NVIDIA Corporation (NVDA)

News Summary

• NVIDIA Data Center revenue hit $35.6B in Q4 FY2025, up 93% YoY and 16% QoQ, beating consensus by a meaningful margin, with Blackwell alone generating $11B in its first major shipping quarter—described as the fastest product ramp in NVIDIA's history, with 100,000+ GPU clusters already deployed at CoreWeave, Microsoft, and OpenAI.
• Jensen Huang articulated three concurrent AI scaling laws—pretraining, post-training via reinforcement learning, and inference-time compute—arguing reasoning models already consume 100x more compute per query than one-shot models, with future demand potentially requiring millions of times more compute, directly sustaining the long-duration accelerator demand thesis.
• On custom ASICs, Jensen offered a four-part rebuttal citing generality across model architectures, full-stack software coverage, performance-per-watt economics tied directly to token revenue, and the compounding difficulty of replicating NVIDIA's software ecosystem, concluding that chip design alone does not guarantee deployment.
• Gross margins are guiding to 70.6-71% in Q1 FY2026 and expected to stay in the low seventies through the Blackwell ramp before recovering to mid-seventies later in the fiscal year, with manufacturing expedite costs, system complexity, and tariff exposure cited as the primary headwinds.
• Agentic AI and physical AI were framed as the next major demand vectors beyond hyperscaler buildout, with Jensen describing enterprises as the larger long-term opportunity and outlining a three-compute-system model per industrial company as the structural driver of the next GPU consumption wave.

Financial Highlights

Earnings Call Highlights

NVIDIA's Q4 FY2025 call was a triumphant, forward-leaning performance with Jensen Huang projecting near-messianic confidence about AI's trajectory, while analysts probed durability of demand, gross margin pressure, and competitive threats from custom ASICs. Management's posture was unapologetically bullish, framing every potential headwind—DeepSeek, ASICs, export controls, gross margin compression—as either a tailwind in disguise or a manageable transition cost. The dominant theme was the emergence of multiple simultaneous AI scaling laws (pretraining, post-training, and inference-time scaling), each independently driving massive incremental compute demand, with Blackwell positioned as the singular platform architected for all three. Analysts were largely receptive but pressed hard on gross margin recovery trajectory and geographic concentration risk, reflecting residual skepticism about whether the demand surge is durable or front-loaded.

• Blackwell generated $11 billion in revenue in its first meaningful quarter of shipment, described as 'the fastest product ramp in our company's history,' with Q1 FY2026 guidance of $43 billion implying continued steep acceleration; clusters of 100,000+ GPUs are already shipping to multiple customers including CoreWeave, Microsoft, and OpenAI.
• Jensen articulated three concurrent AI scaling laws—pretraining, post-training via reinforcement learning, and inference-time/test-time scaling—arguing that reasoning models like DeepSeek R1 and OpenAI O3 already consume 100x more compute per query than one-shot models, and future models could require 'hundreds of thousands, millions of times more compute than today,' directly validating the long-duration demand thesis for AI accelerators.
• On custom ASICs, Jensen offered a pointed four-part rebuttal—generality across model architectures, end-to-end stack coverage, superior performance-per-watt translating directly to token revenue for AI factories, and the compounding difficulty of replicating NVIDIA's software ecosystem—concluding that 'just because the chip is designed doesn't mean it gets deployed,' a direct shot at TPU and Trainium narratives.
• Gross margins are guiding to 70.6-71% in Q1 and expected to remain in the 'low seventies' through the Blackwell ramp before recovering to 'mid-seventies late this fiscal year,' with Colette citing manufacturing expedite costs and system complexity as the primary drivers; tariff impact was flagged as a material unknown, adding downside risk to the recovery timeline.
• Agentic AI and physical AI were cited as the next major demand vectors beyond current hyperscaler buildout, with Jensen describing enterprises as 'by far larger long term' than CSPs, pointing to a 'three computer' model per industrial company—one for employees, one for AI factory training, and one embedded in physical machinery—as the structural demand case for the next wave of GPU consumption.