Big Tech's AI Spending Has a Free Cash Flow Problem

The six largest American tech companies — Amazon, Apple, Alphabet, Microsoft, Meta, and Nvidia — carried a combined valuation of roughly $8 trillion at the end of 2020. By the end of 2025, according to ColdFusion's Dagogo Altraide, that figure had reached approximately $20 trillion, and has since climbed past $23 trillion. Those are the kinds of numbers that make any capital markets analyst pause, not because growth is suspicious in itself, but because valuation eventually has to connect to something — cash flow, earnings, tangible productive output. Right now, that connection is getting harder to trace.

That is the animating tension in Altraide's recent video essay on Big Tech's financial behavior, and it is a tension worth taking seriously regardless of where you land on the bubble question.

The Free Cash Flow Cliff

The forensic accounting argument at the center of Altraide's piece comes from accountants John Weil and Kevin Koharki, whom he features at length. Their framework is not complicated, but it is clarifying. Start with reported net income — impressive. Follow the cash flow — still looks reasonable. Then account for capital expenditures, particularly the data center buildout, and the picture dims. Layer in the cash cost of stock-based compensation and the buybacks used to offset dilution from those awards, and the residual free cash flow approaches zero for some of the largest names in the index.

As Weil puts it: "They used to be free cash flow machines, they're not free cash flow machines anymore."

This matters structurally. Free cash flow is the oxygen of a capital-efficient business. It funds dividends, supports buybacks, enables acquisitions without leverage, and — critically — gives management the flexibility to weather downturns without hitting the debt markets. When it evaporates in pursuit of a capital expenditure cycle that has not yet demonstrated its return, the entire risk profile of these companies shifts. That shift may be temporary and justified by the eventual payoff, or it may not. What it is not, is irrelevant.

The flat-fee pricing collapse already underway at companies like Anthropic and GitHub suggests the infrastructure cost is outpacing monetization in at least some corners of the ecosystem.

What 18 Cents Tells You

Altraide cites a survey of nearly 2,500 companies finding that for every dollar spent on AI, only 18 cents reaches production. The remainder disappears into error correction, rework, and operational friction. He attributes this finding to the survey itself but does not identify the publisher, so treat the specific figure as directionally informative rather than precise. Even so, the pattern it describes — that enterprise AI deployment is substantially more expensive in practice than in pitch — is consistent with what operators across industries have reported publicly.

Microsoft CEO Satya Nadella has acknowledged that large language models alone are insufficient and require human judgment layered on top. Robert Half's Meagan Slapinsky told Fast Company that organizations have had to reassess expectations, recognizing that AI "is not the be-all and end-all solution some initially believed it would be." Box CEO Aaron Levie put it more pointedly: "You can get by for a while by being non-technical about building software, but eventually someone has to understand the thing that got built, has to maintain it, has to fix security issues that come up, upgrade the systems beneath it, and so on. That's all jobs."

That last observation deserves to sit on its own for a moment. The workforce calculus that justified laying off software engineers in anticipation of AI replacement is now being reversed by the same technology leaders who promoted it. The AI economic exposure concentrated in technical roles turns out to be more complicated than the initial displacement narrative allowed.

The Circular Financing Question

The more structurally interesting — and contested — argument in Altraide's piece concerns what he calls circular financing between hyperscalers and the AI companies they both invest in and sell cloud services to.

The Google-Anthropic arrangement is the clearest example in the public record. Google has made substantial investments in Anthropic across multiple funding rounds, both directly and through its cloud division. Anthropic, in turn, has committed to using Google Cloud infrastructure. This creates an accounting situation where Google's investment flows out as capital and a portion returns as cloud revenue — revenue that feeds Google's income statement.

Sasha Yanxin, described in the video as someone with experience creating financial products for major banks, frames the mechanics bluntly: "Google goes and gives Anthropic $10 billion as an investment, and Google receives the $10 billion back as revenue from Anthropic for using their data centers. Google gets a massive profit without actually having to do nothing."

