AI Civilization Map Node: The Thermal Limits of AI Compute
Primary Map Layer: Energy & Physical Infrastructure — Foundation
Primary Map Branch: Cooling, Water & Site Capacity
Secondary Map Layer: Space & Orbital Infrastructure — Orbital Frontier
Supporting Map Layer: Semiconductors, Compute & Packaging — Machine Substrate
Structural Function: Explains how heat capture, environmental heat rejection, site resources, and lifecycle economics constrain scalable, reliable AI computation.
Overview
An AI data center is usually described by what it contains: accelerators, high-bandwidth memory, networking equipment, and an electrical connection measured in megawatts. A more revealing description is what it must continuously do. It must convert electricity into useful computation while moving the resulting heat through a chain of materials, fluids, machines, and environmental interfaces. A failure anywhere along that chain can turn installed hardware into unavailable capacity.
This distinction is becoming consequential as artificial intelligence moves from individual servers toward tightly integrated computing systems and much larger facilities. In its April 2026 Key Questions on Energy and AI assessment, the International Energy Agency (IEA) estimated global data-center electricity consumption at 485 TWh in 2025 and projected 945 TWh in 2030 in its Base Case (rounded to approximately 950 TWh in its executive summary). The IEA's alternative 2030 scenarios range from 833 TWh in Headwinds to 1,008 TWh in Lift-Off; its High Efficiency case is 868 TWh. These figures cover all data centers, not AI-only facilities. The IEA separately expected AI-focused demand to grow much faster than total data-center demand. [1]
The central test is whether cooling, power delivery, manufacturing, and maintenance improve together. A component-level breakthrough matters only if it enables reliable, useful computing over a financially credible service life.
This article reflects evidence reviewed through November 22, 2026. Company specifications, measured experiments, operating claims and analytical estimates are distinguished by evidentiary status. Patent counts track research activity, not installed market share.
Key Takeaways
- Chip heat capture and final environmental heat rejection are separate constraints; a more powerful cooling loop cannot make waste heat disappear.
- Water use depends on the facility's final heat sink, not merely on liquid at the chip. Microsoft's designs and Google's Hamina project show different resource trade-offs.
- Microsoft's Natick program ended despite useful testing; Shanghai reports a 2.3 MW first phase against a 24 MW two-phase plan, not 24 MW of measured IT load.
- SpaceX AI1's 110 m² two-face radiator has temperature-dependent thermal headroom: the ideal 120 kW average and 150 kW peak cases require different operating temperatures.
- At an assumed 250 W/m² electrical output, an ideal continuously illuminated 1 GW solar array needs 4 km² before system losses, storage or auxiliary demand.
- A cooling breakthrough becomes scalable infrastructure only when reliability, maintenance, networks and full lifecycle cost support useful compute at the intended scale.
In this article
- Overview
- The Thermal Wall of AI Civilization
- The Cooling Technologies Powering AI Today
- Beyond Conventional Liquid Cooling
- Underwater Data Centers: The Ocean as a Heat Sink
- Space-Based Data Centers: The Promise and Limits of Radiative Cooling
- The Economics of Cooling at Scale
- The Physical Geography and Economic Limits of AI Civilization
- Conclusion
- Five Proof Gates for Scalable AI Cooling
- Frequently Asked Questions
The Thermal Wall of AI Civilization
A chain of constraints, not a single temperature limit
The phrase thermal wall can suggest a fixed barrier beyond which computing stops. A more useful interpretation is a set of operating boundaries. At chip level, the boundary depends on local heat flux, the path through the package, and allowable junction temperatures. At rack level, it depends on how equipment shares airflow or coolant, how much heat is captured, and what happens when a pump, fan, or connection fails. At facility level, it depends on the ability to transfer that heat into an external environment under the conditions that actually occur. The International Telecommunication Union's cooling-selection framework likewise separates indoor heat exchange, coolant transport, outdoor heat dissipation, and refrigeration. [2]
These boundaries should not be collapsed into a single megawatt figure. A building can possess adequate aggregate cooling capacity but still contain an accelerator whose local hotspot cannot be cooled through its package. Conversely, an excellent cold plate can keep one processor within its thermal limit while the facility lacks the pipe capacity, outdoor equipment, or permitted water supply required to operate thousands of similar processors simultaneously. Each successful component is necessary only within the architecture that connects it to the others.
A simple thermal-resistance relationship illustrates the first boundary: the temperature rise between a heat-generating junction and a coolant reference point is approximately the heat flow multiplied by the effective thermal resistance of the intervening path. This is an engineering simplification, not a complete model of a spatially complex processor. Nevertheless, it reveals why a lower coolant temperature is not the only route to higher power. Reducing the resistance of interfaces or bringing coolant closer to the heat source can preserve temperature margin without making the entire building colder.
Heat removal is not heat elimination
For operating computers, nearly all the electrical energy entering the IT equipment ultimately becomes heat. This is an operational energy balance, not a claim that every bit of information processing has a single immutable energy cost. Heat may be transferred to a useful heating network, stored temporarily, or rejected outside the facility. None of these routes makes energy disappear. Cooling equipment also consumes electricity, and that consumption must be accounted for within the appropriate boundary. The U.S. Department of Energy's data-center design guide treats cooling, electrical systems, IT efficiency, and heat recovery as parts of the same design problem. [3]
Consider an illustrative facility with a continuous 100 MW IT load and an annual power usage effectiveness, or PUE, of 1.20. Its total electrical demand would be approximately 120 MW under that simplified steady-load assumption. The 20 MW difference includes cooling and other infrastructure overhead; it is not automatically all chiller power. Reducing PUE to 1.10 would remove 10 MW of overhead at the same IT load, but it would not remove the approximately 100 MW of heat generated by the computers themselves. This arithmetic explains why efficiency gains and the need for substantial heat rejection can remain true simultaneously.
There is also a difference between peak power and energy over time. A cooling plant must survive demanding operating conditions, whereas its annual operating cost depends on the sequence of loads and weather it experiences. A nominally large cluster that operates intermittently does not have the same yearly heat burden as an equally large cluster running continuously. Yet a short, intense workload can still encounter a local thermal limit before the site's annual energy total appears especially large.
Density is not the same as efficiency
Higher rack density concentrates valuable compute and heat in less space. Even when each accelerator delivers more work per watt, a denser rack may demand more cooling. Efficiency measures useful output per input; density measures how much power and heat occupy a given area or volume.
Hi K Robot's structural interpretation is that cooling makes advertised capacity conditional. A purchase order specifies hardware; a functioning thermal system determines how much of that hardware can operate together, for how long, and under which environmental conditions. The gap between those two quantities is one reason that AI infrastructure cannot be understood through chip shipments alone.
