Going Out With A Bang

In 1.5C Here We Come I criticized people like Eric Schmidt who said that:

the artificial intelligence boom was too powerful, and had too much potential, to let concerns about climate change get in the way.

Schmidt, somewhat fatalistically, said that “we’re not going to hit the climate goals anyway,”

Salomé Balthus
Uwe Hauth, CC BY-SA 4.0

In January for a Daily Mail article, Miriam Kuepper interviewed Salomé Balthus a "high-end escort and author from Berlin" who works the World Economic Forum. Balthus reported attitudes that clarify why "3C Here We Come" is more likely. The article's full title is:

What the global elite reveal to Davos sex workers: High-class escort spills the beans on what happens behind closed doors - and how wealthy 'know the world is doomed, so may as well go out with a bang'

Below the fold I look into a wide range of evidence that Balthus' clients were telling her the truth.

Kuepper quotes Balthus:

'The elephant in the room is climate change. Everyone knows it can't be prevented any more,' she said, adding that the 'super rich' could generally be split into two groups on the topic.

'The one group thinks it only affects the poor, the "not-white race", while the others fear that it could get worse but there's no sense in trying to do anything about it so they just enjoy themselves,' she told MailOnline.

'The one half is in despair and the other, dumber, half is celebrating future mass deaths.

Salome elaborated that some of the uber wealthy people fitting into the first group were saying that those in third world countries 'might all die but us in the North, we're fine'.

She said: 'They say that in a democracy you have to sell it, to lie to people and tell them "we didn't know better and didn't think it would get this bad", not admitting that they know.

'Then there's the other group that thinks it might not be so easy, maybe it will also affect us due to unforeseeable chain reactions.

'But they say they can't do anything against the others so they live following the mantra "after us, the deluge".

'They say they will enjoy a few more nice years on earth and know that there's no future. They are very cynical and somehow deeply sad.'

This attitude matches Schmidt's fatalism — we're doomed but we might as well make money/have fun until then. What it misses is that everything they're doing is bringing "until then" closer. As I wrote about Schmidt:

He is right that “we’re not going to hit the climate goals anyway", but that is partly his fault. Even assuming that he's right and AI is capable of magically "solving the problem", the magic solution won't be in place until long after 2027, which is when at the current rate we will pass 1.5C. And everything that the tech giants are doing right now is moving the 1.5C date closer.

After all, how bad could it be? The economists' models predict a fairly small decrease in global GDP, but Graham Readfearn reports that Average person will be 40% poorer if world warms by 4C, new research shows:

Economic models have systematically underestimated how global heating will affect people’s wealth, according to a new study that finds 4C warming will make the average person 40% poorer – an almost four-fold increase on some estimates.

Neal et al Fig 1

The reason for the underestimation is that their model of the effect of climate change on a country's GDP accounts only for in-country effects. Reconsidering the macroeconomic damage of severe warming by Timothy Neal et al from the University of New South Wales instead models global weather's effects on a world of interconnected supply chains:

Figure 1 shows the projected percentage reduction in global GDP from a high emissions future (SSP5-8.5) relative to a lower emissions future (SSP1-2.6), for the three models outlined in section 2.2. Each economic model is run with and without global weather to determine its impact on the projections. Without the inclusion of global weather (blue line), all three models project more mild economic losses with a median loss at 2100 of −28% for the Burke15 model, −4% for the Kahn21 model, and −11% for the Kotz24 model. Projected losses from the latter two models are more optimistic than in the original articles, likely due to the variations in data and exact assumptions made.

Readfearn notes that:

The study by Australian scientists suggests average per person GDP across the globe will be reduced by 16% even if warming is kept to 2C above pre-industrial levels. This is a much greater reduction than previous estimates, which found the reduction would be 1.4%.

Scientists now estimate global temperatures will rise by 2.1C even if countries hit short-term and long-term climate targets.

If global GDP drops 16% "us in the North" and specifically in the US are unlikely to be fine. Not that "us" in the US are that fine right now, as Aria Bendix' Not even wealth is saving Americans from dying at rates seen among some of the poorest Europeans shows:

Today, the wealthiest middle-aged and older adults in the U.S. have roughly the same likelihood of dying over a 12-year period as the poorest adults in northern and western Europe, according to a study published Wednesday in The New England Journal of Medicine.

