Everything on NVIDIA: GPU architectures and launches, CUDA and software updates, AI platform partnerships, data center wins, and the earnings that keep moving the industry.
Topic brief
Brief updated Aug 6, 2026
NVIDIA is the dominant designer of GPUs and AI accelerators, spanning data center chips such as the Blackwell and now-ramping Vera Rubin architectures, networking systems including the Spectrum Ethernet and NVLink families, software platforms such as CUDA, NIM model hosting and Omniverse simulation, and application-specific stacks for robotics, physical AI, healthcare and edge computing. Led by CEO Jensen Huang, the company sits at the center of the AI hardware supply chain, working with foundries, memory makers, packaging houses and contract manufacturers to turn chip designs into deployable racks that hyperscalers and cloud providers can actually run workloads on.
For data scientists and ML engineers, NVIDIA's roadmap and supply chain health are leading indicators of what compute will be available, at what cost, and on what timeline. A change in a rack system's memory configuration, a new packaging agreement, or a shift in export policy can matter as much to a team's training and inference plans as a new model release. The software layer, from CUDA and the cuFile storage APIs to NIM microservices, Nemotron models, Isaac, Cosmos and Jetson, also shapes which frameworks practitioners build on by default, and increasingly which agent and robotics stacks they inherit.
A third dimension now matters as much as silicon: NVIDIA has become a financier and counterparty to the buildout it supplies. It takes equity stakes in the startups and clouds that consume its hardware and prepays partners to expand packaging capacity, which makes the company both a supplier and an underwriter of demand. Its access to the Chinese market is gated by government decisions on both sides, it now argues a public policy line on open-weight models, and its CEO sits on a multilateral AI governance body alongside heads of state.
Vera Rubin has moved from announcement into industrial ramp, and customers are now committing to it in public. NVIDIA said on July 21 that Vera Rubin NVL72 production is ramping across more than 350 factory sites in 30 countries with racks running at four cloud providers, and Wccftech reported hardware engineering chief Andrew Bell saying about a dozen partners could produce up to 1,000 Vera racks a day, which is stated capacity rather than documented output. CoreWeave says it was the first cloud provider to bring up and validate the NVL72 platform, built from 72 Rubin GPUs, 36 Vera CPUs, ConnectX-9 SuperNICs, BlueField-4 DPUs and NVLink 6. On August 4 Elon Musk said SpaceX had committed to NVIDIA GPUs exclusively, with Wccftech reporting Vera Rubin as the platform for its AI data centers and company targets of 2 gigawatts of compute by year-end and 10 gigawatts by the end of 2027; SpaceX separately said it is working with NVIDIA on the compute payload for its Starmind AI1 satellite, using Rubin GPUs and Vera CPUs. Those capacity numbers are forward-looking targets, not disclosed orders or deployed capacity. The configuration is less settled than the ramp suggests: Wccftech reported on July 26, citing GF Securities, that NVIDIA may halve the CPU-side LPDDR5X memory planned for NVL72 systems, following a supply-driven SOCAMM reduction TrendForce reported in June, while NVIDIA's public specifications still list 54TB of CPU memory per rack and the 20.7TB of GPU-side HBM4 is unchanged in that scenario. Further out, CNBC reported that Kyber NVL144 has been pushed to 2028 because the specialized PCB midplane remains difficult to manufacture, with SemiAnalysis also saying NVL576 could be delayed or limited to small volumes and the NVL72x2 back-to-back architecture canceled; NVIDIA did not comment. For anyone sizing a 2027 or 2028 cluster, host-memory, rack-cost and delivery assumptions should be modeled as ranges: the $21 million Rubin Ultra rack figure and the roughly $1.5 million HBM4e component in it are BofA Global Research and Morgan Stanley estimates, not NVIDIA list prices.