The Cooling Technologies Powering AI Today
Air cooling: the incumbent is not a single architecture
Air cooling remains an important baseline because it is compatible with familiar server enclosures and established maintenance practices. Heat passes from electronic components into heat sinks and then into moving air. What happens next varies: the warm air can meet a room air conditioner, a chilled-water coil, a close-coupled unit, or a rack-door heat exchanger. An air-cooled server can therefore sit inside a facility that uses substantial liquid circulation elsewhere. Describing the server as air cooled does not identify the ultimate heat sink. [2]
Moving more heat with air requires more airflow for a given temperature rise. Better containment, faster fans, warmer exhaust or wider rack spacing can help, but each has costs. Air cooling remains appropriate where density permits; liquid approaches become attractive when airflow, fan power or floor area becomes the binding constraint.
Hybrid retrofits are one practical response to mixed rack loads. A rear-door heat exchanger intercepts heat after air passes through the servers, while other parts of a rack may use direct-to-chip liquid loops. Existing facilities can therefore adopt liquid-assisted heat rejection without replacing every air-cooled component at once.
Direct-to-chip liquid cooling: move the difficult heat first
Direct-to-chip liquid cooling uses a cold plate attached to a high-heat component to transfer energy into a circulating fluid. In common single-phase arrangements, the fluid remains liquid while it absorbs heat. A coolant distribution unit, or CDU, can regulate flow and interface the server technology-cooling loop with the facility loop. Cold plates, manifolds, tubing, connectors and service procedures must function together as a maintainable installation. [2] [4]
Liquid circulation is attractive because it can capture concentrated heat with less transport volume than air and can deliver that heat to the facility at a useful temperature. The operational trade-off is that fluid quality, leakage control, pressure, flow balancing, connectors, and maintenance procedures become part of computer availability. An otherwise efficient loop is inadequate if one common component can interrupt the cooling of an economically significant share of the cluster.
This is why cooling adoption is also a coordination problem. Server manufacturers, component suppliers, facilities engineers, and operators must agree on temperature limits, fluid compatibility, monitoring, and responsibility at the interface. The cost of a cold plate alone says little about the cost of qualifying and operating that complete chain. Conversely, interoperability can make repeated deployment cheaper by reducing the number of one-off integrations required for each hardware generation.
Direct-to-chip liquid cooling may still leave memory, networking, power electronics or other parts on an airflow path. The fraction of rack heat captured by liquid must be specified before attributing savings in fan or room-cooling electricity.
Immersion cooling: expand contact, change operations
Immersion cooling places compatible electronic equipment inside a dielectric fluid. In a single-phase system, it remains liquid while moving heat to a heat exchanger; in a two-phase system, evaporation absorbs heat and condensation releases it elsewhere. Both approaches require fluid and material compatibility, component qualification and a service method suited to the working liquid. Their relative merit depends on the workload and operational setting. [2]
Immersion can reduce dependence on server fans and capture heat from a broader set of components. It can also change how a technician removes, inspects, cleans, replaces, and returns a unit to service. The useful comparison is therefore not simply liquid versus air. It is one operating system against another: tank layout, fluid handling, hardware qualification, contamination control, warranties, safety requirements, and the recovery process following a fault.
Two-phase immersion transfers heat through boiling and condensation, but still needs a condenser and a final sink. Fluid loss, pressure management, material compatibility and environmental properties vary by fluid. A tank's ability to boil coolant is not proof that the complete installation can reject the resulting heat efficiently.
For an operator, the strongest case for immersion would emerge when its density or energy advantages are large enough to outweigh integration and servicing costs over the actual equipment life. The weakest case would rely on an exceptional demonstration temperature while ignoring the labor, replacement, and ecosystem changes required to reproduce that result across a fleet. Neither conclusion can be drawn from a technology label alone.
What the patent evidence actually establishes
Two companion studies published by Taiwan's Intellectual Property Office in Intellectual Property Monthly, No. 335 (November 2026), examine cooling patents published between 2015 and 2024. Their defined 2015–2024 patent search identified 8,449 patent families, with annual publications increasing from 386 in 2015 to 1,878 in 2024 (a reported compound annual growth rate of about 19%). The companion classification assigned 7,546 families to server cooling. Within that classification, single-phase indirect liquid cooling accounts for the largest group. The results measure invention and filing activity, not the fraction of data centers already operating on each technology. [5]
| Server-cooling category in the study | Patent families, 2015-2024 | Approximate share |
|---|---|---|
| Single-phase indirect liquid cooling | 3,371 | 44.7% |
| Air cooling | 2,175 | 28.8% |
| Two-phase indirect liquid cooling | 1,246 | 16.5% |
| Single-phase immersion cooling | 531 | 7.0% |
| Two-phase immersion cooling | 223 | 3.0% |
The study reports that liquid-cooling categories accounted for 83% of server-cooling patent families published in 2024, compared with 57% in 2015. That is evidence of a shift in the sampled patent portfolio. It does not establish that 83% of operating servers were liquid cooled, that patent ownership equals manufacturing leadership, or that one country controls equivalent commercial capacity. Publication timing and the study's search boundaries also limit causal claims about what triggered the increase. [5]
Beyond Conventional Liquid Cooling
Warmer coolant changes the outside of the building
NVIDIA's June 21, 2026 description of its Rubin-generation architecture specifies coolant entering at up to 45 degrees Celsius and an illustrative exit temperature of roughly 55 degrees Celsius. The company describes a fully liquid-cooled system and explains that favorable climates can permit dry-cooler operation with little or no evaporative cooling water. These are attributed platform and reference-design statements, not evidence that every installed AI facility already operates this way or achieves identical annual savings. [6]
The structural significance lies outside the chip. If heat arrives at the facility's outdoor equipment at a higher usable temperature, ambient air has more opportunity to absorb it without a refrigeration cycle. The same principle can increase the number of hours in which a facility avoids compressor-based cooling. It does not abolish the need for pumps, fans, heat exchangers, or contingency capacity. Nor does it establish that 45-degree inlet water can be produced by a dry cooler under every outdoor condition.
Every heat exchanger needs a temperature difference. Hot weather can erode dry-cooling capacity and require larger equipment, more airflow or an alternate mode. Warmer coolant helps only when the semiconductor, fluid loop and outdoor system are designed for compatible operating temperatures.
Hi K Robot's inference is not that water scarcity disappears. It is that an architecture can change which scarce resource governs expansion. A design may exchange lower evaporative consumption for more equipment area, greater peak electrical demand, or higher initial investment. That exchange can be favorable where water is constrained and the electricity and site requirements remain manageable. The same exchange may be less attractive elsewhere.
A closed loop is not the same as a water-free facility
This distinction clarifies three commonly conflated quantities. Water withdrawal is water taken from a source. Water consumption is the portion not returned within the relevant accounting boundary, often because it evaporates. Circulating volume is the liquid present in a cooling loop. A facility can circulate a large volume repeatedly while consuming relatively little replacement water, or circulate an enclosed chip-side loop while consuming water in a separate evaporative system outdoors. A photograph of liquid pipes cannot settle the water-footprint question.
The accounting boundary extends beyond the property. Electricity generation and hardware production may consume water even if a site uses no evaporative cooling. On-site WUE, regional withdrawals and lifecycle water impacts therefore answer different questions.
