This issue shows AI infrastructure demand extending further into the future and deeper into the semiconductor supply chain. Nvidia and AWS delivered strong near- and multi-year compute signals; SK hynix, Kioxia and Sandisk committed to long-cycle memory and packaging investment; Marvell reinforced the custom-silicon and networking opportunity; and possible U.S. tariffs added a new policy variable. The central procurement question is no longer only whether AI demand will continue, but which capacity is contractually secured, when new facilities can deliver qualified output, and how trade rules may change landed cost.
01Nvidia posts USD 96.2 billion quarterly revenue as AI infrastructure demand more than doubles year on year
02AWS plans to deploy 2 million additional Nvidia GPUs across its global infrastructure in 2027–2028
03SK hynix breaks ground on a USD 4 billion Indiana HBM4E packaging base and sees memory tightness through 2030
04Kioxia and Sandisk plan more than USD 31 billion of Japan investment as AI expands long-term NAND demand
05Marvell reports record revenue as custom AI silicon and data-center networking continue to accelerate
06U.S. reportedly considers a new semiconductor tariff round tied to domestic manufacturing investment
07Anthropic’s reported USD 45 billion compute lease adds another long-duration demand signal for Vera Rubin systems
08Intel maps agentic AI from rack to edge with Diamond Rapids, Crescent Island and Wildcat Lake
01
Nvidia posts USD 96.2 billion quarterly revenue as AI infrastructure demand more than doubles year on year
NvidiaAI AcceleratorsData Center
Nvidia reported fiscal second-quarter revenue of USD 96.2 billion, up 18% from the previous quarter and 106% from a year earlier. Data Center revenue reached USD 89.0 billion, up 117% year on year.
The result, together with management’s outlook for continued strong growth, indicates that hyperscaler and enterprise AI infrastructure investment remains on an expansion path rather than moving into a broad near-term slowdown.
R&A View:Demand visibility remains strongest around complete AI platforms—not only accelerators. Programs tied to Blackwell, Rubin and alternative AI ASICs should monitor HBM, advanced packaging, high-speed networking, power delivery, cooling, server boards and rack-level integration as one connected supply chain.
02
AWS plans to deploy 2 million additional Nvidia GPUs across its global infrastructure in 2027–2028
AWSGPU DeploymentCloud Infrastructure
AWS and Nvidia announced a major expansion that includes 2 million additional Blackwell Ultra, Rubin and Rubin Ultra GPUs across AWS infrastructure in 2027–2028. The companies also plan to bring Nvidia Vera CPU-based systems, advanced networking and new AI-factory capacity to AWS.
The scale of the commitment extends component-demand visibility beyond the current procurement cycle and signals that cloud operators are locking in future compute, networking, memory, power and data-center capacity earlier.
R&A View:A deployment commitment of this size affects far more than GPU availability. It can pull forward demand for HBM, server CPUs, NICs, switches, optical modules, retimers, connectors, power semiconductors, liquid-cooling hardware and high-layer-count PCBs. Suppliers should evaluate capacity by deployment year and platform generation.
03
SK hynix breaks ground on a USD 4 billion Indiana HBM4E packaging base and sees memory tightness through 2030
HBM4EAdvanced PackagingU.S. Manufacturing
SK hynix began construction of its advanced packaging and R&D facility in West Lafayette, Indiana. The site is scheduled to begin volume production of next-generation HBM4E products in the second half of 2029 and is intended to become a core U.S. AI-memory hub by 2030.
The company also said it expects the global memory shortage to persist through 2030 as AI demand remains structurally strong. The project is supported by U.S. CHIPS Act grants and loans and will work closely with Purdue University and a wider packaging ecosystem.
R&A View:The HBM bottleneck is increasingly a packaging, yield and system-integration issue as well as a wafer-capacity issue. Customers should track HBM4/HBM4E qualification, hybrid bonding, TSV and packaging capacity, substrate supply, thermal requirements and the timing difference between new-site construction and usable commercial output.
04
Kioxia and Sandisk plan more than USD 31 billion of Japan investment as AI expands long-term NAND demand
NAND FlashJapanCapacity Investment
Kioxia and Sandisk plan to invest more than USD 31 billion in Japan through 2032 to advance memory technology and expand manufacturing capacity. The program is expected to include production infrastructure and technology development tied to the partners’ Japanese operations.
The investment reflects growing expectations that AI infrastructure will require not only HBM and DRAM, but also much larger volumes of high-capacity, energy-efficient NAND for enterprise storage, data pipelines and model-serving workloads.