The second shift is how much of NVIDIA's activity is now capital and supply commitment rather than product launch. It signed a $1.5 billion multi-year advanced packaging and test agreement with Amkor on July 23 that includes an NVIDIA prepayment supporting Amkor's U.S. packaging expansion; announced a long-term partnership with Safe Superintelligence on July 27 giving Ilya Sutskever's lab Vera Rubin access and a tenfold increase in compute, with an investment NVIDIA disclosed but did not size and Reuters reported at $5 billion citing a person briefed on the deal; joined NAVER and Brookfield on July 25 in a proposed expansion of the GAK Sejong AI factory from 55 MW to 200 MW by 2028, funded by up to $9 billion from Brookfield and a planned $1 billion NVIDIA investment; and signed letters of intent with SK Group on July 24 covering a $500-billion-plus initiative including an SK Telecom AI factory of up to 2 gigawatts on Vera Rubin DSX with a first phase targeted for 2027 and long-term HBM4 supply and co-development with SK hynix. Seoul valued the broader July 24 summit package at a reported $950 billion, a figure that combines multi-year cooperation agreements rather than completed purchases. Smaller checks are going out too, including a reported 237.9 crore rupee allocation in Sarvam AI's roughly $74 million Series B extension, following Sarvam's plan for a trillion-plus-parameter model on 10,000 Blackwell GPUs. On the software side, NVIDIA said on August 4 that it is open-sourcing the cuFile APIs and the storage stack beneath them, giving developers a contribution path to the technology that lets GPUs initiate storage reads and writes directly. Policy is the last thread: NVIDIA launched the Open Secure AI Alliance with more than 30 partners on July 27 and signed a July 24 letter with 24 other organizations urging U.S. policymakers to preserve access to open-weight models, which Jensen Huang promoted in his first post on X. Its China access is still decided elsewhere; Reuters-syndicated reporting on July 8 said Beijing was weighing limited H200 approvals for Alibaba, ByteDance and DeepSeek, potentially fewer than 200,000 chips. Investors have separately marked the stock down about 16% from its May 14 peak, removing roughly $1 trillion of market value and putting it near 18 times forward earnings, a rotation toward memory and storage names rather than reported evidence that GPU demand has weakened.
On the hardware roadmap, watch whether NVIDIA confirms or contradicts the reported halving of CPU-side LPDDR5X memory in Vera Rubin NVL72 against its published 54TB-per-rack specification, whether OEM notices or an updated roadmap statement confirm the Kyber NVL144 slip to 2028 that NVIDIA has not addressed publicly, whether NVL576 is delayed or limited in volume and whether the NVL72x2 cancellation is confirmed, and whether the reported Rosa CPU choice of TSMC A16 for a 2028 platform is ever acknowledged by either company. On commitments, the checkable questions are whether the SK Group letters of intent convert into definitive agreements and whether SK Telecom's first phase lands in 2027, whether the conditions attached to Brookfield's up-to-$9 billion and NVIDIA's planned $1 billion NAVER investment are satisfied, whether Amkor discloses the prepayment amount and qualification schedule, whether NVIDIA or Safe Superintelligence ever confirms the investment size that Reuters reported at $5 billion, and whether SpaceX's stated targets of 2 gigawatts by year-end and 10 gigawatts by the end of 2027 are ever backed by a disclosed order or deployed capacity. On policy and supply, track whether Beijing settles the H200 volume it would allow Alibaba, ByteDance and DeepSeek, reported as potentially fewer than 200,000 chips, whether the Open Secure AI Alliance produces maintained artifacts and reproducible security evidence rather than announcements, whether the newly open-sourced cuFile effort ships a conformance suite and multi-vendor implementations, and whether U.S. advanced packaging capacity arrives ahead of the 2028 timing Tom's Hardware described.
NVIDIA says NVL72 production is ramping across more than 350 factory sites in 30 countries with racks running at four cloud providers, and CoreWeave says it was the first cloud provider to bring up and validate the platform. But the host-memory configuration is contested: Wccftech reported on July 26, citing GF Securities, that NVIDIA may halve the CPU-side LPDDR5X memory, following a supply-driven SOCAMM reduction TrendForce reported in June, while NVIDIA's published specifications still list 54TB of CPU memory per NVL72 rack. The 20.7TB of GPU-side HBM4 is unchanged in that scenario, so the disputed figure is host memory, not accelerator memory.
CNBC reported that Kyber NVL144 has been pushed to 2028 because its specialized PCB midplane remains difficult to manufacture, citing SemiAnalysis. The same reporting says NVL576 could be delayed or limited to small volumes, and a SemiAnalysis post said the NVL72x2 back-to-back rack architecture was canceled. NVIDIA did not respond to CNBC's request for comment, so treat all of this as analyst-reported rather than vendor guidance, and track OEM notices and updated roadmap statements before rebuilding 2028 capacity plans.
