Nvidia’s astounding recent growth leveled off in the fourth quarter of its 2025 fiscal year, the 12 months to January 26, but the GPU titan is still producing enviable numbers.
Q4 2025 revenue was $39.3 billion, $11 billion from Blackwell parts alone, earlier defect be damned, and a 12 percent sequential jump from Q3 2025. Profit for the quarter reached $22 billion, 14 percent up from Q3. That kind of low double-digit growth from one quarter to the next is Nvidia’s new normal.
A year-over-year view produced bigger numbers, as Q4 2025 revenue was 78 percent higher than the same period last year. Profit rose 80 percent.
Full-year revenue for FY 2025 grew 114 percent to $130 billion, less than the 126 percent reported in the previous year.
In FY 2024, profits leapt 581 percent from the $4.36 billion recorded in FY2023 to nearly $30 billion. In FY 2025, profits grew a mere 145 percent to $72.9 billion.
To be clear, Nvidia is still making plenty of dough. It’s just not growing as quickly as it has in past quarters and years. The Silicon Valley goliath forecast a bumper Q1 2026, though: A 9.4 percent quarter-on-quarter revenue increase, or about 65 percent year-over-year to roughly $43 billion.
CFO Colette Kress and CEO Jensen Huang remain optimistic that demand for AI infrastructure will continue as Nvidia ramps production of its Blackwell accelerators.
Looking ahead, Kress expects orders for Blackwell to drive considerable growth as customers look to deploy ever larger, more complex clusters requiring not just the corporation’s AI accelerators, but its high-speed NVLink switch interconnects, and Quantum InfiniBand and Spectrum-X networking kit.
“With Blackwell it will be common for these clusters to start with 100,000 GPUs or more,” Kress predicted on a call earlier today with Wall Street analysts, adding that the chip giant is already working with a customer on a 200,000 GPU cluster using its NVLink and InfiniBand interconnects.
Nvidia is counting on next-gen AI models and a shift from training to widespread inferencing deployments to drive these sales.
“The amount of tokens generated, the amount of inference, compute needed is already 100 times more than the one-shot examples and the one-shot capabilities of large language models in the beginning,” Huang said of models like OpenAI’s o3-mini or DeepSeek’s R1.
He also highlighted the continued need for compute to drive pre- and post-training workloads, such as reinforcement learning — the technique used by DeepSeek to give R1 its “thinking” capabilities, for one.
However, Nvidia also faces ongoing geopolitical pressures due in large part to US export controls on AI technologies and the specter of a hefty semiconductor tariff on foreign-made silicon imported into America. Nvidia has long relied on South Korean and Taiwanese fabs to build its chips. While chipmakers in both nations have or are in the process of bringing US-based plants online, it isn’t clear just how much demand those stateside plants can sat