A hopeful path through the 2030s
The words you are reading have travelled through an infrastructure that most of us barely think about. Somewhere, computers have processed a request, equipment has moved information between networks, and electricity has kept the whole arrangement available. The internet feels ordinary because we encounter its finished surface. Its history is decades of people working out how to deliver more communication, more knowledge and more possibility through machines that gradually became better at their jobs. I think that history deserves a place in our conversation about artificial intelligence, particularly when the conversation turns to the environment.
We have become accustomed to expecting improvement from the things we build. A dishwasher should clean effectively while wasting less water. A car manufacturer should find ways to move people using less energy. Both technologies have environmental consequences, and cars remain the subject of considerable argument, but the engineering ambition is familiar. We want greater usefulness for a smaller expenditure of resources. It seems reasonable to expect the same ambition to shape AI, provided we examine what is actually improving and insist that the benefits reach beyond a company's operating costs.
My optimism begins there. I can see a plausible future in which AI delivers vastly more useful work while the growth of its environmental burden slows, perhaps eventually stabilising. Getting there will take deliberate choices. The historical evidence gives us something substantial to work with, and AI brings an additional possibility: the technology consuming the resources can also help us discover how to use them better.
What the internet actually achieved
The internet's origins reach well behind the public web. ARPANET connected its first computers in 1969; the transition to TCP/IP in January 1983 helped establish the networking foundation of the modern internet. These developments depended on sustained research, public investment and collaboration among institutions. The system was being improved long before most households had any reason to know its name. [1]
Tim Berners-Lee proposed the World Wide Web at CERN in 1989. CERN placed the web software in the public domain on 30 April 1993, helping make widespread adoption possible. The web gave people an accessible way to navigate information over the underlying internet. From there came the familiar progression through home connections, commercial websites and increasingly capable online services. If we choose 1993 as a useful public starting point, we are already looking at more than three decades of development. [2]
The efficiency episode that interests me most arrived considerably later. In February 2020, Eric Masanet and colleagues published a study in Science estimating global data centre electricity use between 2010 and 2018. Their analysis put consumption at approximately 194 terawatt-hours in 2010 and 205 in 2018, an increase of about 6%. During the same period, compute instances increased by 550%, meaning the final level was around 6.5 times the starting level. [3]
That arithmetic is worth remembering accurately. A 550% increase means adding five and a half times the original quantity to the original itself. Across those eight years, the world's data centres handled several times as much computing activity while using only slightly more electricity. The researchers linked this achievement to more efficient equipment, better utilisation and a shift towards larger cloud facilities, including hyperscale centres. Dividing the electricity change by the workload change implies roughly 84% less electricity per compute instance across the period. [3][10]
There is our historical precedent. Between the web's public release in 1993 and the beginning of that study period lay seventeen years; by its endpoint in 2018, twenty-five years had passed. Those dates describe an observed period of exceptional efficiency improvement. They do not identify a moment when the internet became carbon neutral. The study concerned data centre electricity, one component of an internet that also depends on communication networks, personal devices and manufacturing. Keeping the boundaries clear makes the achievement easier to defend.
Masanet himself warned against complacency when the research appeared. Growing demand could still produce a sharp increase in consumption later in the decade unless operators and policymakers kept working on efficiency. [12] I find that warning useful because it describes progress as something maintained through effort. An achievement can be real and still be vulnerable. We can admire what happened between 2010 and 2018 while recognising that the next generation of services brings a new set of demands. The precedent gives us a reason to keep engineering, with our attention on whether each improvement survives contact with a larger market.
I use the phrase "hyperscale efficiency threshold" to describe a possible condition in which improvements become strong enough to absorb much of the demand for additional service. Hyperscale itself refers to very large computing operations; the threshold is my shorthand, rather than an established scientific milestone. The useful question is whether we can reproduce that relationship between expanding capability and restrained resource growth.