There is no NVIDIA list price in the public record. Wccftech reported analyst models putting a Rubin Ultra rack near $21 million with HBM4e memory alone around $1.5 million, attributed to BofA Global Research and Morgan Stanley, and a separate Bernstein estimate of roughly $9.1 million for a Vera Rubin rack that depends on future HBM4 pricing. Model these as sensitivity cases, not quotes. The practical implication is that memory pricing can dominate rack economics for this generation, which flows straight into cost-per-token assumptions and cloud contract negotiations.
NVIDIA announced at the Future of Memory and Storage conference on August 4 that it is open-sourcing the cuFile APIs and the storage software stack beneath them, the technology that lets GPUs initiate storage reads and writes directly rather than routing through host CPU staging. The effort includes an API, conformance testing, a user-level implementation and supporting Linux work, giving storage vendors and framework maintainers a contribution path. NVIDIA's separate SCADA and Storage-Next work continues. Note that the frequently cited 230 million IOPS figure is a configuration-specific 2025 benchmark, not a new cuFile result.
Partially and slowly. A senior U.S. Commerce official told Congress that a small number of H200 accelerators have shipped to China or Hong Kong under export licenses, without disclosing recipients, volume, deployment status or revenue. Reuters-syndicated reporting on July 8 said Beijing was planning limited H200 access for firms including Alibaba, ByteDance and DeepSeek, with the total possibly fewer than 200,000 chips, less than half of what companies requested earlier in the year. License approval, shipment, customs clearance, delivery, cluster deployment and sustained utilization are separate milestones and should not be treated as interchangeable.
Because in agentic systems, tool calls, retrieval, code execution and sandbox startup sit on the request path and are CPU-bound. NVIDIA's July 7 positioning describes Vera as a data-center CPU with 88 Olympus cores, up to 1.2 TB/s of LPDDR5X memory bandwidth and a monolithic mesh, and reports more than 1.8x sandbox performance versus x86 under loaded agentic workloads, with a Perplexity example completing a repository clone-and-test workflow about 1.5x faster and starting concurrent sandboxes up to 1.9x faster. Those are vendor benchmarks. The durable point for infrastructure planning is that CPU latency, memory bandwidth and scheduler behavior are now inference cost levers alongside GPU selection.
Latest coverage
Showing newest first with full-topic pagination
Some of Nvidia's largest customers have reportedly been told that prices for AI servers containing its chips could rise by more than 15% in many cases as memory costs climb. The reported increases would affect selected Vera Rubin and Grace Blackwell systems scheduled for delivery early next year, potentially raising data-center expansion costs.
NVIDIA made a minority investment in Cloverleaf Infrastructure on August 21 as part of a partnership to develop U.S. sites for AI data centers. Cloverleaf said it will apply NVIDIA's DSX platform to weigh site, power, cooling, computing, and facility choices earlier in design; neither company disclosed the investment's financial terms.
Starcloud announced a $250 million Series A extension on August 21, 2026, at a $2.3 billion post-money valuation, with Manhattan West leading and NVIDIA participating. The round brings the company’s total funding to $450 million and will support manufacturing, engineering work with NVIDIA, and launch capacity for its orbital AI-compute plans.
Reports published August 20 and 21 say Nvidia agreed to pay $6 billion for a non-exclusive license to Poolside’s model-development software. Newcomer, citing a Poolside investor letter, also reported a separate $1 billion investment at a $12 billion pre-money valuation. The reported structure could give Nvidia access to Poolside’s tooling while leaving the startup a separate company.
NVIDIA's AVO agent achieved a 100.00 RHAE score on the public ARC-AGI-3 benchmark on August 21, completing all 183 levels across 25 environments. NVIDIA reports that AVO uses Claude Opus 5 as its primary model and that the full agent system outperformed the model's reported 30% baseline. The result applies to the public set, not a hidden evaluation set.
