The fundamental question before us is not whether artificial intelligence compute will continue its exponential expansion, but rather: by what mechanisms will the physical constraints of heat dissipation be overcome? This inquiry is not merely technical—it is a matter of capital allocation and the very possibility of continued productive expansion in the semiconductor industry. The issue presents itself as a fork in two domains: the terrestrial data center, where GPU thermal design power has now surpassed 2 kW per module 26 and CPU thermal design power approaches 1 kW per socket 26; and the orbital domain, where SpaceX's emerging "Starmind" constellation concept 31 promises to bypass terrestrial constraints entirely by placing AI compute directly into low Earth orbit.
This cluster of claims reveals that NVIDIA stands at an inflection point. The company's GPU revenue expansion faces a binding constraint not on the supply side—silicon yields remain robust—but on the demand side, where the physical capacity to dissipate heat has become the primary bottleneck to data center densification. Simultaneously, an entirely new customer segment is emerging: space-qualified AI infrastructure operators who must solve thermal management problems in an environment where convection is impossible and radiative cooling dominates.
II. Terrestrial Domain: The Cooling Crisis and Its Implications
A. The Hard Ceiling on Air Cooling
The primary empirical foundation for understanding terrestrial constraints is straightforward: perimeter air cooling is effective only up to 20–25 kW per rack in optimized systems 12, a figure that has not materially advanced despite decades of engineering effort. This is not a design failure but a consequence of thermodynamic law—air, as a thermal medium, possesses limited specific heat capacity and density. When GPU module TDP has crossed 2 kW and next-generation systems demand 30–40 kW per rack, traditional cooling approaches face insurmountable physical limits.
This divergence between compute demand and cooling capacity represents what I term the "expediency gap"—a temporary inefficiency in the allocation of resources that markets correct through innovation or constraint. The data provides clear evidence that the gap is being corrected through liquid cooling architectures. The KAIST manifold microchannel technology exemplifies this innovation: by using ordinary room-temperature water and reducing pumping energy to one-tenth of prior approaches, the system maintains chip temperatures below 100°C under extreme heat flux 30,35. This represents a material improvement in the utility function—the same computational capacity may now be deployed with superior thermal properties and lower parasitic power losses.
B. Systemic Solutions: Cold Storage and Waste Heat Utilization
The problem of inquiry deepens when one considers the total system perspective. Cold Underground Thermal Energy Storage (Cold UTES) offers a second vector of improvement: by leveraging subsurface thermal stability, such systems have demonstrated the capacity to increase IT power capacity by 9.8–11% in simulated deployments across Virginia, Arizona, and Texas 36. The mechanism is elegant—thermal energy is stored during off-peak hours and released during peak compute periods, effectively decoupling instantaneous generation from instantaneous demand.
More provocatively, Microsoft's Puget Sound Thermal Energy Center exemplifies the principle that waste heat is not truly waste but a resource improperly valued. The facility is designed to be approximately 50% more efficient than a conventional utility plant, and the company is exploring repurposing 1 MW of waste heat for direct air capture, potentially removing 5,000 metric tons of CO₂ annually 24. This represents a form of productive recycling—the externality of heat becomes a productive input to carbon sequestration. For NVIDIA, this signals that ecosystem partners capable of capturing, storing, and repurposing waste heat are essential to the company's long-term growth trajectory.
C. Environmental Externalities and Regulatory Headwinds
However, the Method of Difference reveals a critical constraint that market prices may not yet fully reflect. The discharge of warmer water from AI data center cooling systems into surface waters reduces dissolved oxygen, promotes algal blooms, and causes long-term ecological damage 20,21. These negative externalities are not reflected in the private cost of capital but will eventually impose regulatory constraints.
The evidence of regulatory pressure is already visible. In Virginia, educational institutions have reportedly been instructed to turn off lighting to accommodate AI data center electricity demand 22. This anecdote—striking in its implications—suggests that the social willingness to subsidize AI compute expansion through reduced public services is finite. The Jevons Paradox applies directly here: increased cooling efficiency leads to higher total resource consumption rather than decreased usage 14,34. Each improvement in thermal management capacity unlocks additional rack density, which accelerates power consumption, which intensifies regulatory and environmental pressure.