R&A View:NAND investment is long-cycle and does not translate into immediate market relief. Procurement teams should continue to monitor enterprise SSD demand, high-capacity NAND allocation, controller availability, qualification cycles and contract pricing while distinguishing near-term tightness from capacity scheduled for the end of the decade.
05
Marvell reports record revenue as custom AI silicon and data-center networking continue to accelerate
Custom ASICNetworkingMarvell
Marvell reported record fiscal second-quarter revenue of USD 2.739 billion, up 37% year on year. Data-center demand remained the main growth engine, and the company guided the following quarter to approximately USD 3.15 billion in revenue.
Marvell also raised longer-term revenue targets and expects custom-chip revenue to more than double next year, although the larger impact from its expanded Google relationship is expected to build over a multi-year period rather than immediately.
R&A View:Custom AI silicon is becoming a durable second route alongside merchant GPUs. The fastest-growing BOM areas include custom ASIC wafers, HBM and DDR5, advanced packaging, SerDes, Ethernet switching, DSPs, optical interconnects, retimers and high-speed cabling. Revenue timing still depends on design milestones and volume qualification.
06
U.S. reportedly considers a new semiconductor tariff round tied to domestic manufacturing investment
TariffsTrade PolicyElectronics
The U.S. administration is reportedly considering broader tariffs on semiconductors and selected technology products that contain chips, potentially including laptops, gaming consoles and data-center servers. Possible relief may be linked to foreign companies’ U.S. manufacturing commitments.
The proposal remains under discussion and may change before any formal announcement. Its current significance is the wider scope: trade exposure could extend from standalone chips to finished electronics and complete infrastructure systems.
R&A View:Do not price this policy discussion as a confirmed duty. Instead, map country of origin, wafer fabrication, assembly and test locations, finished-product classification, U.S. investment exemptions and Incoterms now. Quotes for long-lead programs should include validity limits and tariff-change clauses.
07
Anthropic’s reported USD 45 billion compute lease adds another long-duration demand signal for Vera Rubin systems
AI ComputeVera RubinData Centers
Anthropic plans to rent AI computing capacity from Nscale in a reported six-year agreement valued at roughly USD 45 billion. The West Virginia campus is expected to use Nvidia Vera Rubin systems to support expanding Claude workloads.
The agreement adds to a growing group of multi-year compute commitments from AI developers and cloud operators, showing how future chip demand is increasingly being secured through infrastructure leases, financing structures and long-term capacity contracts.
R&A View:Large compute leases can improve demand visibility but also increase financing, construction and customer-concentration risk. Supply-chain planning should confirm project funding, delivery milestones, power availability and staged hardware acceptance before treating headline contract value as immediate component demand.
08
Intel maps agentic AI from rack to edge with Diamond Rapids, Crescent Island and Wildcat Lake
IntelAgentic AIEdge Computing
At Hot Chips 2026, Intel outlined three architectures for agentic AI: Diamond Rapids for high-performance orchestration, the Crescent Island GPU for efficient inference, and the Wildcat Lake SoC for client and edge computing.
The roadmap reflects a broader industry shift from treating AI as a single-accelerator problem toward designing complete systems that combine CPUs, GPUs, memory, interconnects, packaging and edge devices for different workloads.
R&A View:Agentic AI is likely to create a more distributed semiconductor demand profile. In addition to data-center accelerators, watch server CPUs, inference GPUs, DDR5, LPDDR, PCIe/CXL connectivity, networking, edge SoCs, power management, sensors and embedded storage. Qualification requirements will vary sharply by deployment environment.
Key Takeaways
The clearest signal this week is that AI demand visibility is extending into 2027–2030. Nvidia’s results, AWS’s two-million-GPU plan and Anthropic’s long-duration compute agreement all support continued infrastructure expansion.
Memory investment is moving in two directions at once: HBM and advanced packaging for high-performance compute, and NAND and enterprise storage for data-intensive AI workloads. Announced investments are large, but much of the output will arrive only near the end of the decade.
Marvell and Intel show AI architectures broadening into custom ASICs, networking, CPUs, inference GPUs and edge systems. Supply planning increasingly needs to follow complete platforms rather than one headline chip category.
The reported U.S. tariff discussion adds uncertainty around country of origin, finished-system classification and landed cost. It is not yet a confirmed policy, but it is already relevant to long-lead quotations and sourcing scenarios.
