The conversation is no longer about whether artificial intelligence (AI) will transform the economy, but at what financial cost. In the last 48 hours, several influential voices — from macro consulting firms to Wall Street bankers and big tech — have sounded alarm bells about a possible speculative bubble surrounding AI. For some, it’s not an early warning sign: it’s a “red flag”.
The most forceful warning came in a report by MacroStrategy Partnership, authored by a team including Julien Garran, a former UBS analyst. The report claims that the bubble fueled by AI would be 17 times larger than the dot-com bubble at the end of the nineties and four times bigger than the global real estate bubble that burst in 2008. This comparison is based on a “wicksellian” approach: after years of artificially low interest rates and quantitative easing, the cost of corporate debt remained below the threshold described by Knut Wicksell as healthy, encouraging an inefficient allocation of capital. This assessment doesn’t just cover AI but also housing, offices, NFTs, and venture capital.
Have we reached the peak of LLMs?
MacroStrategy also points to practical limits in language models (LLMs):
- One software company reported task completion rates ranging from 1.5% to 34%, with inconsistent results even in the best case.
- Data from the US Department of Commerce, shared by economist Torsten Slok (Apollo) and cited in the report, indicate that AI adoption in large corporations has already begun to decline.
- In ad hoc tests — such as asking an image generator to produce a chessboard one move from checkmate — the results fell far short of expectations.
The core argument: LLMs are nearing scaling limits. “We don’t know exactly when a LLM hits diminishing returns because we don’t measure the statistical complexity of language. But if the next model costs 10 times more, uses 20 times more compute, and doesn’t improve significantly, then we’ve hit a wall,” Garran summarizes.
MacroStrategy illustrates this with a cost staircase: ChatGPT-3 reportedly cost $50 million, GPT-4 $500 million, and GPT-5 — $5 billion, according to the report — was delayed and, upon release, was not substantially better than its predecessor. In such a context, long-term competitive moats are unlikely, and margin pressures are mounting: “Business cases are either generic (games) and don’t sell, or public domain data is recycled (tasks, homework), or copyright issues emerge. Advertising is hard to optimize, and training costs grow exponentially with marginal gains. There’s no pricing power in the model, and heavy users consume more compute than the cost of their monthly subscription,” Garran concludes.
Macro risk: “zone 4” of the investment clock
On a macro level, the firm warns of a potential “deflationary collapse” if the wealth effect from AI platforms and the data center build cycle slows down… and then reverses. The problem, they add, would be twofold: “We would not only enter zone 4 of the investment clock; it would also be difficult for the Fed and the Trump Administration to quickly reflate the economy.” MacroStrategy’s portfolio approach recommends underweighting platforms and AI, overweighting resources and emerging markets (highlighting India and Vietnam), and holding long positions in gold, short-term US Treasury bonds, volatility, and yen (against non-USD currencies).
Debt in data centers: another warning sign
The leverage undertaken by big tech to develop AI data centers also concerns Dario Perkins (TS Lombard). In Axios, he compared the current frenzy to the leverage seen during the dot-com and subprime eras: “They say they don’t care if their investment returns or not because they’re in a race. That alone is a red flag.”
Wall Street: warning of “drawdown” (and big tech caution)
At Italian Tech Week, Goldman Sachs CEO David Solomon spoke of a probable “drawdown” in stocks within 12–24 months: too much capital has been deployed too quickly, and some of it won’t pay off. “When that happens, people won’t feel good,” he told CNBC. He didn’t explicitly say it’s a bubble, but admitted that some investors are “at the risk curve” driven by excitement, a classic sign of euphoria.
Similarly, Jeff Bezos (Amazon) acknowledged at the same forum that there is a bubble in AI — “investors have a hard time differentiating good from bad ideas amid the hype” — but argued that the technology will bring great long-term benefits to humanity.
The market ignores (for now) the sirens
Despite these warnings, indices remained strong at week’s end. The S&P 500 hit its 30th record of 2025 (6,715.35), futures opened higher, and gold — another risk indicator — surged to $3,887.6 (+47.3% year-to-date). The Nasdaq Composite reached 22,844.05; the yield on the 10-year Treasury hovered around 4.094%; oil dropped to $60.9. The most active tickers included Tesla (TSLA), NVIDIA (NVDA), GameStop (GME), TSMC (TSM), Palantir (PLTR), NIO (NIO), Amazon (AMZN), Intel (INTC), Apple (AAPL), and AMD (AMD). The ISM Services Index and Federal Reserve appearances (John Williams, Stephen Miran) painted a picture of an estimated unemployment rate holding at 4.3% (according to a Chicago Fed nowcast).
Mixed signals on the corporate front
The flow of news supports the thesis of “a lot of money, very quickly”:
- Applied Materials forecasted an impact of $710 million in revenue over the next five quarters from new export controls (BIS).
- Jefferies downgraded Apple to “underweight” due to inflated expectations around a foldable iPhone.
