Nvidia introduced Vera Rubin and RTX Spark in its keynote at GTC Taipei; a day later, in a media Q&A, Jensen Huang tied it all to an “agent computing pattern” that runs from the data center to your next laptop. The consumer chip will ship this fall; pricing hasn’t been announced and the only gaming figures so far are from Nvidia.
The hardware landed at Nvidia’s keynote at GTC Taipei: Vera Rubin, the company’s next-generation multi-rack system, is now in full production, and a new chip called RTX Spark brings the same idea to the Windows PC. But it was at the next day’s media Q&A where Jensen Huang made the broader argument: that computing is shifting to an “agencial” pattern in which AI agents, not people, are the primary users of software, and that every kind of machine (PC, car, data center, robot) must be rebuilt around it. He kept coming back to five growth drivers; Here’s what each means to a buyer in the GCC.
An “agent computing pattern” and the PC as its agent
Huang’s argument is that the old model (application code running within an application within an operating system) is giving way to one in which a large language model (or several) sits within a “harness” that calls tools and runs across a distributed infrastructure. It describes an agent as a combination of model, harness, tools, and runtime, where the model does the reasoning and the harness connects everything. The practical consequence that Nvidia wants you to take is that the PC stops being a tool that you operate and becomes an assistant that operates tools on your behalf.
That’s a vision, not a shipping feature, and it’s worth keeping both things separate. The agent’s PC first is a direction of travel; what’s really shipping this year is silicon.
Vera Rubin reaches volume and squeezes the supply chain
Huang calls Vera Rubin the most ambitious project in the company’s history and emphasizes that it is not a single GPU but a pod-scale system. Its NVL72 configuration unites 36 Vera CPUs and 72 Rubin GPUs via sixth-generation NVLink, with BlueField, NVLink, and Spectrum-X DPUs integrated into a cable-free rack that Nvidia says now assembles in about five minutes, versus two hours for the Grace Blackwell racks it replaces.
For a regional reader, the most immediate signal is the offer. Huang flew to Taiwan to meet with the president of TSMC about Vera Rubin’s production capacity, and the ramp is testing the Taiwanese supply chain. Add to that the confirmed global memory shortage and it’s easy to read: when the world’s most profitable customers are snapping up next-generation wafers and DRAM, consumer GPU and RAM prices tend to tighten rather than decrease. Expect that pressure to show up at retail, in graphics cards and memory kits, before it subsides.
Vera CPU pursues local x86 territory
The most direct competitive move was the Vera CPU, billed as “the agent CPU”: a monolithic 88-core Arm design with high bandwidth per core and LPDDR5X memory, built to eliminate CPU bottlenecks that limit GPU utilization. Huang’s argument is that CPUs were historically built “for humans,” while agents operate in nanoseconds and are impatient with tool calls, and that the orchestration layer of agent AI is CPU-heavy territory that Intel and AMD have owned. Nvidia claimed that SQL workloads run approximately three times faster on Vera. The open question is whether hyperscalers really displace x86 at scale; This is a credible first attempt, not a definitive result.
RTX Spark: the agent PC and the part you will actually buy
This is the consumer face of the strategy. RTX Spark is based on the N1X superchip: a 20-core Arm Grace CPU co-designed with MediaTek and a Blackwell RTX GPU in a single 3nm class TSMC package, joined by Nvidia’s NVLink-C2C coherent interconnect. It carries up to 128 GB of unified LPDDR5X memory shared between CPU and GPU, and Nvidia quotes a maximum bandwidth of around 600 GB/s and 6,144 CUDA cores, the same number as the desktop RTX 5070, the part that Nvidia uses as a performance criterion. The pitch for gamers and creators is compatibility: Huang said Nvidia’s entire software stack and every app Windows has ever run will run on the chip, and Adobe Photoshop and Premiere are planned to run at launch.
Two precautions. First, the gaming claim is only from Nvidia. The company says the platform is good for “100 FPS 1440p gaming,” relying on DLSS upscaling and multi-frame rendering, and at launch, there are no separate benchmarks for RTX Spark shipping laptops. We’ll withhold any verdict on real-world performance until PCMag Labs has a unit. Second, the Arm angle is quite the gamble: Nvidia is entering Windows-on-Arm territory just as Qualcomm’s exclusivity expired, and loopholes in Arm software have sunk this idea before.
In the roadmap, Nvidia framed it based on GPU architecture rather than a clean “N1X → N2X → N3X” sequence: Grace Blackwell (the N1X part) launches in 2026, a Vera Rubin generation with LPDDR6 follows in 2027/28, and a Feynman generation is scheduled for 2030.
RTX Spark devices (more than 30 laptops and 10 desktops from Microsoft, Dell, HP, ASUS, Lenovo, and MSI) are expected this fall. The price is set by the OEM and is not announced, so we don’t have anything concrete yet. For now, the only honest performance framework is the “RTX 5070 class”, so if you’re buying a machine for gaming today, the Gaming Laptops We Recommend in the UAE and Saudi Arabiaalong with our tests of the current RTX 50 seriesthey are still safer benchmarks than an uncompared Arm piece.
Physical AI: cars and robots that reason
The fifth driver extended the idea of the agent to the physical world. Nvidia unveiled new models, including Alpamayo 2 for autonomous driving, along with platform and CPU news, and Huang repeatedly returned to humanoid robots and “reasoning cars” as the next surfaces for agent computing. The report’s “skills archive” concept—robots that carry packaged, transferable skills rather than being reprogrammed by task—fits Huang’s definition of agents as portable workers, although it’s a directional idea rather than a shipping product.
The conclusion
Setting aside the staging, Computex 2026 was a single, coherent bet: that the agent is the new computing unit and that Nvidia intends to sell the silicon at each layer: rack, desktop, car and robot. Half of the data center is real and shipped; Vera Rubin in volume and a serious Arm server CPU are concrete, and the supply tension behind them is the part that buyers will feel first, as upward pressure on GPU and memory prices.
The consumer half is the real bet. RTX Spark is the most credible Windows attempt on Arm yet, backed by a multi-year roadmap and real OEM commitment that previous efforts lacked, but it asks gamers to rely on an Arm GPU on its own numbers, without independent benchmarks or pricing. Our thinking is to look at the data center and wait for a tested RTX Spark unit before believing the 1440p story. The vision is the most ambitious that Nvidia has presented; Whether your next PC will actually work with it is a question for this fall’s review queue.
