AI & Big Tech infrastructure has a severe environmental cost.
We built the alternative.
Peer-reviewed research now confirms what the physics always implied: conventional AI and Cloud infrastructure measurably warms the areas around it, consumes water at scale, and is incompatible with the net zero commitments most organisations have already made.
Autonomy Cloud was designed from first principles to resolve that conflict — not mitigate it.
The numbers at a glance
All calculations on this page compare one fully populated 42RU Autonomy Rack equipped with 70 AI Accelerators and capable of concurrently supporting 4,600 mixed performance and standard VM workloads, against the equivalent standard equipment from a major supplier using equivalent performance NVIDIA GPU's + VMWare based virtualised machines built exactly to the vendor specifications and using published power profiles.
* Marinoni et al., University of Cambridge / NTU / NUS / CityUHK, March 2026. Full methodology and report is available on request.
AI datacentres are measurably warming the areas around them.
This is no longer a projection. It is documented, peer-reviewed, present-tense reality — with consequences for communities, public health, energy demand, and the regulatory frameworks governing data centre siting.
Research published in 2026 by scientists at the University of Cambridge, Nanyang Technological University, the National University of Singapore, and City University of Hong Kong analysed twenty years of satellite temperature data across more than 6,700 AI data centre locations globally.
The findings are unambiguous: AI hyperscalers raise locally measured land surface temperatures by an average of 2°C — and up to 9.1°C in some cases; with the effect measurable up to 10 km from each facility. An estimated 340 million people are within the zone of influence of existing facilities.
This is not a future risk to be managed at planning stage. It is an ongoing, measured consequence of infrastructure decisions already made — and it is accelerating as AI deployment scales.
Source: Marinoni et al., "The data heat island effect: quantifying the impact of AI data centers in a warming world," arXiv:2603.20897, March 2026. University of Cambridge, Nanyang Technological University, National University of Singapore, City University of Hong Kong.
The Autonomy response
Autonomy Cloud's 23.1 kW maximum draw versus approximately 256 kW for an equivalent conventional estate is not only an energy cost story. It is the difference between infrastructure that measurably creates a data heat island and infrastructure that does not. At that power envelope, Autonomy deployments are too small to generate the thermal signature the research identifies — by a wide margin.
No heat island risk
Clear heat island risk
The same output.
Less than a ninth of the power.
We use maximum component-verified draw throughout — it is the only figure that is independently verifiable, defensible in procurement, and honest about what a facility needs to support. Average figures are unknowable at design time and allow vendors to present numbers that will rarely be achieved in production.
Autonomy Cloud's maximum verified draw of 23.1 kW covers both AI inference and full cloud workloads simultaneously on a single 42RU rack. The equivalent conventional deployment — a separate GPU estate for AI and a separate server estate for cloud — draws approximately 256 kW at maximum. That is not an efficiency gap. It is a different order of magnitude.
The 0–50°C ambient operating range means Autonomy deployments require no dedicated cooling infrastructure — no chiller plant, no raised-floor CRAC units, no water-cooled door systems. Standard data hall airflow is sufficient throughout. These facility footprint savings are a sustainability benefit, not just an operational convenience.
Sources: Component TDP specifications from AMD and SoftIron datasheets. Conventional figures from Dell PowerEdge XE9680 and R760 published specifications. Power cost calculated at £0.25/kWh commercial rate, PUE 1.2 (Autonomy) / 1.35 (conventional). Full methodology available on request.
How Autonomy and conventional estates compare
VM density per kW — like-for-like workload
Autonomy delivers 8.3× higher VM density per kW for the same workloads.
VM density figures calculated from component specifications — full methodology available on request.
Net zero targets & AI ambitions are on a collision course.
Autonomy resolves it.
For any organisation that has made a Science Based Targets initiative commitment, deploying conventional AI infrastructure is not a sustainability challenge to be managed. It is a structural incompatibility with that commitment.
SBTi 1.5°C pathways require participating organisations to achieve a 90% reduction in Scope 1 and Scope 2 emissions by 2030. For any organisation deploying or planning to deploy AI infrastructure on a conventional GPU estate, that commitment and that infrastructure are fundamentally incompatible.
A conventional AI and cloud estate drawing approximately 256 kW continuously cannot be reconciled with a 90% emissions reduction target on any realistic renewable energy or carbon offset strategy at the required scale. The numbers do not close — not with offsets, not with PPAs, not with efficiency improvements to the conventional stack.
