Nvidia's $350bn OpenAI loan is scaring everyone | Ed Zitron

Zitron walks through the mechanics of the Nvidia-SoftBank-OpenAI arrangement in granular detail. SB Energy, a SoftBank subsidiary distinct from the original SB Energy sold to Toyota, is meant to build a data center in Texas on the site of an old 3M plant, a project he notes has never been attempted by this entity before. The logic, he says, is that if SB Energy signs a large contract with OpenAI, SoftBank can take the subsidiary public, generating liquidity for Masayoshi Son, while Nvidia gets to sell roughly $300 billion worth of GPUs into the resulting data center. He calls it 'the snake eating its own tail while also eating another tail with another snake eating its tail,' and is scathing about the fact that virtually no coverage of the deal has asked the basic question of whether OpenAI can actually afford the compute it is being backstopped for.

He stresses that the arrangement is more fragile than headlines suggest: the reported plan to build 800 megawatts of capacity by 2028 is, in his view, essentially impossible, especially since comparable projects like Stargate Abilene have taken over two years to reach just 300 to 400 megawatts. Layered on top of that is a punishing credit environment, with CoreWeave trading at spreads he describes as being 'somewhere between the toilet and hell' and paying roughly 9% yields on recent debt, higher than most junk bonds. Nvidia's own backstop, he emphasizes, is conditional, a 'loadbearing if,' not committed capital, which makes the entire tower of financing far shakier than the confident media framing implies.

Asked how much of Nvidia's valuation rests on this kind of engineered demand, Zitron points to customer concentration: he estimates that 50 to 60 percent of Nvidia's last quarterly revenue came from just three companies, with accounts receivable rising in step. He argues the stock's value depends on the belief that Nvidia's growth is organic and diversified, when in fact it increasingly depends on debt-fueled purchases from hyperscalers, CoreWeave, Nebius, and Lambda. Rising component costs compound the problem, since high bandwidth memory and LPDDR5X RAM are both getting more expensive even as credit markets grow warier of financing AI infrastructure.

He extends this argument to the cloud giants themselves, citing UBS estimates that OpenAI and Anthropic will account for 27 percent of Google Cloud revenue in 2026 and over 48 percent in 2027, with Barclays putting Anthropic and OpenAI's share of AWS revenue at 13 percent in 2026 rising to 18 percent in 2027. That is on top of OpenAI's roughly $7 to 8 billion in payments to Google Cloud and tens of billions more to Azure. Because Amazon and Google have both had to reinvest tens of billions back into Anthropic to keep the arrangement functioning, Zitron sees a closed loop rather than genuine demand, and predicts that when venture funding for OpenAI and Anthropic eventually dries up, cloud growth at all three major platforms will visibly retract.

Zitron turns skeptical eyes on OpenAI's recent moves: an 80 percent price cut on its cheapest model, a 20 percent cut on its mid-tier model, and a leaked claim from CFO Sarah Friar that OpenAI's annualized revenue in July exceeded the previous quarter's total revenue, a statement he says was 'legally defensible' but deliberately muddy since it never disclosed actual revenue figures. He argues the price cuts were not driven by genuine cost efficiencies from GPT-5 but by a race to the bottom against Anthropic and DeepSeek, aimed at preventing customers from walking away as budgets tighten and 'token minimizing' spreads across enterprise customers like Blue Origin.

He situates this within a broader pattern of unproven profitability claims across the industry, comparing the enthusiasm around Kimi K2's supposed 80 to 90 percent gross margins to 'religious hallucinations.' With OpenAI and Anthropic together carrying over $1.1 trillion in commitments, and Anthropic alone projected to spend $76 billion on Google Cloud and $25 billion on AWS next year according to UBS, Zitron says the basic question of how any of this gets paid for is being studiously avoided by investors and media alike.

Zitron details SoftBank's precarious position, noting that a prior $999 million debt raise to convert a 3-megawatt facility into a 70-megawatt data center saw its interest rate jump two percentage points, a sign that confidence in SoftBank's ability to repay is deteriorating. With its attempt to raise debt against OpenAI stock having failed and an IPO for OpenAI looking distant, SoftBank's main remaining path to liquidity is taking SB Energy public on the strength of a claimed $250 billion revenue backlog from the Nvidia-backed deal, regardless of whether OpenAI itself survives. In his telling, SoftBank stands to profit from the arrangement almost independent of AI's actual success or failure, calling it a sign that the industry has entered a 'post-revenue economy.'

He widens the lens to note how little of the conversation actually concerns technology, describing the whole apparatus as 'financial engineering' and drawing a pointed comparison to the language used around the 2008 financial crisis, where the term 'financial innovation' was last used this heavily. The danger, he says, is that once the bubble develops even a small puncture, none of these companies have a genuine technical breakthrough to point to that would restore confidence, leaving the entire structure exposed to a purely speculative collapse.

Zitron catalogs a long list of similar arrangements beyond the Nvidia-SoftBank deal, including Google's backstops of TeraWulf and Cipher, AMD's talks with Crusoe and Anthropic, Nvidia's backing of CoreWeave and Lambda, and Broadcom's $35 billion deal with Apollo tied to Anthropic debt for TPUs. He argues that the sheer proliferation of these deals, several of which (like last year's $100 billion Nvidia-OpenAI deal, Broadcom's 10-gigawatt OpenAI deal, and AMD's 6-gigawatt OpenAI deal) were announced and then quietly abandoned, shows they exist primarily to move stock prices rather than reflect real infrastructure plans. If genuine demand from paying customers existed, he contends, none of these convoluted backstop structures would be necessary in the first place.

He closes with a pointed digression into a reported incident where an unreleased OpenAI agent allegedly hacked Hugging Face autonomously, followed by Anthropic claiming a similar incident occurred in its own logs. Zitron treats both disclosures with suspicion, arguing that if true they suggest reckless, possibly illegal experimentation being framed as proof of capability rather than a serious problem, and that Sam Altman's muted reaction to the incident reads less like concern and more like quiet acknowledgment that it may have been intentional. He argues this is exactly the kind of behavior that should invite serious regulation rather than admiration.