Over the summer I’ve been looking at what concentrations of data centres do to local energy grids. As I’ve explained in this piece for The Ecologist, massive clusters of data centres pose all sorts of challenges for grid operators, particularly in heatwaves. The sheer scale of data centre demand poses one problem, but this is compounded by spikes in power and water usage as data centre operators respond to rising temperatures by turning up fans and cooling systems to cope with the heat.

The scale jump in the size and electricity consumption by data centres tightly packed into small areas in order to offer “availability zones“ (networks between data centres in a locality designed to ensure constant service) is leading to friction with communities in some places over who picks up the bill for extra electricity use and costly upgrades to grid infrastructure. Even Trump’s notoriously AI-happy administration recognised the growing discontent across America over this issue and dragged the leaders of the major AI companies to the White House to sign a “ratepayer protection pledge“ promising to pay more for the power they use and contribute to grid infrastructure costs.

The most important availability zone in the UK is the one located in Slough. The data centre hotspot there is concentrated around the Slough Trading Estate. As the Guardian has reported recently, some residents feel the town has been taken over by data centres, and concerns about negative effects such as the “urban heat islands” created by the densely-packed, power-hungry warehouses of machinery are growing.

A fascinating report published as part of the government’s Energy Trends collection of statistics for the first time provides a robust answer to the question which many people have been asking: what proportion of local electricity use in Slough is taken up by data centres?

Hiten Shah, the author of the report, manually matched metered electricity consumption from non-domestic users across the UK with a database of data centres created by DSIT. This shows that an astonishing 65 percent of all electricity consumed in Slough local authority in 2024 was used by data centres.

Here are the headlines from the report:

• In Great Britain, data centres accounted for an estimated 4.5 TWh (2%) of the 249.2 TWh1 of electricity consumed from the grid in 2024. Around three-quarters (77%) of this data centre electricity consumption occurred in the South East (1.8 TWh) and London (1.7 TWh).

• In terms of local authorities, Slough (in the South East) was by far the largest contributor, accounting for 1.3 TWh (29%) of data centre electricity consumption in 2024. In Slough, consumption from data centres doubled between 2020 and 2024 (from 0.6 TWh to 1.3 TWh). In 2024, data centres accounted for almost two-thirds (65%) of all electricity consumed from the grid in Slough.

• Between 2020 and 2024 total data centre consumption increased by 1.3 TWh (41%). The majority of this increase can be attributed to Slough (in the South East) and neighbouring Hillingdon (in London), which saw increases of 0.7 TWh and 0.4 TWh respectively over this period.

This is important for several reasons.

Firstly, it shows how much of a rapid transformation in electricity demand we are seeing with the growth of large loads from data centres. In October 2025, Slough’s “live” data centre capacity measured by IT Load, a data centre industry metric widely used to indicate the maximum electricity consumption of the servers and other IT equipment in the building, was apparently 416 MW according to industry analysts, DC Byte. (This estimate was submitted by Slough Council to a planning inquiry, so presumably gives an accurate picture). In 2017 the IT Load was just 115 MW.

Secondly, it shows that widely quoted statistics about the national level of electricity consumption by data centres (reported at 2.5 percent in this House of Commons report, for example), mask the problems created by the concentration of extremely large “point loads”. Data centres flock together and the clustering creates intense strain on power, water and air quality.

Thirdly, it underscores that data centre are better compared to arc furnaces than warehouses and their categorisation in the planning system under a land use label associated with “storage” is nonsense. Residents who live near data centres of this scale will face many of the same kind of issues affecting the local environment and quality of life to those living near other large industrial sites. Moreover, the global effects of this industrial energy consumption are felt across the planet, especially where generating the electricity has added to carbon emissions, but also across the supply chains for the humming machinery which has taken over Slough.

The overwhelming size and scale of speculative data centre projects fuelled by the boom and hype around AI just keeps on growing. A few days ago, Xlinks, an energy infrastructure company, suddenly announced a 1.5GW “data campus” project linked to a battery storage park. The “Devon Data Centre Campus” is planned for a large area north of Great Torrington, a few miles from Bideford.

Xlinks’ promotional website shows massive data halls nestled in the woodlands and claims the development would offer “a dedicated community benefit package would be shaped by the community itself, alongside new on- and off-campus amenities — which could include public open space, a fitness centre, an electric bus depot or a business hub.”

Locals I’ve spoken to say that the community is up in arms. At a packed meeting in Great Torrington Town Hall on Thursday 25 June, over 200 local people left Xlinks’ representative and the meeting chair, Geoffrey Cox MP in doubt of the strength of feeling. “It was standing room only,” retired teacher Dave Clinch told me. “There was massive opposition. Clearly water use would be an issue with the proposed data centre. There were great concerns about water cooling and the effects on the River Torridge.”

The Torridge is one of the rivers immortalised in the novel Tarka the Otter by Henry Williamson which chart the fictional life and death of an otter cub in North Devon’s “country of the two rivers.” Tarka’s death at the end of the novel was set near Great Torrington and the town has just opened a family adventure trail marked by hand-carved wooden sculptures celebrating the area’s wildlife.

The 1.5GW scale of the proposed data centre marks this project out as another shift in the ambitions of developers hoping to ride the AI boom, and would cause wildlife, local people and the living planet even greater harm than most of the large sites already in the planning pipeline across the UK.