Paddy Manning's The Death Toll From Extreme Weather Is Becoming Clearer shows it will get much worse:

Heat waves, wildfires, floods, tropical storms and hurricanes are all increasing in scale, frequency and intensity, and the World Health Organization forecasts that climate change will cause 250,000 additional deaths each year by the end of this decade from undernutrition, malaria, diarrhea and heat stress alone. Even so, the impact on human health and the body count attributed to extreme weather remain massively underreported — resulting in a damaging feedback loop of policy inaction. Meanwhile, the very people who might fix that problem, at least in the US, are being fired en masse amid the Trump administration’s war on science.

Source

But we know countries aren't going to "hit short-term and long-term climate targets" because, among other reasons, it would prevent us achieving the wonderful benefits of AI such as generated images in the style of Studio Ghibli.

Kate Aranoff explains that people like Schmidt aren't talking 2C any longer in The Bleak, Defeatist Rise of “Climate Realism”:

Owing to “recent setbacks to global decarbonization efforts,” Morgan Stanley analysts wrote in a research report last month, they “now expect a 3°C world.” The “baseline” scenario that JP Morgan Chase uses to assess its own transition risk—essentially, the economic impact that decarbonization could have on its high-carbon investments—similarly “assumes that no additional emissions reduction policies are implemented by governments” and that the world could reach “3°C or more of warming” by 2100. The Climate Realism Initiative launched on Monday by the Council on Foreign Relations similarly presumes that the world is likely on track to warm on average by three degrees or more this century. The essay announcing the initiative calls the prospect of reaching net-zero global emissions by 2050 “utterly implausible.”
...
Bleak as warming projections are, a planet where governments and businesses fight to the death for their own profitable share of a hotter, more chaotic planet is bleaker still. The only thing worse than a right wing that doesn’t take climate change seriously might be one that does, and can muster support from both sides of the aisle to put “America First” in a warming, warring world.

Of course, we might as well "enjoy a few more nice years on earth", because the 70-year-old Schmidt (and I) will be dead long before 2100. Our grandchildren will just have to figure something out. In the meantime we need to make as much money as possible so the grandchildren can afford their bunkers.

Lets look at some of the actions of the global elite instead of the words they use when not talking to their escorts, starting with Nvidia the picks and shovels provider to the AI boom.

Tobias Mann's Nvidia GPU roadmap confirms it: Moore’s Law is dead and buried sets up the plan:

Nvidia's path forward is clear: its compute platforms are only going to get bigger, denser, hotter and more power hungry from here on out. As a calorie deprived Huang put it during his press Q&A last week, the practical limit for a rack is however much power you can feed it.

"A datacenter is now 250 megawatts. That's kind of the limit per rack. I think the rest of it is just details," Huang said. "If you said that a datacenter is a gigawatt, and I would say a gigawatt per rack sounds like a good limit."

Most current Nvidia racks are 60KW but:

The NVL72 is a rackscale design inspired heavily by the hyperscalers with DC bus bars, power sleds, and networking out the front. And at 120kW of liquid cooled compute, deploying more than a few of these things in existing facilities gets problematic in a hurry. And this is only going to get even more difficult once Nvidia's 600kW monster racks make their debut in late 2027.

This is where those "AI factories" Huang keeps rattling on about come into play — purpose built datacenters designed in collaboration with partners like Schneider Electric to cope with the power and thermal demands of AI.

So Nvidia plans to increase the power draw per rack by 10x. The funds to build the "AI factories" to house them are being raised right now as David Gerard reports in a16z raising fresh $20b to keep the AI bubble pumping:

Venture capital firm Andreessen Horowitz, affectionately known as a16z, is looking for investors to put a fresh $20 billion into AI startups. [Reuters]

For perspective, that’s more than all US venture capital funding in the first three months of 2025, which was $17 billion. [PitchBook]

This means that a16z think there’s at least this much money sloshing about, sufficiently desperate for a return.

PitchBook says a16z is talking up its links to the Trump administration to try to recruit investors — the pitch is to get on the inside of the trading! This may imply a less than robustly and strictly rule-of-law investment environment.

The power to supply them is natural gas on a huge scale. See for example, Reid Frazier's LARGEST NATURAL GAS POWER PLANT IN THE COUNTRY, DATA CENTER COMING TO FORMER HOMER CITY COAL PLANT:

The owners of a recently demolished coal-fired power plant announced the site will become a data center powered by the largest natural gas plant in the country.

The Homer City Generating Station in Indiana County was decommissioned in 2023 and parts of it were imploded last month. It had been at one time the largest coal-fired power plant in Pennsylvania.

