the theses and their beneficiaries
ARK Investment Management LLC (CIK 0001697748) manages approximately $15B in assets across multiple ETFs. ARKX, ARK's Space Exploration & Innovation ETF, holds stakes in companies whose valuations benefit directly from a high-demand narrative for orbital infrastructure. The same firm published a research document that simultaneously makes the case for orbital compute and for the ground-level distributed energy that makes it unnecessary.
two theses. one document.
ARK Big Ideas 2026 makes a compelling case for cheap, distributed ground-level energy, and then separately argues that orbital datacenters are the next computing frontier. The document also states (p.15) that distributed energy generation "should address surging power demand from AI data centers," while simultaneously claiming (p.7) that AI compute is "running into earthly scaling constraints." Both cite AI data centers. Both cannot be true. Neither section cross-references the other.
ARK's p.93 projects electricity prices resuming their Wright's Law decline. ARK's p.15 states distributed energy generation "should address surging power demand from AI data centers." If ARK's energy thesis is correct, the economic case for accepting the physics penalties of orbital compute (thermal radiation constraints, +40ms latency, no maintenance access, radiator mass that must be launched regardless of launch cost) weakens as terrestrial electricity gets cheaper. ARK's own energy projections are the most powerful argument against ARK's orbital thesis. The document does not address this contradiction.
what ARK claims on page 7
ARK's Big Ideas 2026 projects that falling launch costs, driven primarily by SpaceX Starship, will enable commercially viable orbital datacenters. The thesis rests on three interconnected claims: launch cost compression, orbital solar advantage, and global reach without terrestrial network dependencies.
| ARK Claim | ARK Projection | Current Reality (2026) | Status |
|---|---|---|---|
| Starship launch cost | ~$100/kg to LEO at scale (p.80: "Starship can extend that trajectory to $100/kg") | ~$1,000/kg current estimate (Falcon 9 heritage; Starship operational but not at scale) | ~10× above ARK's projected target |
| Orbital solar efficiency | No atmospheric loss · continuous illumination in SSO | Technically accurate — but power must be converted and dissipated as heat on-orbit | Partial — thermal problem not addressed |
| Global compute reach | Orbital infrastructure = no terrestrial cable dependency | Starlink provides connectivity, not compute; terrestrial edge nodes already global | Alternative solutions exist |
| AI training in orbit | Training workloads could migrate to orbital infrastructure | AI training requires high-bandwidth, low-latency GPU clusters; orbital adds latency penalty | Contradicts AI infrastructure requirements |
| Maintenance & reliability | Not addressed in orbital thesis | On-orbit hardware failures are permanent; no repair access below GEO | Critical omission |
Source: ARK Big Ideas 2026 · p. 7 (orbital datacenter thesis). Launch cost context: SpaceX Starship program public statements; industry analyst estimates. On-orbit maintenance: NASA ISS operations data.
what ARK claims on pages 15 and 93
ARK's distributed energy thesis is among their most historically consistent projections, and the most rigorously supported. Solar costs have followed a predictable Wright's Law decline, battery storage costs have tracked similarly, and ARK's directional projections in this domain have largely proven accurate. This makes the internal contradiction sharper: the energy thesis is the one ARK should believe most.
ARK Big Ideas 2026, p.93 (direct quote): "Informed by Wright's Law, ARK's research indicates that, aside from WWII, US electricity prices fell steadily from the late 1800s to 1974, at which time regulation interrupted the decline in nuclear construction costs. Had regulation not intensified, ARK's research suggests that electricity prices would be ~40% lower today. As low-cost power generation scales and serves power-hungry AI data centers, retail electricity prices should resume their decline, following Wright's Law, after 50 years of stagnation."
