If you ask me what the single most underappreciated bottleneck in India’s AI defence ambitions is in 2026, I would not say compute. I would not say software. I would say heat. Specifically, the inability to dissipate heat fast enough from high-performance AI chips in environments where conventional cooling simply cannot function. Power densities in leading-edge AI processors have already been compared to the surface of the sun. That is not hyperbole. Modern AI accelerators consume upwards of 600 to 1,000 watts per chip. The heat flux generated per square millimetre of silicon now exceeds that of a nuclear reactor core. Every drone, every autonomous weapons system, every edge computing node on a fighter aircraft, every command centre running real-time intelligence analysis, is fundamentally a heat management problem wearing a software problem’s clothes. India has been solving the software problem. The heat problem has been waiting for someone to take it seriously.
Why Heat Is Now a Strategic Variable, Not Just an Engineering Problem
The way I read the current state of AI hardware in defence, the thermal bottleneck has crossed from a technical inconvenience into a strategic vulnerability. Three specific mechanisms explain why. The first is thermal throttling in combat. Consider a reconnaissance drone operating in contested airspace: the platform must simultaneously process high-resolution camera feeds, synthetic aperture radar returns, electronic signals intelligence, and navigation data. Sending raw data back to a command centre introduces network latency and data-security risks. The solution is onboard edge AI processing. But when the compute load spikes during active operations, chips heat up. If the thermal management system cannot dissipate that heat fast enough, the chip automatically reduces its clock speed to prevent physical damage. In a commercial laptop, this is a minor inconvenience. In a loitering munition over a contested airspace, it is the difference between acquiring a target in 200 milliseconds and acquiring it in 800 milliseconds. At the speeds that modern air defence systems operate, that 600-millisecond gap is a mission failure.
The second is infrared signature exposure. Military electronics face temperature ranges from arctic conditions at minus 65 degrees Celsius to propulsion environments exceeding 1,100 degrees Celsius. High-power AI chips in airborne or terrestrial defence assets emit substantial thermal energy. Without advanced thermal dissipation, the platform’s infrared signature rises sharply. A higher thermal signature makes the vehicle detectable by hostile infrared search and track sensors, thermal imaging satellites, and heat-seeking anti-air missiles. India’s AMCA stealth fighter programme faces this problem directly. A stealth aircraft that leaks its position through an inadequately managed thermal signature is not a stealth aircraft. It is an expensive target.
The third mechanism is the non-linear consequence that connects heat to sovereign AI infrastructure. Cooling accounts for roughly 30 to 40 percent of a data centre’s total energy consumption. India’s sovereign AI ambitions depend on data centres. Those data centres depend on cooling systems. If those cooling systems rely on foreign dielectric fluids, foreign heat exchangers, or foreign liquid cooling hardware, then India’s sovereign AI compute has a non-sovereign supply chain. A supply chain blockade during a geopolitical conflict does not have to target India’s chips. It can target the fluids that keep India’s chips from melting. The strategic chokepoint moves down one layer, and nobody is talking about it.
The Battlefield in Real Time: What Thermal Failure Actually Looks Like
Let me take this from the abstract to the operational. From first principles, a military edge AI system is a closed thermal system with inputs, a computational process, and an output that must go somewhere. In a commercial environment, that output goes into the building’s HVAC system. In a military environment, it must go somewhere else, because there is no HVAC system on a Tejas MK-2 at 50,000 feet. In this context, the solutions available are air cooling, which fails at high chip densities and generates acoustic signatures; liquid cooling loops, which work but rupture under G-forces and vibration; phase-change materials, which absorb heat by changing state but add weight; vapour chambers, which redistribute heat laterally across a surface; and immersion cooling, which submerges electronics in dielectric fluid, the highest-density solution but also the most infrastructure-dependent. Next-generation AI chips stack compute and memory so tightly that heat gets trapped deep in the package. This is pushing cooling directly into the chip design itself, including microchannels, direct-die liquid manifolds, and early two-phase concepts. The chip design and the cooling design are no longer separable engineering problems. You cannot buy a chip optimised for maximum AI performance and then figure out the cooling separately. The cooling architecture has to be designed into the silicon itself, which means if India is designing its own chips under the India Semiconductor Mission, it must simultaneously design their thermal management. Treating cooling as an afterthought is how you get a world-class chip that throttles to 40 percent of its rated performance in Ladakh in August.
