Big Tech earnings put the AI bubble under pressure

Big Tech’s latest earnings have delivered a volatile mix of warnings, rallies, and enormous spending commitments—prompting questions about whether the AI investment boom is already starting to unwind.

In the latest episode of The Register’s The Kettle, editor-in-chief Matt Rosoff, systems editor Tobias Mann, and host Brandon Vigliarolo examined what recent quarterly results reveal about the AI sector and what IT teams should do in an uncertain market. Their central advice was implicit but clear: companies should not commit their entire enterprise to frontier AI products.

Big Tech stocks are behaving like penny stocks

Rosoff said the sharp market swings resemble the late stages of a speculative bubble. Apple fell 10 percent after warning that its next quarter could be weaker than expected because memory and other component prices were becoming more expensive. IBM lost more value in a single day after its latest earnings warning than it had lost since Black Monday in 1987.

Recommended reading

SpaceX shares plunge 46% as Musk loses over $600 billion

“You have these massive tech stocks, which represent most of the value of the stock market, acting like penny stocks, and that tends not to be a great sign.”

The reaction to Meta and Amazon showed how difficult it is for investors to interpret the numbers. Meta’s free cash flow dropped to under $1 billion, compared with $8.5 billion at the same time last year, as the company spends heavily on data centers. Investors responded with a major sell-off.

Amazon, by contrast, rose 15 percent after AWS grew faster than expected and the company increased its capital-expenditure estimates. But the discussion pointed to a complication: part of AWS’s reported margin improvement came from a successful hedge against energy prices. That boost is not expected to repeat in the third quarter, according to comments made on Amazon’s earnings call.

The accounting and capex pressure

The speakers also discussed analysis by AWS analyst Corey Quinn, who said Amazon had conflated chip revenue with EC2 instance revenue. Quinn also argued that Amazon’s investment and commercial arrangements with Anthropic created overlapping references to the same money across its reporting.

As presented in the podcast, Amazon accounted for $53.4 billion related to deals with Anthropic in the previous quarter. The claim was that a single Anthropic dollar could appear in references to AI business revenue, the chip business, and AWS segment revenue. The podcast did not provide a full accounting breakdown or independent verification of those assertions.

Quinn’s broader warning was that the demand supporting $220 billion in capital expenditure is currently concentrated among a small number of AI labs, while wider enterprise adoption remains a forecast.

“The demand underwriting two hundred and twenty billion in capex is concentrated today in a handful of AI labs, one of which Amazon happens to own a meaningful piece of, when the broader menterprise adoption wave remains a forecast.”

Rosoff characterized the spending less as an attempt to artificially inflate the market than as a bet that the companies cannot safely abandon. Mann described it as a sunk-cost fallacy: after committing so much money, stopping could trigger an even worse market reaction.

Data centers face physical constraints

The discussion emphasized that building AI infrastructure is not as simple as announcing a data-center project. Companies must secure sites and permits, obtain water and electricity, build substations, arrange networking, and confirm that GPUs and CPUs will be available.

Meta is reportedly building data centers that could cost $50 billion to complete, yet Mann said the company has relatively little to show beyond improved recommender systems. Meta’s Llama models were described as adequate but not on par with the work of OpenAI or Anthropic. The speakers also questioned Meta’s ability to consistently produce models that users want, citing disappointing releases and an uncertain record.

Mann said hyperscalers use leased capacity as a hedge because leases are easier to abandon than owned facilities. He pointed to earlier cases in which Microsoft withdrew from data-center leases because the facilities could not support the hardware it ultimately wanted to deploy—not because the company had simply lost interest in infrastructure.

AI-focused data centers also differ from traditional air-cooled facilities. Their power consumption makes evaporative cooling cost-effective, increasing the need for water, backup generation, and utility agreements before construction begins. Liquid cooling introduces another bottleneck: even after a project is otherwise ready, companies may wait a year for plumbing and manifolds needed to connect the systems.

Demand remains the central uncertainty

The infrastructure challenges are only half of the risk. Rosoff said he remains unconvinced that large language models will replace most economic activity. He described current systems as useful for prototyping and for helping experienced programmers do more with fewer resources, but not as reliable replacements for production coders or medical professionals.

The discussion also cited reports of companies rehiring workers previously laid off because of expected AI substitution. Some businesses have found that chatbots remain inferior to human customer-service representatives when customers need complex problems resolved. The episode transcript ends while the speakers are still discussing enterprise AI return on investment, so it does not provide a final forecast or a specific action plan for IT departments beyond avoiding an all-in bet on frontier-lab products.