Across the globe, artificial intelligence has outgrown the limits of code, it now consumes physical space, electricity, and water on a planetary scale. Google’s Ironwood TPU and Europe’s JUPITER supercomputer represent two parallel visions of the future of computation: one driven by private innovation, the other by public purpose. Together, they mark a turning point in how humanity defines progress in technology and sustainability.
The sheer energy draw of these systems, roughly 10 to 11 megawatts per facility, puts them in the same league as industrial power plants. TechRadar confirmed that Google’s Ironwood supercomputer, its most powerful to date, sets a world record for shared memory with 1.77 petabytes across 9,216 chips. Meanwhile, NVIDIA’s announcement of JUPITER celebrated Europe’s first exascale machine, a facility capable of more than a quintillion calculations per second, or one exaflop, while ranking among the world’s most energy-efficient systems.
These projects are not just technical triumphs. They are also symbols of power, politics, and environmental accountability. Whoever controls access to this level of compute now has as much influence as those who once controlled oil reserves or semiconductor supply chains. And as nations race to achieve “compute sovereignty,” the ecological footprint of that race grows heavier by the megawatt.
Inside the Exascale Arms Race: The AI Datacenter Boom No One Can Contain
- Ironwood TPU (Google): 9,216 chips, 1.77 PB shared memory, ~42.5 AI exaflops (FP8 precision), estimated 10 MW power draw.
- JUPITER (Europe/NVIDIA): 1 FP64 exaflop for scientific computing, >90 AI exaflops for deep learning workloads, average 11 MW power consumption, using NVIDIA Grace Hopper (GH200) chips.
- Colossus (xAI): ~100,000 NVIDIA H100 GPUs for training Grok models, reportedly facing emissions and power sourcing concerns.
- DGX SuperPOD/Gefion: Denmark’s state-backed DGX cluster powered by NVIDIA H100s, representing the public-private hybrid model.
Each system highlights a different philosophy: Google’s Ironwood pursues private dominance in inference; JUPITER builds a sovereign foundation for Europe’s science and industry; and Colossus aims to compete with both under Elon Musk’s xAI brand. What unites them is a common tension between limitless computational ambition and finite planetary resources.
Who Owns the World’s Fastest AI? Compute Sovereignty vs Big Tech Domination
In the twenty-first century, compute sovereignty is emerging as a new form of geopolitical power. It refers to a nation’s ability to own, operate, and secure its own advanced computing infrastructure without relying on foreign cloud providers or private corporations. Europe’s JUPITER project, installed at Germany’s Jülich Supercomputing Centre, is designed precisely for that reason, to safeguard European research and innovation against dependence on U.S. hyperscalers. According to NVIDIA, JUPITER’s modular architecture connects more than 24,000 Grace Hopper superchips, allowing it to run both scientific simulations and artificial intelligence at scale.
By contrast, Google’s Ironwood TPU reflects the concentrated compute power of corporate ownership. It is a closed system, hosted within Google Cloud, optimized for serving its proprietary models and client workloads. The company’s latest Cloud Next presentation described Ironwood as the “first TPU built for the age of inference,” purpose-engineered to serve generative AI at scale rather than open scientific workloads.
Meanwhile, Elon Musk’s Colossus cluster at xAI shows what happens when private ownership scales beyond industrial norms. With up to 100,000 H100 GPUs and plans to double that, its construction is reportedly fueled by on-site gas turbines to meet energy demand. Environmental critics have questioned whether this approach represents progress or ecological regression.
Europe’s public strategy and Silicon Valley’s private approach reveal a growing divide: one builds compute for the common good, the other for competitive control. Yet both depend on the same physical reality, power grids, cooling systems, and natural resources, which no amount of cloud abstraction can erase.
Inside the Machines: Ironwood TPU, JUPITER, and the Architecture of AI Power
Scaling artificial intelligence no longer depends solely on algorithms or model size. It depends on how efficiently massive clusters of chips communicate, share memory, and dissipate heat. The details beneath the surface tell the true story of progress.
Google Ironwood TPU
Ironwood’s architecture connects 9,216 Tensor Processing Units through high-bandwidth optical circuit switches, forming a unified computing fabric that acts almost like one giant processor. Its 1.77 petabytes of shared memory (a new world record for a multi-accelerator setup) allows enormous reasoning models to operate without fragmenting data across nodes). This configuration not only accelerates inference but also reduces latency between chips, one of the biggest efficiency barriers in large-scale AI systems.
Ironwood is also liquid-cooled, with each pod consuming roughly 10 megawatts of power, emphasizing that physical infrastructure is now a limiting factor for algorithmic innovation. The system’s flexibility, its ability to reconfigure interconnect topologies, points to the next frontier in sustainable AI: doing more compute with fewer wasted watts.