Where AI stands today
AI has a long research history, but November 2022 provides a recognisable marker for the present consumer boom: OpenAI released ChatGPT on 30 November that year. The subsequent expansion belongs to the early and middle 2020s. People began encountering generative AI directly in everyday work, and the conversation about computing infrastructure became considerably more visible. [4]
The latest evidence offers encouragement alongside a demanding construction schedule. In its April 2026 update, the International Energy Agency reported that electricity per individual AI task had recently been falling by a factor of ten or more each year. It also estimated that global data centre electricity demand grew 17% in 2025, with consumption at AI-focused facilities increasing 50%. Its central outlook puts total data centre demand at about 950 terawatt-hours in 2030, up from 485 in 2025. [5]
Those findings can coexist perfectly well. Each task can become cheaper to perform while the number of tasks grows faster still. The mix also changes: the IEA notes that reasoning, agentic work and video generation can require hundreds or thousands of times the energy of simple text generation. A forecast built around yesterday's short chatbot exchange will struggle if tomorrow's typical request involves producing video or running extended sequences of actions. [5]
I find the everyday distinction helpful. Asking for a brief explanation and asking a system to investigate a problem for an hour are different purchases of computation. We should expect the second to require more work. Equally, we should expect engineers to keep finding ways to perform both more efficiently. Comparing services sensibly requires attention to their quality and complexity, as well as the energy recorded at the meter.
There are already tangible improvements to study. FlashAttention, introduced by Tri Dao and colleagues in 2022, reorganised a widely used AI calculation to reduce movement between types of GPU memory and accelerate training. It illustrates how software can improve the use of existing hardware. A speed improvement alone does not establish a matching percentage reduction in electricity, but it gives engineers another measurable opportunity to investigate. [6]
That leaves plenty of practical questions. Can an ordinary task use a smaller model? Can several requests share resources efficiently? Can useful intermediate results be reused? Which calculations improve the answer enough to justify their cost? I would like those questions to become routine product decisions, with energy measurements alongside assessments of accuracy. Much of the hopeful future depends on making that kind of ordinary discipline widespread.
Imagine a designer who previously commissioned one draft and can now generate fifty alternatives. Each alternative might require much less energy than an earlier system used, yet the whole session could consume more. Some of those additional attempts may produce worthwhile discoveries. Others may add little. The same freedom that makes AI exciting therefore gives us a responsibility to judge its use thoughtfully. A tool that lets someone communicate, learn or create something previously beyond their reach has a value worth discussing. Counting requests alone will never tell us enough about what society is receiving in return for its resources.
A plausible path through the 2030s
There is no credible universal date for AI to cross my proposed threshold. Different services will mature at different speeds, and some already benefit from economies unavailable to others. We can, however, examine published scenarios and then state the conditions under which a broader plateau might emerge. That gives us a useful time map without pretending the calendar determines the outcome.
The IEA's 2025 report explored this uncertainty through several cases. Its central projection reached roughly 1,200 terawatt-hours of global data centre electricity demand in 2035. A High Efficiency Case reached about 970 terawatt-hours, serving the same demand with stronger technical improvements. A Headwinds Case plateaued near 700, partly because adoption and infrastructure expansion encountered constraints. A Lift-Off Case exceeded 1,700. These are scenarios for all data centres, with AI a major driver, and the 2026 update retained a similar central trajectory to 2030 while flagging possible longer-term upside. [7][5]
The High Efficiency Case is especially relevant to an optimistic argument because it preserves useful service while reducing the resources required relative to the central case. Even so, its 2035 consumption remains well above the present level. The plateau in the Headwinds Case has a different explanation, including slower expansion. We should distinguish success through better engineering from a slowdown caused by shortages or disappointing demand. Both can restrain electricity growth; their implications for the usefulness of AI differ.
My own favourable scenario places the possibility of broader stabilisation in the early to middle 2030s, conditional on efficiency improvements catching up with growth in comparable services. That period is a reasonable window to investigate because it follows the current expansion and overlaps the published scenario horizon. It is an interpretation of the evidence, with no assigned probability. It would require deployment of better hardware and software, together with demand that becomes easier to anticipate and manage.
The underlying arithmetic is straightforward. Imagine comparable AI service grows 20% in a year while electricity required per unit falls 20%. Multiplying 1.2 by 0.8 gives 0.96, so total electricity falls 4%. If service instead grows 50% with the same efficiency improvement, electricity rises 20%. These are illustrative calculations. They show exactly what a threshold would require: efficiency gains large enough to keep pace with the combined growth in use and computational complexity.
