NVIDIA News

As statements of intent at different levels of firmness. Yonhap, citing presidential policy chief Kim Yong-beom, reported that participants at the July 24 San Francisco summit agreed to pursue cooperation projects worth a combined $950 billion, a figure that combines separate supply, development and infrastructure plans rather than one purchase. Inside it sit a long-term arrangement for SK Group to supply $750 billion in high-performance semiconductors to global technology companies including Nvidia, and a Samsung memorandum of understanding with Broadcom that Samsung says the companies estimate at more than $200 billion over five years through 2030. The $500-billion-plus SK Group and NVIDIA relationship rests on letters of intent signed July 24 and carries no published project-by-project capital schedule, while the $10 billion NAVER expansion is a financing plan in which Brookfield's up-to-$9 billion sits on a nonbinding term sheet and NVIDIA's $1 billion is subject to customary closing conditions and to NAVER securing at least $9 billion of committed financing separately. None of these figures should be treated as booked revenue or as capacity you can schedule against.

NVIDIA said on July 21 that Vera Rubin NVL72 production is ramping across more than 350 factory sites in 30 countries, supported by more than 300 global partners, with racks running at CoreWeave, Google Cloud, Microsoft Azure and Oracle Cloud Infrastructure. CoreWeave says it was the first cloud provider to bring up and validate the platform, which it describes as 72 Rubin GPUs, 36 Vera CPUs, ConnectX-9 SuperNICs, BlueField-4 DPUs and NVLink 6. Wccftech, attributing the statement through The Information, reported that hardware engineering senior vice president Andrew Bell said about a dozen manufacturing partners could produce up to 1,000 Vera racks per day, which describes potential partner capacity rather than evidence that 1,000 completed racks leave factories each day. NVIDIA has not published daily completed-rack output, order volume, delivery schedules or generally available customer capacity.

It is reported, not confirmed by NVIDIA. TrendForce reported on June 10 that NVIDIA had decided to halve the SOCAMM memory capacity of Vera Rubin Superchip modules, describing the change as a response to limited LPDRAM allocation under suppliers' preliminary 2027 production plans rather than a reduction in NVIDIA's overall memory demand. On July 26, Wccftech cited a GF Securities analysis putting the revised SOCAMM module at 96GB instead of 192GB and projecting about 28TB of Vera CPU memory per NVL72 rack rather than the 54TB to 55TB previously expected. The reported cut affects CPU-side LPDDR5X; the same analysis leaves the rack's 20.7TB of GPU-side HBM4 unchanged. NVIDIA's Vera Rubin product page still lists 1.5TB of LPDDR5X per Vera CPU and 54TB across an NVL72 rack, with the specifications labeled preliminary and subject to change.

It is arguing against broad restrictions while acknowledging the tradeoff. NVIDIA was among the 25 signatories of Open Weights and American AI Leadership, published July 24 alongside Microsoft, Meta, IBM, Mozilla, Hugging Face, Mistral, Palantir and the Linux Foundation, and Jensen Huang amplified the document in his first post on X. The letter asks U.S. policymakers to expand compute access for startups and researchers, invest in shared training assets such as datasets and evaluation tools, avoid broad restrictions on open models, and distinguish common distillation practices from unlawful extraction of value from closed systems, while explicitly stating that weights become difficult to trace or reverse after release. On July 27 NVIDIA also launched the Open Secure AI Alliance with more than 30 inaugural partners, contributing open models, model weights, data and agent-harness research plus the NOOA research framework; Axios reported that the alliance is also a policy intervention aimed at getting regulators to treat open models and security tooling as defensive assets. Both are industry advocacy rather than law or an independently established safety result.

CNBC reported on July 6 that research firm SemiAnalysis says NVIDIA's Kyber NVL144 rack architecture has slipped to 2028 because the specialized PCB midplane remains difficult to manufacture. CNBC also reported SemiAnalysis' view that NVL576, the larger configuration intended to connect eight racks through optical links, is likely delayed or limited to small volumes, and a SemiAnalysis post summarized by secondary outlets said the NVL72x2 back-to-back rack architecture was canceled. CNBC said NVIDIA did not respond to a request for comment, so these remain analyst-reported rather than official guidance. For procurement teams the practical response is to track OEM notices and updated roadmap statements rather than replan against the reported dates.

A senior U.S. Commerce official, under secretary for industry and security Jeffrey Kessler, told Congress that a small number of Nvidia H200 accelerators have been shipped to China or Hong Kong under export licenses, as reported independently by CNBC and Reuters. The testimony did not identify recipients, disclose a unit count, confirm customer acceptance or show the accelerators operating in production clusters, and Nvidia declined to comment to CNBC. The policy backdrop is a January 2026 Bureau of Industry and Security change allowing applications for H200 and comparable accelerators to be reviewed case by case when specified security conditions are met, including compliance procedures, customer screening and independent testing in the United States. On the other side, Reuters-syndicated reporting citing The Information said China is planning to allow top AI companies to buy a limited number of H200 chips, with officials telling Alibaba, ByteDance and DeepSeek they may soon receive permission and the total possibly fewer than 200,000 chips, less than half of what the companies requested earlier this year.