- BlackRock’s Global Infrastructure Partners (GIP) is reported to be in advanced negotiations to acquire Aligned Data Centers (backed by Macquarie) for approximately $40 billion, according to Bloomberg: a sign that colocation assets continue to attract mega-capital.
- AI also drives rallies in China, revitalizing sectors related to computing and networks.
Bubble or healthy correction? What the numbers—and the unknowns—say
The 17x and 4x figures from MacroStrategy include more than just pure AI: they also incorporate real estate, NFTs, and venture capital into their “Wicksellian deficit,” thus capturing the decade’s ultra-loose monetary policy. This doesn’t invalidate the warning but nuances its focus: low interest rates have inflated multiple assets simultaneously; AI is the most visible driver in 2024–2025, with massive capex in chips, memory, data centers, and energy.
The technology side is also more granular. While general LLMs show fatigue (quality vs. cost), the vertical AI pipeline — narrow-task copilots, industrial vision, biosimulation — continues to deliver incremental value. Part of the hyperscale capex is justified by fleet replacements, power efficiency, and latency improvements for edge/serving. The risk isn’t AI as a concept, but the investment pace relative to tangible returns and physical limits (power, cooling, memory).
A key point is the early signals from corporate profits versus S&P 500 earnings. Research by Ned Davis Research, cited by MarketWatch, points to negative aggregate corporate earnings (NIPA) in Q1 and Q2 despite S&P advances. Historically, such divergence doesn’t last long: if NIPA leads the S&P by 1–2 quarters, index estimates could become overly optimistic—implying downside risk in the coming quarters.
What are the “Top of AI” insiders doing?
In this landscape, leaders —NVIDIA, TSMC, AMD, Intel, Amazon, Microsoft, Alphabet— navigate a paradox: they are winners of the cycle (demand for GPUs, HBM, cloud, and colocation), but also the most exposed if capex appetite cools or return timelines stretch. NVIDIA’s rally dominates the story; TSMC acts as a supply thermostat (N2/A16 nodes, advanced packaging); AMD and Intel compete in GPU/CPU AI chips; hyperscalers spread spending across GPUs, custom silicon, and data centers. Increasing debt for infrastructure—highlighted by TS Lombard—adds leverage to the mix.
What could break (or save) the cycle
Could disrupt it:
- Fading hyperscale capex if returns and adoption don’t keep pace;
- Energy constraints that slow data center deliveries;
- More persistent memory/flash and HBM limits than expected;
- Regulatory hurdles (data, copyright, export controls) that increase costs and delay deployments.
Could save the cycle:
- Software efficiency improvements (quantization, sparsity, distillation) that reduce cost per token;
- Custom silicon (NPUs, accelerators) that boost performance per watt;
- Vertical demand (healthcare, industry, finance) with clear ROI cases;
- Industrial and energy policies that streamline megawatt deployment and renewables.
Two ideas for the investor/corporate reader
- Separate “narrative AI” from “accounting AI”. Ask about €/result (€/epoch, €/million tokens, €/inference P95) instead of €/hour. And about kWh/task. If companies measure these, the risk of “hype” diminishes.
- Focus on physical timelines: MW contracted, substation dates, GPU/HBM supply, liquid cooling. The real time-to-compute takes precedence over the storytelling.
Conclusion: today’s noise doesn’t erase the long-term view, but it demands fine filters
AI doesn’t disappear with a “drawdown”: like the Internet after 2000, it will survive and mature. What could get cut short is the expectation curve —and with it, valuations discounting immediate returns and doing so without friction. In the meantime, it’s important to distinguish transformative technology from costly-funded raceways. And as Bezos reminded us, to accept that “amid excitement it’s hard to diferentiate good from bad ideas.” That’s exactly the task for investors and prudent managers in 2025.
Frequently Asked Questions
Where do the “17× dot-com” and “4× subprime” figures come from in the supposed AI bubble?
They originate from a note by MacroStrategy Partnership estimating a “Wicksellian deficit” after years of artificially low interest rates, adding not just AI but also housing, offices, NFTs, and venture capital. Their thesis is that excess liquidity has massively overfinanced assets.
What data points suggest limits to scaling in LLMs?
The note cites task completion rates in real tasks (1.5–34%), a slowdown in adoption among big companies (from US Department of Commerce data via Torsten Slok), and a cost staircase (GPT-3 $50M, GPT-4 $500M, GPT-5 $5B) with only marginal improvements.
What do analysts recommend in the face of a potential burst?
MacroStrategy suggests underweight AI and platforms, overweight resources and emerging markets (India, Vietnam), and holding long positions in gold, short-term US Treasuries, volatility, and yen against most non-USD currencies.
What do big tech and investment banks think?
David Solomon (Goldman Sachs) predicts a “drawdown” in stocks within 12–24 months due to excess capital; Jeff Bezos recognizes a bubble but advocates long-term benefit; Dario Perkins warns of accelerated debt for AI data centers as a “red flag.”
Sources: MarketWatch (10/03/2025); Axios; CNBC; Common Dreams; Bloomberg (GIP–Aligned).