The plant’s owners, Homer City Redevelopment, announced the site will become a 3,200-acre data center campus for artificial intelligence and other computing needs

The "largest natural gas plant in the country" will be pumping out carbon dioxide for its predicted service life of 75 years, into the 3C period of 2100.

Alas, it isn't just natural gas that will be pumping out carbon dioxide, as Ari Natter and Jennifer A Dlouhy report in Trump Signs Orders to Expand Coal Power, Invoking AI Boom:

Taken together, the measures represent a sweeping attempt to ensure coal remains part of the US electricity mix, despite its higher greenhouse gas emissions and frequently greater cost when compared to natural gas or solar power.

The effort also underscores Trump’s commitment to tapping America’s coal resources as a source of both electricity to run data centers and heat to forge steel. The president and administration officials have made clear boosting coal-fired power is a priority, one they see as intertwined with national security and the US standing in a global competition to dominate the artificial intelligence industry. 

Carbon emissions aren't the only problem this mania causes, as Luke Barratt et al report in Big Tech’s data centre push will take water from the world’s driest areas:

Amazon, Microsoft and Google are operating data centres that use vast amounts of water in some of the world’s driest areas and are building many more, an investigation by SourceMaterial and The Guardian has found.

With US President Donald Trump pledging to support them, the three technology giants are planning hundreds of data centres in the US and across the globe, with a potentially huge impact on populations already living with water scarcity.

“The question of water is going to become crucial,” said Lorena Jaume-Palasí, founder of The Ethical Tech Society. “Resilience from a resource perspective is going to be very difficult for those communities.”

Efforts by Amazon, the world’s biggest online retailer, to mitigate its water use have sparked opposition from inside the company, SourceMaterial’s investigation found, with one of its own sustainability experts warning that its plans are “not ethical”.

And:

Amazon’s three proposed data centres in Aragon, northern Spain—each next to an existing Amazon data centre—are licensed to use an estimated 755,720 cubic metres of water a year, enough to irrigate more than 200 hectares (500 acres) of corn, one of the region’s main crops.

In practice, the water usage will be even higher as that figure doesn’t take into account water used in generating electricity to power the new installations, said Aaron Wemhoff, an energy efficiency specialist at Villanova University in Pennsylvania.

Between them, Amazon’s planned Aragon data centres will use more electricity than the entire region currently consumes. Meanwhile, Amazon in December asked the regional government for permission to increase water consumption at its three existing data centres by 48 per cent.

Opponents have accused the company of being undemocratic by trying to rush through its application over the Christmas period. More water is needed because “climate change will lead to an increase in global temperatures and the frequency of extreme weather events, including heat waves”, Amazon wrote in its application.

Right. We need to use more water to cope with the "extreme weather events, including heat waves" we are causing, which will allow us to cause more "extreme weather events" which will mean we need more water! It is a vicious cycle.

Is there really a demand for these monsters? One of Nvidia's big customers is CoreWeave:

NVIDIA, which has given CoreWeave priority access to its chips. As CoreWeave’s own S-1 notes, it was “the first cloud provider to make NVIDIA GB200 NVL72-based instances generally available,” and “among the first cloud providers to deploy high-performance infrastructure with NVIDIA H100, H200, and GH200.”

CoreWeave owns over 250,000 NVIDIA GPUs across 32 data centers, supported by more than 260MW of active power

Before CoreWeave's recent IPO Ed Zitron wrote CoreWeave Is A Time Bomb:

In my years writing this newsletter I have come across few companies as rotten as CoreWeave — an "AI cloud provider" that sells GPU compute to AI companies looking to run or train their models.

CoreWeave had intended to go public last week, with an initial valuation of $35bn. While it’s hardly a recognizable name — like, say, OpenAI, or Microsoft, or Nvidia — this company is worth observing, if not for the fact that it’s arguably the first major IPO that we’ve seen from the current generative AI hype bubble, and undoubtedly the biggest.

It seems a lot of investors agreed with Zitron. The IPO went so badly that Nvidia had to step in and support it. Craig Coben has the story in CoreWeave’s IPO tested the waters — and missed the mark:

The initial public offering of AI infrastructure firm CoreWeave, initially targeting a $2.7bn raise at $47-55 per share, was slashed to $1.5bn at $40 per share. Even then, the deal barely limped across the finish line, thanks to a last-minute $250mn “anchor” order from Nvidia. The offering reportedly ended up with just three investors holding 50 per cent of the stock, and it seems to have required some stabilisation from lead bank Morgan Stanley to avoid a first-day drop. Hardly a textbook success.