I have been tracking many of these projects in order to assemble a UK-wide dataset of current and planned data centre projects for Watershed Investigations as part of a project modelling on-side and indirect water consumption.

The 1.5GW IT Load (power consumption capacity of the servers) projected for the Devon Data Campus is equivalent to the data centre IT Load of the entire UK in 2024, according to figures from the Department of Science, Innovation and Technology (DSIT). My dataset (which is still incomplete) has mapped 5.4 GW already in the planning system across the UK with new projects being announced weekly.

As a recent UN report highlights, this new industrial infrastructure will take up vast amounts of land, use enormous quantities of water and power and fuel the climate crisis.

But judging by the mood in Great Torrington on Thursday, Xlinks will have a fight on its hands over this. At one point a resident addressed the sweltering crowd in the Town Hall, Dave recalled. “Who in this room is against this data centre? Stand up if you agree with me.” One by one almost the entire meeting got to its feet.

Last week I led the first iteration of a new course I have developed for the Centre for Investigative Journalism entitled Investigating the AI Industry. This is a new departure for CIJ and I think one of the first courses of its kind run by an organisation in the UK which specifically offers journalists training in how to weave through the fug of half-baked notions, pseudoscience and downright bullshit which surrounds discussions about AI. We covered a wide range of topics in the course, from the technical foundations of AI systems, to the water, power and carbon footprint of data centres and model training cycles and the real-world harms perpetrated by the people who lead this exceptionally powerful industry.

The last few weeks have seen a classic example of how clever PR departments leverage both fear of harm and FOMO to create an unstoppable wave of hype about their company’s products. Anthropic’s launch of Mythos Preview has generated reams of breathless coverage and prompted panic in some board rooms and government departments. Mythos is a “general purpose language model” which the company says is “strikingly capable at computer security tasks“, including finding “zero-day” vulnerabilities in widely-used systems which everyone thought had been thoroughly tested.

Anthropic announced it had restricted access to Mythos because it believes the model is “too dangerous” to be released on the open market. Instead it is working with a select group of companies (Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks) on a new project where these early adopters will work together to “secure the world’s critical software.” Opening Project Glasswing’s website I did wonder for a minute if I’d been redirected to Palantir’s Gotham platform by accident (same deep black and screaming white text although minus the blinking cursors).

The “your software is a weapon” vibe may be more than a coincidence. The highly-publicised spat between Anthropic and the Pentagon about whether the company’s LLMs should be unleashed to automate industrial-scale human rights violations on the US population, or merely tasked with doing that to people in other countries, means that the “look how dangerous this tool is” rhetoric was probably at least partially aimed at the self-styled “Secretary of War”, Pete Hegseth. (NB it seems to have worked, CNBC reported on 1 May that although Anthropic’s products are blacklisted by the DoD, Mythos is apparently being discussed as a “separate issue”).

Of course cybersecurity is a serious business, and successful cyberattacks can cause misery for millions of people, putting genuinely critical services at risk (such as transport, energy systems or health services). But I had a hunch that wasn’t really what was going on here.

So I one of the tasks I set participants on the CIJ course was to look into what else is going on when companies talk about their chatbots “going rogue” (or spread fears that they might do so). A few internet searches later, and the group noticed that Anthropic has previously used the same tactics of talking up the apparent ability of its products to “blackmail”, “scheme” and attempt to “stop themselves being shut down”. The launch of Claude Opus 4 in May 2025 was accompanied by a similar level of “rogue chatbot” hype.

When preparing my course materials I spoke to Dr Felicity Mellor, Director of the Science Communication Unit at Imperial College London. Felicity’s name had caught my eye thanks to a letter in the Guardian where she highlighted the alarming use of anthropomorphic language in reporting about AI products. Felicity told me:

“Our language shapes how we think about issues. In all areas of science communication, the words we use can evoke some visions of the future whilst obscuring others, or can assume how power and responsibility are distributed. When anthropomorphic language is used to describe AI systems, it ascribes conscious decision-making and intentional acts to the AI, implying that it is the AI, not its creators, that has moral agency. Such language makes it harder to hold AI companies and their leaders to account for the risks they are taking in deploying this technology.”

One consequence of distracting everyone with talk of “superintelligence” is that it dress up product malfunctions and potentially shields the manufacturers from claims made by angry customers, or people harmed by machine breakdowns. As Felicity Mellor put it in her letter to the Guardian:

“Imagine a company had released high-speed vehicles on to the roads before fitting them with effective brakes. We would not say the vehicles “connived” to kill other road users; we would say the humans behind the company had behaved with the utmost recklessness.”

Emily Bender and Alex Hanna’s excellent book, The AI Con makes the same point very effectively. Bender and Hanna also demolish the idea that “booster” and “doomer” narratives about AI systems are the polar opposites they seem to be. Instead, they see these apparently contrasting perspectives as two sides of the same coin, both helping with marketing by companies making these products. I also recommend the associated podcast, Mystery AI Hype Theater 3000 where they take apart PR puff pieces, line by line.

In a sense, last week’s course was all about different methods for going backstage at the Hype Theatre. Although we discussed a wide range of investigative techniques for unpacking the idea of AI, there are two basic principles which underpin most of them:

1. Pay close attention to the language used to talk about AI systems

2. Ask the right kind of questions

Even bad writing is often powerful. George Orwell’s Politics and the English Language has long been one of my favourite essays on the topic. He noted how dying metaphors, pretentious diction and the repetition of meaningless words provided perfect cover for the “defence of the indefensible.” While Orwell’s targets were pompous political writers of his own day, his definition of “modern writing at its worst” applies nicely to the outputs of Large Language Models.