ARK also applies Wright's Law to solar and battery costs (p.92 charts showing log-scale decline with cumulative capacity) and to nuclear construction costs (p.92: "recent Executive Orders in the US should push nuclear energy back onto its previous cost-decline trajectory"). Solar LCOE has tracked this curve, falling from over $400/MWh in 2010 to under $30/MWh in the best 2025 markets. ARK applies Wright's Law to electricity, solar, nuclear, and batteries. The implication for orbital compute is addressed in §05 below.
applied to everything — which proves too much
Wright's Law states that for every cumulative doubling of production, costs fall by a fixed percentage. ARK explicitly invokes it throughout Big Ideas 2026, not selectively but comprehensively. The problem is what follows when you apply ARK's own framework consistently: the orbital compute thesis weakens over time, not strengthens.
| Technology | ARK's Wright's Law Application | Source | Rate |
|---|---|---|---|
| Rocket launch costs | "launch costs should decline by ~17% for every cumulative doubling in upmass to orbit" | p. 80 | 17% / doubling · $1,000/kg → $100/kg (Starship at scale) |
| Satellite bandwidth | "satellite bandwidth costs should decline by ~44% for every cumulative doubling of gigabits per second" | p. 81 | 44% / doubling |
| Electricity prices | "retail electricity prices should resume their decline, following Wright's Law, after 50 years of stagnation" | p. 93 | ~40%+ lower · continuing to fall |
| Nuclear construction | "recent Executive Orders… should push nuclear energy back onto its previous cost-decline trajectory" | p. 92 | Resuming historical Wright's Law curve |
| Solar installation | Log-scale cost decline vs. cumulative capacity (Wright's Law chart) | p. 92 | ~30–40% / doubling (historical) |
| Battery storage | Log-scale cost decline vs. cumulative capacity (Wright's Law chart) | p. 92 | ~28% / doubling (historical) |
ARK's p.7 "Space AI Compute vs. Terrestrial AI Compute" analysis projects orbital compute becoming 25% cheaper than terrestrial once launch costs fall to ARK's projected target. But this comparison is asymmetric in an important way: launch costs improve via Wright's Law, yet terrestrial compute's cost inputs (electricity, cooling) are also following Wright's Law (pp. 92–93).
When both sides improve simultaneously, the question is which has the larger residual fixed cost that cannot be reduced. For terrestrial compute: atmospheric cooling is permanently free. It does not follow Wright's Law because it requires no production or manufacturing. For orbital compute: thermal radiation panels, radiation shielding, and launch overhead cannot be reduced to zero regardless of production scale. These are physics floors, not engineering gaps.
ARK's own data (pp. 92–93) shows electricity and solar costs continuing to fall. As terrestrial electricity gets cheaper and orbital cooling costs remain fixed, the crossover point (if it exists) moves further away, not closer. ARK presents the crossover as a near-term milestone without accounting for the Wright's Law improvements it projects for terrestrial energy in the same document.
ARK Big Ideas 2026 makes two directly contradictory claims about terrestrial AI data center scaling:
p.7 (Space Compute thesis): "Neural network demand for next gen cloud compute is running into earthly scaling constraints." ARK's justification for orbital compute is that terrestrial infrastructure cannot keep up with AI demand.
p.15 (Energy Storage thesis): "These innovations combined with large-scale stationary batteries and Distributed Energy Generation, notably solar and small-scale fission, should address surging power demand from AI data centers." ARK's energy thesis explicitly states terrestrial distributed power WILL solve AI's power requirements.
Both claims cite AI data centers. Both cannot be true. If distributed energy will address surging AI power demand (p.15), then terrestrial compute's scaling constraints resolve without going to orbit. The document does not reconcile these claims.
ARK Big Ideas 2026, p.18 (Robotics section) states:
"Reusable Rockets should continue to reduce the cost of launching satellite constellations… and open the frontier for space-based compute infrastructure unconstrained by terrestrial power and cooling limitations."