This creates a curious inversion: the very innovations that solve NVIDIA's terrestrial thermal problems will, paradoxically, accelerate demand for orbital alternatives by making clear to policymakers the unsustainability of continued terrestrial expansion.
III. Orbital Domain: The Engineering Frontier
A. SpaceX's AI1 Satellite: Proof of Concept and Design Parameters
The most rigorously corroborated data point in this inquiry is SpaceX's nascent orbital data-center satellite program, exemplified by the AI1 design. This satellite is confirmed across four independent sources to feature a 110 m² deployable liquid radiator system 17,18. Thermodynamic modeling yields the following equilibrium conditions under continuous-peak load of 150 kW, assuming a 220 m² double-sided emitting area: equilibrium radiator temperatures of approximately 353 K 18.
Under thermal stress scenarios—where radiator surface emissivity degrades from 0.91 to 0.80 and environmental sink temperatures rise from 220 K to 260 K—equilibrium temperatures climb by roughly 22 K, pushing the system toward 391–412 K 17,18. This range is significant because it approaches and potentially exceeds the operational limits of conventional heat-transfer fluid chemistries.
The sustained compute capacity is estimated at 120 kW with a peak of 150 kW, utilizing approximately 72 NVIDIA B200 GPUs and weighing approximately 2 tons 10. The heat removal capacity is approximately 40 kW with a peak-to-sustained thermal headroom of 30 kW 17. While SpaceX has not publicly disclosed the coolant chemistry 17,18, technical analysis suggests ammonia, though a radiator temperature requirement of 391–412 K may present challenges for a conventional subcritical ammonia loop 17,18.
B. The Physics of Orbital Radiative Cooling
The fundamental constraint of orbital thermal management is non-negotiable: in a vacuum environment, convection is absent entirely 25. Heat dissipation occurs exclusively through radiation—the emission of infrared photons. This necessitates surface areas several times larger than the satellite itself 25, a factor that creates a design tension between thermal performance and orbital mechanics (mass, cross-sectional area for drag, structural rigidity).
This tension is distinct from terrestrial thermal challenges. On Earth, the cooling medium (air or water) convects away heat through bulk fluid motion. In orbit, each photon emitted by the radiator surface must carry away energy individually. The radiative heat flux is proportional to the fourth power of absolute temperature (Stefan-Boltzmann law), meaning that modest increases in radiator temperature compress operational margins severely.
C. Radiation Damage and the Forced Replacement Cycle
A second orbital-specific constraint emerges from the space environment itself: radiation. The Van Allen belts and solar particle events degrade solar panel efficiency, radiator surface properties, and semiconductor chips 23,25. This necessitates radiation-hardened custom silicon—SpaceX claims its custom designs meet aerospace and defense radiation constraints 23,25.
The consequence is a forced 3.5-to-5-year replacement cycle driven not by market demand for newer chips but by accumulated radiation damage 4,9,37. The hardware replacement cycle of 3.5 years 37 is faster than terrestrial GPU refresh cycles, creating a high-velocity replacement market. However, this also means that NVIDIA's orbital revenue stream will be characterized by shorter hardware lifespans and higher replacement rates—a phenomenon that, while improving revenue velocity, compresses margins due to the technical complexity of space-qualification.
D. Reliability, Redundancy, and the Economics of In-Orbit Repair
A final constraint reveals itself through the data: in-orbit repair costs far exceed manufacturing costs 25. This necessitates onboard redundancy—spare GPUs, redundant power systems, and fault-tolerant architectures. The requirement for redundancy increases the weight and cost per unit of deployed compute, reducing the economic utility relative to terrestrial data centers where repair is straightforward.
Moreover, the Reflection-SpaceX contractual arrangement—slated to run through 2029 with termination provisions after an initial 3-month period 2,13,38—illustrates the nascent and unsettled nature of these commercial relationships. Conflicting reports on cancellation notice periods (60, 90 days, or 3 months) 13 suggest that legal frameworks for orbital compute services have not yet crystallized.
