1,400 W/m² with both radiator faces emitting. Contemporaneous reporting did not resolve the area convention.

One scenario shows why the area definitions matter. If 1,400 W/m² is a combined, two-sided planform figure and 110 m² is radiator planform, each emitting face averages about 700 W/m². At emissivity 0.85, that corresponds to an idealized radiator temperature of 347 K, or about 74°C. A single-sided radiator at 259 W/m² would require about 463 m² to reject AI1's reported 120 kW average load, which is 4.2 times the reported 110 m² planform. If 110 m² instead means total emitting surface, or 1,400 W/m² applies per emitting face, the apparent planform comparison moves by about a factor of two. SpaceX has not defined both conventions in its public draft presentation, so the 4.2 figure cannot be verified.

The worked examples favor inference

Sophia's claims are not limited to an AI workload. Its worked example does use low-power AI100 processors. The separate University of Austin concept study is more explicit: Stephen Gaalema, Samuel Indyk, and Clinton Staley model an Integrated Solar Compute Radiator, or ISCR, for distributed inference. The authors argue that tightly synchronized training is incompatible with the latency between their panels. Their April 2026 arXiv preprint has not been peer reviewed.

The modeled system is 16 MW and 150 metric tons, with a 20 meter by 2.2 kilometer grid of roughly 16,000 panels. The authors estimate more than 100 kW of compute per launched metric ton when deployment and station-keeping mass are included. They also estimate about 500 W/kg for an integrated radiator-and-solar structure, compared with less than 100 W/kg for separated conventional implementations. These are model outputs built on assumed custom components. The paper lists structural dynamics, three-dimensional thermal simulation, radiation qualification, and ASIC definition among the work still required.

SCN's inference is narrower than the authors' training statement. Inference is the most plausible initial workload for a surface divided into semi-autonomous compute islands because jobs can require less frequent collective communication. That does not make orbital training categorically impossible; a different topology, accelerator, or optical fabric could change the trade.

Structural integration brings another systems consequence, and this one is again SCN's inference rather than a demonstrated result: if, as the model assumes, the radiator is the mechanical substrate for solar cells and compute, a tear, hot spot, or deployment fault can propagate across power, processing, and thermal functions. The same integration that removes structure also couples failure domains.

A fixed tile locks in radiator margin

A one-to-one tile geometry makes lifecycle margin visible at design time. ABI Research principal analyst Andrew Cavalier used a general orbital-radiator model for IEEE Spectrum and estimated that emissivity degradation could increase required area by about 40 percent after five years. He was modeling orbital radiators in general, without reference to Sophia's design. Beginning-of-life area and temperature do not settle end-of-life capacity.

The fixed pairing makes it difficult to enlarge radiator area independently after deployment. Margin could instead come from overbuilding the initial surface, operating at a higher temperature, throttling compute, choosing more durable coatings, replacing modules, servicing the array, or setting a shorter mission life. The patent's thermal break and optional heat pipes address transport within a tile. They do not supply an end-of-life thermal budget. The patent specification describes those local mechanisms without presenting a lifecycle validation.

Deployment remains the scale test

Caltech's DOLCE experiment provides flown evidence for an ultralight deployable structure at 1.8 by 1.8 meters. During deployment, a wire connecting diagonal booms to the structure became snagged and damaged a boom-to-structure connection. After the structure uncoiled, another element jammed beneath the deployment mechanism. The team ultimately completed deployment by vibrating the structure with its actuators. Caltech's mission review records the two anomalies and the completed deployment.

DOLCE covered 3.24 m². Sophia's 60 by 60 meter worked embodiment covers 3,600 m², about 1,100 times as much area before multiple modules are connected. DOLCE shows that the deployment principle can work in orbit and that a snag or jam can escape ground testing. The scale gap remains open.

Sophia Space says it plans a 2027 on-orbit demonstration aboard an Apex Nova satellite bus. The announcement does not describe that mission as a full-scale data center test. It is a company-stated next step, with the thermal and deployment program still ahead.

The unresolved denominator

Across Sophia's disclosure, SpaceX's AI1 page, and the University of Austin model, no source reports sustained end-of-life compute watts per delivered kilogram on a common basis. A useful result would have to include the spacecraft bus, deployment hardware, thermal coatings and margin, radiation mitigation, communications, propulsion, and spares. It would also need to distinguish solar generation from payload input and payload input from sustained useful compute.

That is the measurement that can connect patent geometry to supercomputing infrastructure. The Sophia architecture is coherent: distribute power consumption and heat rejection across the same surface, keep paths local, and assemble scale from repeated modules. Its next evidence should report deployment yield, end-of-life thermal capacity, radiation-adjusted workload performance, and delivered mass together. Without those denominators, megawatts remain design intent.