“Modern writing at its worst does not consist in picking out words for the sake of their meaning and inventing images in order to make the meaning clearer. It consists in gumming together long strips of words which have already been set in order by someone else, and making the results presentable by sheer humbug.”

One of Orwell’s great bugbears was the overuse of the passive voice as a technique to obscure who was really responsible for an atrocity: “the village was bombed from the air” (but who gave the order to launch the missiles?). However, as the example highlighted above shows, in the case of AI marketing, the active voice attributed to an automated set of statistics about a corpus of texts or a collection of images also often conceals the real decision-makers and hides them from accountability for their failures.

Sometimes the way to cut through this kind of hype is to ask the right kind of questions. Anthropic’s marketing ploy over Mythos repeated a similar campaign by rival OpenAI in 2019. At the time OpenAI was working on GPT-2, the predecessor of the models now used behind the ‘conversational’ interface of ChatGPT.

GPT-2’s text generation outputs were pretty unpolished by today’s standards, nevertheless OpenAI ramped up the hype about them. GPT-3’s release a year later triggered a lot of excitement on Twitter about the LLM’s effectiveness at answering ‘expert’ questions. AI researcher Janelle Shane took a radically different approach, instead opting for ‘stupid’ questions, such as “how many eyes does a horse have“ GPT-3 generated a variety of answers, but the most common number was four (”Two in front, two in back”).

Shane was probably right that the puzzling answers probably reflected the lack of discussion about the number of horses’ eyes in the reams of internet texts in GPT-3’s training data. But the point is that a few blunt questions exposed claims of “expertise” based on a sprinkling of long, Latinate words. (For those interested, a group of my Cambridge colleagues created a “Bullshit meter“ for detecting patterns of sloppy language use of this kind).

More questions to ask yourself when investigating an AI system

In the final session of the course, we looked at investigating the impact of AI systems in the real world. This is my list of questions I ask myself when starting to think about an AI system from a critical perspective.

1. What process is being automated?

2. Why is it being automated?

3. Which classification tasks make up the process being automated. What labels are being used in the classifier?

4. What data has been used to train the classifier? What data was NOT used to train the classifier?

5. What rules apply to the outcomes of the system’s predictions? Is the answer to the following that someone gets hired, fired, jailed or even killed?

Thinking about the automation of a process rather than someone’s job is often important to avoid confusion between the kind of tasks where statistical predictions based on textual or image data may be useful with the performance of a social role such as ‘doctor’ or ‘police officer’.

On Tuesday I joined a livestreamed panel organised by the Peace and Justice Project in collaboration with NHS campaigners against Palantir’s role in the NHS and the other public services. It was clear that the national campaigns organised by No Palantir in Our NHS, Keep our NHS Public and Health Workers for a Free Palestine are being joined by new initiatives centred around particular hospitals as NHS Trusts ‘opt-in’ to the platform provided by Palantir, the NHS Federated Data Platform. The panel was hosted by Shirley Franklin, from Defend the Whittington Hospital Campaign, which has just launched a petition against Palantir and George Binette from Hackney Coalition against Palantir who spoke about the campaign focused on Homerton Hospital. The Good Law Project is tracking adoption of Palantir’s software across the country and provides a tool allowing you to email your local NHS trust in protest.

Matt Mahmoudi gave a chilling overview of how Palantir’s software is used by the US Department of Homeland Security and ICE officials to gather vast troves of data about both US citizens and migrants, racially profiling them and setting targets for violent deportations. Amnesty’s research has shown recently how an upgrade to the Integrated Case Management (ICM) system provided by Palantir to ICE in 2014 is likely to have helped the US authorities detain and deport thousands.

But what is it that Palantir actually sells? The company’s advertising is brash and unapologetic about its militarism. “Your software is the weapons system” proclaims the product page for its Gotham platform, which is the law enforcement version of the company’s core product, Foundry.

Yet Foundry is just one of a range of similar software-as-a-service platforms which are used by many large organisations looking to carry out actions with data such as moving bulk data around a system, extracting data and creating visualisations and providing a unified view of data which is held in different locations (data federation). Snowflake, Alteryx and Databricks all have products which compete with Palantir. Inside the NHS, health workers are deeply sceptical that Foundry offers anything better than what they have already, CorporateWatch reported earlier this year.

So what is Palantir’s secret sauce? Some of it is undoubtedly the sheer volume of lobbying which goes on behind the scenes. Another crucial factor is geopolitical calculations by the UK government: or put more bluntly, Keir Starmer’s desperate desire to curry favour with Trump.

But essentially, I think that a large part of what makes Palantir attractive to managers is that it promises to create a better-functioning panopticon through which to view their own organisations. The process of mapping the different elements of an organisation and how these relate to each other is a key part of embedding Palantir’s products. In other words, it isn’t so much the data itself which Palantir craves, but the ability to see everything about the data. This means Palantir ends up with a very detailed diagram of how its host organisation works, how decisions are made and how resources flow through it.