On Aug. 19, Nvidia discussed joining a funding round valuing data-labeling provider Mercor at a reported $20 billion, according to The Information as reported by PYMNTS. Neither the total round size nor Nvidia's potential commitment was known. PYMNTS also reported that Mercor's Nvidia revenue was growing as the chipmaker developed its Nemotron open-source models.
Groq announced a $350 million Series A on August 17 at a $3.5 billion valuation. Disruptive led the round, with Nvidia's participation planned; Groq says the financing is subject to customary closing conditions. The company expects to expand operating capacity from 54 megawatts to more than 200 megawatts in 2027 as it builds out training and inference infrastructure.
Nvidia announced on August 17 that it will invest $1.5 billion in SB Energy and provide credit support for the first 4.25 IT-gigawatts of the PORTS-Pike campus in Ohio. OpenAI will use the site under a 20-year lease, while Nvidia will be its exclusive AI-compute provider and may take the remaining 3.75 IT-gigawatts. The first capacity is planned from 2028.
Trane Technologies and Eaton introduced a joint reference design on August 17 that coordinates medium-voltage power distribution with cooling for NVIDIA DSX-aligned AI data centers. The companies estimate that the approach can improve combined energy efficiency by up to 15%, reduce installation costs by up to 30%, and cut copper use by up to 80% versus conventional low-voltage designs, while cautioning that the figures are projected benefits rather than deployment results.
LG said on August 14 that it plans to unveil a next-generation bipedal humanoid in Q1 2027 after signing a strategic MOU with NVIDIA. The robot is planned around NVIDIA Jetson Thor, Isaac GR00T, and Halos for Robotics, with actuators, sensors, and batteries from LG affiliates; LG has not disclosed performance benchmarks, production targets, or commercial availability.
Nvidia's US Marketplace now lists the RTX PRO 6000 Blackwell Workstation Edition at $16,000 and out of stock, confirming the price captured by three reports on August 12. Tom's Hardware said the same listing was $13,250 in June, while early 2025 listings were near $8,500; Nvidia has not published an explanation for the increases.
A new proof-of-concept report describes distributed NanoChat pretraining across two NVIDIA DGX Spark systems linked by dedicated 200 Gb/s fiber and administered through Tailscale. The authors report processing about 653 million tokens over four days at roughly 1,890 tokens per second, while explicitly presenting the setup as a reproducibility testbed rather than a matched scaling-efficiency benchmark.
Nvidia released Nemotron 3.5 Lightning, a 30-billion-parameter open mixture-of-experts model, and NeMo Switchyard, an open-source routing library for agentic AI workloads, on August 11. Nvidia describes Lightning as a specialized execution model for high-volume agent tasks, while Switchyard routes requests among models without requiring application rewrites. Nvidia reports up to 4x faster output versus comparable models in its class.
Nvidia signed memoranda of understanding with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to create independent financing platforms for AI infrastructure. The August 10 announcement says the platforms aim to mobilize more than $500 billion in third-party capital over time for Nvidia customers, but the partnerships remain subject to final agreements.
Nvidia agreed on August 8, 2026, to invest up to $3 billion in Blackstone-backed power infrastructure developer Lancium, according to The Information. The reported deal includes an initial $2 billion equity investment for roughly a 20% stake, with another $1 billion tied to milestones including grid hookups. Reuters reported that Lancium's Abilene, Texas, campus is the first operational site of the Stargate data center initiative.
Indosat Ooredoo Hutchison launched Zankore on August 6 with Ooredoo Group, Nokia, and NVIDIA, targeting 1 GW of NVIDIA DSX AI Factory capacity in Indonesia. ANTARA reports that construction is scheduled to start in Batang, Central Java, in the first half of 2027, beginning with roughly 200 MW. The platform is designed to serve Asia-Pacific enterprises, governments, and hyperscalers.
Firmus confirmed a $2 billion equity round on August 7, 2026, valuing the AI infrastructure company at more than $10.5 billion. Coatue and NVIDIA made follow-on investments, while Blackstone-managed funds and Jane Street also participated; the company said the capital will support Project Southgate in Australia and preparations for wider Asia-Pacific expansion.
Ooredoo Group committed an initial $800 million over five years for a 49% stake in Zankore, an Indonesian AI compute platform launched with Indosat Ooredoo Hutchison, Nokia, and NVIDIA on August 6, 2026. The partners target 1 GW of NVIDIA DSX AI Factory capacity, with an initial 200 MW deployment planned for the first half of 2027.