This is factually incorrect. Orbital compute is more constrained by cooling than ground compute, not less. On Earth, water and air cooling reject heat at ~$0.01–0.05/W. In space, the only available mechanism is thermal radiation (Stefan-Boltzmann law: P = εσAT⁴). For a 1 MW compute pod, this requires approximately 2,000–4,000 m² of radiator area: mass that must be launched to orbit at substantial cost regardless of how cheap Starship becomes. Cooling constraints do not diminish with lower launch costs; they scale with compute power and are governed by physics, not engineering. The p.18 claim inverts the actual physical situation.
what orbital compute actually requires
The physics of operating computers in space create constraints that no launch cost reduction can eliminate. These are not engineering challenges awaiting a breakthrough. They are thermodynamic, orbital mechanical, and economic facts. ARK Big Ideas 2026 does not address any of them in the orbital datacenter section.
The International Space Station uses 2,500 m² of radiator panels to reject approximately 120 kW of average thermal load, the only large-scale operational space hardware deployment in history. Source: NASA ISS Thermal Control System documentation (public).
A single modest 10 MW AI training cluster produces ~9 MW of waste heat. Scaled from the ISS rejection ratio (a conservative baseline): ~187,500 m² of radiator area, or 26 football fields. A purpose-built compute radiator operating at 350 K could theoretically achieve ~300–500 W/m² net rejection, reducing the requirement to roughly 20,000–30,000 m² (3–4 football fields). Both figures are enormous; every m² must be launched. ARK p.18 describes this environment as "unconstrained by terrestrial power and cooling limitations."
| Physical Constraint | Ground Datacenter | Orbital Datacenter | Winner |
|---|---|---|---|
| Heat rejection | Water / air cooling · ~$0.01–0.05/W | Thermal radiation only · massive radiator panels required | Ground |
| Network latency | <1ms (local) · <5ms (regional) | ~25–45ms practical (LEO round-trip) · physics floor ~4ms but not operationally achievable | Ground |
| Hardware maintenance | Minutes to replace failed components | Impossible: permanent failure | Ground |
| Energy cost (ARK projection) | Electricity prices resume Wright's Law decline: "~40% lower" and continuing (p.93) | Solar available on-orbit but thermal rejection mass must still be launched | Ground (per ARK's p.93 projection) |
| Infrastructure cost | $5–20M per MW of compute capacity | $50–200M+ per MW (est., launch + radiators + redundancy) | Ground (10–40× cheaper) |
| Density | Unlimited expansion; add more buildings | Constrained by orbital mechanics and debris environment | Ground |
| Radiation environment | Commercial silicon (GPUs, HBM), normal operation | LEO cosmic rays + South Atlantic Anomaly → rad-hard chips 10–100× cost, 5–10× slower; or heavy shielding mass | Ground |
| Hardware lifecycle | Component-level replacement every 5–7 years | Full rack re-launch every 5–10 years; electric propulsion delays but can't eliminate deorbit. No LEO refueling exists | Ground |
the numbers don't close
ARK's p.80 projects Starship reaching $100/kg at scale, a 10× reduction from today's ~$1,000/kg. Even at that target, the economics of orbital compute do not compete with ground alternatives. The analysis below uses a standard server rack (approximately 500 kg including structure and thermal management) as the unit of comparison.
Rack weight assumptions: ~500 kg including structure, power conditioning, and thermal management. Radiator mass: ~40 kg/m² × 400 m² per 1MW heat load (conservative). Ground deployment: standard colo rack pricing. ARK launch cost projection from Big Ideas 2026 p. 80 ($1,000/kg → $100/kg at scale). ARK electricity projection: p.93 ("resume their decline, following Wright's Law"). Ground rack pricing: standard colocation rates.
At ARK's projected $100/kg at scale (p.80), not yet achieved and requiring a ~10× reduction from today's ~$1,000/kg, a 500 kg server rack still costs $50,000 in launch costs alone, plus an additional $20,000–$50,000 for radiator panel launch mass required for thermal management. Total orbital deployment: ~$70,000–$100,000 per rack minimum, before hardware cost. The same rack deployed in a ground facility with atmospheric cooling: ~$500–1,000 in infrastructure cost, standard energy OpEx. The orbital option is approximately 70–200× more expensive at ARK's own best-case projections. ARK also projects terrestrial electricity costs to continue falling (p.93), widening this gap further over time. There is no scenario in Big Ideas 2026 where this math closes.