The choice between AI capability and sustainability commitments is a product of the conventional infrastructure model — not an inherent tension in deploying AI. Autonomy Cloud resolves it by making the AI infrastructure itself the sustainability solution, rather than a problem requiring mitigation.
Continuous load reduction per Autonomy rack versus the conventional equivalent.
Every deployment makes the net zero arithmetic achievable.
Sustainability isn't just about the planet.
It's also about the numbers on your P&L.
Reducing power consumption at this scale doesn't only lower your carbon footprint — it directly reduces operating expenditure, procurement risk, and the facility overhead that conventional AI infrastructure demands. For any organisation running a business case for AI infrastructure, the power arithmetic is as much a financial argument as an environmental one.
The 232.9 kW continuous load reduction that Autonomy Cloud delivers versus an equivalent conventional estate translates directly into cash. The workings are straightforward — and the outcome is not a marginal improvement. It is a structural difference in the cost of operating AI at scale.
How we calculated it
The figures above refer to a single 42RU rack of Autonomy Cloud equipment configured with 70 AI Accelerators, concurrently running over 4,600 mixed VM workloads: AI + Cloud in one 23.1 kW rack. The comparison is against a vendor-specification build-out of conventional servers: an NVIDIA H200 GPU estate for AI, plus a parallel VMware virtualisation estate sized to host the same number of VM workloads. All systems are built to vendor specifications and follow recommended deployment patterns.
Every hour. Every day.
This is not just a projected efficiency gain contingent on workload optimisation. It's the direct consequence of deploying infrastructure with a verified maximum draw of 23.1 kW in place of an estate that draws 256 kW. The saving begins on day one of operation and compounds with every year of the asset lifecycle.
Power figures based on component-verified maximum TDP specifications: AMD and SoftIron datasheets (Autonomy Cloud); Dell PowerEdge XE9680 and R760 published specifications (conventional estate). PUE of 1.3 is applied as a conservative modern UK data centre figure — the UK DC average is higher. Tariff of £0.25/kWh is a representative standard commercial rate; actual rates vary by contract, location, and market conditions. Full methodology available on request.
Benchmarked against NVIDIA's best. Better on efficiency. Competitive on throughput.
The AI100 Ultra cards in Autonomy Cloud are purpose-built inference accelerators. In 2025, researchers at UC San Diego's National Research Platform benchmarked them directly against NVIDIA A100, H200, and AMD MI300A hardware across 15 open-source LLMs. Results were published at ACM PEARC '25 and are available at arXiv:2507.00418.
We show the data as it is. The AI100 Ultra wins on energy efficiency across the majority of models, and wins outright on raw throughput for most models up to ~32B parameters. Only at 70B+ parameters does the H200 become competitive on throughput — though the AI100 Ultra still matches or beats it on efficiency for some models at that scale.
Energy efficiency — tokens per second per watt vs NVIDIA H200
tok/s/W calculated from Table 1, Sada et al. (2025). H200 figures use single-GPU configuration. Qualcomm AI100 Ultra figures use minimum required SoC configuration per model. Higher is better.
Raw throughput — up to ~32B params
AI100 Ultra vs single H200 GPU, tokens per second
70B+ models — throughput and efficiency
A more mixed position...
Llama3.3-70B
Throughput (tok/s):
H200 5,366 vs AI100 4,528
Efficiency (tok/s/W):
H200 wins (16.9 vs 13.1)
DeepSeek-70B
Throughput (tok/s):
AI100 4,528 vs H200 4,333
Efficiency (tok/s/W):
H200 wins (11.4 vs 10.3)
Llama3.3-90B Vision
Throughput (tok/s):
AI100 5,961 vs H200 3,556
Efficiency (tok/s/W):
AI100 wins (13.6 vs 7.46)
What the data shows
Where the AI100 Ultra consistently leads
Energy efficiency (tok/s/W) for models up to ~32B parameters — a factor of 1.2× to 2.6× versus H200.
Raw throughput for the same model range — typically 1.6× to 2.1× more tokens per second per card.
NB: This is the sweet spot for enterprise LLM inference:
coding assistants, RAG pipelines, document analysis, and real-time classification workloads.
