Orbital Data Centers: The Power Architecture Problem

1. The short answer

In the public architectures reviewed below, the compute payload, the spacecraft, the communication link and the launch vehicle set the initial requirements, and the power subsystem must then deliver the required electrical power within given mass and thermal limits. For this analysis, therefore, electrical power is treated not only as a supporting resource but as an architectural variable: where it comes from, what it weighs, and where the heat associated with it goes.

One of the main factors pushing computing infrastructure toward orbit is not a shortage of annual energy as such, but the speed at which new firm power can be brought to a specific point on the ground. It is not the only one: for processing data generated directly in orbit the motive is different altogether and is unrelated to terrestrial interconnection timelines. But in both cases orbit, while offering energy, immediately takes it back through three accounts: mass per kilowatt, radiator area, and the capacity of the space-to-ground link.

The first account is one of the largest and at the same time one of the most manageable at the architectural level. In the Jet Propulsion Laboratory calculation for a node with 1 MW of useful compute power in a high-illumination orbit, the photovoltaics + storage + radiator subsystems amount to about 29.4 kg per kilowatt, of which 16.9 kg/kW is photovoltaics — more than half. In a shadowed low orbit the photovoltaics-with-storage share rises to roughly 32 kg/kW. Total spacecraft mass is estimated at 34–59 kg/kW.

Hence the precise form of the question addressed to any onboard source that does not depend on solar geometry:

The answer is derived, not postulated. Equality of generation-subsystem mass is reached when the specific power of the replacing source amounts to a fraction \( R \) of the specific power of the photovoltaic system it replaces, where \( R \) is set by four parameters — conversion and distribution efficiency, end-of-life power retention, orbital illumination fraction and storage efficiency. For the baseline case of the cited model \( R \approx 0.74 \); within the parameter ranges declared by that same model \( R \) varies from roughly 0.37 to 0.88. The meaning of the quantity is stable: photovoltaics pays for margin against eclipse, conversion and degradation, whereas an onboard source does not carry that particular form of margin.

Numerically, for the baseline case this is of the order of 22 W/kg against the 30 W/kg typical of the flown fleet, 74 W/kg against the model’s 100 W/kg, and 148 W/kg against the historical maximum of 200 W/kg.

To this is added a correction that almost everyone omits: the losses of an onboard source do not go overboard, they go into the same radiator. Unconverted solar energy is to a large extent reflected or re-radiated directly by the surface of the photovoltaic array and therefore does not enter the thermal circuit of the computing equipment; the internal losses of an onboard converter, by contrast, are removed by the spacecraft thermal control system. With this taken into account the threshold rises by 8–45 % depending on efficiency.

Specific power alone, however, is not sufficient, and this becomes clear at the first attempt to refute the statement. There are two requirements and they are independent: specific power and mission-specific energy — the ability to deliver that power for years without consuming stored mass. The non-solar solutions reviewed pass one requirement and fail the other: reactor and radioisotope systems satisfy the mission-energy requirement but remain at only a few watts per kilogram; electrochemical sources meet the specific-power threshold but consume reactants and, over a multi-year horizon, lose to photovoltaics by three orders of magnitude.

The region in which both requirements are met simultaneously remains unoccupied. That is precisely why it is worth looking into — and precisely why it has to be looked into with a full list of what may fail.

2. What this document covers

The document answers the following questions:

• what economic and physical necessity is pushing computation toward orbit;
• what has already been launched, funded and announced with a date;
• where the real limits of existing solutions lie;
• by what measurable indicator the onboard power source of an orbital node is assessed;
• what risks, constraints and threats arise for a discharge-resonant architecture in vacuum;
• under what conditions the hypothesis turns out to be false;
• what test programme has to be completed before a statement of applicability becomes legitimate.

The following remain outside this analysis:

• the results of the independent metrological validation of VENDOR.Max at the Frame 0 boundary provided for by the project programme — these belong to a separate protocol and its own schedule;
• flight mass, layout and efficiency of a space implementation — these parameters are obtained for the derived architecture;
• the commercial model of a space application.

3. The pain: why computation started looking upward

3.1 Electricity

Global electricity consumption by data centers was estimated at roughly 415 TWh in 2024 — about 1.5 % of global consumption, growing at some 12 % per year over the preceding five years. The base case of the International Energy Agency gives a doubling by 2030, to approximately 945 TWh, and around 1,200 TWh by 2035. In 2024 the United States accounts for about 45 % of global consumption, China about 25 %, Europe about 15 %.

The national picture is harsher. According to the estimate of Lawrence Berkeley National Laboratory, data center consumption in the United States grew from 176 TWh in 2023 and is projected into the 325–580 TWh range by 2028: average load rises from 20.1 GW to 37.1–66.2 GW. In construction terms this means bringing 3.4–9.2 GW of average load online every year.

3.2 The bottleneck is not energy, it is time to power

The key point lost in the discussion: the constraint is not annual energy but the ability to deliver firm power to a specific point within a commercial timeframe. Interconnection queues in Northern Virginia, California and Germany are measured in years. In testimony to the United States Congress, Eric Schmidt put the additional data center requirement by 2030 at 67 GW, comparing it to more than sixty power-generating units.

It is exactly this parameter — time to power — that orbit is trying to bypass. It does not promise a cheaper electron; it promises an electron out of the queue.

3.3 Water, land, permits

Evaporative cooling consumes fresh water in volumes that become a political question in arid regions. Land for a campus, water rights, permitting procedures and interconnection agreements form a multi-year front that capital cannot accelerate. A separate factor is the growing resistance of local communities.

3.4 Money

Planned spending by the four largest hyperscalers on terrestrial infrastructure in 2026 approaches USD 400 billion. The European data center market is valued at roughly EUR 535 billion by 2030. Individual contracts are already measured in hundreds of megawatts of contracted future compute capacity.

Against that background even an expensive orbital alternative stops looking absurd. The question moves from “is it possible” to “at what unit cost”.

4. What has already been built: the map of players, money and dates

The discussion has left the mode of futurological presentations. The correct characterisation of the state of the industry as of August 2026 is that demonstration is turning into early infrastructure formation.

4.1 Spacecraft already operating in orbit

Starcloud (Redmond, Washington; formerly Lumen Orbit) launched the Starcloud-1 spacecraft on 2 November 2025, a vehicle of about 60 kg carrying an NVIDIA H100 accelerator — the first GPU of that class in space. In December 2025 the company reported the first training of a language model in orbit and the running of a Gemma-family model there. In March 2026 it raised USD 170 million at a valuation of USD 1.1 billion — the fastest ascent to unicorn status in the history of Y Combinator. On 21 August 2026 an extension of the round by USD 250 million at a valuation of USD 2.3 billion was announced, led by Manhattan West with participation from NVIDIA and Cisco Investments; total funding reached approximately USD 420–450 million.

As of the date of this source compilation, the company’s nearest plan is as follows: two next-generation spacecraft, Starcloud-2 with 8 kW of compute power each, are to be launched as rideshare payloads in 2027; a larger Starcloud-3 for Starship is being developed in parallel; the company expects to fly the dedicated NVIDIA Space-1 Vera Rubin module around the end of 2028 — the module itself has not yet been built. An FCC filing of 3 February 2026 covers a constellation of up to 88,000 spacecraft; the long-term concept is a 5 GW cluster.

The reason for raising capital named by management is telling: not production, but a shortage of available launch capacity. Falcon 9 is planned for retirement by 2028, rideshare slots beyond late 2028 and early 2029 are unavailable, and competing vehicles fly irregularly. Against its announced long-term targets the company has reserved on the order of 16 kW of compute capacity on 2027 rideshare launches. The bottleneck of the industry for the coming years is not the cost of a kilogram but the availability of a launch slot.

Kepler Communications launched the first tranche of 10 optical relay spacecraft of about 300 kg each on 11 January 2026, every one of them carrying several compute modules and terabytes of memory. The declared next step is optics in the 100 Gbit/s class.

Axiom Space placed two orbital data center nodes on the same launch, developed from the AxDCU-1 prototype that operated on the ISS in the autumn of 2025. The declared roadmap is a transition from kilowatt-class to megawatt-class compute capacity after 2030. Thermal tiles are being developed jointly with Spacebilt.

4.2 Announced with dates and technical characteristics

SpaceX Starmind. The programme has received its own name and a public page. The first generation of the spacecraft is AI1. Declared characteristics: deployed height 30 m, span 75 m, compute load up to 250 kW peak and 175 kW average power, double-sided deployable liquid radiators, sun-synchronous orbit in the altitude range of about 600–800 km, optical links integrated into the Starlink network. One figure published separately matters more than all the others for our subject: a published spacecraft-level compute power-to-mass ratio of 75 kW per tonne.

The June revision of the same spacecraft was more modest — 120 kW average and 150 kW peak power with 110 m² of radiators; the structure was enlarged after the processor was selected. On 4 August 2026 a partnership with NVIDIA was announced: each spacecraft carries Rubin graphics processors and Vera central processors, the Vera Rubin NVL72 platform. Prototype testing is scheduled for early 2027, series production later the same year at the Gigasat facility, and commercial launches at scale from the fourth quarter of 2027. The declared public capacity targets — a gigawatt of orbital compute by the end of 2027, 100 GW by the end of 2029 and a terawatt by the end of 2030 — should be read as a direction rather than a plan: at 150–250 kW per spacecraft they are inconsistent even with the million-spacecraft filing already submitted.

NVIDIA made space computing a separate line of business in 2026 and presented the Space-1 Vera Rubin module with a claimed advantage of up to 25× in inference performance relative to the H100. Named partners include Axiom Space, Planet and Starcloud.

Google Project Suncatcher was announced on 4 November 2025: clusters of spacecraft carrying Trillium TPU v6e tensor processors, connected by free-space optical links. Radiation testing confirmed tolerance to the total dose of a five-year mission in low orbit; the most sensitive element proved to be high-bandwidth memory. A laboratory result of 1.6 Tbit/s per transceiver pair was achieved. Groups of up to 81 spacecraft were modelled at an altitude of about 650 km with separations of 100–200 m. Two prototypes with Planet Labs are planned for launch in early 2027.

Aetherflux raised USD 50 million in a Series A and combines orbital computing with infrared laser power transmission to Earth.