Under that favourable scenario, progress through the late 2020s would establish increasingly efficient routine services. During the 2030s, a larger share of demand might move onto mature systems with predictable costs, allowing efficiency to absorb more growth. A less favourable outcome would see intensive new applications continually outrun those gains. Both outcomes are compatible with rapidly improving technology. The question is how the improvement is spent, and whether the resulting services justify their expanding physical demands.
Starting the clock in 2022 also reveals why copying the internet's timeline would be misleading. AI inherits extensive networks, cloud infrastructure and engineering expertise. It also introduces different workloads and an unusually rapid commercial expansion. A twenty-five-year countdown borrowed from the web would tell us very little. The historical analogy supplies evidence that large efficiency gains are possible; current measurements must tell us whether AI is approaching them at the necessary scale.
The technology can help improve its own infrastructure
There is a further reason I remain hopeful. AI can participate in the work of improving the systems on which it runs. We already have examples that can be examined without assuming artificial general intelligence, consciousness or an autonomous machine deciding to repair the world.
In July 2016, DeepMind reported reducing the energy used for cooling a Google data centre by up to 40% through machine learning. The models used operating data to help identify more efficient settings within the facility's constraints. The percentage concerned cooling, a portion of total consumption. Even with that qualification, it demonstrated a practical feedback process: computing was being used to improve the conditions under which further computing happened. [8] The date also complicates a neat separation between an internet we improved ourselves and an AI that helps improve itself. Machine learning was contributing to internet infrastructure before today's generative systems became familiar. The more persuasive argument is that this capacity is expanding into additional engineering tasks, giving human specialists more ways to search for improvements.
Google DeepMind's AlphaEvolve provides a newer example. Announced in May 2025, it combines language models with automated evaluation to develop algorithms. Google reported that a scheduling improvement recovered an average 0.7% of its worldwide computing resources. Another improvement accelerated a calculation used in Gemini and reduced its training time by 1%. These are company-reported results concerning capacity and time; translating them into electricity or emissions savings requires further measurement. [9] Still, the possibility is wonderfully concrete. An algorithm can suggest a better way to organise work, and engineers can test whether it succeeds. Once deployed widely, a small improvement can apply repeatedly across a large operation. The system helps search; people establish objectives, verify correctness and decide what belongs in production. That arrangement can shorten the distance between noticing waste and testing a remedy.
How many times have you wished the dishes would help put themselves away? There is something of that feeling here. A technology creates new work and new demands, while also giving us assistance with the work of managing them. The analogy is domestic, but the underlying idea is serious. An efficiency problem looks more manageable when the tools available to investigate it are themselves improving.
I would still want the accounting to include the cost of that search. An optimisation process consumes electricity, and its proposed improvement must repay the expenditure over its useful life. The same question applies to a new chip or a replacement server. We should measure the whole decision closely enough to know whether an apparent advance delivers an actual benefit.
Making the environmental gains count
Electricity efficiency is only part of the environmental story. Lawrence Berkeley National Laboratory's 2024 US report examined water consumption alongside energy and emissions, including water used directly at facilities and indirectly through electricity generation. It estimated direct US data centre water consumption at around 66 billion litres in 2023. Local conditions matter: an identical volume can have very different consequences depending on the water available and the competing needs of a community. [10]
We also need to know how electricity is produced. A facility can become more efficient while expanding its reliance on fossil generation. Conversely, cleaner electricity can reduce operating emissions even while consumption rises. Corporate purchasing claims need careful interpretation here: the Greenhouse Gas Protocol distinguishes emissions allocated through electricity contracts from those associated with the local grid, and explains that contractual accounting does not itself establish the wider impact of a purchase. [11]
For me, a convincing threshold would therefore involve several observations together. Useful work would keep growing while total electricity demand levelled off, and the electricity supply would become cleaner. Water pressure would decline where it matters locally. Equipment decisions would account for manufacturing and disposal. The precise evidence would vary by place, but the ambition should be understandable to the people living near the infrastructure.
