Nvidia’s Vera Rubin represents a big bet on real-time AI reasoning

Welcome to AI Decoded, Fast Company’s weekly newsletter that breaks down the most important news in the world of AI. You can sign up to receive this newsletter every week here. Nvidia’s Jensen Huang: The world got DeepSeek totally wrong At Nvidia’s GTC developer event in San Jose this week, CEO Jensen Huang called it AI’s “Super Bowl.” Indeed, Nvidia is the chip supplier of the AI boom, and arguably the most influential (and certainly most profitable) company in the growing AI industry. As it does every year, the company used Huang’s keynote to announce its next-generation chips, which will power the training and operation of the AI models of the near future. Huang introduced a new platform called “Vera Rubin,” which includes Nvidia’s first custom-designed CPU, “Vera,” and two new “Rubin” GPUs. (The platform is named after astronomer Vera Rubin, who discovered evidence of dark matter.) Nvidia says Vera is twice as fast as the Arm-based CPU in last year’s Blackwell platform. Together, Vera and Rubin deliver more than double the inference performance of Blackwell. Huang also revealed Blackwell Ultra, a more powerful version of the current flagship GPU, with increased compute and memory. It’s expected to launch in the second half of 2025. These platforms—especially Vera Rubin—are designed to boost performance during inference, the real-time reasoning AI models perform to generate answers. Huang stressed the growing need for chips that can rapidly process and store massive amounts of data during inference. In a pointed comment, he said “almost the entire world got it wrong” about the DeepSeek phenomenon, referencing the Chinese startup that claimed it could train top-tier models with fewer, weaker GPUs. “The amount of computation we need as a result of agentic AI, as a result of reasoning, is easily 100 times more than we thought we needed this time last year,” he said. In other words, inference demand is set to soar—and with it, demand for Nvidia hardware. Huang also updated Project Digits, a desktop AI platform for developers, researchers, and students. It will ship in two versions: a compact GTX Spark, and a larger GTX Station capable of running more complex models. Nvidia says both will reach the market this summer. As Nvidia’s role in generative AI deepens, it’s pushing beyond infrastructure and into the application layer. To that end, the company announced a new family of open-source, mid-size reasoning models called Nemotron. Built on Meta’s Llama, these models can run on the Spark or Station hardware or in the cloud. While not the most advanced, they’re optimized for lighter enterprise tasks and run efficiently on Nvidia chips—offering yet another incentive for companies to stay within the Nvidia ecosystem. Big AI gives the Trump administration pointers on its ‘AI Action Plan’ As the U.S. government rethinks its approach to artificial intelligence, the emerging consensus around this second Trump administration is clear: minimal oversight, maximum freedom for industry. The working theory is that Trump and Vice President JD Vance would take a hands-off approach to AI regulation, favoring rapid innovation over safety mandates. Both Trump and Vance have emphasized the need for U.S. AI companies to operate without transparency mandates or safety guidelines, arguing that this freedom is essential for global leadership. “The AI future is not going to be won by hand-wringing about safety,” Vance said during a February speech at the Artificial Intelligence Action Summit in Paris. On day one, the administration rescinded the Biden-era AI safety and transparency guidelines, which were largely voluntary. The Trump team is reportedly developing its own AI policy. As part of that effort, the Office of Science and Technology Policy (OSTP) solicited public input on an “AI Action Plan.” The plan, according to the OSTP request, aims to define “priority policy actions needed to sustain and enhance America’s AI dominance” while avoiding “unnecessarily burdensome requirements” on private-sector innovation. Comments were accepted through March 15, and several major AI companies—OpenAI, Anthropic, Meta, and Google—posted their submissions publicly. Across the board, these companies warned that overly aggressive safety rules could cause the U.S. to fall behind China in the AI race. Most supported maintaining export bans on advanced AI chips to adversarial nations like China. OpenAI, Anthropic, and Google also called for protecting the Copyright Act’s “fair use” provision, which allows AI labs to train models on publicly available data. Still, some comments signaled a more nuanced view of risk. Anthropic urged the government to create mechanisms for assessing whether private-sector models could pose national security threats, and suggested opening dedicated communication channels between intelligence agencies and major AI labs. Interestingly, OpenAI seems to favor a ban on models from

Mar 20, 2025 - 17:10

Nvidia’s Vera Rubin represents a big bet on real-time AI reasoning

Welcome to AI Decoded, Fast Company’s weekly newsletter that breaks down the most important news in the world of AI. You can sign up to receive this newsletter every week here.

Nvidia’s Jensen Huang: The world got DeepSeek totally wrong

At Nvidia’s GTC developer event in San Jose this week, CEO Jensen Huang called it AI’s “Super Bowl.” Indeed, Nvidia is the chip supplier of the AI boom, and arguably the most influential (and certainly most profitable) company in the growing AI industry. As it does every year, the company used Huang’s keynote to announce its next-generation chips, which will power the training and operation of the AI models of the near future.

Huang introduced a new platform called “Vera Rubin,” which includes Nvidia’s first custom-designed CPU, “Vera,” and two new “Rubin” GPUs. (The platform is named after astronomer Vera Rubin, who discovered evidence of dark matter.) Nvidia says Vera is twice as fast as the Arm-based CPU in last year’s Blackwell platform. Together, Vera and Rubin deliver more than double the inference performance of Blackwell.