Bryce Elder lists reasons why investors might be skeptical in Eight odd things in CoreWeave’s IPO prospectus. Elder starts with an analogy:

Imagine a caravan maker. It sells caravans to a caravan park that only buys one type of caravan. The caravan park leases much of its land from another caravan park. The first caravan park has two big customers. One of the big customers is the caravan maker. The other big customer is the caravan maker’s biggest customer. The biggest customer of the second caravan park is the first caravan park.

Sorry, not caravans. GPUs.

Another reason investors might be skeptical is that technologies advance in S-curves. Frank Landymore's Majority of AI Researchers Say Tech Industry Is Pouring Billions Into a Dead End suggests that generative AI has hit the flat part of the S-curve:

You can only throw so much money at a problem.

This, more or less, is the line being taken by AI researchers in a recent survey. Asked whether "scaling up" current AI approaches could lead to achieving artificial general intelligence (AGI), or a general purpose AI that matches or surpasses human cognition, an overwhelming 76 percent of respondents said it was "unlikely" or "very unlikely" to succeed.
"The vast investments in scaling, unaccompanied by any comparable efforts to understand what was going on, always seemed to me to be misplaced," Stuart Russel, a computer scientist at UC Berkeley who helped organize the report, told NewScientist. "I think that, about a year ago, it started to become obvious to everyone that the benefits of scaling in the conventional sense had plateaued."

Stanford shares Berkeley's concerns as Dan Robinson reports in Billions pour into AI as emissions rise, returns stay pitiful, say Stanford boffins:

AI continues to improve – at least according to benchmarks. But the promised benefits have largely yet to materialize while models are increasing in size and becoming more computationally demanding, and greenhouse gas emissions from AI training continue to rise.

These are some of the takeaways from the AI Index Report 2025 [PDF], a lengthy and in-depth publication from Stanford University's Institute for Human-Centered AI (HAI) that covers development, investment, adoption, governance and even global attitudes towards artificial intelligence, giving a snapshot of the current state of play.
...
However, HAI highlights the enormous level of investment still being pumped into the sector, with global corporate AI investment reaching $252.3 billion in 2024, up 26 percent for the year. Most of this is in the US, which hit $109.1 billion, nearly 12 times higher than China's $9.3 billion and 24 times the UK's $4.5 billion, it says.

Despite all this investment, "most companies that report financial impacts from using AI within a business function estimate the benefits as being at low levels," the report writes.

It says that 49 percent of organizations using AI in service operations reported cost savings, followed by supply chain management (43 percent) and software engineering (41 percent), but in most cases, the cost savings are less than 10 percent.

When it comes to revenue gains, 71 percent of respondents using AI in marketing and sales reported gains, while 63 percent in supply chain management and 57 percent in service operations, but the most common level of revenue increase is less than 5 percent.

Meanwhile, despite the modest returns, the HAI report warns that the amount of compute used to train top-notch AI models is doubling approximately every 5 months, the size of datasets required for LLM training is doubling every eight months, and the energy consumed for training is doubling annually.

This is leading to rapidly increasing greenhouse gas emissions resulting from AI training, the report finds. It says that early AI models such as AlexNet over a decade ago caused only modest CO₂ emissions of 0.01 tons, while GPT-4 (2023) was responsible for emitting 5,184 tons, and Llama 3.1 405B (2024) pumping out 8,930 tons. This compares with about 18 tons of carbon a year the average American emits, it claims.

Recent developments in China have validated the researchers concerns:

The premise that AI could be indefinitely improved by scaling was always on shaky ground. Case in point, the tech sector's recent existential crisis precipitated by the Chinese startup DeepSeek, whose AI model could go toe-to-toe with the West's flagship, multibillion-dollar chatbots at purportedly a fraction of the training cost and power.

Of course, the writing had been on the wall before that. In November last year, reports indicated that OpenAI researchers discovered that the upcoming version of its GPT large language model displayed significantly less improvement, and in some cases, no improvements at all than previous versions did over their predecessors.

In fact, the models can get worse, as Maxwell Zeff reports in OpenAI’s new reasoning AI models hallucinate more:

According to OpenAI’s internal tests, o3 and o4-mini, which are so-called reasoning models, hallucinate more often than the company’s previous reasoning models — o1, o1-mini, and o3-mini — as well as OpenAI’s traditional, “non-reasoning” models, such as GPT-4o.

