Nvidia Is Now Five Companies in One

In any other industry, Nvidia's balance sheet would describe five separate companies. There is the memory supplier: $279 billion in multi-year supply commitments, more than double the $119 billion on the books three months earlier ^[1]. There is the bank: over $750 billion in customer financing arranged with partners like BlackRock and Goldman Sachs ^[2]. There is the investor: a multi-billion-dollar equity portfolio that includes a $3.96 billion stake in the GPU cloud provider CoreWeave, plus positions in Arm, Nebius, and Recursion ^[3]. There is the software-ecosystem owner, whose CUDA libraries now run inside SK Hynix's own chip-design workflows ^[4]. And there is the would-be distributor, pursuing Hugging Face — the open-source model repository exploring a sale at a $13 billion valuation ^[5] — in a deal aimed at keeping the open-source developer community tethered to Nvidia's compute ^[6]. Jensen Huang has a three-word explanation for why a chip company would become all five.

Compute is revenue — Jensen Huang

Nvidia no longer sees itself as selling hardware; it monetizes compute itself, which is why it is willing to finance the very purchases that generate its revenue. The roles reinforce one another. SK Hynix can't design the memory for Nvidia's next chips without Nvidia's software running in its own design tools ^[4]. The financing keeps customers who can't afford to build custom silicon on the platform ^[7]. The equity stakes tie infrastructure companies to the ecosystem ^[3]. The $500 billion deal with the cloud giants is, in the reporting's own words, aimed at countering their efforts to design their own chips ^[8]. Then the hinge. The same tissue that makes escape hard links all five roles to a single demand driver: AI compute spend. If that driver slows, the memory commitments turn overpriced — CFO Colette Kress already expects gross margins to dip into the low 70s as those memory costs land ^[1]. The customer loans default. The equity stakes fall. The chip revenue falls. All at once. Morgan Stanley has warned of "wrong-way risk if AI monetization slows," and Michael Burry has shorted the stock, citing parabolic credit default swap spreads ^[6][2]. This is not a coronation. It is a race. Amazon's Trainium chip business has hit a $20 billion revenue run rate ^[9]. Google has custom-chip deals running through 2031, Meta through 2029 ^[10]. Meta is co-developing AMD's Helios platform even as it remains a major Nvidia customer ^[11]. And the CUDA moat itself is eroding: a startup used AI coding agents to recreate CUDA-like software in ten hours ^[12]. One rival chipmaker's chief argues the shift to inference workloads breaks the lock-in outright.

That's where the CUDA moat from Nvidia gets broken because CUDA is no longer a factor in the inference side. — Marshall Choy

The annual chip cycle — Blackwell to Vera Rubin — forces customers to re-commit or fall behind ^[13]. And the regulatory risk is already real: reporting indicates Nvidia paused some deals under its compute-partnership program amid antitrust concerns ^[14], though the company has denied pausing the take-or-pay partnerships ^[15]. Two clocks are running against each other. Nvidia is verticalizing fast enough to make escape from its platform impossible, while its largest customers are diversifying fast enough to avoid becoming dependents. The balance sheet is what makes this existential rather than merely strategic. If the customers' exit routes mature first, the lock-in holds nothing. If the race stops before either completes, the interconnection turns a downturn into a cascade across all five roles at once.