Vera Rubin Fully Enters Production and Links Korea, Taiwan, and Hyperscalers to Nvidia’s New Capex Cycle

Nvidia announced on 31 May, during the keynote at GTC Taipei, that the Vera Rubin platform has entered full production. Jensen Huang’s reading of the scale involved was straightforward: "the biggest product launch, probably, in the history of Taiwan". The data that gives size to the number is the industrial footprint: 350 factories in 30 countries are in the production chain, of which 150 are located solely in Taiwan, according to the official release. Compared to Grace Blackwell, the system delivers 10 times more agent throughput at scale. The first wave of shipments begins in the third quarter, with a ramp-up in volume in the fourth quarter and continuation into the first half of 2027.

The reference system is the Vera Rubin NVL72: 72 Rubin GPUs and 36 Vera CPUs in a rack, achieving 3.6 exaflops of inference at NVFP4 precision and 2.5 exaflops of training, with 20.7 TB of HBM4 connected to 1.6 PB per second of bandwidth. The rack comes fully equipped with critical complements: BlueField-4 as DPU and Spectrum-X Ethernet Photonics, also declared in production, to weave the fabric between thousands of racks with co-packaged optics.

The Korean Bill and the Asian Power Structure

Memory is the piece that ties Korea to Nvidia’s roadmap until 2028. Samsung Electronics, SK hynix, and Micron have been named as suppliers of HBM4 for the platform, according to Huang’s confirmation reported by Korea Herald. For SK hynix, the agreement provides multi-year visibility to factories in Cheongju and Wuxi. For Samsung, it confirms that its aggressive bet on HBM4 with 12 Hi I/O compensates for the delay until 2024. The effect is symmetrical in Taiwan: TSMC produces the Rubin compute die in N3P, MediaTek designs the Vera CPU die, and the ASIC backend is split among Foxconn, Quanta Cloud Technology, Wistron, Wiwynn, and Pegatron.

Huang’s trip to Seoul on 5 June, with a confirmed agenda with SK Group, Hyundai Motor, LG, and Naver, according to the Seoul Economic Daily, seals the political aspect of the equation. The alliance has an internal name in Korea, "kkanbu", a word that designates close friendship and which has gained geopolitical weight since the first APEC summit in October.

Who is Already Using Vera Rubin in Engineering Racks

The release lists the customers who have signed contracts. The hyperscalers present are AWS, Google Cloud, Microsoft, and Oracle Cloud Infrastructure. The named Neo Cloud Providers are CoreWeave, Lambda, Nebius, and Nscale. The cited AI labs include Anthropic, OpenAI, xAI, and ByteDance. Microsoft, Dell, and CoreWeave already operate engineering racks, according to supplementary materials from the release. For the CIO of a bank, consultancy, or insurance company that has purchased capacity reservations in Bedrock, Azure OpenAI, or Vertex AI, the window for moving critical inference to Rubin hardware opens in the fourth quarter.

The impact on unit inference cost is the figure that matters. Nvidia had projected, at CES in January, a 10-fold reduction in cost per token in transitioning from Blackwell to Rubin, alongside a potential 5-fold gain in inference performance. A typical chat load with frontier models would currently cost around $1.50 per million tokens in Sonnet or GPT-5 class models; with Rubin, according to the company’s thesis, the equation drops below $0.30 per million. Wedbush, via Matt Bryson, raised the price target for NVDA to $300 from $230. Bernstein maintained outperform status with a target of $315, citing "a trajectory of $1 trillion in revenue from Blackwell and Rubin by CY2027" and more than $20 billion in revenue solely from Vera CPU in 2026.

The Outlook for Brazil, Japan, and Europe

The Vera Rubin reinforces the move that SoftBank has just made in France. The Son Group announced €75 billion in AI datacentres on 30 May, and the missing piece of equipment is already in existence. Japan enters the equation twice: SoftBank raises European tickets while Mizuho, MUFG, and Nomura analyse currency exposure to the cycle, and funds like Vision Fund 3 seek co-investment with Korea Investment Corporation. For Europe, the thesis only works if SAP, Siemens, and Airbus unlock corporate demand that justifies the consumption of petaflops in the new French GW.

In Brazil, Itaú, Bradesco, and Petrobras depend on the capacity that Microsoft, Google, and AWS intend to allocate in São Paulo. The rule is simple: an engineering rack Rubin operational in the third quarter in the United States means corporate inference load migration between late 2026 and mid-2027, with a narrow window for the CIO to negotiate a multi-year contract before price pressures end.

The American Securities and Exchange Commission and the Brazilian CADE still have work to do with Rubin. The system concentrates more computational power than any previous generation in a fraction of suppliers, and the next decade of the AI market will depend on who signs the capacity contract first.