Lonestar Data Holdings pursues a dual strategy — low orbit and the Moon, aiming at placing equipment in lunar lava tubes, where the environment naturally shields against temperature swings and cosmic rays.

4.3 The European line

ASCEND (Advanced Space Cloud for European Net zero emission and Data sovereignty) is a feasibility study coordinated by Thales Alenia Space and funded by the European Commission under Horizon Europe. The consortium: Carbone 4, VITO, Orange Business, CloudFerro, Hewlett Packard Enterprise, ArianeGroup, DLR, Airbus Defence & Space.

Results presented in June 2024: the data center market by 2030 was estimated at 23 GW of installed capacity; the target is 1 GW in orbit by 2050; substantially reducing the carbon footprint would require a launch vehicle with a tenfold lower life-cycle emission. The preliminary conclusions describe a paradigm shift involving several thousand tonnes of infrastructure and hundreds of launches per year.

The architecture is published on the official programme site and deserves a detailed account, since it is the only fully disclosed system architecture of an orbital data center in the public domain.

The minimum viable product is defined at 10 MW of compute power. It is divided into building blocks of 800 kW, each including computing equipment, generation, thermal control, intra-block and inter-block communications, attitude control and robotics. Each block is launched by a single heavy partially reusable vehicle directly into a target orbit at 1,400 km and is deployed by robotic assembly in orbit. Thirteen blocks form the minimum viable product, operating in formation and linked by high-speed optical links. Deployment of the ten-megawatt product is placed at approximately 2035–2036; the gigawatt horizon of 2050 would require on the order of 1,300 blocks. Block geometry is given in public materials as roughly 200 × 80 m.

The programme system analysis gives per block on the order of 4,000 m² of solar panels and 2,000 m² of radiators. In specific form this is 5.0 m² of panels and 2.5 m² of radiator per kilowatt.

ESPI Report 98 (November 2025, European Space Policy Institute together with the Technical University of Munich) contains a cost model of two gigawatt-scale orbital data center architectures: a monolithic one of 96 blocks of 10 MW and a constellation of 6,700 spacecraft of 150 kW. The conclusion: the monolithic architecture reaches comparability with high-end terrestrial data centers provided that computing equipment survives three years or more in orbit and launch cost falls substantially; the constellation architecture remains prohibitively expensive. Three economic barriers are named directly: the cost of computing equipment, its service life under radiation, and launch availability.

4.4 The sceptics — and they matter more

In October 2025, at Italian Tech Week, Jeff Bezos suggested that gigawatt-scale data centers in space would appear within 10 years, and hardly more than 20.

Andrew McCalip, an engineer at Varda Space Industries, published an open calculator that in its base scenario gives roughly a threefold excess in the cost of orbital computing per watt compared with terrestrial.

Wood Mackenzie estimates a gigawatt-scale orbital data center at about USD 170 billion — more than three times the cost of a terrestrial equivalent, with about 60 % of the sum attributable to launch and spacecraft cost.

Boston Consulting Group, in its 2026 analysis, gives a twenty-year total cost of ownership of roughly USD 660–750 million per megawatt for the orbital solution against USD 230–300 million per megawatt for the terrestrial one — a premium of 2.5–3×. Something else matters more for us: BCG builds a sensitivity analysis around four variables, and the second of them is spacecraft mass per 100 kW of compute power. In other words, mass per kilowatt is named as the decisive variable independently by both the academic model and the consulting analysis.

Interim conclusion of this section. Orbital computing has already moved from a purely conceptual field into the domain of funded demonstrators, announced infrastructure programmes and independent techno-economic studies. The dispute has shifted to scale, timelines and the classes of workload capable of closing the economics: BCG, for example, speaks not of replacing terrestrial data centers but of complementary scenarios given the 2.5–3× premium that exists today. And all serious models run into the same triad: mass per kilowatt, heat, downlink.

What is absent from those models matters too. In all the architectures reviewed, the power subsystem is sized to a given computer: compute power is chosen first, and the panel, the storage and the radiator are then calculated for it. In none of the public architectures reviewed here is the power layer introduced as an independent comparison variable whose change itself becomes grounds for rebuilding the platform. The sections that follow test the quantitative conditions under which such a formulation acquires physical meaning at all.

Two levels of description are distinguished from here on. A source is a physical subsystem producing or providing electrical power at a defined boundary; the term is used when the device and its specific characteristics are under discussion. A power layer is the position of that subsystem within the architecture together with the system couplings its parameters create for conversion, distribution, heat rejection, structure and load; the term is used when the effect on the spacecraft mass equation is under discussion.

5. Literature review I: where the physical limit runs

The most complete closed model is the work of S. G. Turyshev (Jet Propulsion Laboratory), preprint arXiv:2604.27197 of 1 May 2026. It formulates the competitiveness condition through three coupled quantities: deployed mass per kilowatt of useful power \( m_{kW} \), the admissible exchange intensity with the ground \( \Gamma \), and the loss of delivered compute-years \( \Pi_{life} \). Its structure is used below as a framework.

5.1 Energy: the real advantage of orbit

The solar constant in orbit is about 1,361 W/m², with no atmospheric absorption or scattering. In a dawn–dusk sun-synchronous orbit, illumination is achieved almost continuously. Google estimates the annual energy yield of panels in a suitable orbit at up to eight times the terrestrial value; independent comparisons give a range of five to thirteen times the annual output.

This advantage is real and is not disputed. What is disputed is its price.

5.2 Heat: the principal design constraint

Here the most widespread myth of the subject has to be destroyed.

“Space is cold, so cooling is easy” is wrong.

Vacuum is an excellent final thermal reservoir and a useless immediate coolant. There is no convection. Inside the spacecraft heat is transported predominantly by conduction and is removed outward exclusively by radiation. A hot board without air stays hotter, temperature gradients become steeper, and different parts of the spacecraft find themselves in very different thermal states at the same time. This is exactly why thermal vacuum testing always combines vacuum with repeated cycling between temperature extremes.

The full heat path in an orbital compute node:

die → package and spreader → cold plate → heat pipe or coolant loop → radiator → photons → space.

And the higher the compute density, the heavier the last segment becomes.

Steady-state rejection is set by the relations:

\[ \dot{Q}_{rej} \simeq P_{tot} = \alpha_{OH} P_{IT} \] \[ A_{rad} \simeq \frac{P_{tot}}{\varepsilon_{rad}\,\eta_{view}\,\sigma_{SB}\,T_{rad}^{4} – q_{env}} \]

The key property is the fourth power of temperature. The only freely controlled variable is area; it comes from Earth and costs mass. At an emissivity of 0.90, a view factor of 0.85 and a radiator temperature of 350 K the specific flux is about 651 W/m²: a megawatt of rejection requires on the order of 1,500 m² even before absorbed external fluxes are taken into account. With a realistic absorption of 150 W/m² the required area grows to 2,500 m².

The scale is easiest to grasp through the ISS: the external active thermal control system rejects about 70 kW through radiator wings of some 7 tonnes. Linear scaling to 1 MW gives about 100 tonnes of radiators alone — while a megawatt of modern compute racks weighs on the order of 10 tonnes. The radiator turns out to be an order of magnitude heavier than what it cools.

The order of magnitude for smaller nodes: according to the BCG estimate, 100 kW of orbital computing requires on the order of 400 m² of radiators. The June revision of AI1 published by SpaceX assumed 110 m² for 120–150 kW; in the August revision, at 250 kW peak power, radiator span was taken to 30 m. Starcloud’s own estimate for a double-sided radiator at about 20 °C is on the order of 633 W/m², roughly a thousand times slower than liquid cooling of accelerators on the ground.

Three additional factors rarely reach presentations:

Surface degradation. Radiator coatings degrade under ultraviolet exposure and atomic oxygen. Over five years in low orbit the optical properties deteriorate, and area has to be sized to the end-of-life condition.

Micrometeoroid penetration. The probability of no penetration is described by a Poisson model, \(P_0 = \exp(-\phi_p A_{rad} T)\). At megawatt-scale areas a single undivided loop is unacceptable; segmentation with isolation adds the mass back.

Local overheating. Recent work from 2026 shows that the limitation of orbital computing is set not only by integral radiator area but with the distribution of hot spots: a conventional accelerator architecture with high-bandwidth memory can enter thermal throttling long before the design compute power is reached. This lifts the thermal problem from subsystem level to die layout level.

5.3 Downlink: a hard filter on workloads

For workloads serving a terrestrial user, the exchange intensity \( \Gamma \equiv D_{sg}/E_{IT} \) is introduced — bits per joule of consumed compute energy. The ceiling is set not by the peak link rate but by the average rate, accounting for visibility geometry, weather, overhead losses and ground station queues.

At a terminal peak rate of 200 Gbit/s, four ground sites and a clear-sky probability of 0.5, the ceiling is about 14.8 GB per kilowatt-hour of compute energy for a 1 MW node — and only 1.48 GB/kWh for a 10 MW node. Interactive services tightly coupled to terrestrial storage are excluded by architecture before capital cost is even discussed.

5.4 Radiation and delivered compute-years

Radiation enters the economics twice: through single events, which reduce effective utilisation via checkpointing and recovery, and through accumulated dose, which limits life and raises failure rates.

Illustrative five-year doses behind a shield of a few millimetres of aluminium equivalent: 3–6 krad(Si) at 550 km mid inclinations, about 5 krad(Si) at 600 km sun-synchronous, 12–20 krad(Si) at 1,200 km high inclinations, 20–30 krad(Si) in geostationary orbit.

The encouraging side: testing of commercial, non-radiation-hardened graphics accelerators showed no irreversible failures at doses up to 60 Gy (6 krad).

The economic translation is given by the life multiplier \( \Pi_{life} = \lambda T / (1 – e^{-\lambda T}) \). At \( \lambda = 0.10 \) per year and a five-year life it equals about 1.27; at \( \lambda = 0.20 \), about 1.58. With equipment obsolescence taken into account the multiplier reaches 1.93.

5.5 Serviceability as an economic quantity

On the ground a faulty unit is replaced. In orbit, failed equipment most often becomes irrecoverable mass.

