NVIDIA (NVDA) Deep-Dive: A Chain of Heavy-Asset Bets on the Compute Throne

Thesis verdict: holds, with conditions — compute shortages and the Blackwell ramp are enough to sustain a 75% gross margin, provided downstream model monetization is delivered on schedule; otherwise hundred-billion-scale off-balance-sheet guarantees and purchase commitments turn into fatal impairments.

NVIDIA is, at heart, a fabless full-stack AI infrastructure provider. Its core way of making money is selling data-center accelerated computing chips (GPUs, CPUs, DPUs), networking interconnects (NVLink and InfiniBand) and the full system-level software stack (such as CUDA and AI Enterprise). Its customers fall into three groups: the world's leading hyperscale cloud providers (CSPs) such as Amazon and Microsoft; frontier-model start-ups such as OpenAI and Anthropic; and multinational groups across industries plus sovereign compute investors. The business model is still dominated by one-time outright hardware purchases, while enterprise software subscriptions and cloud-compute revenue-share models are being layered on at an accelerating pace.

By source of revenue and growth, the business structure is overwhelmingly single-polar. Of the $96.221B of total revenue in Q2 FY27, the two segments break down as follows:

• Compute & Networking: quarterly revenue of $88.299B, a full 91.8% of total revenue, up $46.968B year over year, contributing an astonishing 94.9% of the company's total revenue increase. This business is not only the absolute cash generator but also the carrier of the company's huge contractual obligations — supply-chain and capacity purchase commitments have jumped from $119B a quarter earlier to $279B, alongside $36B of cloud-service purchase commitments and $25B of data-center leases not yet commenced. On the news side, Nvidia acquires open-source AI platform Hugging Face for $13B (2026-09-03) further cements the segment's open-source developer barrier.

• Graphics: quarterly revenue of $7.922B, only 8.2% of the total, up $2.510B year over year with just 5.1% of the increment. Though now second in scale, its specific contractual obligation is tied tightly to inventory turnover; the $2.123B of inventory and excess purchase-commitment losses the company provisioned in the first half partly reflect adjustment pressure in consumer and specific inventory.

Viewed through the industry's value stack (the division of labor from bottom-layer hardware to top-layer applications), NVIDIA is far more than a design-and-manufacture vendor selling silicon:

• At L4, orchestration and workflow execution, CUDA (the parallel-computing platform — the underlying software that lets graphics cards compute general-purpose data) has penetrated deep into the daily R&D of 7.5 million developers worldwide and fully controls the compilation standard for model training and inference;

• At L3, proprietary assets and system wiring, NVLink and InfiniBand form high-speed physical interconnects between hyperscale GPU clusters that are extremely hard to replace;

• Even where L2, manufacturing and physical conversion (the chip hardware itself) faces erosion from peers and downstream custom ASICs (application-specific integrated circuits), the absolute stickiness of L4 and L3 has let the company resist the trend of hardware being commoditized and discounted. Once developers are used to a CUDA-based workflow, the all-in cost of rewriting code to migrate to another hardware platform is very high — a deep workflow lock-in.