NVIDIA is moving on three fronts at once this year: rolling out its next-generation Rubin AI chips for data centers, pushing AI hardware into ordinary Windows PCs for the first time with RTX Spark, and navigating a fragile deal that allows limited chip sales back into China after more than a year of export restrictions.
What Happened?
NVIDIA CEO Jensen Huang used back to back events this year, including GTC and Taiwan’s Computex conference, to lay out the company’s roadmap across every layer of the AI hardware stack. In the data center, NVIDIA moved its next-generation Vera Rubin platform into full production, with seven new chips designed to power what the company calls agentic AI, the shift from AI models that answer questions to AI systems that carry out multi-step tasks on their own. On the consumer side, NVIDIA unveiled RTX Spark, a new superchip built for Windows laptops and small desktop PCs, marking the company’s first serious push beyond the data center and into personal computing hardware.
At the same time, NVIDIA has spent much of 2026 navigating a shifting export policy that determines which of its chips it can legally sell in China, the company’s largest market before U.S. restrictions cut off most sales in 2025. A revised licensing framework announced by the Trump administration in December 2025 reopened a narrow path for NVIDIA’s H200 chip to reach approved Chinese customers, though the rollout has moved slowly and remains far more limited than NVIDIA’s sales in China before the restrictions began.
Key Details
Vera Rubin: the next generation of AI data center chips
NVIDIA describes Vera Rubin as a generational leap over its current Blackwell platform, built specifically for what the company calls agentic AI, systems designed to reason through multi-step tasks rather than simply generate a single response. Key specifications and claims include:
More Read
- Seven chips now in full production, including the NVIDIA Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet switch, and a newly integrated NVIDIA Groq 3 LPU.
- NVIDIA says the platform delivers up to a 10x reduction in inference token cost and a 4x reduction in the number of GPUs needed to train mixture of experts models, compared with the current Blackwell platform.
- New Spectrum-X Ethernet Photonics switch systems are designed to deliver 5x improved power efficiency and uptime for massive AI data centers.
- A new NVIDIA Inference Context Memory Storage Platform, paired with the BlueField-4 storage processor, is built specifically to accelerate agentic AI reasoning.
- Rubin based products are expected from partners in the second half of 2026, with AWS, Google Cloud, Microsoft, and Oracle Cloud Infrastructure among the first major cloud providers deploying Vera Rubin based instances, alongside NVIDIA Cloud Partners CoreWeave, Lambda, Nebius, and Nscale.
- Microsoft plans to deploy Vera Rubin NVL72 rack-scale systems in its next-generation Fairwater AI superfactories, scaling to hundreds of thousands of Vera Rubin Superchips.
- AI labs including Anthropic, OpenAI, Meta, Mistral AI, and xAI are among the companies NVIDIA says are looking to Rubin to train larger models and serve growing demand for AI reasoning at scale.
RTX Spark: bringing AI chips to ordinary PCs
Unveiled at Computex 2026, RTX Spark represents NVIDIA’s push into the personal computer market for the first time in a serious way, an area historically dominated by Intel, AMD, and more recently Qualcomm and Apple.
- RTX Spark combines a Blackwell RTX GPU with 6,144 CUDA cores and fifth-generation Tensor Cores, connected to a 20-core NVIDIA Grace CPU built in collaboration with MediaTek.
- The chip delivers up to 1 petaflop of AI compute and 128GB of unified memory, aimed at running AI agents directly on a device rather than relying entirely on cloud processing.
- NVIDIA and Microsoft are jointly building new Windows security features and an NVIDIA runtime called OpenShell specifically so AI agents can run locally with strong security and privacy protections.
- Huang has framed the pitch around cost: an agent running locally on an RTX Spark powered PC can operate continuously without the ongoing usage costs of a cloud hosted agent.
- RTX Spark chips are expected to appear first in premium laptops and desktops, with more affordable options following later, and analysts see the launch as a direct challenge to Apple’s MacBook Pro in the premium Windows laptop category.
The China export saga
NVIDIA’s ability to sell advanced AI chips in China has swung back and forth throughout 2025 and 2026 as U.S. policy shifted.
- In April 2025, the U.S. government informed NVIDIA that its H20 chip, specifically designed to comply with earlier restrictions, would now require an export license for China, a change that cost the company a $4.5 billion charge tied to excess inventory.
- On December 8, 2025, President Trump announced the administration would allow NVIDIA’s H200 chip and AMD’s MI325X to be sold to approved customers in China, contingent on a 25 percent cut of related revenue going to the U.S. government.
- On January 13, 2026, the Commerce Department’s Bureau of Industry and Security issued a rule allowing case-by-case review of H200 and similar chip export license applications, requiring applicants to show the sales would not reduce chip supply available to U.S. customers and that Chinese purchasers have adopted export compliance and customer screening procedures.
- By May 2026, the Commerce Department had cleared roughly ten Chinese companies, including Alibaba, Tencent, ByteDance, and JD.com, to purchase H200 chips, though actual shipments have moved slowly and remain far below pre-restriction volumes.
- NVIDIA’s most advanced chips, including the Blackwell and now Rubin platforms, remain off limits for China under the current framework, which is designed to keep NVIDIA’s newest technology exclusive to U.S. and allied customers.
- In June 2026, the Commerce Department issued additional guidance clarifying that its licensing requirements apply to subsidiaries of Chinese companies located outside China, aimed at closing loopholes in the export control regime.