Chart shows the launch cost ($/kg) required for orbital compute to reach cost parity with ground infrastructure, as a function of terrestrial electricity price. ARK's p.7 crossover is based on current electricity prices (~$0.06–0.07/kWh). ARK's p.93 projects electricity resuming its Wright's Law decline. As electricity falls, the required launch threshold falls with it, reaching below the estimated chemical propulsion physics floor (~$50/kg) before ARK's projected electricity range. Crossover threshold scales linearly with energy cost differential (derived from ARK's own p.7 orbital vs. terrestrial cost comparison methodology).
ARK's p.7 crossover (where orbital compute becomes 25% cheaper than terrestrial) is calculated at current electricity prices. ARK's p.93 separately projects terrestrial electricity resuming its Wright's Law decline. Applied symmetrically: as electricity falls toward ARK's projected range, the launch cost required for orbital to break even drops below the estimated physics minimum for any chemical propulsion system. Under ARK's own projections applied to both sides of the cost comparison, there is no achievable launch cost at which orbital compute becomes economically competitive with ground infrastructure. The two theses cannot both be correct.
ARKX and the demand narrative
ARK Investment Management manages ARKX, the ARK Space Exploration & Innovation ETF. ARKX holds positions in companies whose revenues and valuations are directly tied to space launch demand. A high-conviction orbital datacenter thesis in ARK's flagship annual report creates a demand narrative that benefits ARKX holdings, including SpaceX-adjacent companies and satellite infrastructure providers.
| Entity | ARK Connection | Orbital Thesis Benefit | Conflict |
|---|---|---|---|
| ARKX ETF | ARK-managed · space exploration fund | Higher space infrastructure demand → higher ARKX holdings valuations | Direct financial interest |
| SpaceX (Starlink) | ARKX position · Starship cited as cost-reduction driver | Orbital datacenter demand = massive launch volume for SpaceX | ARK thesis creates demand for ARK holding's primary customer |
| Rocket Lab | ARKX holding | Orbital infrastructure build-out benefits small-launch providers | Indirect benefit |
| Planet Labs / Satellogic | ARKX / ARK holdings | Normalized orbital infrastructure reduces barrier for satellite-compute models | Indirect benefit |
| ARK Innovation ETF (ARKK) | ARK-managed · Tesla top holding | Orbital compute narrative supports Tesla Energy storage as power source for ground compute (alternative) | Circular narrative |
Space-sector positions from ARK's combined portfolio (all ARK-managed ETFs including ARKX). Combined AUM: $15.07B across 197 positions.
| Company (SEC 13F Exact Name) | % of Combined AUM | Primary Business | Orbital Thesis Benefit |
|---|---|---|---|
| Kratos Defense & Security Solutions Inc. | 1.74% | Satellite systems, space propulsion, directed energy | Direct: orbital infrastructure build-out |
| Rocket Lab Corp | 1.03% | Launch provider, satellite constellation deployment | Direct: orbital compute demand = launch volume |
| AeroVironment Inc | 0.70% | Aerospace defense systems, tactical UAVs, space-adjacent programs | Indirect: aerospace sector expansion |
| L3Harris Technologies Inc | 0.60% | Space systems, satellite communications, signal intelligence | Indirect: satellite infrastructure growth |
| Iridium Communications Inc | 0.33% | LEO satellite constellation (voice, data, IoT) | Direct: normalizes LEO as operational compute environment |
| Intuitive Machines Inc | 0.25% | Lunar landers, commercial space mission services (NASA partner) | Indirect — commercial space market expansion |
Source: SEC EDGAR, ARK Investment Management LLC Q4 2025 13F-HR (accession 0001104659-26-013538). Percentages of combined AUM ($15.07B). ARK does not separately itemize per-ETF positions in 13F-HR; table shows space-sector companies benefiting from the orbital thesis.
