When we talk about hyperscalers, we’re talking about big ‘clouds’ like Microsoft, Amazon, Google or Meta. These companies have been multiplying their investment in AI infrastructure for months, and Nvidia has translated that investment into record sales figures. Not so clear is how many of these chips are actually working.
More and more voices are appearing that large companies are looting GPUs that they can’t plug in anywhere because they lack data centers and electricity. And while some see signs of bubbles, others say absolutely not.
In short. A few weeks ago, analyst Ed Zitron published an analysis based on the accounts of major technology companies. It estimates that Nvidia and Broadcom have sold more than $560 billion in AI chips and equipment since the start of 2023, but that half of them will be uninstalled or outright, without power. These are estimates from Zitron, but Bloomberg already reported that Microsoft had a total capacity of about 12 gigawatts, of which only two correspond to AI chips, and Satya Nadella himself said in November 2025 that they have chips that cannot be plugged in.
Again, according to Zitron, Microsoft is talking about $50,000 to $100,000 million in GPUs that they have in their boxes. This is something that has already been discussed before because, basically, hyperscalers are buying like crazy at the idea of being able to connect GPUs at some point in data centers that they haven’t built yet. But, while this is an analyst’s guess, last October 19, an investigation by Business Insider pointed in the same direction for Amazon.
Amazon is estimated to spend about $220 billion this year, and it has negative cash flow because of those expenses. According to the media, internal emails from this summer showed that they expected about 5,000 million racks with GB300 chips from Nvidia in August, and about 3,000 million more per month since then. And it looks like Amazon has data centers, but not enough power, so equipment is stored in intermediate warehouses waiting to be worked on.
In the long run. The problem is, based on this investigation, Amazon doesn’t expect the problem to be resolved anytime soon. In an internal meeting this September, AWS’ infrastructure chief said he expects to maintain negative cash flow for “several years.” In 2003, AWS expected to overcome the capacity gap in 2027, but now they expect demand to still exceed capacity that year.
Despite everything, it seems worth it, as AWS grew 37% in the second quarter of this year, its fastest pace in the last 18 quarters. Let us remember that this year, the investment is astronomical, with the mentioned 190,000 million dollars from Microsoft, 145,000 million from Meta, 205,000 from Google and 220,000 from Amazon. In total, and with only these four without further consideration, 760,000 million dollars.
bubbles. A few days after the Business Insider article, Jennifer M. Harris published an opinion piece in the New York Times in which he explained that the current state of hyperscalar is reminiscent of the 2008 financial crisis. It’s not the first time it’s been said, even by voices that already saw the 2008 crisis coming, but it turns out that Harris was White House economic officer in the Biden era.
To support the thesis, Harris cites data from the Wall Street Journal from this August, when the nine largest tech companies were charged with nearly $3 trillion in off-balance sheet debt. He also noted that the Financial Times put OpenAI’s estimated annual revenue at about $50 billion, which, yes, is a lot, but is about 30% lower than expected.
And the big problem is that the debt companies are taking on is financed with personal credit tied to insurance and retirement savings. He argues that AI would need to generate up to three trillion a year to return invested capital, leaving little room for deficit bailouts. The only solution is to embrace frontier AI as an integral part of government with a national laboratory and a computing infrastructure that is a public service.
“The companies responsible for protecting our pensions and insurance are once again risking our money by buying complex products backed by assets of uncertain value” – Jennifer M. Harris
bubbles? Let’s face it, credit reliance on AI means that if it explodes, it will wipe out a large part of the American economic structure, and we already know the consequences for the rest of the West. But of course, then there are voices like the one most interested in the infinite growth of AI: Jensen Huang. Nvidia’s CEO commented in July of this year that “this time is different” and that there is no such bubble because what is happening is an industry change and in addition, there is a healthy brake that did not exist in the 2008 bubble.
While there’s plenty of money that’s not landing, that healthy brake is a lack of energy and slow deployment of data centers. It’s the physical barriers that prevent companies from spending and spending like crazy, and so, this time is different. Once data centers and power are in place, things will change. Another interested voice is Masayoshi Son, CEO of SoftBank and OpenAI’s largest shareholder, who has commented that talk of a bubble is “blasphemy against AI”.
Among the companies very interested in this technology, there are also profiles that are more moderate. Nadella, in fact, although he does not consider that there is a bubble as there was more than a decade ago, he does consider that the sector needs to attract consumers in order for things to end well.
consequences. For the moment, in the larger sphere, things seem to be going as they have been for months, with Nvidia’s unstoppable progress in hardware that continues to sell billions of GPUs to the original hyperscaler. For now the most visible consequence is one we know all too well: the cost of ingredients
In recent months we’ve seen how prices of both RAM and storage have skyrocketed, and all consumer technology has been running out of stock for both users and companies. This has pushed consoles, but also budget mobile phones, and we’re already seeing desktop computers, laptops and premium mobile phones that are more expensive than a year ago without necessarily having better specifications.
And the worst thing is that no one knows how long this situation is left or if the prices will normalize once the situation with the data centers stabilizes. According to Lenovo, when the storm passes, prices should not return to pre-crisis levels. And, for companies, it makes perfect sense: if we need a computer, even if it costs 500 euros more, we’re going to buy it and they don’t have to lower the price.
Picture | NVIDIA (edited)
In Xataka | All AI companies want to make their own chips, so Nvidia now wants to have its own AI