JUPITER Exascale
Europe’s JUPITER supercomputer uses a very different architectural philosophy. It combines CPU and GPU resources within each Grace Hopper superchip, creating a tight memory coupling that eliminates traditional data transfer bottlenecks. Each module has 120 GB of LPDDR5X CPU memory linked with HBM3 on the GPU, providing both high bandwidth and energy efficiency. The entire system operates across NVIDIA Quantum-2 InfiniBand, supporting in-network computing; a technique that processes data directly within the network fabric to minimize energy-intensive transfers.
JUPITER’s goal is not only raw performance but also sustainability. Its designers prioritized liquid cooling and heat reuse, feeding waste heat into district heating networks to support nearby communities. According to NVIDIA’s energy report, it ranks among the five most energy-efficient supercomputers in the world, averaging around 60 gigaflops per watt.
NVIDIA SuperPOD and Colossus
NVIDIA’s SuperPOD model, seen in deployments like Denmark’s Gefion AI system, offers a middle ground: modular, high-performance clusters that governments and enterprises can co-own. Colossus, by comparison, is built on similar NVIDIA hardware but at a staggering private scale. Using Spectrum-X Ethernet networking and liquid-cooled Supermicro 4U servers, it pushes beyond 100,000 GPUs in one deployment. While technically impressive, this concentration of compute power in a single entity raises new questions about energy accountability and emissions oversight.
Together, these architectures illustrate an uncomfortable truth: achieving exascale computing today means engineering not just for performance but for planetary scale, balancing power, politics, and sustainability in every design choice.
The 10-Megawatt Wall: How AI Datacenters Battle Power and Cooling Limits
AI datacenters have entered an era where the real bottleneck is no longer compute performance: it is power. Once a facility exceeds roughly 10 megawatts (MW) of sustained electrical load, it faces physical, environmental, and regulatory limits that redefine how supercomputers are designed and located. Both Google’s Ironwood TPU and Europe’s JUPITER supercomputer sit precisely at this threshold, serving as case studies for how data infrastructure must now adapt to energy realities.
Power Density and Location Constraints
Google’s Ironwood cluster reportedly consumes about 10 MW per superpod, a figure derived from liquid cooling requirements and chip density as documented by ServeTheHome. To put that into perspective, this is roughly equivalent to the electricity consumption of 10,000 American homes operating simultaneously. The amount of waste heat generated is immense (measured in megawatts of thermal energy) which forces operators to carefully plan facility placement near reliable water and energy sources.
Europe’s JUPITER, operating at approximately 11 MW, is not simply a machine, but a coordinated regional energy project. The German site integrates with the national grid through partnerships that ensure renewable electricity supply while maintaining stability for surrounding communities. According to NVIDIA’s system briefing, the supercomputer’s liquid-cooled racks channel heat into a local district heating system, turning an inevitable byproduct into an environmental advantage.
The Cooling Revolution
The jump to double-digit megawatts has forced a revolution in cooling. Traditional air cooling cannot efficiently extract heat at these scales; it simply wastes too much energy moving air. Both Ironwood and JUPITER employ direct liquid cooling, where coolant flows through pipes that touch each chip or memory stack. This method removes heat more efficiently, reducing the overall power usage effectiveness (PUE) rating, a key industry metric that compares total facility energy to actual computational energy.
Newer architectures are experimenting with immersion cooling, where entire server boards are submerged in non-conductive fluid. Though not yet mainstream, this approach may soon be necessary as transistor densities increase further. Meanwhile, other systems, such as Denmark’s Gefion DGX SuperPOD, demonstrate hybrid models that combine air, water, and refrigerant loops to balance cost and sustainability.
Heat As A Resource
In sustainability-focused designs like JUPITER, heat is no longer treated as waste. District heating integration allows supercomputers to serve surrounding residential or industrial zones, effectively turning computational exhaust into usable warmth. This idea has also been piloted by hyperscale datacenters in Finland and Denmark, where recycled heat supports local green initiatives. Such measures are crucial because they transform an ecological liability into an opportunity for energy circularity, a principle central to sustainable technology ecosystems.
In essence, the megawatt threshold represents more than a power statistic. It is a dividing line between conventional data processing and a new paradigm where engineering, climate science, and public policy must converge to make frontier AI infrastructure viable.
Can Exascale AI Go Green? Inside the Race for Sustainable Supercomputing
The global rush to exascale computing has triggered an equally urgent search for sustainability. As supercomputers multiply, they are projected to consume up to 5% of total global electricity by 2030, according to several independent studies. This trend has made sustainable AI innovation not just desirable but indispensable.
Energy Efficiency and Material Limits
JUPITER stands as a model of how to balance extreme performance with environmental accountability. The system’s architecture, using Grace Hopper (GH200) superchips, minimizes energy waste through close coupling between CPUs and GPUs. Each superchip eliminates redundant data transfers, reducing latency and lowering total energy per operation. NVIDIA’s documentation highlights this as a major contributor to its Green500 ranking for efficiency.