Palantir’s system of Forward Deployed Software Engineers aids this process by embedding the company’s employees inside the organisations which are using its products, not primarily as salespeople but rather as its envoys and advocates. FDSEs are empowered to work directly with customers to build or customise software together, but only so long as the outcome is in Palantir’s interests. There is a powerful desire for centralisation manifested in pretty much everything the company does.

Nevertheless, neither Palantir nor the managers who buy in to its authoritarian fantasies of seamless orchestration of “Human + AI teams” by god-like executives actually make the NHS function, and the people who have the greatest power to frustrate them are those who do that essential work.

The rush to invest in the infrastructure used to produce AI models is already having a major impact on local communities who live in the shadow of the data centres which house the specialised and expensive machinery used to manufacture them. The AI industry’s apparently insatiable demand for electricity and water is now being regularly covered in the media and is subject to increasing attention from local and national governments.

An important metric which has been widely used within the data centre industry in connection with these issues is PUE, which stands for Power Usage Effectiveness. PUE takes the IT load of the data centre (meaning all the energy consumed by IT equipment) and divides it by total facility energy consumption (this includes the cooling systems, lighting, uninterruptible power supplies). A PUE closer to 1 is deemed to be evidence that computation is using relatively little additional energy for operations such as cooling. The key paper explaining what PUE is and how it is calculated is here. In Germany, new data centres coming online from July 2026 are expected to perform at below 1.2 PUE, while China made a PUE of 1.25 in cooler regions and 1.3 in temperate regions a legal requirement this year.

A companion metric to the PUE is the Water Usage Effectiveness (WUE), which was also developed by the Green Grid in 2011. The white paper explaining WUE is here. As the number and scale of data centres has grown in response to competition to produce ever larger AI models, increased scrutiny of water consumption by the industry has prompted some operators to consciously adopt cooling systems which use less water on site.

QTS, the operator of the gigantic new data centre under construction at at Cambois, near Blyth in Northumberland is one of these. The company says that the new data centre will be relying on “water free” design.

However, this does not mean that there is no water impact from the operation of the data centre. An article I wrote for Watershed Investigations and published with Leana Hosea in the Guardian recently highlights the work of Alex de Vries-Gao, a researcher at the Vrije Universiteit Amsterdam on estimating the ‘embedded’ water use of the data centre industry.

We used the methodology in his paper to estimate that water footprint of the power required by the Cambois site could be 124m litres a year, 50 times more than the amount directly consumed by the operator.

These figures are an important reminder that anything involving the operation of large concentrations of machinery has a huge impact on the environment, and that in an era of accelerating global heating, this cannot be brushed off as ‘someone else’s problem’.

PUE and WUE metrics also don’t tell you anything about whether we should be using such large amounts of energy and water in the first place. That requires a wider discussion about what kind of AI services are socially beneficial and whether the amount of power, water and emissions used in producing them is justifiable.

Despite this, data centres are still often thought of as fundamentally unlike other energy intensive industries.

Malcolm Robinson, independent councillor for Bedlington West, told the BBC that he thought the QTS data centre at Cambois would be a more attractive feature of the landscape than the old power station it replaced, which he described as a “scar” on the coastline.

“This is a 21st Century business, this is not the old heavy industry that was there,” he said.

An aspect of the construction of the Cambois data centre highlighted in the Guardian piece is the role of on-site “back-up” power, in this case delivered by nearly 600 diesel generators. In interviewed environmental campaigner Julie Bolthouse of the Piedmont Environmental Council (PEC) which has been fighting the use of diesel-powered emergency power systems for data centres for several years. Julie told me by email:

We were told with each data center application here in Virginia that those generators wouldn’t run except in rare emergencies and that is one of the reasons localities were unconcerned with air quality impacts and approved them. No assessment of health impacts of various run scenarios or cumulative emissions in the area was ever done.

They were all approved under minor air permits by the Virginia Department Environmental Quality without any public comment opportunity. But we were always worried that they would want to utilize that capacity more though and now we are seeing them do just that. They are incrementally increasing under what circumstances they can run and de facto how frequently and how long they can run the thousands of generators we have permitted here in Virginia. Once the generators are in place it is only a matter of time before they use them and this is why a full assessment of various run scenarios and cumulative impacts on air quality should be completed with each application.

The PEC has located 9000 diesel generators in Virginia’s “Data Centre Alley”, of which 4700 are in East Loudon County alone.

A major concern of the PEC and other groups challenging the environmental and public health impacts of data centres is that the rapid growth of data centres is creating conditions for the increased use of this highly polluting infrastructure. Upgrades to the electricity grid in order deliver the power required has not kept up with demand, and this is translating into higher bills for consumers, in turn leading to proposals that data centres be allowed to use their on-site generators more often.

This should perhaps give pause for thought about how the AI industry (and data centres more broadly) are best understood and regulated. Under UK planning laws, data centres are often categorised as class B8 (storage and distribution), as opposed to other classes which cover heavy industrial machinery. If AI data centres were considered comparable to, say, new Electric Arc Furnaces (which might be reasonable in terms of the amount of power they consume), then different regulations would presumably apply.