We would also need a sensible definition of useful work. A thousand inaccurate answers are a poor substitute for a smaller number someone can trust. An efficiency claim should preserve the standard of the service being compared, or explain clearly what changed. Otherwise, a company could appear to make progress simply by delivering less. I would welcome comparisons that ask whether a task was completed successfully, how much checking it required and what the total process consumed. That brings the discussion closer to ordinary experience, where the point of a helpful machine is that it helps us finish something properly.
There is a reasonable role for public expectations in bringing this about. Operators can publish comparable measurements. Customers can choose services appropriate to the task. Governments and utilities can require new demand to fit credible plans for electricity and water provision. I would rather support an industry whose achievements can be checked than have to defend it through estimates nobody outside the business can inspect.
AI might also help reduce waste elsewhere, through better planning or improvements to industrial processes. Any claimed saving needs a fair comparison with what would otherwise have happened, including the resources used to produce it. I am comfortable being enthusiastic about those possibilities while waiting for evidence of their scale. Useful applications deserve the same curiosity and scrutiny we apply to the infrastructure itself.
Choosing hope and keeping our manners
The genie is out of the bottle. Millions of people have already found uses for AI, and the institutions developing it are making consequential decisions now. That leaves us with work to do: shape the systems, measure their effects and keep improving the choices around them. I want to take part in that work with some confidence that human beings can respond to a difficult engineering problem creatively.
I also say please and thank you to my AI. I know that can sound faintly ridiculous when placed beside electricity projections, but it belongs to the same personal outlook. When something helps me understand a problem, organise my thoughts or make something I care about, gratitude comes naturally. I am making no claim about a chatbot having feelings. I am expressing the way I prefer to behave, and acknowledging the human effort behind an unusually helpful invention.
If future systems become more capable, I hope we approach them with care and intellectual honesty. Being polite is a habit I am happy to preserve as the technology changes. It offers no guarantee about safety, consciousness or what artificial general intelligence might become. Those questions require their own evidence and serious decisions. Courtesy can sit alongside that work without carrying the burden of answering it.
The precedent I keep returning to is simple enough to say aloud: between 2010 and 2018, data centre computing activity grew around 6.5 times while electricity use rose about 6%. [3] That relationship shows what sustained improvement can achieve. AI now gives us additional ways to investigate the next round of improvements. I want us to pursue them carefully, with enough openness to recognise failure and enough imagination to keep trying. There is a hopeful path through the coming decade, and I think it deserves our effort. That is a future I would like my daughter to inherit.
Sources and further reading
Research checked 12 November 2026. Article body 3,000 words. The 2030s stabilisation scenario is the author’s conditional interpretation; the IEA cases describe total data centre electricity demand.
[1] Barry M Leiner and colleagues. A Brief History of the Internet. Internet Society, 1997.
[2] CERN. The birth of the Web. Institutional history.
[3] Eric Masanet and colleagues. Recalibrating global data center energy-use estimates. Science 367, 984-986, 28 February 2020. DOI 10.1126/science.aba3758.
[4] OpenAI. Introducing ChatGPT. 30 November 2022.
[5] International Energy Agency. Key Questions on Energy and AI. 16 April 2026, executive summary.
[6] Tri Dao and colleagues. FlashAttention Fast and Memory-Efficient Exact Attention with IO-Awareness. 2022. arXiv 2205.14135.
[7] International Energy Agency. Energy and AI. 10 April 2025, chapter on energy demand and sensitivity cases.
[8] Google DeepMind. DeepMind AI Reduces Google Data Centre Cooling Bill by 40%. 20 July 2016.
[9] Google DeepMind. AlphaEvolve A Gemini-powered coding agent for designing advanced algorithms. 14 May 2025.
[10] Arman Shehabi and colleagues. 2024 United States Data Center Energy Usage Report. Lawrence Berkeley National Laboratory, December 2024. Water and emissions discussion on printed pages 55-57.
[11] Greenhouse Gas Protocol. Top Ten Questions about the Scope 2 Guidance. Explanation of location-based and market-based accounting, especially questions 3 and 8.
[12] Harrison Tasoff. The World’s Data Plan. University of California Santa Barbara, 27 February 2020. Research announcement including comments from lead author Eric Masanet.
