GOOG, MSFT, AMZN, NVDA: Sven Carlin Warns on AI Risk–Reward

Sven Carlin’s video examines the risk–reward trade-off for the dominant AI and cloud "hyperscalers" — Alphabet (GOOG), Microsoft (MSFT), Amazon (AMZN), Nvidia (NVDA), Meta, Apple and Oracle — as of early August 2026.

According to Carlin, headline numbers look excellent: strong cloud growth at Google, big revenue gains at Nvidia, booming operating income at Amazon, and powerful AI narratives almost everywhere. Yet he argues that beneath the surface much of this AI-driven growth is being engineered through what he calls circular financing, not through proven, durable profitability.

Carlin’s overarching claim is that these companies have morphed from cash-generating, capital-light machines into high-capex, highly leveraged AI gamblers whose valuations already price in more-than-perfect outcomes. He stresses that the video is an exploration of those dynamics, not a forecast of short-term stock prices.

Carlin centers his analysis on the big cloud and AI players: Alphabet (GOOG), Microsoft (MSFT), Amazon (AMZN), Nvidia (NVDA), Meta (Facebook), Apple, Oracle and their key AI customers Anthropic and OpenAI. He argues that the apparent strength in cloud and AI revenues rests on massive capital expenditure, vendor financing, and a narrow customer base.

For Alphabet, Carlin notes that Google Cloud reported strong growth and positive share-price reaction, but he highlights that, according to his query to Google’s own AI, about 40% of Google Cloud business is tied to Anthropic. He states that Anthropic and OpenAI together account for roughly half of what he describes as a two-trillion revenue backlog across Amazon, Microsoft, Google and Oracle.

On Nvidia, Carlin points to very strong revenue growth and a price-to-earnings ratio he cites as around 30, but contends that Nvidia increasingly needs to finance customers so that those same customers can buy Nvidia’s chips at high prices. The same theme extends, in his telling, to Microsoft, Amazon and Meta: huge AI capex justified by backlog commitments that, he argues, ultimately rely on a few unprofitable AI labs.

To support his concerns, Carlin walks through a series of figures and contractual structures he sees as evidence of an AI capex bubble.

On the cloud and AI side, he says Anthropic has a 200 million dollar commitment from Google and that hyperscalers collectively are planning to invest roughly 200 billion per year each, resulting in what he labels a 750 billion AI bet and a circular tech bubble reminiscent of the dot-com era. He asserts that Anthropic and OpenAI together sit behind about half of the roughly two trillion in revenue backlogs he attributes to Amazon, Microsoft, Google and Oracle.

For Microsoft (MSFT), Carlin cites property, plant and equipment rising from 109 billion to 313 billion, while depreciation and amortization only increased by about 10 billion. He argues that if this 313 billion were depreciated over 5 years, annual cost would be about 60 billion instead of the roughly 40 billion he references, implying rising future expense as Microsoft continues to add, in his words, around 50 billion in capex in a single quarter.

Carlin also highlights what he calls hidden liabilities: he cites estimated 1.65 trillion of balance-sheet debt and off-balance sheet commitments at hyperscalers via special purpose vehicles, capacity off-take agreements, leases, and other structures. He mentions around 800 billion in contractual purchase commitments for Alphabet, Meta, Microsoft and Amazon, and describes additional off-balance sheet investments such as Meta’s 80/20 fund partnership with BlackRock.

Carlin repeatedly returns to the idea of circular financing: hyperscalers and Nvidia (NVDA) providing capital, credits, or long-term commitments to AI startups and cloud customers, which in turn use that capital to purchase the hyperscalers’ own chips and services. In his view, this makes a large share of current AI revenue growth not yet economically real and highly dependent on continued funding.

He stresses that core AI customers like Anthropic and OpenAI, as he understands it, are not profitable and may never be because of fierce competition and massive spending. He notes that Meta’s costs and expenses grew 55% against 28% revenue growth in one period he cites, with income from operations and margins declining, which he uses as an example of AI investments that are not cushioned by circular customer arrangements.

Carlin further argues that hyperscalers are spending multiples of their current cloud revenue on capex, calling out Meta as spending around 10x, with Google and Microsoft at roughly 2x and 3x respectively in his chart discussion. To him, this “spend now, profits maybe later” profile is a sharp break from what these businesses used to be: low-capex, high-margin, buyback-heavy cash generators like Apple’s older model.

Looking ahead, Carlin emphasizes several markers he believes will determine whether the AI spending wave pays off or unravels.

First, he argues that the biggest risk is not AI failing, but AI succeeding to the point that someone can ask an advanced system how to replicate the same capabilities at 10% of the current cost. In that scenario, he suggests, customers could quickly seek cheaper AI compute, possibly from new entrants or from China, which he notes is also investing heavily in AI infrastructure. He believes most users will not care where their AI comes from if it is cheaper and works.

Second, he points to the long-dated cost recognition. He references Goldman Sachs work implying that by 2041 there could be almost a trillion per year in depreciation and amortization related to AI compute investments. In his framing, revenue is being recognized today, while a large portion of the costs will hit income statements 5–10 years later, leaving the true return on investment unknown.

Finally, Carlin flags valuations as a key signal. Using Alphabet (GOOG) as an example, he claims that to justify a future valuation of around 9 trillion within 5 years, Alphabet would need to earn about 400 billion in net profits annually — more than its then-current revenue level, by his calculation. He makes similar comments about Microsoft and Meta, and recalls how Microsoft shares went nowhere for roughly 15 years after the dot-com bubble, suggesting to viewers that a long period of flat returns for hyperscalers is, in his words, "very likely" if expectations are not met.