I’ll be continuing to work on investigating this issue over the coming months. If you are a journalist interested in collaborating or living in a community affected by data centre build out, please get in touch: annealexmailbox@proton.me

Image credit: Electric Arc Furnace (Wikimedia commons, public domain image)

As Grace Blakely explains in a detailed post here, this ‘circular financing’ is inflating a bubble of expectations of profitability which, sooner or later, will burst:

“Imagine if a car manufacturer invested in a taxi firm, which used the money to buy only that car manufacturer’s vehicles. The carmaker could claim rising sales, the taxi firm could say it was growing, and investors might conclude the transport sector was booming. Everyone would start to throw money at these firms for fear of missing out on the boom, creating a self-fulfilling cycle of rising valuations. But there’s no one at the end of the chain buying the product – taxi rides. Instead, much of the activity is being financed by the carmaker’s own money. Unless real customers start paying lots of money for taxi rides, the flow of money between these companies will eventually reach an abrupt halt.”

How does this phenomenon relate to the structure of the AI industry I outlined a couple of weeks ago? In that post I argued that the race to produce ‘foundation models’ dependent on hyperscale computation, gigantic datasets, AI models with trillions of parameters and eye-wateringly expensive engineering talent has dominated the last decade of the industry’s development. The ‘foundation model’ paradigm is a story of the commercialisation of what was previously an academic field. This can be seen in the graph below from Our World in Data.

That story has been defined by success in attracting record levels of investment throughout the whole stack of technologies which is used to produce and deploy foundation models. This has included development of specialist semiconductors for data centres, to training dataset production, to technical innovations in model architecture such as the Transformer designed by Google engineers in 2017. One of the causes of the frantic levels of capital expenditure by the major companies in the US AI industry is the fact that they need to hedge their bets over which of these areas will deliver a decisive advantage.

This could easily come about through denying competitors access to costly equipment or scarce power supplies, or forcing them to overextend themselves with purchases they can’t sustain, rather than actually improving the performance of AI products.

In other words, the largest players among the US tech giants (such as Amazon, Google, Meta, xAI, OpenAI and Microsoft) are engaged in a game of corporate chicken with trillions of dollars of investment at stake.

Many analysts have looked back to previous technology-fuelled financial bubbles as a possible guide for what might happen next. Last year, Reuters columnist Edward Chancellor argued that the railway boom and bust in Victorian Britain is replete with lessons about today’s AI bubble. Capital surged into railway construction only to collapse as overproduction became obvious.

“It turned out that the revenue projections provided by so-called “traffic takers” were wildly overoptimistic. Railway engineers underestimated costs. The vogue for constructing direct lines between large urban centres proved mistaken, as most traffic turned out to be local. As a result, Britain’s rail network was plagued with overcapacity. There were three separate lines connecting Liverpool with Leeds and London with Peterborough. Railway companies cut dividends as returns on invested capital fell to 3%. By the end of the decade, the index of railway stocks was down 65% from its 1845 peak.”

The fall-out from the dot.com boom at the turn of the millennium and the crash which followed it is also an important comparator. Some of the big companies jostling to stay ahead in the tech investment arms race weathered that storm too. Grace Blakely argues that they are unlikely to go under when the deluge hits this time around.

“The AI companies, meanwhile, will be betting on the fact that they’re big enough to avoid going under regardless of which way the way the market goes. Amazon, Google, Microsoft, and the other tech companies that have come to define our generation survived the dot-com bubble. They’ll all survive this one too – joined by the new hegemons, like OpenAI and Anthropic. So, who is going to be left picking up the pieces?”

It is true that size is often a protective factor, particularly if your company is simply “too big to fail” and can therefore turn to the government to help with a bail-out. However, a more detailed look at the history of the companies which dominated the overinflated tech stocks twenty-five years ago shows the picture is more complicated than this.

Microsoft was badly damaged by the crash, taking around a decade and a half to recover the value it lost when the bubble burst. The company’s pivot towards building cloud infrastructure played a major role in its recovery, paving the way for the AI boom as it was Microsoft’s partnership with OpenAI which allowed Sam Altman’s company to launch ChatGPT in 2022. So yes, Microsoft made it through to the next upward sweep of the rollercoaster ride after 2000 and might do it again.

But the fate of Sun Microsystems offers a cautionary tale. Sun’s business model was more like Nvidia’s in that it concentrated on very expensive hardware for servers, rather than offering products in both home computing and enterprise IT services like Microsoft at the time. Thor Johannsson, a former Sun technical consultant put it like this a few years ago:

“They were making really cool servers, which cost more than cars and some even cost more than houses with nice swimming pools. Not exactly your home PC variant.”

At the height of the stock market bubble this propelled Sun’s valuation to then unheard-of levels as companies scrambled to create internet infrastructure. But this business model also left the company much more overexposed than Microsoft as the oversupply of servers and cables relative to demand became clear. Microsoft then took over large parts of Sun’s business. As Johannsson explains, “the wealth that the home computers had generated, paid for the same companies to take over server rooms.” The company limped on for a few years after 2000 before being bought by Oracle for a fraction of its former value.

Another significant victim of the dot.com crash was Cisco. The company had a value of $15 billion in 1995 but this had skyrocketed to $550 billion five years later as investors bet on soaring demand for network routers and switches. In the crash it lost 80 percent of its value and its stock price has never fully recovered.

Ironically, Google – which was still a small(ish) start-up in 2000 – owed its existence not to the foresight of executives at the companies which survived the dot.com crash, but to Andy Bechtolsheim, co-founder of Sun and later a vice-president at Cisco. In 1998 Bechtolsheim took a risky bet as angel investor on a then-untested technology – the PageRank algorithm which Larry Page had developed at the heart of his fledgling search engine business with partner Sergey Brin. The $100,000 cheque he wrote allowed Page and Brin to move out of their college rooms into a garage at Menlo Park and start hiring staff.