Microsoft provides a design-level test of the water-versus-power trade-off. In December 2024 it announced next-generation AI data-center designs using recirculating, chip-level liquid cooling to eliminate on-site evaporative water use for cooling, not all facility or upstream water consumption. The company estimated that each such facility could avoid more than 125 million liters of water per year, with the calculation based on its FY2024 global-average water-withdrawal WUE of 0.30 L/kWh; this is a modeled comparison, not measured savings at newly operating sites. Microsoft stated that the change would nominally increase annual electricity use relative to its evaporative data-center designs across the global fleet. In its June 24, 2026 update, Microsoft reported a lower FY2025 global-average WUE of 0.27 L/kWh for its fully owned data centers meeting the annual reporting criteria, compared with 0.30 L/kWh in FY2024. The newer fleet-wide figure does not retroactively change the FY2024 baseline used for the modeled 125-million-liter savings estimate. In the same 2024 announcement it scheduled pilot projects in Phoenix, Arizona, and Mount Pleasant, Wisconsin, for 2026, with facilities expected to begin coming online in late 2027. Microsoft's June 2026 water-efficiency update describes the continuing design transition but does not publicly verify completion or measured performance of those named pilots. [7]
Google's land-based Hamina data center in Finland supplies a different site-specific example: Google reports using seawater from the Gulf of Finland for cooling, while the connected district-heating recovery system began supplying heat to the local network in Q4 2025, according to operator and investor disclosures published by 2026. This is neither server immersion nor evidence of zero freshwater use across Google's fleet. It shows how a marine heat sink and a real district-energy customer can alter the site-level economics without eliminating intake, discharge, infrastructure and environmental constraints. [8]
Power usage effectiveness (PUE) compares facility energy with IT energy; site water usage effectiveness (WUE) relates operational water use to IT electricity. The definitions need explicit measurement boundaries. A facility with low PUE can still consume substantial cooling water, while low on-site water use says nothing by itself about water used upstream for electricity generation.
Move coolant closer to silicon, but count what changes
A second frontier lies inside the package. Microsoft's 2025 microfluidics demonstration etched channels into the back of silicon so that coolant could reach closer to the heat source. The company reported up to three times the heat-removal performance of cold plates in laboratory-scale tests and a reduction of up to 65% in maximum silicon temperature rise, with results dependent on configuration and chip type. These are relative results under specified tests, not a claim of three times the computing output or a universal 65% reduction in absolute operating temperature. [9]
Microchannels, two-phase cold plates and impinging jets seek to reduce local thermal resistance or increase heat removal at concentrated hot spots. They differ in pumping pressure, phase control, materials, serviceability and validated operating range, so a favorable chip-level heat-flux result must be matched to the complete system.
For example, shrinking a flow passage can increase contact with a heat-transfer surface, but it can also change pumping requirements and sensitivity to contamination. Introducing boiling can exploit latent heat, but it introduces a need to manage vapor and avoid operating beyond the stable heat-transfer regime. Integrating a cooling feature more closely with the chip can reduce intermediate resistance while making a defect or service problem harder to separate from the expensive electronic package. These are engineering trade-offs, not arguments that advanced cooling cannot succeed. [2][9]
Recover heat where there is a real customer for it
Waste-heat recovery becomes valuable when a nearby customer can use the heat at the required temperature and time. Google's Hamina facility uses seawater for cooling and transfers recoverable heat to the local district-heating system operated by Haminan Energia. First district-heat delivery occurred in Q4 2025, according to the energy operator and CapMan Infra, its infrastructure investor. In 2024, Google and Haminan Energia projected up to 80% of the local network's annual heating demand; in May 2026 CapMan instead described approximately 70–75% once the facility is fully operational. Neither percentage should be reported as an independently verified annual share already achieved. The operational evidence establishes initial heat delivery, not the ultimate revenue or energy-recovery performance. [8]
Economically useful heat requires the right temperature, timing, distance, and commercial relationship. A nearby heating network can make warm return water valuable; a distant network or a seasonal mismatch can leave much less recoverable value. A heat pump may increase the delivered temperature, but it adds electrical consumption and capital equipment. Thermal storage can shift when heat is used or rejected, but it cannot indefinitely absorb a persistent imbalance.
Underwater Data Centers: The Ocean as a Heat Sink
Coastal cooling and submerged computing are different decisions
Coastal seawater cooling does not require a submerged server enclosure. A land-based facility can use seawater through an engineered intake, heat exchanger and return arrangement, while keeping normal service access. A submerged module adds pressure containment, marine installation and recovery obligations; these are separate choices from whether the ocean is the final heat sink.
A land-based coastal facility can use seawater through an intermediate exchanger while retaining conventional service access. A submerged enclosure adds marine installation, sealing, cabling and recovery requirements, so its business case must demonstrate benefits beyond those available from seawater cooling alone.
What Project Natick demonstrated
Microsoft deployed the Northern Isles module off Scotland in 2018 and recovered it in 2020. Its September 2020 report described servers that were eight times more reliable than those in a land-based replica during the experiment. Microsoft associated the result with the controlled, sealed environment and the absence of routine human intervention, while investigating the mechanisms responsible. [10]
Natick provided useful evidence about the reliability of a sealed underwater module, but the recovered experiment did not demonstrate a repeatable commercial fleet. Deployment, maintenance and technology refresh are distinct tests from a successful operating interval.
Microsoft later ended this particular deployment path. In June 2024, the head of its Cloud Operations + Innovation division told Data Center Dynamics that the company was no longer building subsea data centers, while applying Natick's lessons elsewhere. That commercial decision does not erase the reliability experiment. It does show why proof of operation must be followed by evidence that lower cooling or failure costs outweigh marine deployment, replacement and servicing premiums. [11]
A lower component failure rate may still produce higher system cost if repairing a fault requires a vessel, retrieval and recommissioning. What matters is the combination of failure probability, service time, lost computing output and marine logistics across a full hardware refresh cycle.
From Hainan to Shanghai: commercial deployment claims
A February 19, 2025 report carried by China's State Council website described a new underwater computing module connected to an existing facility near Lingshui, Hainan. It stated that the module could house more than 400 high-performance servers and that the underlying project had entered commercial operation in March 2023. These are reported project milestones and enclosure capacity, not independently audited utilization, power density, or lifecycle cost. [12]
A newer example is the Shanghai Lingang offshore-wind-powered underwater data center. Official accounts differ on the exact start month: China's Ministry of Transport reported the first phase already operating in February 2026, while Shanghai municipal coverage places its commencement in May 2026 with 2.3 MW of installed phase capacity, within a 24 MW total two-phase plan. The first phase comprises one subsea data module, an onshore control center, and submarine power connections; reported electricity is supplied predominantly from renewable sources. Its site-specific cooling and land-use reductions are developer/government claims rather than independently audited comparative economics. The 2.3 MW phase rating is not a measured continuous IT load, and the full 24 MW remains a project-scale plan. [13]
The ocean is large; the usable interface is local
The ocean is a large thermal reservoir, but local discharge temperatures, currents and ecological requirements still constrain the exchange point. A permit and engineering design must address the actual site, not the average temperature of an entire sea.