For Edge AI, optimised models, speed, and power efficiency are everything — and the AI100 Ultra on Autonomy Cloud has no Edge AI parallel.
Where H200 is still competitive
Raw throughput for 70B parameter models, where H200's large memory bandwidth and NVLink scaling give it some advantages for latency-sensitive high-volume processing.
If your primary workload is serving a single 70B model at maximum throughput, H200 may be faster. But at the cloud core, where H200 can make sense, it is both undeployable at the edge and often inferior on efficiency.
At 23.1 kW maximum draw, Autonomy Cloud fits into environments where an H200 cluster cannot operate at all.
Source: Sada et al. "Serving LLMs in HPC Clusters: A Comparative Study of Qualcomm Cloud AI 100 Ultra and High-Performance GPUs," ACM PEARC '25, Columbus OH, July 25. arXiv:2507.00418v1.
tok/s/W figures calculated from Tables 1 and 2. Full methodology available on request.
The sustainability benefit no one wants to talk about:
infrastructure you don't need to build.
Conventional AI data centres require dedicated and increasingly complex cooling infrastructure — chiller plants, cooling towers, water supply, immersion units, raised floor environments — that carries its own energy load, capital cost, and in many markets a water consumption footprint that has become a planning and community risk. Autonomy requires none of that.
Because we have low power loading and no significant cooling requirements, we can use existing datacentres and network POPs as AI + Cloud hosting hubs. Hugely expensive upgrades to national power generation and grid infrastructure become unnecessary; massive multi-GW mega-datacentres need not be built.
Existing DC assets, their power connections and network links can all be used as-is — vastly accelerating AI roll-out, reducing cost and environmental impact, delivering Sovereign AI at a few percent of the TCO for Big Tech AI build-out.
Zero water consumption
The 0–50°C ambient operating range means Autonomy Cloud operates on standard data hall airflow and HVAC — no water-cooled door systems, no evaporative cooling towers, no direct liquid cooling circuits requiring continuous water supply.
In water-stressed markets — and an increasing number of data centre planning jurisdictions — zero potable water consumption is not a marketing differentiator. It's a necessary planning permission requirement.
Standard DC facilities
Autonomy deployments fit any standard 19″ rack environment — from an 8 RU entry point in a PoP room, scaled up 1RU at a time; to a full 42RU rack in a standard colocation facility.
No specialist raised-floor environment, no dedicated chiller plant room, no high-density power zone. The embodied carbon, concrete, and capital cost of specialist data centre infrastructure is eliminated, not just reduced.
~90% lower colocation costs
At 23.1 kW maximum draw across a single rack, Autonomy Cloud's facility footprint requirement is a fraction of the conventional equivalent. Over five years, colocation costs run approximately 90% lower than the equivalent conventional estate — space and power being the two primary cost drivers for colocation.
The facility saving is a sustainability saving: less space, less standby power, less cooling overhead.
Deployable with renewables
Because Autonomy Cloud's power envelope is compatible with the output of modest renewable generation — rooftop solar, small wind installations, micro-hydro — it can be deployed at locations where renewable power is the primary or only supply. Remote government sites, island territories, development-market sovereign AI deployments, emergency services infrastructure: all compatible. Conventional AI GPU clusters rule out these locations entirely on power grounds alone.
Sustainability isn't a feature we added.
It's what happens when infrastructure is designed sensibly and marketed honestly.
We didn't design Autonomy Cloud to hit a sustainability metric. We designed it to deliver AI inference and cloud workloads in constrained environments where conventional infrastructure simply cannot operate.
— PoP rooms, edge sites, sovereign deployments with no spare power budget -
The sustainability outcomes are a consequence of that design discipline and our commitment to open, transparent presentation of the facts; not a separate engineering or marketing objective.
The result is infrastructure that is compatible with SBTi commitments, deployable in water-stressed markets, operable from renewable power sources, and measurably incapable of generating the thermal footprint that peer-reviewed research now attributes to conventional AI data centres.
This is not because we set out to achieve those things.
Because we set out to build something that respected real-world global constraints.
Are your sustainability commitments and your AI roadmap are pulling in opposite directions?
We should talk.
We work through your obligations first, before anything else. SBTi pathways, Scope 2 targets, water stress constraints, planning considerations — bring your real world constraints and we'll work through them with you.
>Request a briefing