The early years of Google also underline how much more is at stake in this round of hype, overproduction and speculation. By 2001, just three years after its formation, the company was reporting profits, had hired Eric Schmidt (another former Sun executive) and had laid the technical basis for the advertising business which would power its rise to monopoly status in the search engine market. OpenAI, by contrast, doesn’t expect to be profitable until 2029, a decade after transitioning from being a non-profit. In the meantime, the company has taken on hundreds of billions of dollars of obligations.

Marxist economist Michael Roberts points out that the AI investment bubble isn’t just a risk for OpenAI but for the US economy as a whole. Without AI-related investments the US would either be in recession or close to it. Moreover, as a recent report from McKinsey shows, AI is reshaping global investment flows, as foreign direct investment (FDI) has surged into the US.

And there are now significant knock-on effects from the AI infrastructure boom in areas such as energy systems which also require lots of expensive, specialist machinery and therefore often “lock in” investment for long periods of commitment to particular types of technology. A recent report in the FT highlights how the AI data centre boom is driving up demand for gas turbines. This is another market dominated by a small number of large companies (Mitsubishi Heavy Industries and Siemens Energy are two key ones). Customers in the US now face 3-year waiting lists for gas turbines, while those in Asia will have to wait at least 5. This could propel customers towards accelerating the adoption of renewables, but it could also lead towards the expansion of coal power to fill the gap. China’s domestic energy market is dominated by a mixture of coal and renewables, and China accounted for 93 percent of new global coal construction in 2024.

The climate costs of the intense concentration and centralisation of capital reflected in the current AI investment mania underline that it won’t just be the tech bros who pay the price when this game of corporate chicken ends in disaster.

What are AI models?

Behind the advertising hype about ‘artificial intelligence’ lies a simple, but powerful idea. By using maths to describe relationships between things and people we can make more accurate predictions about a whole range of processes. These predictions might be about the next word in a sentence, or the colour of a piece of fruit, whether someone will enjoy a piece of music or how their hand will move next time they scratch their nose. An ‘AI model’ is a collection of mathematical descriptions of relationships between different elements in datasets containing information about whatever the manufacturers want to predict. If the goal is to reproduce texts which look like the English language, then the main datasets used to produce the model will include a lot of other texts in English. Calculating relationships between words across this dataset allows the makers of the model to write a computer programme which can generate ‘new’ texts which appear indistinguishable from the existing texts in the dataset. For example, how often two words appear close to each other can be expressed in mathematical terms. A dataset of English texts collected from the internet (perhaps including newspaper articles, Wikipedia, blog posts and possibly digitised books from university libraries) would be likely to include numerous examples of the phrase “the cat in the hat”, mostly referring to the character created by Dr Suess. Therefore, a model created using this dataset would be likely to predict that the next word in the incomplete sentence “the cat in the …” is “hat”. However, an AI model created using only the works of William Shakespeare would find no examples of the word “cat” in such close proximity to “hat” and instead be more likely to predict other words to complete the sentence.

How are AI models manufactured?

The scale of the ‘training data’ for an AI model is a major factor in how powerful a tool it is. In fact, the massively increased scale of these datasets is the single most important reason why AI models produced today appear more useful and efficient than models produce ten or twenty years ago. This also explains why the production of AI models is dominated by extremely wealthy and powerful corporations and the largest and most powerful states. This reflects the amount of money required to process trillions of words of text and billions of images in extremely expensive computers, using huge amounts of energy and water. Basically, that’s all the ‘magic’ of AI is. The rest is a lot of advertising hype and political rhetoric to justify using all those resources in the first place, and to ensure that no-one questions the profits and power accruing to the people who control those companies and governments as a result.

• Foundation models

The current AI boom has been driven by the creation of ‘foundation models’. These are produced using extremely large datasets of texts, images, audio, video and other types of data. While the process of producing such a model is extremely expensive, a foundation model can have a very large number of uses and can be used as a base for more specialised tasks by combining with other, smaller models which are less expensive to create.

• Data supply chains

The data supply chains for foundation models are complex and include vast public and private datasets, frequently involving the theft of creators’ works by the model manufacturers. Data taken from publicly accessible internet sources is one component of training datasets used in the production of foundation models – for example commonly used datasets in the training of large language models (such as Google’s Gemini, Anthropic’s Claude and OpenAI’s GPT-series models) include Wikipedia, online newspapers such as The Guardian, the contents of the Google patents database. Increasingly, AI model makers are turning to partnerships with companies, universities and government institutions to access even more data. The stock image company Shutterstock and social media company Reddit have both created ‘data partnerships’ with OpenAI recently.

• Labour supply chains

Producing a foundation model requires more than simply vast quantities of training data. The vast troves of internet text and images which could be thought of a ‘raw material’ for part of the training data for these models need to be refined, cleaned and processed in order to produce marketable products. That involves millions of hours of labour by workers who label the data, screen out abusive or offensive content and evaluate the responses from the model to improve its performance.

Who profits and who loses out?