A sealed module also faces the rapid refresh cycle of AI hardware. Removing a unit to replace obsolete accelerators creates a different operating cost from swapping boards in a terrestrial rack. Subsea viability therefore depends on recoverability, modularity and service economics as much as heat transfer.
Space-Based Data Centers: The Promise and Limits of Radiative Cooling
Why the orbital proposition deserves a serious test
The appeal of orbital computing begins with a real asymmetry. On Earth, a large facility needs a suitable site, a dependable power connection or generation system, an environmental heat-rejection route, and access to communications. Moving computation into orbit could avoid some local land and water conflicts and substitute solar generation for part of the terrestrial energy system. Under an appropriate orbital design, sunlight can be available for a much larger portion of the operating cycle than at a single ground location. These are reasons to investigate the architecture, not proof that the resulting electricity is economically free. [14][15]
Google's original Project Suncatcher research, announced in November 2025, proposed solar-powered satellites carrying tensor processing units, or TPUs, connected through optical inter-satellite links. Its design work treated orbital geometry, communications, radiation tolerance, and launch economics as joint requirements. That systems framing is more informative than a picture of a server attached to a solar panel: the payload, thermal system, power system, and network must work together before a useful computing service exists. [14]
Orbital deployment also changes the comparison baseline. Avoiding a terrestrial transmission connection might be valuable at a site where delivery is delayed, but an operator must still manufacture and launch a power source, distribute its output, reject heat, maintain orbit, and communicate with users or other satellites. These new systems are not incidental attachments. They perform functions that terrestrial infrastructure would otherwise provide.
A cold background is not a convective coolant
NASA's thermal-control guidance makes the relevant distinction explicit: a spacecraft exchanges heat with its external environment through radiation, while internal heat can move through solids and engineered fluid systems. The near vacuum does not provide the bulk convective transport available from air or water on Earth. A cold radiative background is useful, but it does not pull heat directly from a chip through empty space at an arbitrary rate. [16]
An orbital processor may still use a cold plate or internal fluid loop; the difference arises at the final boundary. Without an atmosphere or body of water, the spacecraft must send heat to surfaces that radiate it into space.
NASA also identifies sunlight, reflected solar energy, planetary infrared radiation, internal heat generation, and storage as parts of the thermal balance. A radiator looking away from the Sun can avoid one major heat input without necessarily seeing only deep space. Its view of Earth, neighboring structures, and other radiating surfaces matters. Surface coatings matter as well: aluminum is a structural material description, not a complete specification of infrared emissivity or absorbed solar energy. [16]
There is a further distinction between thermal radiation and damaging space radiation. The former is the energy-transfer mechanism used by an infrared radiator. The latter can include energetic particles that affect electronics. A successful heat-rejection system does not establish radiation tolerance, just as a radiation-tested processor does not establish sufficient radiator capacity.
SpaceX AI1: the 1,300–1,400 W/m² design claim
Elon Musk’s late-Monday November 2, 2026 U.S.-time discussion (November 3 on some UTC displays) proposed roughly 1,300 W/m² of radiator rejection. The engineering envelope preceded that post: in a June 9 SEC-filed discussion, SpaceX described its draft AI1 architecture at 120 kW of sustained compute and 150 kW peak power, with 250 W/m² solar-array generation and 1,400 W/m² radiator rejection from both faces of a panel oriented edge-on to the Sun. Contemporaneous technical reporting describes approximately 110 m² of deployable radiator panel. These are draft design targets, not measured orbital performance. [17]
For a uniform radiating surface, the simplified heat-balance model is:
Q = εσA(Trad4 − Tbackground4)
Here Q is net radiation in watts, A is total effective emitting-face area, ε is emissivity, and σ is the Stefan–Boltzmann constant. The calculations below assume ε = 0.90, a near-3 K radiative background and unobstructed faces, without absorbed solar or Earth radiation. NASA’s thermal-control guidance describes the additional terms required for a flight design. [16]
Area convention matters: one square meter of radiator panel planform can provide roughly two square meters of emitting area if both faces radiate effectively. Under the ideal assumptions above, 1,300 W/m² of two-sided planform corresponds to about 336 K (63 °C), and 1,400 W/m² to about 342 K (69 °C). A one-sided panel at the same temperature rejects only approximately half the planform-based flux. These temperatures test physical consistency, not operational readiness.
The 110 m² AI1 radiator makes the distinction concrete. At full utilization of both emitting faces, the claimed 1,400 W/m² corresponds to roughly 154 kW of ideal rejection, close to AI1’s stated 150 kW peak. If the entire cited power appeared as radiator heat, the 120 kW average case requires about 1,091 W/m² and an ideal surface near 322 K (48 °C); the 150 kW peak case requires about 1,364 W/m² and near 340 K (67 °C). This is a simplified Hi K Robot calculation using the June design envelope, not a temperature reported by SpaceX. The spacecraft’s actual heat budget could differ from nominal compute power.
The coolant-temperature implication changes with load. A 55 °C outlet from a terrestrial-style warm-water loop is hotter than the ideal 48 °C average-load radiator, so passive transfer is possible in principle if heat-exchanger and piping temperature differences remain sufficient. At 67–69 °C peak/rated rejection, that same hypothetical 55 °C outlet could not passively heat the radiator: a hotter qualified loop, another thermal architecture or a heat pump would be necessary. SpaceX has not specified 55 °C as its AI1 coolant temperature. Real solar/Earth loading, surface gradients, loop temperature drops and auxiliary waste heat can narrow either margin.
The solar side is not optional. At the stated 250 W/m² electrical output, 120 kW needs a theoretical 480 m² of illuminated arrays and 150 kW needs 600 m², before mission-specific power, degradation, eclipse, storage and conversion margins. The published June configuration therefore combines a large solar array with a smaller, hotter two-sided radiator; neither area figure alone determines spacecraft mass, cost or reliability.
Scaling the radiator from one megawatt to one gigawatt
The following scenarios assume a single emitting face with emissivity 0.90 and show the area needed to reject continuous thermal loads at specified radiator surface temperatures. For an ideal two-sided panel, multiply the tabulated single-face flux by 2 and approximately halve its required physical planform area if both faces are equally unobstructed and at the stated temperature. Auxiliary equipment, absorbed environmental radiation, supports, spacing, protection and gradients require separate allowances.
| Radiator temperature | Net flux, ideal scenario | Area for 1 MW heat | Area for 100 MW heat | Area for 1 GW heat |
|---|---|---|---|---|
| 300 K / 27 °C | 413 W/m² | 2,419 m² | 0.242 km² | 2.419 km² |
| 330 K / 57 °C | 605 W/m² | 1,652 m² | 0.165 km² | 1.652 km² |
| 350 K / 77 °C | 766 W/m² | 1,306 m² | 0.131 km² | 1.306 km² |
| 400 K / 127 °C | 1,306 W/m² | 765 m² | 0.077 km² | 0.765 km² |
Calculation boundary: These are heat-rejection scenarios, not claimed spacecraft designs. A 1 GW IT installation would require accounting for platform and power-system losses in addition to IT heat. The displayed temperatures are radiator surface temperatures, not coolant inlet specifications, semiconductor junction temperatures, or approved operating limits for a named accelerator.
