NVIDIA's New Capital Model Puts 170,000 GPUs Into Play for AI Startups

170,000 NVIDIA GPUs — that's the scale of a single AI factory campus Firmus is building in Batam, Indonesia under NVIDIA's new capital partnership model. Forget waiting months for site selection and power procurement; NVIDIA is restructuring how AI startups rent compute, and it's already signing up partners with real numbers.

Revenue-Sharing Instead of Upfront Leases

NVIDIA's strategy is straightforward: capital partners build DSX AI factories, sell cloud services on NVIDIA's stack, and split the revenue with NVIDIA. This replaces the old model where startups had to commit long-term to unlock financing they couldn't get anyway. For model builders, inference providers, and agent platforms, it means faster access to full-stack accelerated computing without the capital lock-in. Sharon AI is deploying up to 40,000 Grace Blackwell GB300 GPUs; Firmus is scaling its Batam campus to 360 megawatts and up to 170,000 GPUs. These aren't announcements on spec — these are deployments in motion.

Who's the Target Customer?

AI natives like Baseten, Fireworks AI, and Together AI are the early demand signal. They need immediate capacity for training, post-training, fine-tuning, and high-volume agentic inference. Their customers — developers, digital natives, enterprises — want reliable, large-scale compute as usage grows, but they also need commercial flexibility as products move from pilot to production. NVIDIA's model lets these startups scale without betting the company on a multi-year lease.

What This Means for the AI Infrastructure Stack

NVIDIA is effectively becoming a marketplace for GPU time, with capital partners shouldering the physical plant costs and NVIDIA taking a usage-linked earnings stream. This shifts the risk from startups to balance sheets that can handle 360 MW campuses. If the model works, it accelerates the buildout of continuously operating AI factories that manufacture tokens at scale — which is exactly where compute demand is headed as AI moves from training to inference. I expect more capital partners to follow Sharon AI and Firmus, because the economic alignment is cleaner than traditional colocation or hyperscaler resale.

NVIDIA just made it a lot easier for an AI startup to get 40,000 GPUs without a decade of infrastructure baggage.

Source: NVIDIA Unlocks AI Compute at Scale, Inviting Capital Partners to Power the AI Infrastructure Buildout
Domain: blogs.nvidia.com