Industry models of mid-2026 assume a node failure rate of about 9 % per year — that is, replacement of roughly one node in eleven annually. Active constellation projects assume about 20 % spare accelerators and plan not repair but deorbit and replacement of the entire spacecraft. Analytical models attribute the bulk of the cost premium to this line: a five-year asset life against a fifteen-year terrestrial one. Demonstrations of orbital servicing are advancing, but none of them yet addresses component-level replacement in dense compute assemblies.

BCG separately includes failure rate in its sensitivity analysis: even with aggressive reductions in launch cost, the economics remains noticeably sensitive to whether the life-cycle failure share is 30 %, 10 % or 5 %.

5.6 The mass summary — the central table of the review

The baseline calculation case for a node with 1 MW of useful compute power in a high-illumination orbit, with an overhead factor of 1.25, deployed photovoltaic specific power of 100 W/kg, radiator temperature of 350 K and radiator areal density of 5 kg/m²:

Storage falls to zero only because a dawn–dusk orbit without eclipses has been selected. In an ordinary low orbit the picture differs. Calculated on the same relations, with a battery specific energy of 200 Wh/kg, an admissible depth of discharge of 0.8 and a residual capacity of 0.8 at end of life:

5.7 The apparent threefold discrepancy in mass: the boundaries have to be aligned first

Here the most interesting observation of the whole subject has to be recorded — and recorded more carefully than is usually done.

The spread is nearly an order of magnitude. Before declaring it a dispute about achievable specific mass, it is worth checking whether the same quantities are being compared at all.

The JPL model defines the boundary strictly. Mass per kilowatt is the ratio of the entire launched and operated mass of the node to the delivered useful compute power, with all relations interpreted as end-of-life constraints. The boundary includes generation, storage, conversion and distribution, thermal transport and radiators, attitude control, propulsion with station-keeping propellant, communication terminals, avionics, structure and deployment, shielding allowance and compute packaging. That definition is chosen deliberately, so that the metric is set by explicit accounting rather than by silent exclusions.

The SpaceX figure is published without a metrological boundary. It is not publicly stated whether peak or average compute power stands in the numerator; whether dry, fuelled, launch or deployed mass stands in the denominator; whether the figure refers to beginning or end of life; whether redundancy is included; whether propellant is included; what degradation is accounted for; what happens to useful compute power under failures. Arithmetic inversion of the ratio is correct, but it cannot be asserted today that the resulting 13.3 kg/kW is the same quantity that the academic model defines.

And this is not only a question for SpaceX. Within a single published European architecture a discrepancy of the same order is present on its own: block specific mass gives about 41 kg/kW, while the total infrastructure mass of the minimum viable product, exceeding 1,200 tonnes, gives about 115 kg/kW for the same compute power. The cause of the discrepancy is not disclosed in the public mass breakdown. Since the programme materials contain differing descriptions of the transport scheme, the additional hundreds of tonnes cannot be attributed to any specific subsystem without a published mass breakdown. The comparability-of-boundaries problem is general, not directed at anyone.

At the same time, the area estimates converge remarkably well, and that matters. The European architecture gives 5.0 m² of panels and 2.5 m² of radiator per kilowatt; the academic model for its baseline case gives 5.64 and 2.50 m² per kilowatt respectively. Specific radiator area matches exactly, while specific panel area agrees to within twelve per cent. These are two independent sources: a European system architecture and a calculation model built from first principles.

Hence the precise formulation:

Areas converge where they can be compared. Masses diverge by nearly an order of magnitude. The observed discrepancy therefore concerns neither the physics of radiation nor the geometry of collection, but either structural specific mass or the definition of the accounting boundary. Which of the two has not been publicly established.

The academic model and the European system architecture give an order of 40 kg/kW at the full system boundary. SpaceX publishes a substantially more aggressive figure, equivalent to 13.3 kg/kW under literal inversion of the ratio. However, the definitions of mass, power, life and boundary for that metric have not been publicly disclosed. The observed threefold discrepancy is therefore an unresolved question of system boundary comparability, not yet a demonstrated dispute about achievable specific mass.

For our subject this is decisive. If, after boundaries are aligned, 40 kg/kW turns out to be the real level, the niche for an alternative power source is large. If 13 kg/kW turns out to be achievable, the niche contracts almost to zero — not because the alternative is poor, but because the competitor has moved into a region where there is nothing left to catch up with. The question determines whether the problem exists.

5.8 Economic closure

A first approximation of the competitiveness condition:

\[ (L_{\$/kg} + B_{\$/kg})\, m_{kW} + C_{link} + C_{ops} \lesssim C_{terr} \]

At \( m_{kW} \simeq 40 \) kg/kW and a terrestrial infrastructure cost reference of 10,000–40,000 dollars per kilowatt of critical load, the admissible sum of launch and spacecraft construction is only 250–1,000 dollars per kilogram — before communications, operations, utilisation and service life are counted. The published Falcon 9 price list for a dedicated launch to low orbit corresponds to roughly 3,360 dollars per kilogram for launch alone. The gap is a factor of 3.4–13.5, before the spacecraft is even built.

The optimistic side of the same equation is the Google analysis: if launch cost falls below 200 dollars per kilogram by the mid-2030s, the launch cost of a spacecraft with a mass-to-power ratio of the Starlink v2 level, reduced to a kilowatt-year, would be about 810 dollars per kilowatt per year, which falls within the observed terrestrial energy cost range of 570–3,000 dollars per kilowatt per year. The condition for reaching it is sustaining a learning rate on launch cost of about twenty per cent, which requires roughly 180 Starship flights per year.

The intermediate parity references named by different parties diverge: about 500 dollars per kilogram in the model published jointly with Starcloud; about 100 dollars per kilogram in the BCG model; a range of roughly 50–250 dollars per kilogram in summary estimates. BCG also gives an absolute reference: at today’s roughly 1,500 dollars per kilogram, launching a gigawatt of compute would cost about 30 billion dollars for transport alone.

Conclusion of this section. Both sides of the dispute are computing the same equation. Launch cost receives more public attention, yet reducing \( m_{kW} \) is a second independent lever on the same economics — and it is here that academic models, the European system architecture, integrated layouts and the new spacecraft of private operators converge. The derivative is direct: at 40 kg/kW, every 100 dollars per kilogram of reduction in the combined launch and construction cost yields 4,000 dollars per kilowatt. And conversely, every kilogram per kilowatt removed is equivalent to a reduction of the mass load on launch. In the formulation used below, this makes it possible to reduce the first comparison of power architectures to a few measurable coordinates: mass per delivered kilowatt, thermal penalty, and the ability of the system to sustain the required compute work over a given lifetime. The remaining system requirements do not disappear; they return in the full inequality of Section 7.4.

6. Literature review II: what powers orbit today

6.1 Photovoltaics with storage

The only solution in mass use. It is important not to get the comparison base wrong here, so it is given precisely.

The comparison range is set not only by the Jet Propulsion Laboratory calculation model but by the empirical state of flown photovoltaics recorded in NASA reviews. The NASA state-of-the-art review for small spacecraft power gives the following empirical picture: flown missions with solar arrays, regardless of spacecraft mass, cluster tightly around 30 W/kg; no mission has flown below 1 W/kg, and the maximum in the sample is 200 W/kg. Optimised rigid-panel systems approach a plateau of about 70–80 W/kg at beginning of life; flexible blankets promise to go beyond 100 W/kg.

This means that the calculation assumption of 100 W/kg used in the baseline case of the model is already optimistic relative to the flown fleet and corresponds to the upper part of existing practice, not to its middle.

The specific energy of space-grade batteries is 150–265 Wh/kg.

Three structural limitations of photovoltaics:

1. Dependence on geometry. Continuous operation without storage is possible only in a narrow class of orbits. This makes the dawn–dusk sun-synchronous orbit an object of competition: every project wants the same shell.
2. Area. Thousands of square metres per megawatt — a structure that has to be folded, launched, deployed and held in attitude; the same structure sets drag, moment of inertia and the area exposed to debris.
3. Degradation. The ratio of end-of-life to initial power is assumed in the range 0.75–0.95, and oversizing is paid for in mass.

And one non-obvious advantage, critically important for what follows: unconverted solar energy is to a large extent reflected or re-radiated directly by the surface of the array and therefore does not enter the thermal circuit of the computing equipment. The array itself has its own thermal regime, but its own surface balances it. The internal losses of an onboard converter, by contrast, are removed by the spacecraft thermal control system. This distinction is easy to miss in comparison.

6.2 Nuclear sources

Radioisotope thermoelectric generators: about 110 W electrical at a mass of some 45 kg — roughly 2.4 W/kg. Plutonium is becoming more expensive and less available.

Fission reactor systems. The Kilopower project of NASA and the United States Department of Energy: the 1–10 kW electrical class, a solid core of highly enriched uranium, sodium heat pipes, Stirling cycle conversion. The KRUSTY test was completed in March 2018. The estimated mass of a single reactor is on the order of 1.5 tonnes, which gives about 6.7 W/kg at ten kilowatts. The Fission Surface Power programme aims at demonstrating up to 40 kW electrical on the Moon; as of August 2025 the declared goal is to deploy systems of at least 100 kW electrical by 2030 with a life of at least ten years.

The advantage is complete independence from solar geometry and operability through the lunar night and in permanently shadowed craters. The limitations are specific power one to one and a half orders of magnitude below photovoltaics, heavy radiation shielding, nuclear regulation, launch restrictions and a separate thermal load.

6.3 Power transfer in orbit

Star Catcher Industries is building an orbital power grid: spacecraft collect and concentrate sunlight, convert it into a spectrum optimised for the customer’s standard solar panels, and transmit it by beam. In November 2025 a record for optical power transfer was set at the Kennedy Space Center site; delivery of one to ten “suns” onto series single- and triple-junction panels is claimed. In May 2026 a Series A of USD 65 million closed, with USD 88 million raised in total; the first orbital demonstration was planned before the end of 2026. The company’s formulation is precise and worth remembering: today all key space applications — communications, computing, security, sensing — are power-limited.

Aetherflux solves the inverse problem — transmitting power from orbit to Earth by infrared laser.

Both approaches share one limitation: they do not create an autonomous onboard source but redistribute power from external energy infrastructure, and they introduce dependence on a third party, on pointing accuracy and on the state of the network.

