Nvidia announced its latest generation of chips designed for artificial intelligence at its annual GPU Technology Conference (GTC) on Tuesday in San Jose, California.
CEO Jensen Huang introduced "Blackwell Ultra," a new family of AI-focused chips set to ship later this year, and previewed "Vera Rubin," the next generation of GPUs expected to debut in 2026. These launches mark a strategic shift for Nvidia, moving toward an annual release cycle from its previous biennial schedule.
The introduction of these new chips underscores Nvidia's dominant position in the AI market, a sector that has propelled the company's sales more than sixfold since the AI boom sparked by OpenAI's ChatGPT in late 2022. Nvidia’s GPUs are critical in AI model training, and the new hardware aims to sustain the company's leadership by providing enhanced performance and efficiency.
Major cloud companies such as Microsoft, Google, and Amazon, which collectively spend billions building data centers equipped with Nvidia's GPUs, will closely evaluate the performance gains from these new chips.
“This past year has seen global engagement in AI skyrocket," Huang noted during his keynote. "The computational demands and scaling capabilities of AI have accelerated dramatically."
Nvidia also revealed details about the "Vera Rubin" GPU architecture, named after astronomer Vera Rubin, which features two main components: Nvidia's first custom CPU, "Vera," and a new GPU design, "Rubin." Nvidia stated that the custom CPU, built on its proprietary "Olympus" core architecture, delivers twice the speed of the previous "Grace Blackwell" CPUs.
When paired together, the Vera CPU and Rubin GPU can deliver 50 petaflops of AI inference capability, more than double the performance of Nvidia's current-generation Blackwell chips. Rubin GPUs will also support up to 288 GB of high-speed memory, crucial for demanding AI applications.
In a strategic shift, Nvidia explained that the Rubin GPU design consists of two separate GPU chips working in tandem, a change from its previous approach of referring to combined chips as a single GPU. By late 2027, Nvidia plans to introduce "Rubin Next," featuring four combined dies, doubling Rubin's speed and capacity.
Alongside these hardware announcements, Nvidia showcased other significant launches, including new laptops and desktops featuring its chips. Two notable products, DGX Spark and DGX Station, are specialized PCs capable of running large-scale AI models like Llama and DeepSeek.
Additionally, Nvidia introduced updates to its networking technology, enhancing the ability to interconnect thousands of GPUs efficiently, and announced Dynamo, a software tool designed to optimize performance across Nvidia's hardware ecosystem.
Highlighting collaborations, General Motors announced it would utilize Nvidia's services for its next-generation autonomous vehicles. Moreover, Nvidia disclosed plans for its subsequent chip architecture, named after physicist Richard Feynman, scheduled for release in 2028.
The company emphasized that its new Blackwell Ultra chips significantly enhance "tokens per second," allowing cloud providers to offer premium AI services, potentially generating up to 50 times more revenue compared to earlier models.
Addressing recent concerns from Nvidia investors following the introduction of China's DeepSeek R1 model—which operates efficiently with fewer chips—Huang clarified Nvidia's perspective. He indicated that DeepSeek’s success actually validates Nvidia’s strategy, as advanced AI "reasoning" models like DeepSeek increasingly require the kind of high-performance computing power provided by Nvidia's latest chips.
“In recent years, we've witnessed a fundamental breakthrough in AI that we call agentic AI," Huang concluded. "These systems don't just respond; they reason and actively solve complex problems."