At SpaceX's assumed 250 W/m² electrical output, a continuously illuminated 1 GW array requires at least 4 km² of panel planform before losses or auxiliary power. The 330 K single-face radiator scenario requires 1.652 km² of emitting area for a separate 1 GW heat load; an ideal equally exposed two-face panel would halve its physical planform. The comparison concerns area, not verified hardware mass or continuous orbital yield.
Radiative output rises with the fourth power of absolute temperature, but at a fixed temperature and emissivity the required area scales with the heat load. Manufacturing more panels may reduce unit price; it does not eliminate this physical scaling requirement.
Real radiator panels also develop temperature gradients, face obstructions and absorb environmental radiation. These effects reduce net rejection relative to the ideal uniform-temperature table, so spacecraft thermal design requires margin rather than treating the calculated areas as finished specifications.
Area becomes mass, manufacturing, and deployment
An area calculation becomes economically interesting only when connected to a mass model. At 330 K in the table, rejecting 1 GW requires approximately 1.652 million square meters of effective emitting area. As an explicitly hypothetical sensitivity, assume an installed radiator-system mass of either 2 or 10 kilograms per square meter of that effective area. The resulting radiator mass would be approximately 3,305 or 16,523 metric tonnes, respectively. Neither assumption is presented as a measured mass for a commercial orbital data center.
The definition of the mass coefficient matters. If a supplier quotes kilograms per square meter of physical panel planform, a model must first account for whether one or two faces effectively emit. Otherwise, multiplying a planform mass coefficient by a two-sided emitting area can double-count structure. For clarity, the sensitivity above defines its coefficient directly against effective emitting area and does not make a separate claim about physical panel thickness or construction.
These estimates exclude computing equipment, solar arrays, batteries, power conversion, communications, propulsion, radiation protection, and reserve hardware. They are not a spacecraft bill of materials. Their purpose is to expose the leverage of areal mass: a design change that appears modest in kilograms per square meter can become decisive when repeated across a megawatt-to-gigawatt infrastructure program.
Deployment creates another distinction. A structure can be light enough to launch in aggregate and still be difficult to package, unfold, connect, position, or repair. A distributed fleet can reduce the size of each individual structure, but it introduces more interfaces and a network architecture. A centralized platform can share infrastructure, but it may concentrate failures and demand larger assemblies. Neither form escapes the requirement to connect emitting surfaces reliably to their heat sources.
A credible scaling case would therefore connect four measurable quantities: rejected thermal power, effective emitting area, installed system mass, and sustained availability. Quoting only the first two leaves the logistics incomplete. Quoting a launch price without the third leaves the transport cost incomplete. Quoting all three without the fourth leaves the useful-service denominator incomplete.
What has actually operated in orbit
Starcloud states that Starcloud-1 launched in November 2025 carrying an NVIDIA H100 GPU and subsequently performed small-model training in space. This is concrete evidence of high-performance computing hardware operating in orbit. It is not evidence that a megawatt-scale cluster has achieved the same behavior, or that the complete lifecycle cost has been demonstrated against a terrestrial alternative. The company's description is used here for the launch and computing milestones, not as an independent audit of economics. [18]
Google reported on October 30, 2026 that its Project Suncatcher prototype, launched on October 1, had powered its host and TPU trays, run early machine-learning tests and returned an AI-generated response. Reporting on the initial MVP satellite specifies four TPUs and about 1 kW of solar electrical capacity; the test can power only one TPU at a time for limited periods rather than run four processors continuously. Google's own update identified longer thermal-stress and radiation tests as further work and proposed twin interconnected satellites in 2027. Compared with a hypothetical 1 GW orbital power system, a 1 kW prototype is six orders of magnitude smaller in nominal power—not an equivalent measurement of useful compute capacity. [14][19]
Google's Project Suncatcher lead Travis Beals also told U.S. public-radio reporters that radiator hardware was among the heaviest elements of the current mission. This observation gives the thermal-area argument an immediate mass-budget counterpart even at prototype scale: operating a TPU in orbit is a major milestone, while lowering kilograms per rejected kilowatt remains a separate challenge. [19]
SpaceX's ambitions have a more specific documented starting point than a hypothetical billion-satellite constellation: the AI1 draft satellite presented in its June 2026 SEC-filed technical discussion. Its 120 kW sustained-compute target is more than two orders of magnitude above the power budget reported for Google's roughly 1 kW prototype, but it remains a proposal rather than a demonstrated AI1 flight result. [17]
The U.S. Government Accountability Office's April 2026 assessment offers a useful independent boundary. It distinguishes familiar spacecraft components from their unproven integration and scaling for large data centers, and identifies power, cooling, communications, manufacturing, launch cost, and orbital congestion as connected challenges. Its assessment should not be read as denying the later demonstrations, but neither do those demonstrations automatically resolve the system-scale questions it identifies. [15]
Workload and network topology determine what counts as success
A satellite that analyzes data already collected in orbit can reduce the amount of information that must be transmitted to Earth. That can create value even when its computing cost would not compete with a terrestrial cluster on a generic benchmark. Independent inference requests may also fit a distributed architecture differently from training that requires frequent synchronization among many accelerators. These are distinct markets, not merely successive sizes of the same data center. [15]
Google's original design research recognized this coupling by studying high-bandwidth optical links and closely spaced satellite formations. Its laboratory link demonstration and orbital modeling are evidence about particular parts of a proposed network, not measurements of a completed computing constellation. The relevant implication is that thermal modules cannot be placed wherever cooling is easiest without regard to the communication pattern required by the workload. [14]
Tightly coupled AI jobs need predictable interconnect performance. Even a thermally stable satellite cluster can leave processors idle if optical links cannot provide the throughput, latency and availability required by the workload. The most suitable early workloads may be those that already originate in orbit.
Future advances could change that balance. More tolerant training methods, better optical links, different memory architectures, or workloads naturally suited to distributed execution could make orbital systems more attractive. Those are conditional pathways. They should be evaluated as changes to the complete system rather than assumed in advance whenever a cooling estimate requires an optimistic economic outcome.