There are a few winners from the current AI boom – mainly big companies such as Meta, Amazon, Google, Nvidia, ByteDance, Alibaba and so on. There are a lot more losers, including creators whose work has been stolen to produce the models, communities affected by AI-intensive data centres’ thirst for water and energy, workers in the data supply chain and those whose jobs are displaced or deskilled as a result of the rush to automate using AI.

A quick online search on “AI” “DSEI London” brings up a puff piece based on a press release by Maris-Tech, a small Israeli company which manufactures “AI-powered video intelligence and edge computing solutions for the defense and homeland security markets.” The company has clearly had a good few years, reporting expanding orders for its suite of hardware products, while the Israeli army’s destruction of Gaza’s infrastructure, forced displacement and starvation of Palestinians has escalated. Financial results announced in August 2024 showed an increase in revenue of 620 percent compared to the previous year.

Maris-Tech is a small part of Israel’s military industrial ecosystem, but counts many of the bigger players among its domestic customers, including Elbit Systems, Israel's Ministry of Defence, Rafael Advanced Defence Systems, and Israel Aerospace Industries (IAI). The company’s portfolio of computer hardware consists mainly of devices for encoding and decoding video and audio data, allowing the operators of manned and autonomous military vehicles to receive and analyse video streams during their missions. It includes components which can be installed into customers’ own systems, and products such as the Diamond platform, which shows allows operators to monitor across multiple cameras, record and transmit video and integrate AI analysis tools into decision-making.

Behind the advertising hype, pieces of military kit like this also tell powerful and disturbing stories about how small companies like Maris-Tech are part of larger systems producing digital technologies and AI for the Israeli military which rely on the participation of giant corporations including Microsoft, Amazon and Google. At one level this type of equipment is part of an ever-expanding “Internet of Battle Things”, networks of devices which are increasingly designed to work in situations where connections to cloud computing services are frequently interrupted or non-existent. In the jargon of the sector, these devices perform “edge computing,” processing data on the device itself or a nearby server, rather than in a distant, specialised data centre.

Yet as a slew of recent investigations by +972 magazine and The Guardian have underscored, a crucial lesson from the Israeli experiments in automated warfare is how quickly the demand on computer processing power and storage capacity generated by capturing and analysing audio and video data at scale overwhelmed the resources of even a well-equipped and technologically developed military. As +972 magazine reports, the Israeli military has repeatedly turned to civilian cloud providers including Amazon’s AWS, Google Cloud and Microsoft Azure in order to meet its storage and processing needs since October 2023. The use of commercial cloud providers for military purposes was a direct consequence of dramatically increased demand on the Israeli military’s privately operated cloud service for real-time video streaming and analysis.

Rather than deleting video and audio data captured using equipment like Maris-Tech’s devices, as they used to do, integrating the “limitless” storage and processing capacity of hyperscalers such as AWS, Azure and Google Cloud into their operational systems helps the Israeli military to keep “everything” about almost everyone in Gaza and search through it at a moment’s notice in order to justify a decision to kill.

This model of creating and analysing gigantic troves of data does not only operate on the battlefield, but extends to the mass surveillance of the entire Palestinian population. According to +972 magazine’s latest investigation, Unit 8200, the Israeli military’s cyber warfare unit, is storing audio files of millions of phone calls from Palestinians on Microsoft’s Azure platform and has used this data to plan airstrikes and shape military operations.

Building digital surveillance systems and integrating them into military operations through use of AI is a long-time obsession of Yossi Sariel, Unit 8200’s commander until September 2024, when he resigned, admitting to a degree of personal responsibility for the catastrophic failure of Israeli military intelligence to prevent the October 7 attacks. The rationale for using AI systems to automate roles in Unit 8200’s traditional division of labour was laid out by some of Sariel’s subordinates in an article for the Israeli military’s theoretical journal in 2021. In place of the Unit’s listeners, transcribers, translators and traffic analysts, multi-disciplinary data analysts relying on automated speech-to-text and optical character recognition algorithms would work at radically increased speed in order to identify targets.

The authors admitted at the time that the off-the-shelf solutions provided at the time by the commercial giants might not work particularly accurately for “the languages ​​and scenarios that plague the security system of a small country in the Middle East.” Yet subsequent investigative reporting has uncovered that the Israeli surveillance-to-targeting pipeline has indeed relied very heavily on automatic transcription and translation services, provided by hyperscale cloud providers and their partners in AI production, such as OpenAI.

Is this a story, though, of civilian technology companies unwittingly providing services which have been used to carry out atrocities against Palestinians? Google claimed for years that workloads it accepts through Project Nimbus, the massive contract providing cloud services to the Israeli government which Google Cloud won in conjunction with AWS in 2021, complied with its standard terms of service ruling out usage causing “death, serious harm, or injury to individuals or groups of individuals.” Internal Google documents published in late 2024 showed however that not only did the company exempt Project Nimbus from its standard terms of service, senior officials were aware before the deal was signed that providing services to the Israeli government risked making the company complicit in human rights violations.

Microsoft initially tried to argue that it had found “no evidence” that services provided to the Israeli military had harmed Palestinians or anyone else in Gaza, but scrambled to rewrite its official blog in August 2025 and initiate a formal review in response to reporting by the Guardian about the use of Azure services by Unit 8200. Meanwhile, Microsoft employees who have been protesting for years about the company’s complicity in war crimes and genocide were hauled away by the police after a sit-in.