Perhaps more concerning, the ChatGPT maker doesn’t really know why it’s happening.

In its technical report for o3 and o4-mini, OpenAI writes that “more research is needed” to understand why hallucinations are getting worse as it scales up reasoning models. O3 and o4-mini perform better in some areas, including tasks related to coding and math. But because they “make more claims overall,” they’re often led to make “more accurate claims as well as more inaccurate/hallucinated claims,” per the report.

OpenAI found that o3 hallucinated in response to 33% of questions on PersonQA, the company’s in-house benchmark for measuring the accuracy of a model’s knowledge about people. That’s roughly double the hallucination rate of OpenAI’s previous reasoning models, o1 and o3-mini, which scored 16% and 14.8%, respectively. O4-mini did even worse on PersonQA — hallucinating 48% of the time.

It may well turn out that people put more value on being right than being plausible.

Increasingly, there are other signs that the current, costly, proprietary AI approach is coming to an end. For example, we have Matt Asay's DeepSeek’s open source movement:

It’s increasingly common in AI circles to refer to the “DeepSeek moment,” but calling it a moment fundamentally misunderstands its significance. DeepSeek didn’t just have a moment. It’s now very much a movement, one that will frustrate all efforts to contain it. DeepSeek, and the open source AI ecosystem surrounding it, has rapidly evolved from a brief snapshot of technological brilliance into something much bigger—and much harder to stop. Tens of thousands of developers, from seasoned researchers to passionate hobbyists, are now working on enhancing, tuning, and extending these open source models in ways no centralized entity could manage alone.

For example, it’s perhaps not surprising that Hugging Face is actively attempting to reverse engineer and publicly disseminate DeepSeek’s R1 model. Hugging Face, while important, is just one company, just one platform. But Hugging Face has attracted hundreds of thousands of developers who actively contribute to, adapt, and build on open source models, driving AI innovation at a speed and scale unmatched even by the most agile corporate labs.

For another example, there is Kyle Orland's Microsoft’s “1‑bit” AI model runs on a CPU only, while matching larger systems:

Now, researchers at Microsoft's General Artificial Intelligence group have released a new neural network model that works with just three distinct weight values: -1, 0, or 1. Building on top of previous work Microsoft Research published in 2023, the new model's "ternary" architecture reduces overall complexity and "substantial advantages in computational efficiency," the researchers write, allowing it to run effectively on a simple desktop CPU. And despite the massive reduction in weight precision, the researchers claim that the model "can achieve performance comparable to leading open-weight, full-precision models of similar size across a wide range of tasks."

The problem is that the driving force behind the data center boom isn't the AI researchers, it is the real estate people. David Gerard's China massively overbuilds empty AI data centres:

ChatGPT came out in 2022, and the Chinese government declared AI infrastructure a national priority. Over 500 new data centres were announced in 2023 and 2024. Private investors went all-in.

Demand for the data centres turns out not to be there. Around 80% are not actually in use. [MIT Technology Review]

The business model was to rent GPUs. DeepSeek knifed that, much as it did OpenAI. There’s now a lot of cheap GPU in China. Data centre projects are having trouble finding new investment.

The Chinese data centre boom was a real estate deal — many investors pivoted straight from real estate to AI.

Paul Kunert talked to someone with long experience of the data center real estate business in Dot com era crash on the cards for AI datacenter spending? It's a 'risk':

Having lived through the early days of the internet frenzy, Fabrice Coquio, senior veep at Digital Realty, which bills itself as the world's largest provider of cloud and carrier-neutral datacenter, colocation and interconnection services, is perhaps better placed than most to venture an opinion. Is there a bubble?

"I have been in this industry for 25 years, so I've seen some ups and downs. At the moment, definitely that's on the very bullish side, particularly because of what people believe will be required for AI," he tells The Register.

Grabbing a box of Kleenex tissues, he quips that back at the turn of the millennium, if investors were told the internet was inside they would have rushed to buy it. "Today I am telling you there is AI inside. So buy it."

"Is there a bubble? Potentially? I see the risk, because when some of the traditional investments in real estate – like housing, logistics and so on – are not that important, people are looking to invest their amazing capacity of available funds in new segments, and they say, 'Oh, why not datacenters?'"

He adds: "In the UK, in France, in Germany, you've got people coming from nowhere having no experiences… that have no idea about what AI and datacenters are really and still investing in them.