Huang also revealed Blackwell Ultra, a more powerful version of the current flagship GPU, with increased compute and memory. It’s expected to launch in the second half of 2025.

These platforms—especially Vera Rubin—are designed to boost performance during inference, the real-time reasoning AI models perform to generate answers. Huang stressed the growing need for chips that can rapidly process and store massive amounts of data during inference. In a pointed comment, he said “almost the entire world got it wrong” about the DeepSeek phenomenon, referencing the Chinese startup that claimed it could train top-tier models with fewer, weaker GPUs. “The amount of computation we need as a result of agentic AI, as a result of reasoning, is easily 100 times more than we thought we needed this time last year,” he said. In other words, inference demand is set to soar—and with it, demand for Nvidia hardware.

Huang also updated Project Digits, a desktop AI platform for developers, researchers, and students. It will ship in two versions: a compact GTX Spark, and a larger GTX Station capable of running more complex models. Nvidia says both will reach the market this summer.

As Nvidia’s role in generative AI deepens, it’s pushing beyond infrastructure and into the application layer. To that end, the company announced a new family of open-source, mid-size reasoning models called Nemotron. Built on Meta’s Llama, these models can run on the Spark or Station hardware or in the cloud. While not the most advanced, they’re optimized for lighter enterprise tasks and run efficiently on Nvidia chips—offering yet another incentive for companies to stay within the Nvidia ecosystem.

Big AI gives the Trump administration pointers on its ‘AI Action Plan’

As the U.S. government rethinks its approach to artificial intelligence, the emerging consensus around this second Trump administration is clear: minimal oversight, maximum freedom for industry. The working theory is that Trump and Vice President JD Vance would take a hands-off approach to AI regulation, favoring rapid innovation over safety mandates.

Both Trump and Vance have emphasized the need for U.S. AI companies to operate without transparency mandates or safety guidelines, arguing that this freedom is essential for global leadership. “The AI future is not going to be won by hand-wringing about safety,” Vance said during a February speech at the Artificial Intelligence Action Summit in Paris. On day one, the administration rescinded the Biden-era AI safety and transparency guidelines, which were largely voluntary.

The Trump team is reportedly developing its own AI policy. As part of that effort, the Office of Science and Technology Policy (OSTP) solicited public input on an “AI Action Plan.” The plan, according to the OSTP request, aims to define “priority policy actions needed to sustain and enhance America’s AI dominance” while avoiding “unnecessarily burdensome requirements” on private-sector innovation. Comments were accepted through March 15, and several major AI companies—OpenAI, Anthropic, Meta, and Google—posted their submissions publicly.

Across the board, these companies warned that overly aggressive safety rules could cause the U.S. to fall behind China in the AI race. Most supported maintaining export bans on advanced AI chips to adversarial nations like China. OpenAI, Anthropic, and Google also called for protecting the Copyright Act’s “fair use” provision, which allows AI labs to train models on publicly available data.

Still, some comments signaled a more nuanced view of risk. Anthropic urged the government to create mechanisms for assessing whether private-sector models could pose national security threats, and suggested opening dedicated communication channels between intelligence agencies and major AI labs.

Interestingly, OpenAI seems to favor a ban on models from Chinese startup DeepSeek, citing national security concerns similar to those that led to the U.S. ban on Huawei equipment. “DeepSeek faces requirements under Chinese law to comply with demands for user data and uses it to train more capable systems for the CCP’s use,” the company said.

While an AI Action Plan may be written, companies aren’t expecting strict safety regulations to follow—especially under the current administration and Congress. Yet as AI systems grow more powerful, so do the risks of unregulated misuse, potentially on a large scale.

And as Wendy Gonzalez, CEO of the AI training data company Sama, points out, a lack of guardrails may actually be bad for business in the long run. “The comparison to vehicle safety standards is a good example,” she tells Fast Company. “Just as we don’t view seatbelts as ‘over-regulation’ that stifles automotive innovation, thoughtful AI guardrails protect stakeholders while enabling progress.”

Superhuman is a case study in the right way to integrate AI into apps

Much of the post-ChatGPT AI boom has focused on models themselves—but increasingly, the spotlight is shifting to how AI is applied and experienced in real-world apps. Some apps are entirely new, made possible only by generative AI. Others, like Superhuman, stand out for how thoughtfully they’ve integrated the technology into existing products.

Superhuman has been around since 2014, and the addition of AI features hasn’t changed Superhuman’s spare and sleek design very much. Its creators have integrated AI features in a purposeful and understated way. Superhuman founder and CEO Rahul Vohra gave me a run-through of some of the newer AI features when we met at the HumanX conference last week.

The AI works like a behind-the-scenes assistant, aware of the context and content of your inbox. For long or complex threads, it precomputes summaries. It can draft replies in your writing style, prompt you to respond to urgent or high-priority messages, and learn contacts’ communication patterns to time your replies for maximum visibility. It also helps organize your inbox by automatically routing less important emails—like marketing, social updates, or cold pitches—into separate folders.

Vohra claims Superhuman can save users up to four hours per week. One internal study by a major consulting firm found it saved partners 3.3 hours weekly, sped up response times by 3.6 hours, and increased email throughput by 60%.

At $30–$40 a month, Superhuman isn’t cheap. But for professionals overwhelmed by email, its AI-powered productivity boost may be worth the price.

More AI coverage from Fast Company:

Want exclusive reporting and trend analysis on technology, business innovation, future of work, and design? Sign up for Fast Company Premium.