On the corporate side, Google claims that its Ironwood TPU pods achieve significant power-per-compute improvements over prior generations by integrating custom silicon and optical interconnects that reduce electrical resistance. While Google does not disclose its complete carbon footprint, the company has pledged to run all operations on carbon-free energy by 2030, an ambitious commitment given the growing appetite for AI compute.
Cooling Water and Climate Tradeoffs
Cooling efficiency often comes at a cost: water. The global datacenter industry consumes billions of liters of water annually, mostly for evaporative cooling. Some facilities now experiment with closed-loop systems to minimize loss, reusing water multiple times before purification. JUPITER’s liquid-cooling system, for instance, avoids direct evaporation, reducing overall consumption. By comparison, U.S. hyperscalers operating in arid regions face growing scrutiny from environmental groups concerned about the stress on local aquifers.
Innovations in Renewable Integration
A promising direction lies in co-locating datacenters with renewable energy sources. Solar or wind installations can directly supply clean electricity to nearby computing clusters, while excess energy can be stored in grid-scale batteries for peak load balancing. Northern Europe’s moderate climate and stable grid infrastructure make it particularly suitable for these hybrid deployments. Some operators even explore geothermal integration, capturing heat from the Earth to maintain stable cooling conditions.
Designing for Efficiency at Every Scale
Sustainability does not always require mega projects. Compact, modular systems such as NVIDIA’s DGX Spark Mini illustrate how smaller AI deployments can deliver high performance with significantly lower energy demands. The future likely belongs to a mix of centralized exascale systems and distributed, energy-conscious clusters optimized for specific workloads. This diversification could help flatten the global power curve while democratizing access to computation.
The Politics of Power: Why AI Compute Inequality Could Shape the Next Decade
The expansion of AI supercomputing has profound implications for who wields power, both politically and economically. Access to compute is rapidly becoming the new measure of influence, reshaping global hierarchies.
Public vs. Private Control
The contrast between JUPITER and Ironwood epitomizes the divide between public infrastructure and private monopolies. JUPITER represents a collective European effort, funded through the EuroHPC initiative, to ensure that universities, research institutes, and smaller companies have access to advanced computing. In contrast, Ironwood operates within the closed ecosystem of Google Cloud, accessible primarily through commercial arrangements.
This imbalance raises the risk of what some researchers call “compute colonialism,” where a handful of corporations dominate the global supply of computational power. As AI applications increasingly shape public policy, scientific research, and even cultural production, this asymmetry could entrench new forms of inequality.
Environmental and Ethical Oversight
The sheer energy consumption of these systems also demands public accountability. Elon Musk’s Colossus project, for instance, has faced allegations of operating gas-powered generators that bypass environmental review processes. Without independent audits, the environmental cost of private AI infrastructure remains opaque.
Governments are beginning to respond. The European Union is developing energy and transparency standards for high-performance computing centers, requiring sustainability reports and operational disclosures. Similar proposals are being discussed in the United States and Japan. These efforts aim to ensure that the race for exascale computing does not undermine global climate objectives.
Socioeconomic Divide and Knowledge Gaps
Another hazard is the emerging compute divide, a disparity between institutions that can afford access to frontier hardware and those that cannot. This gap mirrors historical inequalities in access to education and technology. Without intervention, smaller nations and underfunded research organizations risk being locked out of the next wave of scientific discovery.
Bridging this divide requires a commitment to open data, cooperative frameworks, and transparent infrastructure. Public investments like JUPITER and Denmark’s Gefion cluster provide pathways toward equitable access, ensuring that compute power serves the collective rather than the few.
From Silicon to Sustainability: Rethinking the Future of AI Power
The next decade will determine whether exascale computing becomes a tool for global advancement or a driver of ecological strain and inequality. The direction depends on the decisions made today about how data centers are powered, governed, and shared.
Toward Planetary-Scale Responsibility
The lesson from both Ironwood and JUPITER is clear: progress in AI is inseparable from environmental stewardship. Sustainable architecture, renewable integration, and transparent governance must move from optional goals to non-negotiable requirements. This shift will require collaboration across governments, corporations, and civil society to align technological growth with planetary limits.
Innovation Paths Ahead
Expect new breakthroughs in chip efficiency, cooling design, and distributed computing. Hybrid optical-electrical interconnects and AI-driven load balancing could dramatically cut power usage. At the same time, modular, regionally distributed compute systems could replace a portion of the monolithic megacenters now dominating the landscape. These trends mirror the same ethos driving Europe’s JUPITER and other open-access clusters worldwide.
A Shared Future of Compute
In the long term, the question is not how powerful our supercomputers become but how responsibly we use them. Energy-positive datacenters that return heat to local grids, zero-water-loss cooling systems, and open-access research infrastructure are not idealistic fantasies, they are achievable with current technology and collective will.
As the world builds the next generation of artificial intelligence, it must also build the frameworks that ensure this power benefits everyone. The true measure of exascale success will not be counted in floating-point operations, but in how harmoniously these machines coexist with the planet that sustains them.