Source: Adjusted close prices via public market data. ARKX launch: March 30, 2021 (~$20.57 NAV). SPY and QQQ shown as comparable benchmarks for the same period through March 2026.
Since its March 2021 launch, ARKX has returned −13.8% while the S&P 500 returned +41.5% over the same period, a ~55 percentage point gap. A high-conviction orbital datacenter thesis in ARK's flagship annual research publication creates a positive demand narrative for ARKX's space-sector holdings. This is not an allegation. It is the documented relationship between fund management and investment research. ARK does not disclose within Big Ideas 2026 that it manages an actively traded space ETF whose holdings directly benefit from the orbital demand thesis it presents.
ARK Investment Management discloses in its 13F-HR filings (CIK 0001697748) the full composition
of ARKX and other ETFs on a quarterly basis. The orbital datacenter thesis in Big Ideas 2026
creates a favorable narrative environment for these holdings. ARK does not disclose within
Big Ideas 2026 the financial relationships between its research theses and its active positions.
The document presents the orbital datacenter thesis as independent analysis.
The distributed energy thesis (which undermines the orbital thesis)
appears in a separate section without cross-reference or reconciliation.
ARK's Works Cited (p. 108) includes: "Maguire, D. et al. 2025. 'ARK's Expected Value For SpaceX
In 2030: ~$2.5 Trillion Enterprise Value.' ARK Investment Management LLC." ARK's orbital demand
narrative cites ARK's own SpaceX valuation as a supporting reference. The loop closes:
Big Ideas 2026 projects orbital compute demand → SpaceX needs that launch volume
to justify its business model → ARK's own analysts valued SpaceX at $2.5T based on that demand →
Big Ideas cites that valuation as corroborating evidence. The circular reference is not flagged
in the document.
what the document doesn't address
Ten anomalies identified across ARK Big Ideas 2026 and related ARKX filings. All findings are based on direct quotes from the document. Click each to expand.
ARK Big Ideas 2026 contains an orbital datacenter thesis (p. 7) and a distributed energy thesis (pp. 15, 93) in the same document. The energy thesis, projecting electricity prices resuming their Wright's Law decline (p.93) and distributed generation solving AI data center power demand (p.15), is the strongest available argument against the economics of orbital computing.
No section of Big Ideas 2026 addresses this contradiction. The two theses are presented as independent positive visions without cross-reference or acknowledgment that one directly undermines the economic foundation of the other.
Source: ARK Big Ideas 2026, pp. 7, 15, 93.
The orbital datacenter section of Big Ideas 2026 does not analyze thermal management, the most fundamental engineering constraint for space computing. Computers convert roughly 90% of power to heat. On Earth, water cooling addresses this at ~$0.01–0.05/W. In space, thermal radiation is the only mechanism, requiring radiator panels of significant mass and area.
For a modest 1 MW compute pod (a small fraction of a modern datacenter campus), orbital thermal rejection requires approximately 2,000–4,000 m² of radiator area. This mass must be launched and adds substantially to per-rack orbital cost even at ARK's projected $100/kg target (p.80). This is not a technology gap. It is thermodynamics.
ARK Big Ideas 2026, p.80 projects Starship reaching approximately $100/kg at scale, based on applying Wright's Law (17% cost decline per doubling of cumulative upmass). This is a forward projection of a trend; p.80's stated baseline is SpaceX reducing costs from ~$15,600/kg to under ~$1,000/kg via Falcon 9 reusability. Starship has launched but is not yet operating at the frequency or payload volume that ARK's Wright's Law model requires to reach $100/kg.
The p.7 orbital vs. terrestrial cost comparison depends on this $100/kg target materializing. ARK does not specify when this is achievable or how many cumulative launches are required to reach it. The crossover economics are predicated on a launch cost that exists only as a projection of a projection. ARK's prior targets (Tesla at $4,000/share by 2024) followed the same structure.