IV. Competitive Landscape and Market Dynamics
The orbital infrastructure market is rapidly fragmenting into competing constellations, each with distinct technical and commercial characteristics. AST SpaceMobile operates 9 BlueBird satellites as of mid-June 2026, each featuring approximately 2,400 square foot phased-array antennas, with FCC authorization for up to 248 satellites 4,6,14. The company targets 45–60 satellites by end of 2026 and approximately 90 for broader coverage 4. AST's direct-to-cell approach 4 and large antenna arrays 4 compete for spectrum and regulatory attention, though this application domain differs from data-center compute.
Eutelsat and Telesat are deploying competing constellations 4, while Spire Global operates over 240 satellites and has received first-light data from its hyperspectral microwave sounder demonstrator 3. Google launched the FireSat proto-satellite in early 2025 for wildfire detection in collaboration with the Earth Fire Alliance and Muon Space 11. SpaceX's Starlink constellation includes over 650 satellites with direct-to-cell capability 4, with Starlink V3 satellites built and scheduled for 2026 launch 4. The Starship launch vehicle can deploy dozens of AI satellites per flight and is projected to launch approximately 1 million tonnes into space within 3–4 years 1,29.
This constellation of competitors suggests that while SpaceX may pioneer orbital AI infrastructure, other operators will rapidly follow, creating a diversified demand base for space-qualified compute hardware.
V. Software and Model Developments: The Fundamental Tailwind
Underlying all thermal management challenges—terrestrial and orbital—is an inexorable fact: artificial intelligence capability continues to expand at exponential rates, driving ever-greater compute demand. OpenAI's Sol model showed approximately 123.2% improvement on ExploitGym 15; OpenAI's Luna model scored 84.7% on Terminal-Bench 15; Poolside's Laguna XS 2.1 achieved 63.1% on SWE-bench Multilingual 27; Moonshot AI released Kimi K2.7-Code, a 1-trillion-parameter MoE coding model 5.
Anthropic reports that code produced per engineer per day is 8× higher than two years ago 28, while Schrödinger used AlphaEvolve to improve its PyTorch implementation 19. These developments are not mere incremental improvements; they represent systemic advancements in the productive capacity of AI systems. Each such advancement increases the utility of additional compute capacity, driving further demand for GPUs. This is the fundamental force that necessitates both terrestrial thermal solutions and orbital infrastructure expansion.
VI. Strategic Implications and the Tendency of the Market
For NVIDIA, the implications of this inquiry are multifold. First, on the terrestrial front, the binding constraint to GPU deployment is no longer supply but the physical capacity to dissipate heat. NVIDIA's ecosystem strategy should prioritize deep partnerships with liquid cooling technology providers (microchannel systems, advanced cold plates like Medusa Technologies' THERMALOCK™ 32), thermal storage operators (Cold UTES deployers), and predictive thermal management firms (such as Omen AI, which provides algorithmic cooling optimization 7,8). These partnerships directly unlock additional rack density per data center facility, accelerating GPU revenue growth.
Second, on the orbital front, SpaceX's AI1 satellite represents a nascent but material new customer segment. The estimated 72 B200 GPUs per satellite 10 imply per-satellite GPU content values measured in the millions of dollars. However, NVIDIA should model orbital compute as a lower-margin, higher-velocity segment due to mandatory redundancy, radiation degradation, and the impossibility of in-orbit repair. The 3.5-year replacement cycle 37 means recurring revenue, but the technical complexity and risk premium will compress returns relative to terrestrial data centers.
Third, environmental and regulatory scrutiny of data center water discharge 21 and power consumption 22 represents a forward-looking indicator of demand constraints. As regulatory pressure intensifies, the economic attractiveness of orbital alternatives will improve, potentially triggering an inflection in orbital compute adoption rates. NVIDIA should monitor regulatory developments at the state and federal levels as a leading indicator of orbital compute demand acceleration.
Fourth, the governance risk surface for autonomous systems is expected to expand significantly between 2027 and 2030 16, and potential AI regulations may restrict biology and climate-related AI research 33. These regulatory uncertainties create a tail risk to GPU demand growth that should be monitored alongside environmental regulatory developments.
In sum, NVIDIA faces a dual-vector growth opportunity constrained by thermal physics on one side and a regulatory and environmental tendency on the other. The company's ability to navigate this landscape—through ecosystem partnerships on the terrestrial front and space-qualified product development on the orbital front—will substantially determine its long-term productive utility in an increasingly heat-constrained global economy.