1. NVIDIA GTC 2026 ① Vera Rubin Enters Full Production
• Core Source
“Vera Rubin provides the foundation for next-generation AI factories at POD scale, delivering 10x higher agent throughput compared to the previous-generation NVIDIA Grace Blackwell platform.”
“Including 150 companies in Taiwan, hundreds of NVIDIA supply chain ecosystem partners are producing Vera Rubin across more than 350 factories in 30 countries worldwide.”
“Agentic AI is a new type of workload. A single prompt can begin a 1,000-step journey spanning reasoning, search, tool use, and response generation.”
“Vera Rubin is scheduled to ship production volumes starting this fall.”
• Expected Impact
The announcement that drew the most market attention at GTC 2026 was the confirmation of Vera Rubin entering full production. Concerns about potential delays in Vera Rubin shipments had been consistently raised among investors, and this announcement officially resolved that risk.
Vera Rubin is the third-generation NVIDIA MGX rack-scale system, structured with five purpose-built racks operating as a single massive AI supercomputer. It integrates the Vera Rubin NVL72 GPU system, Vera CPU, BlueField-4 STX storage, and Spectrum-6 networking into a single end-to-end system, delivering 10x improvement in agent throughput over the previous-generation Grace Blackwell.
The scale of the supply chain was also confirmed. Production is underway across more than 350 factories in 30 countries, including 150 companies in Taiwan, with participation from all major server manufacturers including Dell, HPE, Lenovo, Supermicro, Foxconn, Quanta, and Wiwynn. Shipments are set to begin this fall.
The core logic Jensen Huang emphasized was the structural shift in agentic AI era workloads. While conventional AI operated on simple query-response structures, agentic AI executes a chain of 1,000 sequential steps spanning reasoning, search, tool use, and response generation from a single prompt. This demands far greater computing density and continuous operation compared to existing infrastructure, suggesting that the demand base for Vera Rubin is expanding structurally. The introduction of Spectrum-X Ethernet Photonics (CPO-based) also delivers 5x better power efficiency, 5x longer AI uptime, and 1.3x faster deployment compared to conventional transceiver-based networks.
2. NVIDIA GTC 2026 ② Vera CPU — A Direct Challenge to Intel and AMD
• Core Source
“Vera delivers 1.8x higher performance compared to x86 across common agentic tools such as Python, code analysis, and compilation, enabling data centers to generate more token revenue.”
“It delivers 3x the performance of x86 competitors AMD/Intel on SQL workloads, and 6x on data processing.”
“In the age of AI agents, the CPU has now become the bottleneck for the GPU.”
“CPU supply will be quite substantial. This will be our new major growth driver.”
• Expected Impact
The true protagonist of this GTC was the Vera CPU. Samsung Securities analysis also assessed that “this keynote was centered on the Vera CPU in the agentic AI era.” The competitive significance is clear: NVIDIA is directly entering the data center CPU market with a standalone product, marking a full-scale escalation of competition with Intel and AMD.
The Vera CPU specs are concrete. Built on the NVIDIA custom Olympus core, it features 88 cores and 176 threads, with 1.2 TB/s LPDDR5X ECC memory bandwidth, 40% lower load latency versus x86, and 3.4 TB/s core-to-core bandwidth. TDP ranges from 250W to 450W. On performance, it claims 1.8x advantage over x86 on general agentic workloads, 3x on SQL workloads, and 6x on data processing.
The competitive implications are significant. Historically, NVIDIA sold one CPU for every two GPUs, but in agentic AI environments, the CPU-to-GPU ratio is now approaching nearly 1:1. Standalone CPU racks are also being configured, with 256 CPUs per system. This represents NVIDIA directly challenging Intel Xeon, AMD EPYC, and Amazon Graviton’s dominance in the data center CPU market — a structural threat to Intel and AMD of losing CPU market share to the very company whose GPU customers they serve. The first customers confirmed are Anthropic, OpenAI, and SpaceX, with volume production set for Q3. As JP Morgan projects GPU-to-CPU ratio to shift from 5.4:1 in 2023 to 2.4:1 in 2028, NVIDIA’s Vera CPU is positioned as a direct beneficiary of this CPU demand expansion phase.
3. NVIDIA GTC 2026 ③ RTX Spark — NVIDIA Enters the AI PC Market
• Core Source
“The three-year partnership between Microsoft and NVIDIA is focused on ‘reinventing the PC.'”