ARK Big Ideas 2026 projects orbital datacenters as serving AI compute workloads. The document does not include a latency analysis. The physics floor for a LEO round-trip signal is ~4ms (speed of light at 600 km altitude, satellite directly overhead), but this is not the operational number. Practical LEO latency runs 25–45ms round-trip (consistent with observed Starlink measurements), reflecting beam-switching, on-satellite processing, and ground-station routing. This is comparable to a cross-continental ground connection, not a transatlantic one, but still 25–45× worse than local ground edge compute.
For AI inference endpoints serving interactive applications, this latency is a hard constraint. Ground edge nodes in the same market deliver <1ms round-trip latency: 25–45× lower latency than practical LEO. One scenario where orbital latency is less relevant: in-orbit pre-processing of satellite imagery before downlink to reduce bandwidth requirements. However, this niche use case is not the general AI compute thesis ARK presents on p.7. For AI training, high-bandwidth GPU interconnect is critical; orbital latency and the atmospheric signal path both introduce constraints with no ground equivalent.
ARK Investment Management manages ARKX (ARK Space Exploration & Innovation ETF), which holds positions in space launch, satellite, and space infrastructure companies. ARK's Big Ideas 2026 orbital datacenter thesis creates a positive demand narrative for ARKX holdings.
Big Ideas 2026 does not disclose ARK's active ARKX positions in the sections presenting the orbital datacenter thesis, nor does it note the firm's financial interest in the space infrastructure sector. ARK's Tesla valuation model (Big Ideas 2024) also did not disclose that ARK's ARKK fund held Tesla as its largest position. The pattern of research omitting holdings disclosure is consistent across ARK's published research materials.
ARK Big Ideas 2026, p.93 states: "retail electricity prices should resume their decline, following Wright's Law, after 50 years of stagnation." ARK also presents Wright's Law cost-decline charts for solar and battery costs (p.92). These projections are presented as a positive thesis for distributed energy adoption.
The document does not acknowledge that falling terrestrial electricity costs simultaneously strengthen the economic case for ground compute while orbital compute's fundamental costs (radiator launch mass, radiation shielding, maintenance) are fixed by physics and do not follow Wright's Law. As ARK's p.93 electricity projection materializes, the terrestrial cost advantage compounds. ARK's energy cost projections are the strongest available argument against ARK's orbital compute thesis. The document presents both without connecting them.
ARK Big Ideas 2026 applies Wright's Law comprehensively: to rocket launch costs (p.80: 17% decline per upmass doubling → $100/kg for Starship at scale), satellite bandwidth (p.81: 44% per Gbps doubling), electricity prices (p.93: "should resume their decline, following Wright's Law, after 50 years of stagnation"), nuclear (p.92), solar (p.92), and batteries (p.92).
ARK's p.7 "Space AI Compute vs. Terrestrial AI Compute" analysis projects orbital compute becoming 25% cheaper than terrestrial at ARK's projected launch cost target. This analysis treats the improvement in launch costs (Wright's Law) as the driver of orbital compute competitiveness. But the same Wright's Law framework ARK uses for electricity (p.93) means terrestrial compute costs also continue to fall.
The critical difference is what cannot follow Wright's Law. Atmospheric cooling on Earth is permanently free: it requires no manufacturing and has no production curve. Orbital thermal radiation panels, radiation shielding, and the launch overhead they represent do not approach zero regardless of scale. As ARK's projected improvements in terrestrial electricity materialize (p.93), the fundamental physics advantages of ground compute compound. ARK presents the p.7 cost crossover as a near-term milestone without addressing how its own energy projections shift that crossover.
Source: ARK Big Ideas 2026, pp. 7, 80, 81, 92–93.
ARK Big Ideas 2026, p.18 (Robotics section) states that reusable rockets "open the frontier for space-based compute infrastructure unconstrained by terrestrial power and cooling limitations."