"It's the expression of a typical bubble. At the same time, is the driver of AI a big thing? Yes… [with] AI [there] is a sense of incredible productivity for companies and then for individuals. And this might change drastically the way we work, we operate, and we deliver something in a more efficient way.

The "slow AIs" that run the major AI companies hallucinated a future where scaling continued to work and have already sunk vast sums into data centers. The "slow AIs" can't be wrong:

Nonetheless, if Microsoft's commitment to still spending tens of billions of dollars on data centers is any indication, brute force scaling is still going to be the favored MO for the titans of the industry — while it'll be left to the scrappier startups to scrounge for ways to do more with less.

Sophia Chen's Data centres will use twice as much energy by 2030 — driven by AI shows the International Energy Agency believes the data centers will get built:

The IEA’s models project that data centres will use 945 terawatt-hours (TWh) in 2030, roughly equivalent to the current annual electricity consumption of Japan. By comparison, data centres consumed 415 TWh in 2024, roughly 1.5% of the world’s total electricity consumption (see ‘Global electricity growth’).

The projections focus mostly on data centres in general, which also run computing tasks other than AI — although the agency estimated the proportion of data-centre servers devoted to AI. They found that such servers accounted for 24% of server electricity demand and 15% of total data-centre energy demand in 2024.

Source

AI boosters will point out that there are other demand growth culprits, but:

Alex de Vries, a researcher at the Free University of Amsterdam and the founder of Digiconomist, who was not involved with the report, thinks this is an underestimate. The report “is a bit vague when it comes to AI specifically”, he says.

Even with these uncertainties, “we should be mindful about how much energy is ultimately being consumed by all these data centres”, says de Vries. “Regardless of the exact number, we’re talking several percentage of our global electricity consumption.”

Source

James Temple critiques the International Energy Agency report in Why the climate promises of AI sound a lot like carbon offsets:

There are reasonable arguments to suggest that AI tools may eventually help reduce emissions, as the IEA report underscores. But what we know for sure is that they’re driving up energy demand and emissions today—especially in the regional pockets where data centers are clustering.

So far, these facilities, which generally run around the clock, are substantially powered through natural-gas turbines, which produce significant levels of planet-warming emissions. Electricity demands are rising so fast that developers are proposing to build new gas plants and convert retired coal plants to supply the buzzy industry.

If the data centers get built, they will add to carbon emissions and push us closer to 3C sooner. Of course, this investment in data centers needs to generate a return, but it may well turn out that the market isn't willing to pay enough for Ghibli-style memes to provide it. Ed Zitron has been hammering away at this point, for example in There Is No AI Revolution:

Putting aside the hype and bluster, OpenAI — as with all generative AI model developers — loses money on every single prompt and output. Its products do not scale like traditional software, in that the more users it gets, the more expensive its services are to run because its models are so compute-intensive.

For example, ChatGPT having 400 million weekly active users is not the same thing as a traditional app like Instagram or Facebook having that many users. The cost of serving a regular user of an app like Instagram is significantly smaller, because these are, effectively, websites with connecting APIs, images, videos and user interactions. These platforms aren’t innately compute-heavy, at least to the same extent as generative AI, and so you don’t require the same level of infrastructure to support the same amount of people.

Conversely, generative AI requires expensive-to-buy and expensive-to-run GPUs, both for inference and training the models themselves. The GPUs must be run at full tilt for both inference and training models, which shortens their lifespan, while also consuming ungodly amounts of energy. And surrounding that GPU is the rest of the computer, which is usually highly-specced, and thus, expensive.

Zitron makes an important point:

OpenAI, as I've written before, is effectively the entire generative AI industry, with its nearest competitor being less than five percent of its 500 million weekly active users.

Source

Ed Zitron has been arguing for more than a year that OpenAI's finances simply don't make sense, and in OpenAI Is A Systemic Risk To The Tech Industry he makes the case in exquisite detail and concludes:

Even in a hysterical bubble where everybody is agreeing that this is the future, OpenAI currently requires more money and more compute than is reasonable to acquire. Nobody has ever raised as much as OpenAI needs to, and based on the sheer amount of difficulty that SoftBank is having in raising the funds to meet the lower tranche ($10bn) of its commitment, it may simply not be possible for this company to continue.

Even with extremely preferential payment terms — months-long deferred payments, for example — at some point somebody is going to need to get paid.