“The flagship version features 20 CPU cores, 6,144 GPU cores, and 128GB of LPDDR5X memory.”
“Using 128GB of unified memory, it can run AI agents with up to 120 billion parameters locally.”
“Currently 8 laptop models have been confirmed, with more than 30 laptops and over 10 desktop models in development.”
• Expected Impact
NVIDIA has entered the consumer PC chip market for the first time. RTX Spark is an SoC combining a Blackwell RTX GPU with a 20-core Grace CPU co-designed with MediaTek via NVLink C2C, delivering 1 petaflop of AI compute performance and up to 128GB unified memory (600 GB/s). SK Hynix’s LPDDR5X has been confirmed as the primary memory supplier.
The key differentiator is local AI agent execution. With 128GB unified memory, the device can run LLMs with up to 120 billion parameters locally without cloud dependency, supporting 1 million token context as well. While previous AI PCs relied on cloud API calls, RTX Spark runs large models on-device directly — a fundamentally different paradigm.
The ecosystem scale has also been confirmed. Microsoft Surface Laptop Ultra leads the lineup, with ASUS, Dell, HP, Lenovo, MSI, Acer, and Gigabyte all announcing fall launches. Currently more than 30 laptops and over 10 desktop models are in development. The competitive landscape spans Qualcomm Snapdragon X Elite, Intel, AMD, and Apple’s M-series. Later this year, the DGX Station for Windows powered by the GB300 Grace Blackwell Ultra will also launch, extending NVIDIA’s PC ecosystem from consumer devices to enterprise workstations in a staged rollout.
4. Meta Launches AI Subscription Service — BofA: “1% Conversion Rate Could Add $4.2B in Annual Revenue”
• Core Source
“According to Bloomberg, Meta has launched a paid consumer subscription service for its Meta AI features, introducing two tiers. The basic plan, Meta One Plus, is priced at $7.99 per month, targeting users who generate images and videos or use advanced reasoning features. Meta One Premium is priced at $19.99 per month, offering higher usage limits.”
“Snap’s stronger-than-expected subscription service penetration rate of approximately 5.5% of daily active users suggests Meta could have a meaningful subscription opportunity across its products. Assuming a 1% AI subscription conversion rate — approximately 35 million users — and a monthly ARPU of approximately $10, annual incremental revenue of approximately $4.2 billion could be generated. This represents approximately 1.5% upside to consensus 2027 revenue estimates.”
“The enterprise AI solutions market, including cloud capacity, is expected to exceed $1 trillion by 2028, and even a small share for Meta could be a meaningful contribution.”
• Expected Impact
Meta has officially launched AI subscription services across WhatsApp, Instagram, and Facebook. The basic plan, Meta One Plus, is priced at $7.99 per month, offering image and video generation and advanced reasoning features, while Meta One Premium is priced at $19.99 per month with higher usage limits. The launch began in Singapore, Guatemala, and Bolivia, with broader regional expansion expected within weeks. Meta already operates subscription services at $2.99–$3.99 per month across WhatsApp, Instagram, and Facebook, and this AI subscription layer sits on top of that existing structure.
BofA quantified the monetization potential with specific figures. Using Snap’s subscription penetration rate of approximately 5.5% of daily active users as a reference, and assuming a conservative 1% AI subscription conversion rate of approximately 35 million users with a monthly ARPU of approximately $10, the result is approximately $4.2 billion in annual incremental revenue — representing approximately 1.5% upside to consensus 2027 revenue estimates.
The broader significance of this launch is that Meta has officially formalized a structure for converting AI investment costs directly into revenue. Following Meta’s announcement of dramatically higher operating costs and capex for 2026, its stock had fallen 16% while NASDAQ rose 7% over the same period — a direct reflection of the market’s lack of visibility into a revenue model capable of justifying the AI investment. This subscription launch is the first concrete answer to that question. Meanwhile, Meta’s CEO noted at the annual shareholder meeting that cloud computing market entry remains an option should infrastructure investment lead to excess capacity, citing strong inbound demand from external companies seeking access to Meta’s API and computing resources.
5. Morgan Stanley: “Hyperscaler CapEx Is Not a Cost — It’s a Revenue Accelerator”
• Core Source
“We estimate hyperscalers will add approximately 14GW of new capacity in 2026 and approximately 20GW in 2027.”