Altraide notes that Google's Q1 2026 earnings attributed a significant portion of profit growth to "other income" — and that the company simultaneously laid off staff in its cloud division, citing a reallocation toward AI. Whether this arrangement constitutes genuine economic activity or accounting architecture depends partly on how you weigh the actual infrastructure services being rendered against the optics of two nominally competing companies financially intertwined.

Amazon presents a virtually identical structure through its relationship with Anthropic. The pattern is not unique to any single deal. As Altraide observes, money flowing between companies that are ostensibly rivals has become a defining feature of the current AI investment landscape.

The optimistic read: these are rational expressions of mutual confidence in AI infrastructure, where hyperscalers use investment stakes to lock in anchor cloud customers and AI startups use cloud credits to defer upfront costs. The skeptical read: it is a mechanism for reporting revenue that cycles through a counterparty rather than being earned from end customers in the broader economy. Both readings can be simultaneously true in different proportions, which is precisely what makes it worth examining.

The Competitive Pressure From Below

There is a separate threat to the central business case that does not involve any accounting interpretation at all: cheaper alternatives are working well enough for most applications.

Open-source and open-weight models — many developed by Chinese AI labs — have narrowed the performance gap with proprietary American LLMs substantially. Altraide notes that companies including Coinbase, Shopify, Airbnb, and Siemens have begun routing workloads to these models. The value proposition is straightforward: lower API costs, the ability to run models locally without an internet dependency, and full control over data. For the majority of enterprise use cases, which do not require frontier-model capability, this is a rational procurement decision.

The implication for hyperscaler economics is direct. If a meaningful share of enterprise AI workloads migrate toward open-weight models running on commodity hardware, the revenue thesis for the data center infrastructure buildout weakens considerably. The diverging market reactions to Meta and Microsoft's AI capital expenditure announcements already reflect investor uncertainty about which companies are actually positioned to monetize the infrastructure they are building.

Sam Altman has publicly walked back his earlier predictions about AI-driven job displacement, describing himself as "delighted to be wrong." That is a notable rhetorical retreat from a CEO whose company's valuation depends partly on the market believing in transformative near-term economic impact.

What the Nvidia-Valor Arrangement Illustrates

Michael Burry — whose 2008 credit crisis positioning established his credibility as a structural skeptic — has publicly outlined a chain connecting Nvidia GPU sales, a company called Valor, private credit intermediaries, and xAI's infrastructure. According to Altraide's account of Burry's diagram, the arrangement allowed Nvidia to book revenue on GPU sales while those chips remained operationally tethered to xAI through what Burry characterized as a non-independent intermediary.

The core allegation is that Valor functions as a pass-through entity: Nvidia records a sale, xAI gets compute, and the financing runs through institutional capital — including, per Burry's framing, pension fund exposure via private credit vehicles. These are allegations, not established findings, and Altraide is careful to note that nothing in this structure has been identified as explicitly illegal. The SEC has not brought an enforcement action. But the architectural complexity of the arrangement raises the question that always follows these kinds of structures: who is it complicated for, and why?

The GDP Distortion Problem

Harvard economist Jason Furman — whom Altraide references in the video — has raised a point that applies regardless of one's view on the bubble debate. When a substantial share of GDP growth is concentrated in data center construction and software investment driven by a small cluster of companies, the headline number becomes a poor proxy for broad economic health. Altraide quotes Furman noting that the two categories of information processing and software accounted for an outsized share of demand growth in a recent period, though the precise figure he cites comes from the video and should be verified against Furman's published work before treating it as definitive.

The structural point stands independently: an economy where GDP is being moved by a handful of companies spending heavily on infrastructure whose return on investment remains unproven is not necessarily an economy experiencing the kind of distributed prosperity that the headline number implies. It may be. But the measurement itself cannot tell you.

The question for the next two to three years is whether the subsidy era ending forces a reckoning with those returns before the infrastructure buildout has paid off — or whether the revenue eventually materializes at a scale that makes the free cash flow trough look, in retrospect, like a smart bet. The honest answer is that nobody knows. What we do know is that the companies making the bet have been very selective about how they account for it.

---

Marcus Tate is the Sports Desk Editor at Buzzrag, covering the business of professional and collegiate athletics.