The Economics of Cooling at Scale
Compare two axes before comparing costs
Cooling technology and deployment location are separate design choices. A rack with direct-to-chip cooling can operate in a terrestrial, coastal or orbital system; the external heat sink and maintenance model must be compared independently.
| Heat-capture approach | Principal advantage | Main deployment trade-off |
|---|---|---|
| Air cooling | Familiar hardware and service practices | Airflow and density constraints at high concentrated loads |
| Direct-to-chip liquid cooling | Targeted heat capture with a conventional rack pathway | Fluid-system qualification, interfaces, and residual heat management |
| Immersion cooling | Broad contact with compatible electronics and reduced fan dependence | Fluid compatibility, servicing, hardware adaptation, and ecosystem integration |
| Deployment environment | Potential advantage | What must be included in feasibility |
|---|---|---|
| Terrestrial, including coastal | Accessible maintenance and established infrastructure; possible heat reuse | Site power, climate, water arrangements, network, permits, and expansion space |
| Subsea enclosure | Marine heat sink and a controlled modular environment | Shared power and data links, enclosure reliability, deployment, recovery, and local marine conditions |
| Orbital system | Solar access and reduced dependence on some terrestrial resources | Radiators, arrays, launch mass, communications, orbital operations, replacement, and end-of-life obligations |
Comparison boundary: These tables summarize mechanisms and trade-offs established in the preceding sections; they are not numerical rankings. Commercial maturity differs across individual projects. A demonstrated component, a pilot installation, and a production fleet cannot be assigned equivalent evidentiary weight simply because each has a working prototype.
Capital cost, operating cost, and the cost of waiting
Cooling CAPEX includes cold plates or air systems, pumps, heat exchangers, electrical interfaces, redundancy, installation and facility modifications. At sea it can include vessels and pressure enclosures; in orbit, launch, deployed structure, thermal hardware and spacecraft operations.
OPEX includes energy, water where applicable, fluids, maintenance, replacement and unavailable compute. Cheaper cooling equipment may be a poor trade if it increases downtime or delays a site from becoming operational. Compare the full service delivered, not purchase cost alone.
The IEA’s 2026 data product reports installed capacity, power usage effectiveness and load factors alongside modeled electricity consumption. These measures make clear why nominal installed hardware alone cannot determine energy use or useful annual computation. A full comparison must also include utilization, equipment replacement schedules and the cooling plant’s service life. [1]
A lifecycle comparison should distinguish cash expenditure from accounting depreciation. Counting the full initial hardware cost and then adding depreciation as another cash expense would double-count the same investment. Likewise, equipment refresh must be timed consistently across alternatives. A terrestrial design evaluated over a long building life cannot be compared with an orbital proposal that omits the replacement of short-lived electronics, unless the comparison explicitly explains the different service arrangements.
The denominator is reliable, useful computation
Hi K Robot uses the following analytical expression, not a new industry standard:
Lifecycle cost per useful compute unit = discounted lifetime system cost, net of realizable heat-use value / discounted useful computation delivered at the required service quality.
The system cost includes processors, power, cooling, transport, operation, hardware refresh and retirement; the denominator is useful work at a specified quality of service. Training, inference and space-native processing may produce different denominators and should not be mixed without qualification.
Cooling should be valued by its effect on sustained useful output. A modest rise in auxiliary energy can be justified if it substantially improves uptime or delivered accelerator performance; the converse is also possible.
An illustrative utilization comparison shows the leverage. If two otherwise identical capital assets operate productively for 90% and 60% of the same period, the second spreads its fixed capital cost across only two-thirds as much useful work. Its fixed capital cost per unit of work is therefore 50% higher. This is arithmetic under a controlled assumption, not an empirical claim about air, liquid, subsea, or orbital availability. It explains why maintenance and networking can overturn an apparent energy advantage.
Quantify savings without confusing a scenario with a quotation
For an illustrative constant 100 MW IT load, a 0.05 reduction in annual PUE saves 5 MW of average facility electricity, or 43.8 GWh over 8,760 hours. This is an arithmetic scenario, not an observed savings claim or a value for all weather conditions.
The same discipline applies to orbit. Using the earlier hypothetical 330 K single-emitting-area scenario and an assumed radiator-system coefficient of 2 kilograms per effective emitting square meter, a 1 GW thermal load implies approximately 3.3 million kilograms of radiator-system mass. At hypothetical transport prices of $100 and $1,000 per kilogram, transporting that mass alone would cost around $330 million and $3.3 billion. This sensitivity excludes solar arrays, computing hardware, assembly, spare capacity and orbital operations. Under the same 250 W/m² solar-output assumption used for SpaceX AI1, the minimum 4 km² of arrays would add another potentially substantial mass and deployment requirement.
Google's initial Suncatcher research explored a more specific conditional pathway in which sustained launch-cost learning could bring prices below $200 per kilogram by the mid-2030s. Its comparison concerned launching and operating the proposed system relative to reported terrestrial energy costs on a per-kilowatt-year basis. That is not automatically parity with the full lifecycle cost of a terrestrial computing service. The distinction matters because different denominators can turn a useful research hypothesis into an overstated commercial claim. [14]
A counterweight came from within Google's own data-center organization. The Information reported on November 1, 2026 that Google's data-center energy head, Amanda Peterson Corio, described launching and building computing infrastructure in orbit as potentially up to 100 times more expensive than terrestrial construction, while also identifying maintenance and upgrades as additional challenges. The remark was an attributed executive assessment of current deployment economics, not an audited 100-fold multiplier for comparable lifetime computing services. It contrasts with Project Suncatcher's conditional below-$200/kg future launch-price scenario, which addresses one input to the business case rather than the complete cost of radiators, arrays, replacement, communications, and useful compute. [20]
Scale can reduce unit cost without removing physical requirements
Economies of scale may arise through repeated manufacture, shared infrastructure and learning. Yet physical requirements such as radiator area per megawatt can remain approximately proportional to load. Shared infrastructure can also concentrate failure risk.
There can also be diseconomies. A bigger shared cooling system can create a more consequential common-mode failure. A denser collection of outdoor heat exchangers can require attention to recirculating warm exhaust. A larger marine development can intensify the need to assess local heat dispersion. A larger orbital constellation can require more coordination and increase exposure to collision and communication constraints. GAO identifies orbital congestion and frequency coordination among the issues that accompany growth beyond an individual demonstration. [15]
Modularity can help manage these problems by containing failures and allowing incremental construction. It can also duplicate overhead. The optimal module size depends on the workload, service model, and infrastructure interfaces; it cannot be inferred from the desire to manufacture more identical units. Repetition is a strategy for scaling a design, not proof that the design is already economically scalable.
Reliability includes the failure of shared infrastructure
A useful cooling system needs to be evaluated during maintenance and credible faults, not only at its best operating point. Redundant pumps or cooling modules can preserve service, but redundancy has meaning only when shared dependencies are understood. Two pumps on the same vulnerable power or control path do not necessarily provide the resilience implied by simply counting two units.
Serviceability differs by environment: technicians can reach terrestrial equipment, marine modules may need recovery, and orbital hardware may require redundancy or replacement launches. Each path can work for particular workloads, but maintenance and downtime must appear in the same cost denominator.
The Physical Geography and Economic Limits of AI Civilization
Site capacity is a combination of available conditions
Site capacity is not determined by one resource. Cheap electricity without grid delivery, cool weather without network access, or abundant water without permits can each leave an otherwise viable project stranded. Cooling feasibility is therefore a local engineering and infrastructure question.