6.4 Integrated architectures

A separate line of research in 2026 abandons separate collection, computing and radiating structures in favour of a single tile with photovoltaics on one side, a radiator function on the other and computing placed inside. The approach reduces structural mass and eliminates part of the deployment mechanics. It is a direct competitor to any “separate power source” concept: it attacks the same line of the mass budget.

6.5 Electrochemical sources: why high specific power is not enough

This section was added following a counter-check of the statement about the unoccupied region, and it materially refines the formulation.

The alkaline fuel cell of the Space Shuttle orbiter delivered 12 kW of continuous power and up to 16 kW briefly, at a unit mass of about 113–122 kg. Hardware specific power is about 98–106 W/kg — that is, directly inside the threshold corridor of Section 7.3. Three such units provided all electrical power for the orbiter without backup batteries, at an efficiency above 70 %; cumulative flight time of the family exceeded ninety thousand hours. On specific power, the electrochemical source passes the threshold.

It fails the second requirement. High hardware specific power of a fuel cell does not characterise the complete mission power subsystem: electrical energy is obtained from stored reactants whose mass grows in proportion to the integral of delivered energy. At an efficiency of about 70 % and an output of some 2.7 kWh per kilogram of consumed hydrogen and oxygen, one continuous kilowatt requires:

To this is added service life: the orbiter stack was rated for 2,000–2,600 hours between overhauls, with a prospective target of 5,000 hours — three tenths to six tenths of a year of continuous operation against the ten to fifteen years required.

Hence a refinement of the entire formulation. The selection criterion is not one-dimensional. Two things are required simultaneously:

1. specific power — not below the threshold of Section 7.3 for the selected comparison base;
2. mission-specific energy — the ability to sustain that power over the mission life without consuming stored mass.

Formally, the admissible region is defined by four conditions holding at once: end-of-life specific power not below the threshold; the ratio of energy delivered over the mission to the sum of source mass and consumable inventory mass not below the required value; efficiency not below the minimum of Section 7.3; and service life not below what the mission requires. Up to this point the article has constructed only the first of these conditions; the second is introduced here and is applied to VENDOR.Max on the same terms as to every other architecture — its verification is assigned to Stage 0 of the programme in Section 11.

Photovoltaics satisfies the second condition trivially: the solar flux is not consumed by the spacecraft. Nuclear sources satisfy the second and fail the first. Electrochemical sources satisfy the first and fail the second. Power beaming, at the onboard boundary, is external power supply rather than an onboard source, and transfers both questions to the supplier.

6.6 Summary: the unoccupied region of the parameter space

In the public set of solutions reviewed in Section 6, three families dominate: photovoltaics with storage, nuclear sources — radioisotope and reactor — and power delivery from an external supplier. The review did not identify in that set a family that simultaneously removes dependence on solar geometry, provides specific power of the order of modern photovoltaic systems, and is able to sustain it over a multi-year horizon without a consumable energy inventory whose mass grows with delivered energy. It is precisely this unoccupied region of the parameter space that is examined below.

The bottom row is not a description of VENDOR.Max. In the body of sources reviewed, no flight solution was found that satisfies both conditions at once: the photovoltaic order of specific power and multi-year delivery of energy without a consumable inventory scaling with mission duration. That unoccupied region of the parameter space was outlined by the literature, not by us.

7. Where VENDOR.Max fits into this picture

7.1 What it is

VENDOR.Max is a nonlinear electrodynamic installation operating in a controlled discharge-resonant regime; its generation architecture contains no rotating electromechanical converter and no combustion cycle. Excitation of the coupled resonant regime is assigned, in the functional classification of the project, to discharge-resonant excitation of the Armstrong type; that attribution describes the excitation stage and is not a classification of the installation as a whole.

The protected structural solution concerns the excitation node and is fixed in independent claim 1 of the patent: ES2950176B2 (Granted). The family is continued by WO2024209235A1 (Published) and by national phases: EP4693872A1 (Under examination), US20260088633A1 (Under examination), CN119096463A (Under examination), IN 202547010911 (Under examination).

The architecture is described in its entirety by classical electrodynamics and standard power electronics.

The boundary model is two-level and mandatory.

Frame 0 is the external envelope of the product. Three levels must be strictly distinguished here, and they are not mixed anywhere below.

Observation. In the implemented prototypes of VENDOR.Max, an operating regime was observed in which, after a brief external start — an impulse of about 9 V lasting some ten seconds — the starting source was physically disconnected, and the installation continued to sustain the operating regime and to power the connected load. The regime was reproduced in various laboratory and field conditions and on various loads; the observations are recorded in experimental logs and in photographic and video documentation. This is a result of the assembled system in operation, not a consequence derived from the schematic.

Architectural property. Corresponding to that regime is the absence at Frame 0 of any continuing external power port provided by the design after the start has completed. That, and only that, is a statement about topology.

Metrological task. The next level of validation is the complete quantitative energy balance of Frame 0 and independent attribution of all exchange channels across its external boundary. The question here is not whether the installation operated after the starting source was disconnected: it did. The question is the complete energy attribution of the regime observed. That attribution remains an open metrological task: it has to be closed experimentally at the complete boundary of the installation under a sustained real load. That closure belongs to the verification protocol and is not the subject of this article.

Frame 1 is the boundary of the working core. In steady state a real measurable input flow \( P_{exc,1} > 0 \) passes through it, and its boundary condition is maintained by an internal regime branch. This internal flow must not be conflated with the external supply of Frame 0: the two belong to different boundaries.

Stage. The prototype implemented before the relocation reached the maturity level on which the experimental base of the project rests. After the laboratory was moved, the project is provisionally qualified one level lower, until the prototype has been fully rebuilt on a new component base and the required test cycle has been repeated. This change of status reflects the condition of the current implementation after relocation and does not annul the results and the regimes observed on the previous prototype: cumulative operating time of the previous configurations exceeds one thousand hours and constitutes the experimental base of the project. Project stage: TRL 4 — Prototype Rebuild After Relocation.

The installation is being rebuilt on a new component base; the next cycle provides for restoration of the operating regime, an extended metrological programme, independent validation and subsequent qualification stages. A space implementation, if the programme confirms that it is warranted, is the next separate engineering class of the system, not a transfer of the existing assembly.

The measured integral output in the laboratory configuration is 1.40 kW by load calorimetry; the output interface rating is 3.52 kVA. An end-to-end efficiency cannot be correctly derived from the available data set: the quantities belong to different boundaries and different time classes, and between them lie two storage elements and a closing regime branch.

7.2 Four architectural grounds for inclusion in the conversation

First: the topology of exchange with the environment. The function of fast release of a stored charge state is implemented in a sealed vacuum switching node; the high-voltage switching node of Block 3 is of that type. What matters is that the switching gap is a sealed component with its own internal environment and does not exchange gas with the surrounding atmosphere: it maintains its own sealed vacuum. It does not consume external air as a working medium and does not depend on external barometric pressure as a process parameter.

For the vacuum implementation there is an additional consequence: in space the pressure differential across the envelope of the node falls to zero. On Earth a sealed vacuum device holds the atmosphere outside for its entire operating life; in orbit that load is removed.

Second: the absence of combustion. No oxidiser, no products, no consumable working medium in the conversion path.

Third: the generation architecture contains no rotating electromechanical converter. The conversion path has no rotor, no bearing supports, no dynamic imbalance, no gyroscopic torque acting on spacecraft attitude, and no lubricants with their vacuum incompatibility.

This statement concerns the generation architecture and nothing else. The existing ground implementation contains mechanical components in the cooling system; they take no part in the generation mechanism but are part of the installation. From this follows directly the principal engineering challenge of a space-derived implementation — and it lies not in bearings but in converting the thermal path from convective to conductive-radiative. The task is stated in Section 8.2.

Fourth: direct current as the native form of output. The project canon fixes a regulated DC output as the termination of the physical core; an inverter for alternating current is an additional user interface. The target interface of the telecom implementation is a nominal −48 V DC, normal range −40.5 to −57.0 V, per ETSI EN 300 132-2 V2.8.1. The project doctrine is “48V DC First. 220V AC Later”.

Precedent. Vacuum switching is not exotic for space engineering, and this should be recorded precisely, separating what is confirmed from what is not.

It is confirmed that a triggered vacuum gap was investigated specifically for space power electronics. Report NASA/CR-2002-211562, performed by the Center for Electromechanics of the University of Texas at Austin for Glenn Research Center, established the feasibility of a high-voltage DC-DC converter based on a rod-array triggered vacuum gap for a space solar power system. The motivation is notable in that it coincides with the logic of the present article: the device was selected as a high-voltage element already demonstrated at currents above those required and at substantially higher per-device voltages than semiconductor switches, possessing a significantly higher specific power than any solid-state device, and probably more tolerant of the space environment. An earlier line is Marshall Center contract NAS8-20526, under which a triggered vacuum gap was developed in 1965–1968 with the characteristics of the GL-7703 ignitron but, unlike it, independent of orientation in space.

It is confirmed that the operating point of the node lies inside the experimentally investigated voltage window of the class. The low-voltage firing characteristics of a triggered vacuum gap were investigated by J. Farrall in 1966; the review by J. Lafferty of the same year summarises the behaviour of the class over a wide range. Later experiments with two-sided triggering show stable firing below one hundred volts with a switching time of about half a microsecond in the range from three to eighteen kilovolts. The Marshall Center contract report independently describes a triggering range from several hundred volts to many kilovolts. The operating point of the switching node of Block 3 lies inside that window and is neither anomalous nor alien for the class. This does not confirm correspondence in event energy, repetition rate and service life — those parameters require separate comparison.

It is not confirmed that a production-grade, flight-qualified triggered vacuum gap exists with a rated operating point and parameters comparable to the node of Block 3. Device classes must not be conflated here: sealed aerospace contactors qualified to military standards and used in high-voltage conversion systems cover the required voltage range — but that is an electromechanical contactor, not a triggered vacuum gap.

This is not yet a space architecture. But it is a reasonable initial topology for investigating one — and, as the report cited shows, a topology that space power engineering has already turned to for the same reasons.

7.3 The criterion by which it should be judged

Let us introduce a working analytical coordinate — not an industry standard, but our comparison metric:

\[ m_{power} = \frac{M_{\text{full power subsystem}}}{P_{\text{continuous DC}}} \quad [\text{kg/kW}] \]

The numerator is all the mass that had to be launched so that the bus continuously delivers power: generation, storage, conversion, distribution, structure, deployment, shielding, redundancy. The denominator is continuously available DC power at end of life. The comparison is made at one and the same electrical boundary:

“photovoltaics + storage + power management module” against “power module on the VENDOR architecture”.