Why This Matters
The Rubin platform matters because it signals where the entire AI industry is heading next: away from single-turn chatbot style responses and toward systems that can plan and execute multi-step tasks with minimal supervision, sometimes called agentic AI. NVIDIA’s claim of a 10x reduction in inference cost, if it holds up in real deployments, would meaningfully lower the cost of running the kind of AI agents that companies like Microsoft and OpenAI are racing to build.
More Read
RTX Spark matters for a different reason. NVIDIA’s dominance has been built almost entirely on data center GPUs, and moving into personal computers puts it in direct competition with Intel, AMD, Qualcomm, and Apple in a market NVIDIA has never seriously contested before. If AI agents increasingly run locally on a laptop instead of in the cloud, that could reshape how much computing power everyday consumers need and how much cloud AI companies can charge for the same capability.
The China situation matters because it captures a broader tension in U.S. technology policy: balancing national security concerns about advanced chips reaching Chinese military or surveillance programs against the economic reality that China was, until recently, one of NVIDIA’s largest markets, and against fears that restricting exports simply accelerates China’s push to build its own domestic AI chip industry through companies like Huawei. Members of Congress from both parties have pushed back on the H200 framework, arguing it risks giving Chinese AI development capability it would not otherwise have as quickly.
Background and Timeline
- August 2022: The U.S. government first imposes export restrictions on advanced AI chips, including NVIDIA’s A100 and H100, to China and Russia.
- January 2025: The outgoing Biden administration introduces the AI Diffusion Rule, creating global performance thresholds and a “green zone” for lower-powered chips like NVIDIA’s China-specific H20.
- April 9, 2025: The U.S. government tells NVIDIA that H20 chips now require an export license for China, effectively halting sales and forcing a $4.5 billion inventory charge.
- 2025: Congress and administration officials debate further restrictions as Beijing responds with its own domestic chip investment push.
- December 8, 2025: President Trump announces a new framework allowing H200 sales to China in exchange for a 25 percent revenue share to the U.S. government.
- January 13, 2026: The Commerce Department’s Bureau of Industry and Security formalizes case-by-case licensing rules for H200 and comparable chips.
- May 2026: The Commerce Department approves roughly ten major Chinese companies to purchase H200 chips; Jensen Huang joins President Trump’s state visit to China seeking further progress on stalled shipments.
- May 31 to June 1, 2026: NVIDIA unveils RTX Spark at Computex 2026 in Taipei, alongside DLSS 4.5 and other GeForce updates.
- June 1, 2026: The Commerce Department issues guidance extending chip export licensing requirements to overseas subsidiaries of Chinese companies.
- Mid-2026: NVIDIA announces the Vera Rubin platform is in full production at GTC, with partner deployments expected in the second half of 2026.
- Second half of 2026: Rubin-based products begin shipping through cloud partners including AWS, Google Cloud, Microsoft, and Oracle Cloud Infrastructure.
What Officials Have Said
NVIDIA CEO Jensen Huang described the RTX Spark launch as reinventing the personal computer, saying, “For forty years, you launched apps. Click. Type. With RTX Spark and Microsoft Windows, you ask, and the PC does the work.”
On the China licensing framework, Under Secretary for Industry and Security Jeffrey Kessler said export controls “should evolve with changes in technology, while protecting national security,” adding that permitting controlled H200 sales to China “will strengthen the American technology ecosystem.”
NVIDIA has described the Rubin platform’s arrival as part of the company’s annual cadence of delivering new AI supercomputer generations, stating that extreme codesign across its new chips represents a step toward the next frontier of AI infrastructure, built to support pretraining, post-training, and real-time agentic inference.
Frequently Asked Questions
What is the difference between NVIDIA’s Blackwell and Rubin platforms? Blackwell is NVIDIA’s current generation data center AI chip platform. Rubin is the next generation, now in full production, which NVIDIA says delivers significantly lower inference costs and requires fewer GPUs to train certain types of AI models, with a particular focus on supporting agentic AI systems that carry out multi-step tasks.
More Read
Can NVIDIA sell its newest chips to China? No. The current export framework allows limited, license approved sales of the older H200 chip to a small number of vetted Chinese companies. NVIDIA’s most advanced chips, including Blackwell and Rubin, remain restricted from export to China.
What is RTX Spark and do I need one? RTX Spark is a new NVIDIA chip built for Windows laptops and compact desktop PCs, designed to run AI agents locally on the device instead of relying entirely on cloud processing. It is aimed at premium computers first, with more affordable options expected later, and is most relevant to users who want to run AI tools without ongoing cloud subscription costs.
Why did the U.S. restrict NVIDIA’s chip sales to China in the first place? The restrictions stem from national security concerns that advanced AI chips could accelerate Chinese military applications or otherwise strengthen capabilities the U.S. government considers a strategic risk, a policy that has evolved significantly since the first restrictions in 2022.
When will Rubin-based cloud services be available? NVIDIA and its partners expect Rubin-based products to become available starting in the second half of 2026, with major cloud providers including AWS, Google Cloud, Microsoft, and Oracle Cloud Infrastructure among the first to deploy Vera Rubin based instances.
Conclusion
NVIDIA’s chip news this year reflects a company trying to extend its dominance in three directions simultaneously: staying ahead in data center AI hardware with Rubin, opening an entirely new market in personal AI computing with RTX Spark, and managing a politically fraught relationship with its former largest export market in China. How each of these plays out, particularly whether Rubin’s cost claims hold up at scale and whether the H200 China framework survives further political pressure, will shape not just NVIDIA’s business but the pace of AI development globally for the next several years.