“By 2027, AWS is estimated to generate approximately $14 billion in incremental revenue per new GW, and Google Cloud approximately $11 billion.”
“Morgan Stanley estimates AWS revenue growth at 35% in 2026 and 36% in 2027, and Google Cloud growth at 77% in 2026 and 86% in 2027.”
“Industry conversations also indicate that new capacity contracts are being discussed at revenue benchmarks of over $20 billion per GW.”
• Expected Impact
The central debate surrounding AI infrastructure investment is whether hyperscalers’ astronomical CapEx can actually translate into revenue. In its May 27 report, Morgan Stanley answered: “It can — and current forecasts are likely conservative.”
The core framework is “revenue per GW.” Morgan Stanley estimates hyperscalers will add approximately 14GW in 2026 and approximately 20GW in 2027. Of this, capacity allocated to public cloud businesses is approximately 6GW in 2026 and 8GW in 2027. Translating this into revenue, by 2027 AWS generates approximately $14 billion in incremental revenue per new GW and Google Cloud approximately $11 billion. Industry conversations indicate new capacity contracts are being discussed at over $20 billion per GW, suggesting these estimates may be conservative. Unlike neo-cloud operators who primarily offer bare-metal GPU rental, hyperscalers bundle software and platform services to command premium revenue.
The growth rate forecasts are also strong. AWS is expected to reaccelerate from approximately 20% in 2025 to 35% in 2026 and 36% in 2027, while Google Cloud is forecast at 77% in 2026 and 86% in 2027. Morgan Stanley noted that Google Cloud’s figures exclude external TPU sales, meaning the potential for upside revisions is greatest there. Alphabet’s bull-case price target was raised from $425 to $460, and Amazon’s target stands at $330. The conclusion is that 2026–2027 represents the period when CapEx backlogs begin flowing into income statements, marking a transition from a cost-burden phase to a revenue-acceleration phase for hyperscaler investment cycles.
6. JP Morgan: Global Memory Market to Reach $1.7 Trillion by 2028
• Core Source
“We have raised our FY26E–28E memory TAM by 37–53% versus our March 2026 model, and expect the supply deficit to worsen.”
“This year’s DRAM/NAND prices rising 220–250% year-over-year could add further price upside risk next year as well.”
“We see the HBM allocation share of DRAM wafer capacity continuing to rise from 24% in 2026E to 31% in 2028E.”
“JP Morgan estimates total capex over the next three years at $450 billion, raised from $300 billion in the December 2025 model.”
“While memory accounted for the mid-teens percentage of total CSP hardware capex in the early stages of AI, it is expected to exceed half this year.”
• Expected Impact
JP Morgan’s global memory report published on May 29 projects the 2028 memory market TAM at $1.7 trillion and 2027 at $1.3 trillion — a 37–53% across-the-board upward revision from the March model. The core thesis is a structural shift in the composition of memory demand.
As AI demand spreads from GPUs to CPUs broadly, memory consumption is growing faster than anticipated. As the GPU-to-CPU ratio is expected to shift from 5.4:1 in 2023 to 2.4:1 in 2028, this has become a factor driving an additional 20–22% upward revision to FY27E–28E server-grade DDR/LPDDR5 bit demand. AI CPU DRAM demand is projected at 11 million GB in FY27E and 17 million GB in FY28E, representing 19% and 24% of total market demand respectively.
Supply-side structural bottlenecks are also confirmed. With new fab construction requiring at least three years, 60% of incremental DRAM wafers are allocated to HBM, creating a dual squeeze on commodity DRAM supply. The HBM allocation share is projected to rise from 24% in 2026 to 31% in 2028, and total capex over the next three years has been raised from $300 billion to $450 billion — yet EUV procurement and infrastructure build-out remain the key bottlenecks.
On pricing, DRAM/NAND prices have already risen 220–250% year-over-year this year, with further upside risk projected next year. HBM ASP is forecast to rise 10% on a like-for-like basis and 30% on a blended basis year-over-year. The memory share of CSP hardware capex has surged from the mid-teens in the early stages of AI to over 50% this year. JP Morgan concludes that the memory industry is undergoing a structural transition from a cyclical commodity to a strategic asset of AI infrastructure, with the proliferation of LTAs providing institutional support for this revaluation.
Comment [0]