From the baseline calculation case, photovoltaics gives 16.9 kg per kilowatt of useful compute power at an overhead factor of 1.25. A replacing source is obliged to fit within the same mass.

Equating the masses does not give a constant — it gives a function. The mass of the photovoltaic subsystem per kilowatt of useful power takes the form

\[ m_{kW,PV} = \frac{1000\,\alpha_{OH}}{\eta_{PMAD}\,SP_{PV}}\left(1 + \frac{1-f_\odot}{f_\odot\,\eta_b}\right)\frac{1}{f_{EOL,PV}} \]

and the mass of the replacing subsystem takes the form \( m_{alt} = 1000\,\alpha_{OH}/SP_{alt} \).

The definition of \( SP_{alt} \) is fixed strictly, otherwise the comparison loses meaning: it is the specific power of the complete alternative power subsystem at the agreed DC boundary, at end of life, referred to its total mass — including conversion, distribution, structure, deployment and its own control, but before the separately accounted growth of the overall thermal subsystem, which is introduced in the next step. Conversion losses and life degradation are therefore already inside \( SP_{alt} \); the quantity \( \eta \) below describes not those but the fraction of supplied energy reaching the bus, and it determines the additional load on the radiator.

Equating gives:

\[ SP_{alt,min} = SP_{PV}\,\frac{\eta_{PMAD}\,f_{EOL,PV}}{1 + \dfrac{1-f_\odot}{f_\odot\,\eta_b}} \equiv R \cdot SP_{PV} \]

The condition under which \( \alpha_{OH} \) cancels. With the same overhead factor for the two architectures being compared, \( \alpha_{OH} \) cancels, and the ratio turns out to be independent of how much the platform spends on its own needs. If, however, replacing the source changes the platform’s own demands — pumping power of the thermal loop, conversion and distribution, attitude control requirements, housekeeping consumption, redundancy — the corresponding difference does not disappear but returns into the full system inequality of Section 7.4.

The threshold ratio of 0.74 is the value of the baseline case of the cited model, not an industry constant. It shifts with changes in illumination fraction, conversion efficiency, storage efficiency and end-of-life power retention — that is, with a change of orbit or of technology generation. The meaning of the quantity nevertheless remains stable: the discount is exactly what photovoltaics pays extra for margin against recharging through eclipse, for conversion and distribution losses, and for degradation at end of life. In a shadowed orbit the relative threshold for a non-solar source falls further still — precisely because the photovoltaic architecture there additionally pays for storing energy for the eclipse period.

Even this is not sufficient. A panel rejects its own losses by itself; an onboard converter does not, and its losses go into the same radiator. If the source has an efficiency \( \eta \) from the supplied form of energy to the bus, the radiator grows as \( 1/\eta \):

\[ \Delta m_{kW,rad} = m_{kW,rad}\left(\frac{1}{\eta} – 1\right) \]

Combining both corrections gives the full criterion, sensitive to the comparison base:

Under different assumptions on illumination and life, the columns shift according to the expression given for \( R \).

The table reads in both directions, and both directions matter.

Left to right. The better the competitor’s photovoltaics becomes, the higher the bar. If, after boundaries are aligned, a series system architecture does confirm a level corresponding to 75 kW/t, the mass budget available to an alternative source narrows sharply and may disappear for the architecture under consideration.

Top to bottom. The lower the efficiency of the source, the faster the growth of the radiator consumes the gain. The admissible mass of the source falls to zero at

\[ \eta_{min} = \frac{m_{kW,rad}}{m_{kW,rad} + m_{kW,PV}} \]

This is not a universal cut-off but a function of the comparison base: the lighter the photovoltaics being replaced, the higher the efficiency required of the source.

As efficiency falls, the region of admissible specific power narrows quickly; the specific lower bound is set jointly by source mass, radiator areal density and operating temperature, and the mass of the generation-plus-storage being replaced.

Where this pair of quantities already exists as a formal framework. The pair “efficiency — gravimetric power density” was not invented here. In power electronics it constitutes a standard performance space in which multi-objective optimisation is carried out: the indices \( \eta \), volumetric power density and gravimetric power density \( \gamma \) in kW/kg form an achievability space whose boundary is the Pareto front of best trade-offs. The school of ETH Zurich develops this formalism and publishes demonstrators against it: 15 kW class DC-DC converters on silicon carbide at 44 kW/kg with an average efficiency of 98.8 %, and on gallium nitride at 86 kW/kg, while optimisation estimates give above 62 kW/kg at an average efficiency above 98.5 %.

In those terms, the threshold relation obtained above is a requirement line in the same space in which achievability is conventionally described. The experimentally demonstrated gravimetric density of the conversion stage specifically lies two to three orders of magnitude above a threshold that refers to the complete power subsystem. This shows that the mass of modern power electronics itself potentially has considerable margin relative to the first order of the mass budget under consideration. However, the results of fifteen-kilowatt ground demonstrators cannot be transferred linearly to a megawatt-class flight power management subsystem: scaling, insulation, redundancy, radiation tolerance and thermal integration change the system boundary. What follows from this is therefore not a ready mass for a space converter but a narrower conclusion: the conversion function itself does not appear to be a dominant mass item in advance, and this must be verified at the boundary of a flight implementation.

This is the technical specification. Not “does it work or not”, but “how many watts per kilogram at what efficiency, against what base and at what accounting boundary”. For VENDOR.Max none of these quantities is defined today in a flight formulation. We publish the criterion and its derivation, not compliance with it.

7.4 The full inequality

Strictly speaking, the mass equation of an orbital node has the form:

\[ M_{sys} = M_{compute} + M_{solar} + M_{storage} + M_{power} + M_{thermal} + M_{struct} + M_{comms} + M_{prop} + M_{shield} + M_{red} \]

An architecture built on an onboard source makes sense only if:

\[ \Delta\left(M_{solar} + M_{storage} + M_{power}\right) > M_{source} + M_{inventory}(T_{mission}) + \Delta M_{shield} + \Delta M_{thermal} + \Delta M_{red} \]

None of the terms on the right-hand side may be omitted, and it is precisely these that are most often omitted in reasoning of this kind.

\( M_{inventory}(T_{mission}) \) is the mass of the consumable or finite energy inventory required to deliver the required energy over the given mission duration. For an architecture powered by an external energy flow and not consuming an onboard inventory proportional to delivered energy, this term equals zero. It is this term, and not the mass of the hardware, that settles the question of the electrochemical sources of Section 6.5: their \( M_{source} \) is small, while their \( M_{inventory} \) over a fifteen-year horizon exceeds the mass of the photovoltaics being replaced by three orders of magnitude.

\( \Delta M_{shield} \) is shielding for the source’s control electronics, which is several times more radiation-sensitive than the power physics.

\( \Delta M_{thermal} \) is the radiator growth calculated above.

The inverse relation between environments is examined in Section 7.10.

\( \Delta M_{red} \) is redundancy. With no servicing available, the power source is a single point of failure for the entire spacecraft; a photovoltaic panel degrades gradually and almost never fails instantaneously as a whole. That is a structural advantage of the panel, which will have to be paid for with an N+1 or N+2 scheme.

7.5 What removing the solar tie provides

Four consequences, and all four are architectural.

First — freedom of orbit. The dawn–dusk sun-synchronous orbit is a narrow shell that all projects compete for at once; SpaceX targets it for a million spacecraft, and Google for its constellations. A source indifferent to eclipse removes that constraint: arbitrary inclinations and altitudes become available, and the difference in mass load between orbits disappears.

Second — the disappearance of storage cycling. In a shadowed low orbit an orbital period of about 1.59 h gives some 5,509 revolutions per year and about 82,600 charge-discharge cycles over fifteen years. It is that quantity, rather than nominal capacity, which determines admissible depth of discharge and residual capacity at end of life.

Third — environments locally deprived of solar access. First of all the permanently shadowed polar regions of the Moon, where local photovoltaics does not exist as an immediate source. A separate but related case is the lunar night of about fourteen Earth days: there solar generation exists as a class, but requires storage or power transfer on a scale at which the mass budget is no longer settled by the panel. In both formulations the practically applicable alternative today is fission with a specific power of about 6.7 W/kg. Against such a base the threshold from the table in Section 7.3 falls to single watts per kilogram. That, and not low orbit, is the environment where the bar is lowest.

Fourth — alignment of interfaces. A spacecraft is a distributed direct-current network; the terrestrial data center is moving increasingly toward direct current. For an orbital node there is no sense in first building a terrestrial alternating-current architecture and then turning it back into numerous DC buses inside the server. The correct chain is “source → DC bus → point-of-load conversion → accelerator rails”. The absence of terrestrial AC conversion infrastructure is named directly as one of the factors potentially reducing the cost of an orbital architecture. Our product doctrine coincided with this independently and in advance.

7.6 The double weight of a kilowatt of losses

A separate system effect worth stating explicitly, because it works in our favour and is usually not counted.

On Earth an extra kilowatt of losses in the power path costs one kilowatt of fuel. In space it costs twice:

1. it had to be generated — hence added photovoltaic area and mass;
2. it then has to be removed through the radiator — hence added radiator area and mass.

It follows that a loss eliminated sufficiently high up the energy path is capable of reducing two mass items at once — required generation and required heat rejection. Its system value may therefore exceed the nominal one watt of electrical power; the size of that excess depends on exactly where the loss arises and which system boundary is under consideration.

This does not cancel the principal limitation of the next section, but it explains why a power architecture with fewer conversion stages deserves a separate account.

7.7 What removing the solar tie does not provide

Here it is necessary to be equally precise.

• The radiator does not disappear. It remains in full and grows by the losses of the source. Practically all electrical power consumed by computation ultimately becomes heat: \( P_{heat} \approx P_{compute} + P_{power\,losses} + P_{aux} \). A megawatt of compute load requires roughly a megawatt of thermal power to be radiated into space, plus infrastructure losses. Even an ideal source of zero mass does not repeal the Stefan–Boltzmann law. This has to be said plainly in the text, otherwise the text turns into advertising.
• The downlink does not widen. The constraint \( \Gamma \le \Gamma_{max} \) concerns the workload and the communication link and has nothing to do with the power source.
• Radiation does not decrease. Dose and single-event effects act on the compute payload regardless of where the watts came from.
• The economics does not close. This is the main honest result of the section.