What India Has and What Is Still Missing
India is not starting from zero on this. The ecosystem is more developed than most defence commentary acknowledges, though it is also more fragmented than the strategic situation allows. Bharat Electronics Limited develops defence-grade liquid-cooled avionics and ruggedised tactical computers built to withstand extreme military shock, vibration, and temperature ranges. L&T Defence designs custom liquid-cooling loops and heat exchangers for naval combat systems, airborne radar systems, and missile command units. Tonbo Imaging, which produced the electro-optic systems used in Operation Sindoor, engineers compact thermal management for defence payloads and edge processing hardware on small UAVs and armoured vehicles. On the data centre side, STT GDC India operates dedicated liquid cooling innovation labs in Pune testing direct-to-chip and immersion cooling setups tailored for high-density AI compute.
The leverage point that is currently missing is not capability in any single domain. It is integration across domains, and specifically the absence of a military-grade thermal standard that bridges chip design, system integration, and battlefield deployment into a single coherent framework. My own assessment of the gap: India can cool a data centre. India can cool a radar system. India cannot yet cool a 1,000-watt AI accelerator mounted inside a loitering munition operating in the Thar Desert at 48 degrees ambient temperature under 9G manoeuvring loads, using only domestically sourced materials. That specific capability is the missing piece, and it is the piece that will determine whether India’s autonomous weapons systems can sustain peak performance through an entire mission rather than throttling down at the moment of highest demand.
The Four Things India Must Build Now
The roadmap here is specific and actionable, and it builds directly on what India already has. The first priority is integrating thermal architecture into the India Semiconductor Mission. Domestic chip design initiatives must design microfluidic cooling channels directly into 3D-stacked silicon architectures at the design stage, not as a post-fabrication addition. Co-designing silicon and cooling is how TSMC and Intel are approaching their next-generation products. India’s DIR-V programme and any future indigenous AI accelerator effort must adopt the same philosophy from the start.
The second is developing indigenous dielectric fluids. Immersion cooling, the most effective high-density thermal management approach available, requires non-conductive, low-boiling-point dielectric fluids. Currently, the global market for these fluids is dominated by 3M and Chemours, both American companies. A geopolitical disruption that cuts off dielectric fluid supply does not need to be an export control on chips to cripple India’s AI compute capacity. Developing a domestic production line for these fluids is not glamorous. It is essential.
The third is launching dedicated iDEX challenges for military-grade thermal management. The Innovations for Defence Excellence framework has successfully pushed Indian startups to build combat-proven autonomous systems. The same framework should now define specific thermal challenges: G-tolerant two-phase immersion cooling for fighter aircraft, zero-water closed-loop cooling for tactical edge data centres in Ladakh, and ruggedised vapour chambers for loitering munitions operating at 4,000 metres above sea level. These are engineering challenges with clear military specifications. iDEX has shown it can translate clear specifications into fielded capabilities. Apply it here.
The fourth is mandating closed-loop liquid or immersion cooling for all strategic national security AI data centres. This is not about efficiency. It is about sovereignty. A data centre cooling system that requires continuous supply of foreign components or fluids has a foreign dependency embedded in India’s sovereign compute infrastructure. Closed-loop immersion cooling systems, once installed, operate without continuous material replenishment. They also reduce power consumption by 30 to 40 percent, reduce water usage to near zero, and can be deployed underground or in modular ruggedised pods that are resistant to physical attack and satellite thermal imaging.
The Invisible Problem That Determines Whether Everything Else Works
India has invested substantially in the visible layers of its AI defence capability: the drone programmes, the autonomous weapons platforms, the sovereign data centre ambitions, the India AI Mission compute infrastructure. All of these depend on a layer that receives almost no strategic attention: the ability to move heat out of silicon fast enough that the silicon keeps working at full capacity when it matters most.
As processing speeds increase through AI, and form factors shrink, driven by demand for smaller size, weight, and power, heat concentration becomes a variable needing more control. Every new generation of AI chip runs hotter. Every new operational environment India’s forces deploy into is more demanding. The gap between what the silicon can do at full capacity and what it actually does when throttled by inadequate cooling is the invisible performance deficit that no procurement announcement addresses and no press release measures.
In warfare, the invisible deficits are the dangerous ones. They do not show up in capability demonstrations. They show up in operations. The question India’s defence planners must answer now is whether it holds under the sustained, high-tempo, high-ambient-temperature, high-vibration conditions of a longer and more demanding conflict. The answer to that question is being built, or not built, in thermal engineering labs in Bengaluru, Pune, and Hyderabad right now. It deserves the same strategic urgency as everything running on top of it.
Bio
Dr. Sudhanshu Kumar is a Subject Matter Expert on AI, Cyberwarfare and Cybersecurity at CENJOWS (Centre for Joint Warfare Studies), HQ (IDS), Ministry of Defence, New Delhi. He holds a PhD on “AI and Security policy” from the School of International Studies in JNU. He is also a Visiting Research Fellow at MGIMO, Moscow.