Tom's Hardware reported on August 6 on a Hong Kong-based eBay listing for modified Nvidia RTX 2080 Ti cards with 22GB of VRAM at $499. The listing itself was last updated May 20, so the current event is renewed coverage rather than a new product launch. It offers used blower-style cards from unspecified board partners for local AI workloads.
Nvidia is assembling an AI Safety and Security Engineering team, with job listings for a founding technical leader and engineers focused on evaluation, security research, and platform work. The team is tasked with testing AI agents and building tools to find, validate, and patch software vulnerabilities as Nvidia expands its open-weight model strategy.
Sarvam AI approved a roughly ₹697.8 crore ($74 million) Series B extension at a July 24 shareholder meeting, with Nvidia slated to invest ₹237.9 crore, according to MCA filings reported by Inc42. Reports differ on whether the issuance is proposed or closed; it follows Sarvam's June first close and a separate plan for a trillion-plus-parameter model on 10,000 Nvidia Blackwell GPUs.
SpaceX said on August 4 that it is working with Nvidia to design the compute payload for its Starmind AI1 satellite, with Rubin GPUs and Vera CPUs planned for each spacecraft. On the same day's earnings call, Elon Musk said SpaceX would build its broader AI infrastructure exclusively on Nvidia and projected up to 10 GW of compute by the end of 2027, a forward-looking target rather than deployed orbital capacity.
Elon Musk said on August 4 that SpaceX had committed to using NVIDIA GPUs exclusively, while Wccftech reported that he identified Vera Rubin as the platform for the company's AI data centers. Wccftech also reported company targets of 2 gigawatts of compute by year-end and 10 gigawatts by the end of 2027; those remain forward-looking projections.
NVIDIA announced at the Future of Memory and Storage conference on August 4 that it is open-sourcing its cuFile APIs and the storage software stack beneath them. The move gives developers a contribution path for technology that lets GPUs initiate storage reads and writes directly, while NVIDIA continues its separate SCADA and Storage-Next work.
South Korean companies and U.S. technology firms announced AI semiconductor cooperation projects valued by Seoul at a reported $950 billion during a July 24 San Francisco summit hosted by President Lee Jae Myung. The package combines $750 billion in SK Group supply and infrastructure agreements with a Samsung-Broadcom memorandum of understanding estimated at more than $200 billion over five years, so the headline total represents planned multi-year cooperation rather than a single completed purchase.
Nvidia and Safe Superintelligence Inc. announced a long-term partnership on July 27 that will give Ilya Sutskever's lab access to Vera Rubin systems and increase its compute capacity tenfold. Nvidia disclosed an investment but not its size; Reuters reported a $5 billion equity investment, citing a person briefed on the deal.
Nvidia and more than 30 inaugural partners launched the Open Secure AI Alliance on July 27 to develop and share open tools for securing AI software and agents. Nvidia is contributing models, data and agent-harness research, while the coalition is calling on policymakers to treat open models as defensive assets alongside closed frontier systems.
The Wall Street Journal reported on July 26 that Nvidia was in talks to provide roughly $250 billion in guarantees supporting OpenAI's lease of a proposed 10-gigawatt data center project in southern Ohio. Bloomberg separately described a possible guarantee of as much as $250 billion. The discussions are preliminary, the companies have not confirmed a deal, and the terms could change or collapse.
In a May 30 build log, Oscar Molnar described adapting a used 16GB Tesla V100 SXM2 to run beside an RTX 4080, producing 32GB of combined VRAM for about £200. The mixed-GPU system reportedly ran a 27-billion-parameter model at 32 tokens per second, though it required legacy drivers, an adapter, and custom fan control.
On July 26, Wccftech reported that NVIDIA may halve the CPU-side LPDDR5X memory planned for Vera Rubin NVL72 systems, citing GF Securities. TrendForce had separately reported a supply-driven SOCAMM reduction in June, while NVIDIA's public specifications still list 54TB of CPU memory per rack. The reports concern host memory, not the listed 20.7TB of HBM4 attached to Rubin GPUs.