Take the limiting case: a source of zero mass. Then \( m_{kW} \) falls from 40 to 23.1 kg/kW, and the admissible sum of launch and construction grows by exactly a factor of 1.73 — from 250–1,000 to 433–1,732 dollars per kilogram. The gap against the published Falcon 9 reference narrows from 3.4–13.5 to 1.9–7.8 times.

That is, even a weightless power source does not by itself close the economics of an orbital data center for a terrestrial user. It shifts the boundary by about a factor of 1.7. Closure requires simultaneous movement on launch cost, exchange intensity, effective utilisation and service life. No lever in this problem is sufficient on its own — and ours is no exception.

Where a shift of 1.7× does decide the outcome is in processing data generated in space, and in computing embedded in a relay or broadband constellation. There the orbital location itself creates the value, and the competitiveness threshold is lower. It is precisely these regimes that the literature calls realistic on the near horizon.

7.8 The derived architecture and the localisation of genuinely new development

A space implementation is not the same installation in a different casing. It requires its own component base, its own design, its own qualification and its own documentation. But it does not follow that most technology classes would have to be created anew — and that distinction matters more than the statement itself.

The classes of electrical components from which the VENDOR.Max architecture is built are normal objects of aerospace component engineering. The NASA parts selection list for electrical, electronic and electromechanical components contains them as separate product classes: capacitors, inductors, transformers, relays, connectors, diodes and transistors, resistors, wire and cable. For high-voltage ceramic capacitors there is a dedicated specification, MIL-PRF-49467, with ratings of 600 V, 1, 2 and 3 kV. For power switches the position is even more definite: radiation-hardened superjunction field-effect transistors qualified by the European Space Agency are produced in the range from 60 to 650 V, certified to ESCC-5000, with total dose tolerance up to 100 krad(Si) and single-event testing to a linear energy transfer of 95 MeV·cm²/mg. The declared applications are DC-DC converters, switched-mode power supplies and ion thrusters, with a declared service life above fifteen years. The devices are manufactured in the European Union and are not subject to export control, which has practical significance for a rights holder based in Romania.

The question of a space version therefore begins, in most lines, not with “how do we invent a component for space” but with “does a qualified device exist with the required voltage, current, on-resistance, gate charge, switching speed and radiation characteristics — and if not, how much must the stage change”.

Most lines belong to selection, adaptation and qualification. There are four heavy lines, and they are named in advance: the service life of vacuum switching, the geometry of high-voltage and high-frequency insulation, thermal architecture, and possibly the resonant magnetic system at the required combination of frequency, power and mass.

Hence the correct statement of the task:

derived architecture = existing functional architecture + substitution of aerospace components + rework for the specific application + qualification,

rather than “a new technology from scratch”.

And hence a distinction worth keeping separate. Qualification risk and fundamental technology risk are not the same thing. If a ground capacitor fails on outgassing or radiation while the required pulse class exists in a space implementation, that is a task of substitution and recalculation. If a field-effect transistor fails on single events, a radiation-hardened family is sought or the stage topology is changed. If a transformer varnish is incompatible with vacuum, the material stack is rebuilt. But if it turns out that no existing vacuum switching technology withstands the required combination of event energy, repetition rate and service life — that is component development, and it belongs to a different risk class and a different time scale.

The correct formulation of the initial task therefore reads: the same functional architecture, transferred to an aerospace implementation, with the zones of genuinely new development localised in advance.

7.9 The correct entry point is not a gigawatt

One more consequence, important for strategy.

Starting the conversation with a gigawatt-scale orbital hyperscaler is wrong. The market is actually climbing from below: a demonstrator with a single accelerator; a 100 kW class spacecraft; kilowatt-class Axiom nodes with a declared transition to megawatts after 2030; the European ten-megawatt minimum viable product; the SpaceX spacecraft with 175 kW of average compute power.

Today’s entry point is a node of 10–100 kW, not a megawatt or gigawatt complex. That is the power class in which modular architecture is the industry norm rather than the exception. This circumstance has a methodological consequence as well: verifying a significant part of the decisive parameters does not require a megawatt demonstrator. Specific power, efficiency, thermal penalty, switching life and regime stability can be determined sequentially on ground installations before moving to a full-scale orbital platform.

At the same time, the contradiction recorded by the European cost model must be kept in mind: at gigawatt scale, constellations of many small nodes turn out to be substantially more expensive than monolithic blocks. That is, modularity is advantageous at entry and works against the economics at exit. The resolution of this contradiction is not obvious.

7.10 One source, different system weights

The mass of one and the same power node has unequal significance in a terrestrial and in an orbital system. For a terrestrial architecture the source’s own mass is usually not the first constraint; greater weight may attach to footprint, distribution topology, grid access, cooling and distance to the load. In an orbital system that mass enters the launch budget directly, and the geometry of the source is coupled to deployment, attitude control and load-bearing structure.

The quantity \( m_{kW} \) therefore cannot be interpreted outside the environment of application. A reduction of \( m_{kW} \) that on Earth changes only one characteristic of the equipment is capable, in orbit, of changing several coupled terms of the system inequality of Section 7.4. How large that secondary effect is can be established only for a specific architecture, and it must not be added to the mass gain without a separate calculation.

8. Constraints, risks and threats: the honest account

This section is written to be read by a sceptic. The points are ordered by decreasing severity. Each has a defined means of closure and a required class of experimental work; the correspondence between the points and the programme stages is given in Section 11.

8.1 Switching life — the dominant question

A space mission requires years of continuous operation without scheduled servicing. The service life of a high-voltage switching element is determined not by a single number of operations but by a combination: single-event energy, current loading, condition and material of the electrodes, the medium of the gap, admissible drift of the threshold and admissible displacement of the spectrum relative to resonance.

The nearest class of devices with published life data and real space application is the vacuum arc thruster of small satellites. This is a repetitive vacuum discharge operating in orbit, and its life statistics are instructive. Published results give on the order of a million pulses for configurations with fixed electrodes and above ten million for systems with active cathode feeding. In every case the limiting factor is named as erosion of the electrodes and the insulator, and the accepted industry solution is not increased material resistance but geometry compensation: cathode feeding that maintains an unchanged interelectrode gap throughout the service life. Switching engineering knows a related technique: the multi-rod structure of a triggered vacuum gap reduces current density at the electrode surface and increases the charge transferred per event, which the authors relate directly to service life.

This gives a correct frame of reference. The service life of a repetitive vacuum discharge demonstrated by the industry in space applications lies in the region from a million to tens of millions of events. The requirement of a multi-year unserviced mission for our node is substantially higher, and the gap should be named rather than avoided.

For VENDOR.Max the corresponding life budget has not been established experimentally.

The mitigating circumstance should be named honestly. The protected solution of claim 1 of the patent provides for several parallel channels with mutually differing firing thresholds and a sectioned storage; with N channels the load on each electrode set is divided by N, and the overlap of the displaced spectra passively holds spectral density near resonance as the electrodes wear — without a drift sensor and without a tuning loop. Notably, this is the same principle that thruster engineering arrived at, but implemented passively: there the geometry is maintained by mechanical feeding, here spectral density is held by the overlap of displaced thresholds. In addition, the required number of events scales with the selected operating point.

But mitigation is not a solution.

The service life of the switching node in the absence of orbital servicing is a potentially blocking variable: it is capable of closing the subject entirely, regardless of every other merit of the architecture. Until accelerated life testing with measurement of electrode erosion and threshold drift has been performed, a ten- or fifteen-year service life cannot be claimed.

8.2 The thermal path: the principal task of the space-derived implementation

Among the measurement planes of the installation there is a plane \( \Gamma_{vent} \) — the ventilation path, characterised by mass flow, heat capacity and inlet and outlet temperatures. This states directly: the current laboratory configuration uses a convective air channel for heat removal.

In vacuum that channel disappears completely. The entire heat path must be reworked as conductive: thermal interfaces, spreaders, cold plates, heat pipes or coolant loops, and then a radiator. This is not adaptation but a redesign of the whole thermal part — with its own mass budget, its own failure modes (loop rupture, coolant loss, pump stoppage) and its own pumping power, which returns into the overhead factor.

High-voltage and high-frequency nodes complicate the task: they sit poorly with a metallic conductive path that must provide electrical insulation and low thermal resistance at the same time.

It is also material that space is not a cold heat sink in the everyday sense. Steady-state rejection, written in Section 5.2 through the absorbed external flux \( q_{env} \), admits an equivalent representation through the effective temperature of the radiative sink:

\[ P_{emit} = \varepsilon\,\sigma\,A\left(T_{rad}^{4} – T_{sink,eff}^{4}\right) \]

where \( T_{sink,eff} \) is not the temperature of any physical body but an equivalent representation of the total external radiative environment, already folded from view factor, solar irradiance, Earth’s own emission, albedo, spacecraft attitude and eclipse. It is the same model as in Section 5.2, written in a form convenient for analysing orbital states. One and the same radiator design has different performance on the sunlit and the eclipsed segment of the orbit and in the transitions between them.

The central engineering task of the space-derived implementation is therefore not the removal of mechanical elements from the installation but the transfer of the thermal path to a conductive-radiative architecture and the determination of the admissible temperature regime in all orbital states — under solar illumination, in eclipse and in transient thermal regimes.

The measurable quantities of this task: dissipated power, efficiency, component junction temperatures, thermal resistance of the conductive path, required radiator area and the transient thermal envelope. The link to the mass budget is already established in Section 7.3 by the relation \( \Delta m_{kW,rad} = m_{kW,rad}(1/\eta – 1) \): measuring efficiency in the reworked thermal architecture directly determines the position of the row in the threshold table.

8.3 High voltage and high frequency in vacuum

Here lies the most interesting and the most unpleasant conclusion of the whole analysis.

Intuition suggests: no air, so discharge is harder. For our class of system that is wrong. The combination of high frequency, high field, resonant structures, insulating gaps, surfaces and vacuum creates a new class of electrical risk that does not exist on Earth.

