However, now the widespread talk is less about AI as ‘the’ key to growth and prosperity and more about worries that AI will be yet another ‘bubble’,[15],[16] in the light of the enormous sums invested and even larger future commitments and the absence, so far, of clear pathways to profitability.

It is challenging to estimate just how much money has been raised in the race to develop and maintain leadership in AI. All told, AI took almost 50% of all global startup funding in 2025, rising to around $202 billion, up from $114bn in 2024.[17] Of this, 79% was to US-based companies.

OpenAI, developer of the AI assistant ChatGPT, is now the most valuable private company of all time. Nvidia and Amazon have agreed to invest a combined $80bn in the company alongside another $30bn from Japan’s SoftBank.[18]  OpenAI had a valuation of $500bn at a secondary share sale in October 2025[19] and may have a valuation after further fundraising of $750-830bn, but has about $1.4tr in spending obligations over eight years.[20],[21]

Meanwhile, Meta (owner of Facebook), Google, Microsoft, and Amazon – but notably not Apple, which appears to be sitting out the ‘AI capex arms race’ – have announced plans to spend a total of $660bn during 2026 in AI capital expenditure, a 60% rise over 2025.[22]

Where is all this money going? A large fraction will be spent on data centres. Data centre partners of OpenAI have already committed to spend $800bn.[23] Anthropic, developer of AI assistant Claude, has announced $50bn investment in data centres with its partner Fluidstack.[24] OpenAI CEO Sam Altman is reported to have told journalists over a dinner in San Francisco, “You should expect OpenAI to spend trillions of dollars on data centre construction in the not very distant future”.[25]

Looking forward, a report from McKinsey & Co predicts that data centres will require around $6.7tr investments worldwide by 2030 to keep pace with demand, of which $5.2tr is related to AI demand.[26] However, this is the ‘Continued momentum’ scenario. If growth accelerates higher than this, then AI-related data centres might need up to $7.9tr of investment; if growth is more constrained then this investment might be ‘only’ $3.7tr. This spending will not only be on data centre infrastructure: roughly two-thirds will be on IT equipment, and rather less than one-tenth on electrical power, according to the McKinsey report.

As always, there is considerable uncertainty. New AI models might require less electricity for training and inference, or conversely new use cases and widespread access might increase demand. The supply chain is complex, and disruptions could slow the ability of companies to expand. Geopolitical tensions, and tariffs and trade controls, could impact demand one way or the other.
 

The data centre energy surge

This data centre expansion translates into raw electrical power, measured in gigawatts (GW). OpenAI ended 2025 with around 1.9 GW of data centre power, from 0.6 GW in 2024,[27] and is aiming for 30 GW by 2030, while arch-rival Anthropic is aiming for around 6 GW in three years. These are explosive increases compared with the organic growth of ‘hyperscaler’ cloud providers AWS, Microsoft Azure, and Google which each built around 10-12 GW of capacity in 15-17 years. [28],[29]

The AI behemoths believe that building more and more computing power is essential to maintain customer adoption and revenue. Sarah Friar, chief financial officer of OpenAI, has written that lack of current power is already a constraint on growth.[30]

The McKinsey report cited above[31] suggests that, in the Continued momentum scenario, AI would demand 156 GW of data centre power worldwide by 2030, which is an increase of about 125 GW since 2024. Their higher estimate would demand a whopping 205 GW extra by 2030. And that’s only AI-related demand. Non-AI data centre workload is also expected to continue to grow, albeit more slowly, roughly doubling to 64 GW.

To put this into perspective, it is projected that US data centres could consume 11.7% of the entire US power grid by 2030.[32] This is higher than other countries because the USA has around 45% of the world’s data centres, with particular concentrations in Virginia, Texas, and California,[33],[34] although in Ireland data centres are already around 20% of domestic electricity demand.[35] Across Europe, data centre power demand is projected to reach 4-5% of total electricity consumption by 2030, from around 2-3% in 2024.

Access to electricity has become a critical factor in the drive for new data centre builds. As the power ecosystem struggles to meet this demand, it faces constraints including securing reliable and sustainable energy sources, grid infrastructure, power equipment, and skilled workers to build them. For example, the lead time to power new data centres in Northern Virginia can be more than three years.[36] In Dublin, the state-owned electricity operator imposed a moratorium on data centres in early 2022 and set out conditions to connect new ones to the grid. In places including West London and the south of Sweden, businesses are waiting years to be connected.[37]

Despite this rapid growth, it should be remembered that data centre electricity demand is still far from the largest driver of global growth in final electricity demand, being behind air conditioning and electrification of industries.[38] Globally, around 85% of the projected increase in electricity demand 2024-2027 is projected to come from China, India, and other emerging economies, but demand in the ‘advanced’ economies is expected to grow strongly too, having remained virtually static for some years, driven by the deployment of electric vehicles, air conditioners, and heat pumps, as well as data centres.[39]
 

Embracing nuclear and fossil fuels

To meet the challenge that the US power grid is not able to keep pace with demand, there is  increasing interest in onsite power generation, with the expectation that around 30% of data centres will use at least some onsite power supplemental to the grid by 2030, according to a survey of hyperscaler and colocation developer companies.[40] However, onsite or local power generation can be reliant on fossil fuels and even non-compliant with basic clean air regulations.[41] Under Trump, the Environmental Protection Agency’s enforcement actions are at a record low, while the federal government is actively restraining moves towards decarbonisation.[42]

The rapidly increasing demand also threatens firms’ sustainability targets, outpacing their green agenda.[43] The promise of stable, low-carbon, reliable power has led leading AI companies to turn towards nuclear power, especially so-called Small Modular Reactors (SMRs).[44] Sam Altman was chair of an SMR company called Oklo but stepped down in April 2025 to avoid a potential conflict of interest.[45] In October 2024, Amazon, together with partners, agreed to invest $500m in SMR start-up X-Energy, aiming to bring 5 GW online in the United States by 2039.[46] Google has ordered six to seven SMRs from nuclear energy company Kairos Power specifically to power its AI processing, aimed to come onstream between 2030 and 2035.[47]

However, these timescales are unlikely to be achieved due to numerous technical obstacles.[48] Indeed, the desperation of Big Tech has perhaps been shown most clearly by their attempts to re-start nuclear reactors which have been recently closed down. For example, in September 2024, Microsoft signed a 20-year deal to take electricity generated by a newly re-opened Unit 1 at Three Mile Island. This is next to Unit 2, site of the 1979 nuclear meltdown, the worst industrial disaster in US history. Constellation Energy, the owner of Three Mile Island Unit 1, has felt it necessary to rebrand the facility as the Crane Clean Energy Center.[49] Especially disturbing is the news that the Trump administration has recently watered down nuclear safety regulations to help developers.[50]
 

Pushback

In addition to the initiatives from the UN and international bodies discussed earlier, numerous civil society organisations and others are increasingly speaking out against these poorly regulated developments. New networks and coalitions are forming, including PauseAI[51] and, in the UK, Pull The Plug.[52] In addition, a range of new ‘AI-free’ consumer labels are being piloted in sectors where there are concerns that AI is a threat to jobs.[53] We must all become more engaged in this debate if society is to redirect AI onto a more responsible path.
 

Dr Philip Inglesant spent 20 years teaching and researching social aspects of computing and Responsible Innovation in areas including AI, quantum computing, and information technologies more broadly at universities throughout the UK. He is now retired. He is an Advisor to SGR’s Board of Directors.

Image credit: Ron Lach via Pexels.
 

References