Sovereign AI Runs on Nvidia's Credit

The pursuit of sovereign AI was supposed to be about owning your chips. It is turning into something closer to owing Nvidia. Over the past six weeks, three national AI infrastructure projects have taken shape on Nvidia's DSX platform — a program that does not so much sell hardware as front it. Korea committed 2.08 trillion won to a state AI project anchored on Nvidia Vera Rubin GPUs, with SK Telecom and Naver building gigawatt-scale factories on DSX ^[1]. Indonesia's Firmus Technologies is deploying 170,000 Nvidia GPUs through DSX with revenue-sharing running to 2034 ^[2]. Australia's SharonAI raised $1.6 billion to build one of the country's largest AI factories with 40,000 GB300 GPUs under a six-year strategic compute collaboration with Nvidia ^[3]. The common thread is the financing, not the chips. DSX extends GPUs and token credits to customers upfront and collects a share of product and cloud revenue afterward ^[4]. A country that wants sovereign AI no longer needs to buy Nvidia's hardware outright. It operates Nvidia's hardware on Nvidia's credit terms and pays Nvidia from the revenue that its sovereignty generates. The dependency has migrated from the supply chain to the balance sheet. Firmus CEO Brian Rosenfield described the credit-support structure as a way to level the playing field — Nvidia extending financing to players who lack the capital or credit rating to buy upfront ^[2].

This partnership with Nvidia provides AI-Natives with unprecedented access to the most advanced AI accelerators in the world, with the certainty, scale, and flexibility that best fits their high-growth trajectory. — Tim Rosenfield

This is not a small-business lending program dressed up for export. Nvidia raised $25 billion in bonds while sitting on $50 billion in cash, positioning itself to take equity stakes in the very companies it finances ^[5]. It holds a $30 billion stake in OpenAI, which buys Nvidia chips — a circular arrangement Nvidia cut from $100 billion amid antitrust scrutiny but did not abandon ^[6]. CoreWeave shows the loop in miniature: Nvidia holds 47.2 million shares, CoreWeave buys Nvidia GPUs, CoreWeave sells compute to Meta and Anthropic, and revenue flows back to Nvidia through both equity appreciation and continued hardware orders ^[7]. The financier, the supplier, and the shareholder are the same entity. The structure is viable because Nvidia holds more than 90% of the discrete GPU market ^[8]. Revenue-sharing only works as a financing model when the hardware being financed has no comparable alternative. And it is politically durable because it sits below the line where governance currently looks. The EU's Tech Sovereignty Package, unveiled in June, targets "risky dependencies on single dominant suppliers" and restricts foreign cloud providers from sensitive-sector public tenders ^[9]. But a country could kick out Microsoft's cloud and still build its sovereign AI factory on Nvidia's DSX — shifting the dependency from the cloud layer, which the EU regulates, to the financing layer, which it does not. US antitrust investigators have subpoenaed Nvidia's circular investment pattern ^[6], but the probes target equity stakes and exclusionary conduct, not the credit-allocation terms that determine who gets to build sovereign AI and on whose balance sheet. State-level data center moratoriums in Oklahoma, Alabama, and Michigan address zoning and electricity ^[10] — the physical footprint, not the financial one. None of the legislative or regulatory frameworks cited here — the EU package, US antitrust, or state-level moratoriums — has proposed regulating revenue-sharing or compute-pricing structures. The custom-chip movement is the strongest challenge to this arrangement. Amazon's Trainium hit a $50 billion standalone run rate, OpenAI and Broadcom unveiled the Jalapeño inference chip, and Anthropic is exploring 2nm chips with Samsung ^[11][12][13]. These are real incursions on Nvidia's hardware dominance. But they modulate the dependency rather than breaking it. Custom chips handle inference, not training, where Nvidia still sets the pace ^[13]. The custom-silicon projects require their own financiers — Broadcom's expansion ran through Apollo and Blackstone, and Samsung took a strategic stake in Anthropic's $65 billion Series H — so the escape route substitutes one financier for another ^[12][13]. Nvidia invested $2 billion in Marvell Technology to capture value even on the custom-silicon path ^[11]. And Amazon, which is selling Trainium to third parties to serve "growing international demand for sovereign AI services," is simultaneously buying one million Nvidia GPUs for AWS through 2027 ^[14]. The custom-chip movement is on track to erode Nvidia's share of inference workloads: TrendForce projects ASIC shipments growing 44.6% in 2026 versus 16.1% for GPUs ^[11]. But training remains Nvidia's territory, and training is what determines who can build a frontier model in the first place. The most telling signal is that the one company with the hardware credibility to challenge Nvidia is copying the financing model. Google is providing billions in financial guarantees so developers can raise debt for AI clusters — $3.2 billion for Lake Mariner, $7 billion for River Bend, $1.4 billion for Colorado City — explicitly adopting the circular-financing strategy Nvidia pioneered ^[15]. Google Cloud grew 63% year over year past $20 billion with a $460 billion backlog while selling its own TPUs ^[16]. The most credible hardware alternative to Nvidia is being financed using Nvidia's own playbook. The compute-financier model is becoming the industry standard, not Nvidia's idiosyncrasy. A third-party GPU financing market has already started to form independently. Alpha Compute secured a $31.9 million non-recourse loan using Nvidia GPUs as collateral, with its CEO calling GPU-as-collateral financing a preferred instrument for AI-native companies ^[17]. Nvidia's chips now function as a financial asset class. When the hardware is the collateral, the financing, and the means of production, the question is not whether countries will notice the dependency. It is whether there is a layer of the stack left that is not Nvidia's to allocate.