Multipactor. Resonant secondary-emission electron multiplication: electrons synchronise with the high-frequency field, strike metal surfaces, knock out secondary electrons and form a growing avalanche. The phenomenon is so specific and so dangerous for high-power high-frequency equipment in high vacuum that the European Space Agency maintains a dedicated standard for it, ECSS-E-ST-20-01C “Multipactor design and test”. The Jet Propulsion Laboratory maintains a specialised facility for detecting multipaction and ionisation breakdown in vacuum.

The Paschen curve and the outgassing period. The most dangerous condition is not deep vacuum but intermediate pressures near the Paschen minimum. Materials release gas during the first weeks after launch, locally raising pressure inside cavities. It has been shown experimentally that outgassing reduces breakdown voltage: for cable insulation samples at 125 °C a reduction of nearly 300 V at 0.01 Pa·m was recorded, about 30 % relative to the background value. Standard practice is to delay the application of high voltage until outgassing is complete, to use vented construction without closed cavities, and to bake out on the ground.

Triple points. Metal–solid dielectric–vacuum junctions are the classic initiation site for surface flashover.

Surface flashover along the dielectric replaces bulk breakdown as the dominant mechanism in vacuum.

Spacecraft charging. Low-orbit plasma, differential charging of surfaces and deep charging of dielectrics create potentials superimposed on the installation’s own potentials.

Hence a paradox that should be stated openly:

The absence of dependence on atmospheric air makes space adaptation conceptually admissible. But a high-frequency resonant architecture simultaneously makes vacuum electrical qualification one of the most serious project-specific risks.

This formulation is considerably more valuable than a marketing “works in vacuum” — and considerably closer to the truth.

None of the listed mechanisms is a fundamental prohibition: all of them are routinely handled in high-voltage space engineering. But all of them mean redesign of the insulation system, not transfer of the existing one.

Taking the precedents of Section 7.2 into account, the question of this section should be posed precisely. It is not whether the principle of vacuum discharge switching can exist in space — research and thruster practice show that it can. It is how far the specific geometry, repetition rate, event energy and required service life of the node of Block 3 correspond to the already existing aerospace class of vacuum switching — and which of its solutions, including compensation of electrode geometry, are transferable to our design.

8.4 Discharge into air in the current regime

In the adopted regulation model it is established that below 3.3 kV a streamer does not ignite at a threshold of 6 kV, and at low load the machine operates without discharge into air. This directly means that in the upper part of the amplitude range, in the current ground configuration, discharge into air is present.

Hence a legitimate and unpleasant question that must be asked before any reasoning about space: is that discharge a parasitic loss channel — or does it participate in forming the regime? If the former, its elimination in vacuum improves the balance. If the latter, removing the atmosphere changes the regime, and all ground results require re-acquisition.

The answer cannot be given analytically. It is given by a comparative run of the same assembly in atmosphere and in a thermal vacuum chamber with synchronous recording on all planes. Until that run, nothing can be asserted about the applicability of the architecture in vacuum, in either the positive or the negative direction.

8.5 Thermal cycling and resonance drift

In a shadowed low orbit there are some 82,600 cycles over fifteen years. For an assembly containing sealed switching nodes, ceramics, soldered and pressed joints, a planar transformer and storage elements, this means fatigue loading through coefficients of thermal expansion across the entire service life.

But for a resonant architecture there is an additional, specific mechanism: thermal expansion changes geometry, and geometry changes the resonant frequency. Inductances, capacitances, gaps and the permittivity of insulation are all temperature-dependent. The question reads: does resonance drift over the full range of orbital temperatures remain inside the control range, or does the system leave the limits within which the regulator is able to hold the regime?

A paradox worth noting: removing the solar tie opens shadowed orbits — and by doing so introduces thermal cycling that would not exist in a dawn–dusk orbit. A gain in one place is paid for by a requirement in another.

8.6 Radiation: the weak point is not where it seems

There are three fundamental classes of effect on electronics: single events, total accumulated dose and displacement effects. The consequences range from a bit flip and parameter degradation to destructive latch-up and total loss of the device.

For VENDOR.Max the asymmetry is material. The power physics is able to tolerate radiation well. A small control microcircuit is able to kill the entire node. Google’s experience with tensor processors shows the same structure: the processor itself withstood the dose of a five-year mission, while memory proved the most sensitive element.

The architectural consequence: radiation-hardened control electronics, watchdog circuits, error detection and correction, duplicated controllers, fault isolation, possibly cold redundancy and a reconfigurable topology will be required.

This is a requirement for the derived architecture, not a statement about the ground one. For a long unserviced mission, a space implementation will require diagnosable and fault-tolerant monitoring of the parameters of the resonant regime, including feedback sufficient to compensate temperature, life and component drift, and to bring the node into a safe state on sensor failure.

8.7 Materials: a separate science

Low orbit is not “emptiness” in the everyday sense. Atomic oxygen, ultraviolet and vacuum ultraviolet, charged particles, thermal cycling and debris are all present there.

Atomic oxygen destroys organic materials, changes the thermal and optical properties of surfaces, and causes erosion and cracking. NASA maintains a dedicated current handbook on the atomic-oxygen durability of space polymers, including data from years of material exposure on the external surface of the ISS.

Materially, this is a task of substitution and qualification rather than of creating new material classes: the corresponding space product ranges exist, and non-conformities are localised by the table of Section 7.8. Outgassing requirements: total mass loss and collected volatile condensable material. Polymers, compounds, impregnants and varnishes standard for terrestrial high-voltage engineering largely fail these criteria. Contamination by condensate is doubly critical: it degrades both the optical properties of the radiator and the dielectric strength of the insulation.

8.8 Launch

Vibroacoustic loads, shock during separation, quasi-static accelerations. Separately — depressurisation: the rate of pressure decay in the payload bay constrains the design of any cavity; enclosed volumes are either vented or designed for the differential. For an assembly with sealed nodes and high-voltage insulation this is a direct design requirement.

8.9 Electromagnetic compatibility

A pulsed source with nanosecond edges, placed on a spacecraft next to optical terminals, receivers, the attitude control system and a high-speed digital payload, is a serious compatibility problem. Conducted and radiated emissions, common mode, chassis currents, shielding. On Earth this is solved partly with mass; in space, mass is what is being counted.

8.10 No servicing, and redundancy

Servicing in orbit is not assumed. A space VENDOR must be conceived not as a single unit but as N modules with isolation of a damaged one, automatic reconfiguration, partial loss of power without loss of the spacecraft, power margin for degradation, and localisation of failures.

8.11 Regulation, debris, export

A separate licensing category of “space data center” does not exist. Environmental requirements for mega-constellations at the proposed scale are undefined. Debris mitigation rules were not designed for infrastructure with a five-year equipment replacement cycle. For a spacecraft with a high-voltage power installation, questions of radio-frequency coordination through the ITU are added. For a European rights holder, a dual-use export control loop is added.

8.12 Parameters determined by the space programme

The list below is not a list of gaps but the composition of the next stage of work. Each line has its own measurable criterion and its own pass/fail result.

• The Frame 0 boundary — the metrological basis, fixed before the architecture is transferred to another environment.
• Atmosphere against vacuum — a comparative experiment on one and the same assembly.
• Switching life — the mission life budget at fixed event energy and current loading.
• Flight mass and layout — the design of the derived architecture of a space implementation.
• Radiation tolerance — qualification of the control and power components.
• Temperature drift of resonance — thermal vacuum testing with tuning monitored.
• The restored assembly — the initial configuration of the next test cycle.

The order and content of this work are set out in Section 11.

9. The opportunity–counterforce matrix

This table is what distinguishes a conscientious analysis from a presentation.

10. Conditions that falsify the hypothesis

A hypothesis must be refutable. Below are the conditions under which, if any one of them holds, an architecture of the VENDOR class is not the best solution for orbital power and the work should be stopped.

F-1. The sum “source mass + mission energy inventory + shielding + radiator growth + redundancy” exceeds the mass of the photovoltaic system with storage that it replaces, at the same electrical boundary.

F-2. Multipactor in the resonant structures is not suppressed by design at the operating amplitudes and frequencies.

F-3. Discharge into air in the ground configuration turns out to be regime-forming, and the regime is not reproduced in vacuum.

F-4. Temperature drift of resonance over the full range of orbital temperatures takes the system outside the control range.

F-5. The service life of the switching node does not reach the required number of events even with reasonable channel multiplicity and an admissible reduction of the operating point.

F-6. The efficiency of the source turns out to be below the value \( \eta_{min} \) defined in Section 7.3 for the actual comparison base — that is, radiator growth absorbs the entire admissible source mass.

F-7. After accounting boundaries are aligned, a level of system specific mass of the competing photovoltaic architecture is confirmed at which VENDOR.Max cannot satisfy condition F-1 at the measured values of its specific power, efficiency and required redundancy.

F-8. The redundancy required for the control electronics makes the power node more complex and less reliable than a gradually degrading panel.

F-9. After the Frame 0 energy balance is closed, it turns out that sustaining the required power over the mission horizon requires a consumable or finite energy inventory whose mass grows with delivered energy and eliminates the mass advantage over the baseline architecture.

Each condition corresponds to a specific experiment or calculation stage of the programme of Section 11; the sequential closure of these conditions is what determines the decision to continue or to terminate work on a space implementation.

10.1 The asymmetry of sequential verification

A significant part of the decisive parameters can be determined before space qualification. This creates a useful asymmetry in the programme: an early negative result is able to close the corresponding branch of the hypothesis before the more expensive stages, whereas moving to the next level is permitted only after an already defined criterion has been passed.

A positive result at any single stage does not confirm the applicability of the architecture as a whole. It removes one uncertainty only. System significance arises only when the thresholds are met simultaneously on specific power, mission-specific energy, efficiency and thermal penalty, service life, regime stability and subsequent environmental qualification.

The rationality of the investigation is therefore determined not by an assumption of a positive result but by the ability to exclude unsuitable solutions sequentially, ahead of the most expensive verification stages.

11. The programme for transition to a space implementation

The order follows from the logic: each subsequent stage makes sense only if the previous one has a positive outcome. Each stage removes a defined uncertainty and ends either in a measurement result or in a criterion for transition to the next stage; what is stated below is therefore not only the content of the work but also which quantity ceases to be an assumption after it.