I will give Sam Altman credit. He's found many partners to shoulder the burden of the rotten economics of OpenAI, with Microsoft, Oracle, Crusoe and CoreWeave handling the up-front costs of building the infrastructure, SoftBank finding the investors for its monstrous round, and the tech media mostly handling his marketing for him.

He is, however, over-leveraged. OpenAI has never been forced to stand on its own two feet or focus on efficiency, and I believe the constant enabling of its ugly, nonsensical burnrate has doomed this company. OpenAI has acted like it’ll always have more money and compute, and that people will always believe its bullshit, mostly because up until recently everybody has.

OpenAI cannot "make things cheaper" at this point, because the money has always been there to make things more expensive, as has the compute to make larger language models that burn billions of dollars a year. This company is not built to reduce its footprint in any way, nor is it built for a future in which it wouldn't have access to, as I've said before, infinite resources.

Zitron uses Lehman Brothers as an analogy for the effects of a potential OpenAI failure:

I can see OpenAI’s failure having a similar systemic effect. While there is a vast difference between OpenAI’s involvement in people’s lives compared to the millions of subprime loans issued to real people, the stock market’s dependence on the value of the Magnificent 7 stocks (Apple, Microsoft, Amazon, Alphabet, NVIDIA and Tesla), and in turn the Magnificent 7’s reliance on the stability of the AI boom narrative still threatens material harm to millions of people, and that’s before the ensuing layoffs.

Siddharth Venkataramakrishnan's AI hype is drowning in slopaganda suggests that "the AI boom narrative" isn't self-sustaining:

One hint that we might just be stuck in a hype cycle is the proliferation of what you might call “second-order slop” or “slopaganda”: a tidal wave of newsletters and X threads expressing awe at every press release and product announcement to hoover up some of that sweet, sweet advertising cash.

That AI companies are actively patronising and fanning a cottage economy of self-described educators and influencers to bring in new customers suggests the emperor has no clothes (and six fingers).

There are an awful lot of AI newsletters out there, but the two which kept appearing in my X ads were Superhuman AI run by Zain Kahn, and Rowan Cheung’s The Rundown. Both claim to have more than a million subscribers — an impressive figure, given the FT as of February had 1.6mn subscribers across its newsletters.

If you actually read the AI newsletters, it becomes harder to see why anyone’s staying signed up. They offer a simulacrum of tech reporting, with deeper insights or scepticism stripped out and replaced with techno-euphoria. Often they resemble the kind of press release summaries ChatGPT could have written.

Yet AI companies apparently see enough upside to put money into these endeavours. In a 2023 interview, Zayn claimed that advertising spots on Superhuman pull in “six figures a month”. It currently costs $1,899 for a 150-character write-up as a featured tool in the newsletter.
...
“These are basically content slop on the internet and adding very little upside on content value,” a data scientist at one of the Magnificent Seven told me. “It’s a new version of the Indian ‘news’ regurgitation portals which have gamified the SEO and SEM [search engine optimisation and marketing] playbook.”

But newsletters are only the cream of the crop of slopaganda. X now teems with AI influencers willing to promote AI products for minimal sums (the lowest pricing I got was $40 a retweet). Most appear to be from Bangladesh or India, with a smattering of accounts claiming to be based in Australia or Europe. In apparent contravention of X’s paid partnerships policy, none disclose when they’re getting paid to promote content.
...
In its own way, slopaganda exposes that the AI’s emblem is not the Shoggoth but the Ouroboros. It’s a circle of AI firms, VCs backing those firms, talking shops made up of employees of those firms, and the long tail is the hangers-on, content creators, newsletter writers and ‘marketing experts’ willing to say anything for cash.

The AI bubble bursting would be a whole different and much quicker "going out with a bang". How likely is it? To some extent OpenAI is just a front for Microsoft, which gets a slice of OpenAI's revenue, has access to OpenAI's technology, "owns" a slice of the "non-profit", and provides almost all of OpenAI's compute at discounted prices. Microsoft, therefore, has perhaps the best view of the generative AI industry and its prospects.

That makes David Gerard's Confirmed: Microsoft stops new data centres worldwide very interesting:

In February, stock analysts TD Cowen spotted that Microsoft had cancelled leases for new data centres — 200 megawatts in the US, and one gigawatt of planned leases around the world.

Microsoft denied everything. But TD Cowen kept investigating and found another two gigawatts of cancelled leases in the US and Europe. [Bloomberg, archive]

Bloomberg has now confirmed that Microsoft has halted new data centres in Indonesia, the UK, Australia and the US. [Bloomberg, archive]

The Cambridge, UK site was specifically designed to host Nvidia GPU clusters. Microsoft also pulled out of the new Docklands Data Centre in Canary Wharf, London.