Climate should therefore be interpreted as an operating envelope rather than a geographic slogan. Dry-bulb temperature constrains dry heat rejection; humidity and wet-bulb conditions affect the opportunity for evaporative cooling. Elevation, seasonal extremes, and heat-wave conditions can further change equipment behavior. The engineering comparison is between designs operating through local conditions, not between broad labels such as cold country, desert, coast, or space. [2]
Higher-temperature liquid cooling may broaden the set of terrestrial locations where low-water operation is practical. Coastal heat exchange may provide an attractive option where marine infrastructure and environmental conditions align. Orbital computing could create another option for workloads and hardware designed around its strengths. None of these possibilities implies that infrastructure becomes independent of geography. Each changes the geography and the resources that matter.
The binding constraint can migrate
This article uses sustainable deployable compute capacity as an analytical description: the useful computing service that can be maintained within a specified power, thermal, resource, network, reliability, and economic envelope. It is not a standardized engineering metric, a universal sustainability certification, or a single number that can be read from a facility's nameplate.
Better chip cooling can move the bottleneck to facility distribution; dry cooling may shift it to land and electricity; orbital deployment may trade terrestrial constraints for launch mass and communications. A viable architecture must address the newly binding constraint rather than merely relocating it.
Stakeholders optimize different outcomes: chip suppliers prioritize qualified performance, facility operators lifetime service, communities local resources, and orbital developers mass and launch cadence. No single efficiency metric captures these incentives or their costs.
What would change the assessment
For terrestrial liquid cooling, the decisive evidence would include repeatable operation at high density across realistic climates, with measured water and auxiliary-energy performance and a serviceable supply chain. For immersion, it would include lifecycle advantages that persist after hardware adaptation, fluid management, and maintenance are counted. For subsea systems, it would include multi-cycle operating data that connects thermal performance with availability, recovery, replacement, and marine effects.
For orbital infrastructure, the strongest evidence would be sustained useful computing supported by a measured thermal and power budget, together with demonstrated networking, credible system mass, repeatable deployment, and an economic refresh strategy. A replicated orbital fleet delivering comparable workloads at competitive, fully measured lifecycle cost would materially weaken the present assessment that hyperscale orbital economics remain unverified. A profitable space-native application could be an important milestone without demonstrating parity for all terrestrial workloads; conversely, one failed architecture would not establish that every orbital approach is physically or economically impossible.
The AI Civilization Map makes these dependencies visible. This node sits within Energy & Physical Infrastructure, in the Cooling, Water & Site Capacity branch. It connects the thermal consequences of semiconductor and packaging choices with power delivery, network coordination, physical siting, and capital recovery. These cross-system dependencies explain the article's structural function. The Related K Robot Research section identifies the connected studies, the direction of each relationship, and its causal mechanism.
Conclusion
The thermal limit of AI compute is not a universal temperature ceiling. Better cold plates, warmer liquid loops, immersion tanks, seawater exchange and orbital radiators can each shift a constraint, but none independently supplies the power, reliability, site conditions and capital needed for sustained operation.
Existing evidence separates working mechanisms from commercial scale. Microsoft ended Natick after a useful experiment; Shanghai reports a 2.3 MW first phase; orbital processors have demonstrated early computing milestones, not terrestrial hyperscale equivalence. Each future pathway requires measured performance over its real service life.
For terrestrial, subsea and orbital AI infrastructure alike, the relevant ceiling is not installed processor count but useful computation that can be continuously delivered at an economically sustainable cost. Cooling innovations raise that ceiling only when the full heat-rejection chain can scale alongside them.
Five Proof Gates for Scalable AI Cooling
| Requirement | Evidence needed for scale | Boundary |
|---|---|---|
| Chip-to-site thermal chain | Sustained cooling performance under full workload, design weather and credible equipment faults | Laboratory heat flux is not full-site capacity |
| Water and environment | Site-level heat rejection, electricity and water measurements within a stated resource boundary | Zero on-site evaporative use does not imply zero water footprint |
| Underwater operation | Long-run availability and total cost after vessel deployment, cables, recovery and hardware refresh | Natick's research result and Shanghai's project estimates are different evidence types |
| Orbital thermal architecture | Measured radiator flux, operating temperature, areal mass, array generation and heat-transport performance | 1,400 W/m² is a design parameter, not fleet validation |
| Commercial economics | Repeatable cost per reliable useful compute unit, including replacement and communications | A successful chip demonstration is not a 1 GW business case |
The governing test is how much useful AI computation a system can sustain, for how long, and at what fully accounted physical and economic cost.
Frequently Asked Questions
Why does AI cooling become harder when chips become more energy efficient?
Energy efficiency and heat density describe different quantities. A system can perform more work per watt while concentrating more total power into each chip or rack, making local heat removal more demanding even as efficiency improves.
Does direct-to-chip liquid cooling necessarily consume large amounts of water?
No. Liquid circulation near the chips does not determine whether the facility ultimately uses dry air cooling, evaporative equipment, seawater, or another heat-rejection route. A closed chip-side loop can coexist with either low or substantial water consumption elsewhere in the system.
Why can 45-degree coolant remove heat from AI processors?
Heat flows from a hotter component toward a cooler fluid. A qualified system can therefore use coolant that feels warm to a person, provided the temperature differences and component limits remain adequate throughout the entire thermal chain.
Is an underwater data center the same as immersion cooling?
No. An underwater data center places computing equipment inside a submerged enclosure, while immersion cooling places compatible electronics in a dielectric fluid. A subsea enclosure can use internal cooling arrangements other than immersion and does not require seawater to contact the electronics.
Does the cold of space make data-center cooling easy?
Space provides a cold radiative background but no external convective medium comparable to air or water. A spacecraft must transport internal heat to emitting surfaces whose temperature, emissivity, area, orientation, and environmental exposure determine net rejection.
Is 1,300 W/m² of orbital radiator cooling physically plausible?
Yes, under a specified geometry: an idealized radiator emitting from both sides with emissivity 0.9 needs a uniform surface temperature near 336 K (63 °C) to radiate about 1,300 W/m² of panel planform toward a 3 K background. SpaceX's 1,400 W/m² assumption similarly corresponds to roughly 342 K (69 °C); neither result demonstrates that a particular cooling loop can maintain that surface temperature economically or continuously in orbit.
Have orbital AI data centers already matched terrestrial hyperscale facilities?
The cited Starcloud and Google records establish processor and model-execution milestones in orbit, not equivalent validation of megawatt-to-gigawatt production fleets. Comparisons with terrestrial facilities require sustained output, measured system budgets, and complete lifecycle costs under comparable workload assumptions.
Sources
Sources prioritize original technical records, official data, and attributed reporting; the original Chinese-language patent studies appear alongside them. Claims about planned designs and projected savings remain identified as such.
- [1] International Energy Agency. Key Questions on Energy and AI. April 2026, Annex A, Table A.1, p. 108: the IEA's 2025 estimated 485 TWh and 2030 scenarios of 945 TWh (Base), 1,008 (Lift-Off), 868 (High Efficiency), 833 (Headwinds). IEA data product identifies installed capacity, PUE, load factor and electricity consumption in the underlying dataset.