Stage 0. Restoration and ground closure. Completion of the rebuild. Full inventory of Frame 0 under the extended protocol, with synchronous recording on all planes and closure of the balance residual within the declared uncertainty. Without this stage the space subject does not exist. Determined after this stage: the metrological basis of the Frame 0 boundary, the balance residual with the declared uncertainty and, separately, whether a consumable or finite energy inventory scaling with the integral of delivered energy exists; and if it does, its mission mass cost.

Stage 1. Comparative run, atmosphere against vacuum. The same assembly, two environments, identical methodology. A sweep across the critical pressure range through the Paschen minimum with partial discharge recording. Determined after this stage: whether or not the operating regime is retained when the atmosphere is removed.

Stage 2. High-voltage and high-frequency qualification in vacuum. Multipactor per ECSS-E-ST-20-01C, triple points, surface flashover, outgassing with ground bake-out and delayed voltage application per NASA-HDBK-4007 and ECSS-E-HB-20-05A. Determined after this stage: the dielectric strength margins of the insulation system in vacuum and the applicability limits of the chosen geometry.

Stage 3. Rework of the thermal part. Replacement of the convective path by a conductive one. A thermal model with calculation of the required radiator area at the real efficiency. Determined after this stage: the value of \( \eta \), and hence the radiator growth and the position of the row in the table of Section 7.3.

Stage 4. Switching life testing. Accelerated operation with measurement of electrode erosion, firing threshold drift and spectral density near resonance. The goal is extrapolation to the full number of mission events with a declared confidence interval, at fixed event energy and current loading. Determined after this stage: the life budget of the switching node with a confidence interval.

Stage 5. Thermal cycling with resonance monitoring. Not only mechanical fatigue but retention of tuning and controllability across the full temperature range. Determined after this stage: the temperature drift of the resonant tuning and the limits within which the regulator holds it.

Stage 6. Radiation characterisation. Total dose and single events for supervision and power switches; separately, a check for latch-up. Determined after this stage: the radiation margin of the control and power sections for the selected orbit and lifetime.

Stage 7. Materials and compatibility. Outgassing, atomic oxygen, ultraviolet, insulation ageing; complete rework of the list of materials used. Determined after this stage: an approved list of space-implementation materials and their behaviour in the environment.

Stage 8. Mechanical qualification and compatibility. Vibration, acoustics, shock, depressurisation; electromagnetic compatibility. Determined after this stage: the mechanical margins and compatibility with the accompanying equipment of the spacecraft.

Stage 9. Determination of flight-specific power. Only here does a number appear that can be compared with the threshold from the table in Section 7.3. Not earlier. Determined after this stage: flight-specific power and its position relative to the threshold.

Stage 10. Comparative system modelling. The baseline “photovoltaics + storage + power management module” against the hypothesis, on kg/kW, m²/kW, redundancy, service life, thermal load and replacement requirements. Determined after this stage: whether the hypothesis changes the system inequality of Section 7.4 in a favourable direction.

Stage 11. Demonstration payload in low orbit. Not the power source of the spacecraft, but an experimental module on housekeeping power, recording the behaviour of the node in the real environment. Determined after this stage: whether or not the transfer of ground results into the real environment is confirmed.

The indicative duration is years, not quarters. The competences and equipment required exceed our current capabilities and imply partnership with a qualified space laboratory.

12. What is established and what is determined by experiment

We assert:

1. Orbital computing has moved from the domain of concepts into the domain of working equipment, filed applications and funded programmes. This is confirmed independently of us.
2. The power supply problem of an orbital compute node is not closed by existing solutions: photovoltaics accounts for more than half of subsystem mass and ties the architecture to a narrow class of orbits.
3. The requirement for an alternative onboard source can be stated quantitatively: approximately 74 % of the specific power of the photovoltaic system being replaced, corrected for radiator growth from the source’s own losses.
4. The criterion is two-dimensional: specific power and mission-specific energy. The non-solar solutions reviewed pass one requirement and fail the other; the region where both hold remains unoccupied.
5. The switching function of the VENDOR.Max architecture is implemented in a sealed node and does not depend on the external atmosphere as a working medium; the generation architecture contains no combustion cycle and no rotating converter, and the physical core terminates in a regulated DC output. The mechanical elements of the existing ground implementation belong to the cooling system and are replaced in the space-derived implementation by a conductive-radiative scheme.
6. No physical or circuit-level prohibition against stating such a problem was found in the literature.
7. An architecture of this class is capable of replacing existing onboard power-supply points within an orbital node — provided that the tests of Section 11 are passed successfully and none of the falsification conditions of Section 10 is met.

What must be determined experimentally:

1. The behaviour of the architecture in vacuum, microgravity and a radiation environment.
2. Flight mass, layout, efficiency and specific power of a space implementation.
3. The life budget of the switching node over the mission horizon.
4. The existence and mission mass cost of a consumable or finite energy inventory, should one be found when the Frame 0 balance is closed.
5. Retention of the operating regime when the atmosphere is removed.
6. Vacuum dielectric strength of the resonant structures and the insulation system.
7. Temperature drift of the resonant tuning over the full range of orbital temperatures.

Two conclusions are already established and need no verification: the absence of a requirement for atmospheric air removes one barrier and introduces another, and an onboard power source by itself does not close the economics of an orbital data center for a terrestrial user — it shifts the boundary by about a factor of 1.7. A third: the same change in the mass of a power subsystem has different system consequences in a terrestrial and in an orbital architecture, because in an orbital system mass is directly linked to launch and may change the requirements on coupled structural subsystems.

13. Conclusions

First. Orbit attracts computation not by cheap energy but by bypassing the queue for power connection.

Second. The three accounts of orbit — mass per kilowatt, radiator area, downlink — form a coupled system. None is eliminated separately, and improving one often worsens another. The coupling works in the reverse direction as well: changing the power layer simultaneously affects generation mass, deployed area, attitude control requirements, load-bearing structure and thermal budget. That is precisely why, in the class of systems under consideration, the power system of an orbital node is a matter of architecture rather than of component choice.

Third. Launch cost receives more public attention, but system models show that it is inseparable from the second lever — launched mass per delivered kilowatt. The academic model defines that quantity directly as one of three controlling coordinates; the consulting analysis independently uses spacecraft mass per 100 kW of compute power as a sensitivity variable. Work on this lever is under way: it is being done by the European system architecture, by integrated panel–compute–radiator layouts and, judging by declared characteristics, by operators. The question is not whether anyone is working on mass, but by what means it is being reduced and at what accounting boundary it is measured.

Fourth. Within mass per kilowatt the largest single item is the generation system: 16.9 kg/kW in the best orbit and about 32 kg/kW in an ordinary low one. That is the point of application.

Fifth. Specific power alone is insufficient for selecting an alternative source. The requirement is two-dimensional: a source must simultaneously pass the threshold on specific power and on mission-specific energy — that is, provide the required multi-year delivery without a consumable energy inventory whose mass destroys the mass advantage. The orbiter alkaline fuel cells show why this distinction is mandatory: about a hundred watts per kilogram of hardware power is enough for the first condition, but a multi-year reactant inventory fails the second.

Sixth. The requirement on an alternative is expressed by a threshold ratio dependent on the comparison base: in the baseline case, about 74 % of the specific power of the photovoltaics being replaced, plus a correction for radiator growth. Numerically, from 22 W/kg against the flown-fleet median to 148 W/kg against the historical maximum. The minimum admissible efficiency is likewise not a constant: for the bases considered it varies from about 0.18 to 0.60.

Seventh. The value of the threshold depends on what the competitor turns out to be. Here the industry has an unresolved discrepancy of nearly an order of magnitude: the academic model and the European architecture give about 40 kg/kW, an operator’s public figure gives about 13 kg/kW, and within the European architecture itself the block and system estimates differ threefold. Meanwhile the specific areas of panels and radiators converge across all sources. This is therefore, so far, a question of comparability of accounting boundaries rather than a demonstrated dispute about achievable specific mass. The size of the niche depends on its resolution.

Eighth. We look in this direction because the topology of our architecture’s exchange with the environment contains no dependence on the atmosphere at the level of the switching function, because its generation architecture contains neither combustion nor a rotating converter, and because the target DC output interface coincides with the practice of both the spacecraft and the modern data center.

Ninth. The principal engineering challenge of the space-derived implementation lies in the thermal path: the existing ground implementation removes heat by convection, whereas a space one is obliged to do it by conduction and radiation, with the effective sink temperature changing within the orbit. The absence of a rotating converter neither removes that task nor relates to it.

Tenth. The absence of a requirement for air removes one barrier and at the same time introduces another: a high-frequency resonant architecture makes vacuum electrical qualification one of the most serious project-specific risks. That is the precise formulation, and it is more honest than the convenient one.

Eleventh, and it outweighs all the preceding. The next level of evidence is created no longer by the ground operating history of the installation but by space qualification of the derived architecture. It must determine service life over a ten- to fifteen-year horizon without servicing, vacuum dielectric strength, thermal architecture, radiation tolerance, thermal-cycling drift of resonant tuning and the redundancy scheme. Different materials, a different laboratory, a different class of testing. Each of these parameters has its own measurable criterion and its own pass/fail result, and any one of them is capable of closing the subject on its own. This is not a caveat but the content of the work.

Twelfth. The most probable area of first application is not the replacement of terrestrial data centers but a node of 10–100 kW in regimes where the orbital location itself creates value. Beyond low orbit, environments locally deprived of solar access are of particular interest — above all the permanently shadowed polar regions of the Moon, where local photovoltaics does not exist as an immediate source and the bar for an alternative is substantially lower.

Thirteenth. A significant part of the decisive parameters of the hypothesis is determined on the ground, before the transition to space qualification. A negative result at an early stage closes the corresponding branch ahead of more complex testing; a positive result authorises only the next level of verification. If the entire sequence of criteria is satisfied, a change of the power layer becomes no longer a local substitution of a source but a design parameter of the orbital computing platform.

Fourteenth, as a closing point. Independent industrial activity today is sufficient to consider orbital computing an emerging infrastructure problem. Known physics is sufficient to say that the VENDOR.Max architecture is not excluded from that problem a priori. The next question is engineering, not rhetorical: what will happen to this architecture when it is redesigned for the real space environment and tested against it.