In Wisconsin, US, Microsoft had already spent $262 million on construction — but then just pulled the plug.

Mustafa Suleyman of Microsoft told CNBC that instead of being “the absolute frontier,” Microsoft now prefers AI models that are “three to six months behind.” [CNBC]

Google has taken up some of Microsoft’s abandoned deals in Europe. OpenAI took over Microsoft’s contract with CoreWeave. [Reuters]

Ed Zitron covered this "pullback" a month ago in Power Cut:

As a result, based on TD Cowen's analysis, Microsoft has, through a combination of canceled leases, pullbacks on Statements of Qualifications, cancellations of land parcels and deliberate expiration of Letters of Intent, effectively abandoned data center expansion equivalent to over 14% of its current capacity.
...
In plain English, Microsoft, which arguably has more data than anybody else about the health of the generative AI industry and its potential for growth, has decided that it needs to dramatically slow down its expansion. Expansion which, to hammer the point home, is absolutely necessary for generative AI to continue evolving and expanding.

Microsoft may have got the message, but T.D. Cowen reported others haven't:

While there is a pullback in Microsoft's data center leasing, it’s seen a "commensurate rise in demand from Oracle related to The Stargate Project" — a relatively new partnership of "up to $500 billion" to build massive new data centers for AI, led by SoftBank, and OpenAI, with investment from Oracle and MGX, a $100 billion investment fund backed by the United Arab Emirates.

The data centers will get built, and they will consume power, because even if AI never manages to turn a profit, some other bubble will take its place to use all the idle GPUs. Despite all the rhetoric about renewables and small modular reactors, much of the additional power will come from fossil fuels.

Data center carbon emissions don't just come from power (Scope 1 and 2). In 2023 Jialin Lyu et al from Microsoft published Myths and Misconceptions Around Reducing Carbon Embedded in Cloud Platforms and stressed the importance of embedded carbon (Scope 3) in the total environmental impact of data centers:

For example, 66% of the electricity used at Google datacenters was matched with renewable energy on an hourly basis in 2021. With historic growth rates, this is likely closer to 70% today. Our LCAs indicate that with 70-75% renewable energy, Scope 3 accounts for close to half of datacenter carbon emissions. Therefore, Scope 3 emissions and embodied carbon are important factors both currently and in the near future.

Microsoft is commendably frank about Scope 3 emissions. A year ago Dan Robinson reported that Microsoft's carbon emissions up nearly 30% thanks to AI :

The Redmond IT giant says that its CO2 emissions are up 29.1 percent from the 2020 baseline, and this is largely due to indirect emissions (Scope 3) from the construction and provisioning of more datacenters to meet customer demand for cloud services.

These figures come from Microsoft's 2024 Environmental Sustainability Report [PDF], which covers the corp's FY2023 ended June 30, 2023. This encompasses a period when Microsoft started ramping up AI support following the explosion of interest in OpenAI and ChatGPT.

Microsoft's "pullback" will likely reduce their Scope 3 emissions going forward, but I would expect that their recent build-out will have reduced the proportion of renewables being consumed. If the Stargate build-out goes ahead it will cause enormous Scope 3 emissions.

Source

Here is one final note of gloom. Training AI models requires rapid access to large amounts of data, motivating data centers to use SSDs instead of hard drives. Counterintuitively, research Seagate published in Decarbonizing Data shows that, despite their smaller size, SSDs have much higher embedded carbon emissions than hard drives. 30TB of SSD has over 160 times as much embedded carbon as a 30TB hard drive.

Source

Even more of a suprise, Seagate's research shows that SSDs even have higher operational emissions than hard drives. While actively reading or writing data, a 30TB SSD uses twice as much power as a 30TB hard drive.

The legacy of Balthus' clients attitude that they 'know the world is doomed, so may as well go out with a bang' and the unsustainable AI bubble will be a massive overbuild of data centers, most of which will be incapable of hosting Nvidia's top-of-the-line racks. If the current cryptocurrency-friendly administration succeeds in pumping Bitcoin back these data centers will likely revert to mining. Either way, the Scope 3 emissions from building and equipping and the Scope 1 and 2 emissions from powering them with natural gas and coal, will put megatons of CO2 into the atmosphere, hastening the point where it is unlikely that 'us in the North, we're fine'.