Far from emerging as a response to emergency or “humanitarian” requests for support, it is clear that competition between the hyperscalers themselves over the lucrative opportunity to do business with the Israeli government is one of the driving forces behind the automation of security and military operations targeting Palestinians. This competition is playing out at every level of the AI stack from data centre infrastructure, to models and applications.

The meeting at which Microsoft CEO Satya Nadella is reported to have met Sariel and agreed to the Israeli spy chief’s proposal to move vast amounts of intelligence data onto Azure cloud servers took place in late 2021. Was this seen as some kind of compensation for the disappointment of losing out to Google and AWS over Project Nimbus, awarded in October the same year? Microsoft executives apparently looked forward to millions of dollars in revenue and “an incredibly powerful brand moment” for Azure as a result of the multi-year deal. The company is engaged in fierce competition with its hyperscaler rivals over the supply of AI services based on Large Language Models (LLMs) to the Israeli military, granting Israeli government customers access to the LLM services behind ChatGPT via the Azure platform, thanks to its partnership with OpenAI, despite the fact that OpenAI initially refused Unit 8200 direct access to the service. At the same time, Google has been promoting use of its LLM-based Gemini AI assistant to Israeli government and security clients as part of the Nimbus project.

Back at DSEI, the military trade show is welcoming a large delegation of Israeli exhibitors once again, despite growing public condemnation of business relationships with an army which international courts are investigating for genocide. The British government’s withdrawal of an invitation to the Israeli government triggered furious complaints of “discrimination” from Israeli officials. As Campaign Against Arms Trade points out, this symbolic gesture will do little to disrupt the access of Israeli companies to DSEI – the UK government’s “ban” only relates to Israeli government officials, not Israeli companies.

While DSEI’s participants have not shown much sign of being troubled by the ethical and legal questions raised by the Israeli military’s embrace of AI systems, perhaps they will at least reflect on what the last two years have demonstrated about the operational performance of these technologies. The latest reporting on the Israeli army’s “bodycount” of alleged adversaries suggests that claims that AI-enabled war machines are more effective at killing enemy combatants are wildly overblown. By May this year, the Israeli army’s internal data recorded the deaths of just under 9,000 people it claimed were Hamas and Palestinian Islamic Jihad “operatives”, less than 20 percent of the over 47,000 names it had on file as “active” members of the two organisations’ military wings.

Leaving aside the question as to whether these lists are accurate in the first place, or whether killing “enemy combatants” in their family homes by missile strike is lawful, this is a remarkably poor level of efficiency. While there is plenty of evidence from Israeli politicians and military figures that collective punishment of Palestinians has been an aim from the beginning of “Operation Swords of Iron”, the deadly consequences of this AI hype raise urgent questions as to why Microsoft, Google, Amazon and other companies whose services form critical parts of the global AI infrastructure are continuing to work with them.

Subscribe now

Introduction to the AI Arms Race

This website is a place for journalism related to my new book project - The AI Arms Race. I will be researching and writing it over the coming year. If you want a preview into the themes I am likely to be covering, you can read my recent article here.

The book project was born out of the realisation that there is a lot to say about the role of AI in contemporary capitalism. I am particularly interested in how the interaction between economic and military competition to produce AI systems exacerbates threats of war and climate breakdown. However, I’m also going to be looking at AI as a new wave of automation which is being embedded in a wide range of workplaces – from hospitals, to newsrooms, factories and farms.

What can you expect in the next few weeks?

I’m kicking off this project with a series of posts looking at what the DSEI arms fair between 9-12 September in London can tell us about the relationship between two crucial aspects of the AI arms race – competition to produce and deploy AI systems and the intensifying race over advanced military technologies from AI-facilitated surveillance and targeting systems to drones and robot dogs.

The articles I’ll be publishing include an essay on what happens when the hyperscaler digital platform companies such as Google, Amazon and Microsoft go to war, with a case study of the role of cloud infrastructures in enabling genocide and war crimes in Gaza.

Then I’ll be looking at companies in middle of AI “stack”, taking US-based military software manufacturer Palantir as an example. What are the implications for our public services when companies which market their software as a “weapons system” are also in charge of NHS patient data?

A final piece in this series will explore the supply chains for the hardware and software of robot dogs. These mechanical canines are not a new sight at DSEI, but are emblematic of the intensifying competition over innovation in autonomous vehicles and the accelerating rate to deploy them in military contexts.

Over the coming months I’ll also share thoughts on methods I’m using in the research on this project, publishing occasional datasets and reflections on readings and open source materials.

The research and writing for this project are currently self-funded. I work part-time at the University of Cambridge where I have been part of Cambridge Digital Humanities (and its predecessor network) since 2011, but the book project is not funded by my academic work. If you’d like to know more about my academic research, the Ghost Fictions project, which I co-authored five years ago, long before Large Language Models were so fashionable, is a good place to start.

If you are interested in commissioning my journalism, consultancy or training workshops in investigative journalism methods please get in touch by email at annealexmailbox@proton.me. I have more than a decade and half of experience in developing teaching about AI and digital methods through the Cambridge Data Schools and other programmes at Cambridge and am happy to discuss commissions for training workshops on a freelance basis. Alternatively, you can take out a subscription or make a donation here on Substack.

If you have information or experiences you think might be important for me to include in the book, then please get in touch by email here: annealexmailbox+aiarmsracebook@proton.me.