This statement is factually incorrect. Orbital compute is more severely constrained by cooling than ground compute, not less. On Earth, water and air cooling reject heat at ~$0.01–0.05/W using well-understood engineering. In space, the only available heat rejection mechanism is thermal radiation, governed by the Stefan-Boltzmann law: P = εσAT⁴. For a 1 MW compute pod operating at representative orbital temperatures, thermal rejection requires approximately 2,000–4,000 m² of radiator panel area. This radiator mass must be launched to orbit and scales linearly with compute power regardless of launch cost reductions. The constraint does not diminish as launch costs fall; it is a fixed overhead of operating computers in a vacuum.
The p.18 claim inverts the actual physical situation. The cooling constraint is one of the primary reasons the numbers in §07 (cost comparison) do not close, and the document's own p.18 assertion obscures this by framing space as the unconstrained environment.
Source: ARK Big Ideas 2026, p.18. Stefan-Boltzmann law (P = εσAT⁴). ISS thermal control system data (NASA public records).
ARK's orbital datacenter thesis (p.7) does not address the LEO radiation environment. At 400–600 km equatorial LEO, the dominant radiation hazards are galactic cosmic rays and the South Atlantic Anomaly (SAA), a region where Earth's magnetic field dips, allowing inner-belt protons to reach low altitudes, causing routine memory bit-flips and component failures in commercial-grade electronics on both LEO satellites and the ISS. The inner Van Allen proton belt itself peaks above ~1,000 km at equatorial latitudes; polar-inclination orbits experience higher fluences. At any inclination above ~400 km, the radiation environment is hostile to unshielded commercial silicon.
Modern AI training hardware (the GPUs and HBM memory that make AI compute viable) is fabricated on commercial silicon processes (7nm, 5nm, 3nm) with transistor geometries that are highly sensitive to single-event upsets from high-energy particles. This silicon cannot operate unshielded in LEO for extended periods without failures. The available mitigations both have prohibitive costs: (1) heavy radiation shielding adds mass that must be launched, adding directly to the per-rack cost model; (2) radiation-hardened chip equivalents cost 10–100× more than commercial alternatives and operate 5–10× slower, negating the compute density assumptions that make orbital AI economically interesting.
ARK's p.7 cost comparison makes no reference to this hardware substitution cost or performance penalty. The assumption of unmodified commercial AI silicon in orbit is inconsistent with the known LEO radiation environment.
Source: NASA radiation effects documentation; ESA space environment and effects program; ISS radiation measurement records (public). ARK Big Ideas 2026, p.7.
ARK's p.7 orbital compute cost comparison presents a single deployment cost. It does not model the hardware replacement cycle inherent to LEO operations.
At orbital altitudes of 400–600 km, residual atmospheric drag causes hardware to deorbit within 2–5 years without continuous station-keeping propulsion. The ISS, at ~400 km, requires roughly 7 metric tons of propellant annually for orbital maintenance. Electric propulsion can extend operational life to 5–10 years, but propellant depletion is inevitable: LEO refueling infrastructure does not exist. Regulations also mandate reserving propellant for controlled end-of-life deorbit maneuvers, compressing the usable operational window further. Once propellant is exhausted, the satellite must deorbit regardless of hardware condition.
Ground datacenter server hardware is typically refreshed at the component level (individual drives, GPUs, DIMMs) on 5–7 year cycles, without replacing the rack structure or facility infrastructure. Orbital compute requires full physical re-launch of each rack on each replacement cycle, at whatever the prevailing launch cost is. At ARK's $100/kg target, a 500 kg rack still costs $50,000 per launch, per cycle, every 5–10 years.
This creates an internal contradiction in ARK's own Wright's Law thesis. If launch costs fall with cumulative production volume, frequent replacement launches are required to sustain the cost decline, but frequent launches mean shorter effective satellite lifetimes, which intensify the capex treadmill. ARK cannot simultaneously argue for long satellite lifetimes (to amortize hardware) and high launch frequency (to benefit from Wright's Law cost curves). ARK's p.7 cost model shows a single deployment cost; it does not model replacement cycle economics or propellant depletion.
Source: ISS orbital mechanics public data (NASA). ARK Big Ideas 2026, p.7, p.80.
primary documents
All findings are based on publicly available documents.