Nvidia’s Q3 revenue and Blackwell momentum
Nvidia reported $57.01 billion in total revenue for the third quarter of fiscal 2026, representing a 62.5% increase over the $35.1 billion from the same period last year. This revenue beat the $55.0 billion consensus estimate. Diluted earnings per share reached $1.30, which exceeded the $1.24 consensus estimate by 4.84%. The Data Center segment drove the growth, producing $51.22 billion in revenue, a 66% increase year-over-year. Within this segment, networking revenue surged 162% to $8.2 billion because of NVLink scale-up and Spectrum-X Ethernet adoption. Blackwell Ultra leads all customer categories as the primary architecture. Within this product line, the GB300 model crossed over the GB200, contributing roughly two-thirds of total Blackwell revenue. CEO Jensen Huang stated that Blackwell sales are off the charts and cloud GPUs are sold out. For the fourth quarter, management expects revenue to reach approximately $65.0 billion, implying 54% year-over-year growth. The company expects GAAP and non-GAAP gross margins to be 74.8% and 75.0%, respectively, plus or minus 50 basis points. This growth reflects the high demand for Blackwell and the previously released Hopper architectures. Professional visualization revenue rose 56% to $760 million, driven by DGX Spark and Blackwell sales. Automotive revenue grew 32% to $592 million, while gaming revenue climbed 30% to $4.3 billion across the consumer sector.
Blackwell Ultra performance and specifications
| Feature | NVIDIA Blackwell | AMD Instinct MI325X |
|---|---|---|
| Transistors | 208 billion | Not specified |
| Memory | Not specified | 256GB HBM3E |
| Bandwidth/Link | 10 TB/s chip-to-chip | 8 TB/s memory bandwidth |
The Blackwell architecture uses a custom 4NP TSMC process with two-reticle limit GPU dies. These dies connect using a 10 TB/s chip-to-chip link. The second-generation Transformer Engine uses new micro-tensor scaling support and advanced dynamic range management algorithms integrated into NVIDIA TensorRT-LLM and NeMo Megatron frameworks to support double the compute and model sizes. Fifth-generation NVLink provides 1.8TB/s bidirectional throughput per GPU, allowing communication among 576 GPUs for complex large language models. Blackwell contains a dedicated RAS engine for reliability, availability, and serviceability, while it also provides confidential computing capabilities to protect AI models. The architecture contains a decompression engine to accelerate database queries. In the semi-analysis Inference max benchmark, Blackwell achieved the highest performance and the lowest total cost of ownership across every model and use case. On DeepSeek-R1, Blackwell delivered 10x higher performance per watt and 10x lower cost per token versus the H200. MLPerf training results show Blackwell Ultra delivers 5x faster time to train than Hopper. You should note that while Nvidia dominates the high end, AMD’s MI325X provides a competitive alternative in the cloud. Vultr deployed the AMD Instinct MI325X in its Chicago data center, providing 256GB of HBM3E and 8 TB/s of memory bandwidth.
Hyperscaler spending and the AMD alternative
Hyperscaler capital expenditure for top cloud service providers in 2026 reaches approximately $600 billion, an increase of $200 billion from the start of the year. This spending supports the transition to accelerated computing and generative AI. Nvidia’s networking business, which focuses on AI, generated $8.2 billion in revenue, marking a 162% increase year-over-year because customers adopted NVLink, InfiniBand, and Spectrum-X Ethernet to build their massive data center deployments. Blackwell Ultra wins the inference competition through superior cost-per-token metrics and massive scale. Management sees $500 billion in Blackwell and Rubin revenue visibility through the end of 2026. However, sizable purchase orders failed to materialize this quarter because of geopolitical issues and a competitive market in China. Companies like Anthropic, Mistral, and OpenAI scale compute aggressively to build intelligence, with xAI building the Colossus 2 Gigawatt-scale data center. AWS and HUMAIN expanded their partnership to deploy up to 150,000 AI accelerators, including the GB300 and GB200 systems. While Vultr deployed the AMD Instinct MI325X to offer 256GB of HBM3E and 8 TB/s of memory bandwidth, Nvidia remains the provider of choice for the largest scale deployments. Will AMD’s presence in regional clouds erode Nvidia’s dominance among the largest hyperscalers?