- [2] International Telecommunication Union. Recommendation ITU-T L.1327: Guidelines on the Selection of Cooling Technologies for Data Centres in Multiple Scenarios. August 2024. Cooling-system selection and the distinction between indoor exchange, transport, outdoor dissipation, and refrigeration.
- [3] U.S. Department of Energy, Federal Energy Management Program. Best Practices Guide for Energy-Efficient Data Center Design, July 26, 2024. Covers IT systems, electrical and cooling equipment, air management, heat recovery, and energy-water interactions.
- [4] Open Compute Project. Cooling Environments Project. Ongoing technical workstreams covering cold plate, coolant distribution unit, immersion, door heat exchangers and heat reuse; consulted November 22, 2026.
- [5] Taiwan Intellectual Property Office. Companion studies, Intellectual Property Monthly, No. 335 (November 2026): Part I, Global Patent-Technology Trends (original PDF, pp. 5–25) and Part II, Data-Center Thermal Management and Server Cooling (original PDF, pp. 26–56). These distinct analyses report the 8,449-family global patent dataset and its classified cooling technologies; they are not market-installation statistics.
- [6] NVIDIA. Hotter Than a Hot Tub: The 45°C Breakthrough to Cool AI’s Biggest Machines. June 21, 2026. Company description of warm-water liquid cooling and climate-dependent dry heat rejection.
- [7] Microsoft. Sustainable by Design: Next-Generation Datacenters Consume Zero Water for Cooling, December 9, 2024. The 125-million-liter modeling footnote specifies Microsoft's FY2024 global-average water-withdrawal WUE of 0.30 L/kWh; the company describes a nominal annual energy increase and planned 2026 pilots. See also Microsoft's June 24, 2026 water-efficiency update and its official PUE/WUE table, reporting global WUE of 0.27 L/kWh for FY2025 versus 0.30 L/kWh for FY2024 across the specified owned operational fleet. These sources do not provide verified operating results for the named 2026 pilot sites.
- [8] Google, Haminan Energia and CapMan Infra. Google's Hamina data-center location documents seawater cooling; Google's May 20, 2024 announcement attributes the original 80% district-heating projection to Haminan Energia. Haminan Energia's sustainability statement confirms heat recovery began on the main network in late 2025. CapMan Infra, May 18, 2026, confirms first heat in Q4 2025 and estimates about 70–75% of local district-heating energy needs once fully operational; that share is prospective rather than audited delivery.
- [9] Microsoft. AI Chips Are Getting Hotter. A Microfluidics Breakthrough Goes Straight to the Silicon to Cool Up to Three Times Better. 2025; accessed November 22, 2026. Company-reported laboratory results and their configuration-dependent performance boundaries.
- [10] Microsoft. Microsoft Finds Underwater Datacenters Are Reliable, Practical and Use Energy Sustainably. September 14, 2020. Project Natick deployment, recovery, and a specific experimental reliability comparison.
- [11] Data Center Dynamics. Microsoft Confirms Project Natick Underwater Data Center Is No More. June 17, 2024. Directly quotes Microsoft's cloud operations executive on the cessation of subsea builds.
- [12] State Council of the People’s Republic of China / Xinhua. China Launches Pioneering Underwater Intelligent Computing Cluster in Hainan. February 19, 2025. Reported project milestones and module capacity, rather than independently audited operating economics.
- [13] Shanghai Municipal Government / China Daily. Waves with World's First Wind Power Undersea Data Center. June 1, 2026. Official municipal republication identifying May commercial operation, 24 MW planned capacity and project-attributed cooling/land claims. Shanghai municipal follow-up, August 7, 2026, identifies 2.3 MW first-phase capacity within a 24 MW plan. The Chinese Ministry of Transport, February 14, 2026 reports the same first phase operating by February 2026; the two official accounts differ on the reported start month.
- [14] Google Research / Google. Exploring a Space-Based, Scalable AI Infrastructure System Design, November 4, 2025 (concept, orbital and network model, cost conditions); and We've Received Our First Response from the Project Suncatcher Prototype Satellite, October 30, 2026 (flight, TPU tests and continuing validation). Research assumptions and flight observations are different evidence categories.
- [15] U.S. Government Accountability Office. Science & Tech Spotlight: Data Centers in Space. April 28, 2026; GAO-26-109012. Opportunities, infrastructure maturity, and engineering and coordination challenges.
- [16] NASA. Small Spacecraft Technology State of the Art: 7.0 Thermal Control. May 7, 2026. Spacecraft heat balance, internal heat transport, radiators, optical properties, and environmental heat inputs.
- [17] Space Exploration Technologies Corp. (SpaceX). SEC-filed technical discussion featuring AI1 satellite assumptions. Filed June 9, 2026; 150 kW peak, 120 kW sustained computing, 250 W/m² solar, 1,400 W/m² double-sided radiator. Contemporaneous June 9 DCD report for the additional 110 m² radiator specification; Related late November 2 U.S.-time / November 3 UTC statement on 1,300 W/m².
- [18] Starcloud. Starcloud-1. Mission record; accessed November 22, 2026. Company-reported H100 launch and small-model training milestones.
- [19] Independent prototype reporting. TechCrunch, Google's First Data Center Satellite Is Online and Writing Poetry, October 30, 2026 (four TPUs, approximately 1 kW solar, operating duty cycle); and Oregon Public Broadcasting / NPR, Google Launches Project Suncatcher, a Step Towards AI Data Centers in Space, October 2, 2026 (OPB publication timestamp; the satellite launch was October 1; Travis Beals on radiator mass). Both are attributed reporting, not a Google engineering specification.
- [20] The Information. Google Energy Chief: Space Data Centers Cost Up to 100 Times More Than Those on Earth. November 1, 2026. Its publicly accessible news listing attributes the reported remark to Amanda Peterson Corio at The Information's WTF summit. Data Center Dynamics, November 2, 2026, separately reports the attributed cost and maintenance comments. These statements are estimates of present launch and build economics, not a published matched-workload lifecycle-cost model.
Related K Robot Research
Relationship direction: The Thermal Limits of AI Compute → linked Active Node. Each predicate expresses a causal dependency examined in this research; the linked article is the endpoint, not its Layer or Branch.
- Why Space Becomes the Next Infrastructure Layer of AI Civilization — constrains
- SpaceX (SPCX): What Role Will It Play in the Future of AI Civilization — constrains
- U.S. AI Power Delivery & Grid Coordination — depends on
- AI Energy Infrastructure and the HVDC Bottleneck — forms a feedback loop with
- AI Data Center Power Architecture: Nuclear, Gas & Onsite Power — depends on
- The Fiber Limit of AI Civilization: Why Data Movement, Not Compute, Defines AI Scaling — depends on
- AI’s Next Bottleneck: The Coming Semiconductor Equipment Supercycle — is downstream of