Semiconductor Engineering’s review for the week ending September 18, 2026, brings together a wide range of developments reflecting a single trend in the chip industry: growing demand for artificial intelligence is driving companies and governments to redistribute investments among memory, data centers, advanced manufacturing, packaging, and power. But this expansion is not occurring evenly; the high cost of advanced nodes, shortages of domestic infrastructure, and security and supply-chain risks are placing clear limits on it.
Memory Moves to the Center of Competition
Infineon agreed to sell its NOR flash and F-RAM businesses to Winbond for $1.12 billion in cash, with the deal expected to close in the second half of 2027. The review says the move will allow Infineon to focus its resources on its core semiconductor businesses, while for Winbond it represents an expansion of its memory portfolio.
In the United States, SK hynix is holding exploratory discussions with Intel about manufacturing memory chips, possibly by leasing part of Intel’s Ohio facility or establishing a joint venture with Intel and cloud service providers. However, SK hynix confirmed that there are no final plans or agreements. In another development, Micron introduced a 512-gigabyte DDR5 module for servers, based on vertically stacking DRAM dies and connecting them using TSVs.
These developments indicate that memory capacity and the way it is integrated have become part of AI infrastructure design, rather than merely a component separate from the processor. TrendForce also expects NOR flash supplies to remain tight through the end of the year, with high-capacity product prices potentially rising by between 90% and 110%.
AI Expands the Race from the Chip to the System
Huawei unveiled a portfolio of 11 AI chips covering computing, connectivity, and storage. The company says Ascend 960 processors will double computing performance compared with the 950 series, with two versions arriving in 2027 ahead of schedule. Huawei is relying on system-level scaling to compensate for limitations in individual-chip performance.
At the AI Infra Summit, Cornelis announced an open architecture for scaling networks within and between systems, with programmable computing embedded in the interconnect fabric, along with $205 million in funding and collaboration with Qualcomm. Other companies presented solutions for data interfaces, GPU memory connected directly to the fabric, shared memory, 1.6-terabit optical links, and multi-die interconnects for AI and high-performance computing systems.
McKinsey points to growing demand for chips optimized for inference workloads, while Amazon, Google, Meta, and Microsoft are considering designing custom chips in cooperation with semiconductor companies and possibly selling them commercially. Companies such as Axelera AI and Ambarella have also launched processors targeting continuous inference and edge AI.
Advanced Nodes and Manufacturing: Costly and Uneven Growth
CSIS estimated the cost of designing a chip using 2nm technology at approximately $725 million, compared with $48 million at 28nm. This gap places advanced-chip development beyond the reach of most companies and makes design financing and manufacturing capacity two closely linked issues in semiconductor policy.
MediaTek launched the Dimensity 9600 Pro mobile processor, the company’s first 2nm processor, with an upgraded NPU for generative AI applications. In sub-2nm manufacturing, testing, metrology, inspection, and fabrication have become more closely interconnected because of the smaller dimensions and shrinking error margins.
By contrast, Chinese semiconductor equipment companies increased their global share during 2025, according to CSET, to 12.6% in CMP equipment, 10.2% in etching and cleaning equipment, and 9.8% in deposition. Lithography remains a weak point, while Chinese suppliers’ share of assembly, test, and packaging equipment declined.
A Broader Investment Map Beyond the United States
Expansion plans are distributed across North America and Asia. They include Japanese-American discussions about a potential chip factory in the United States, an expanded agreement between GlobalFoundries and Marvell to increase silicon-germanium capacity in Vermont, and Air Products’ $250 million plan for gas-supply infrastructure in Arizona. IBM also completed a $1 billion federal award for quantum-chip manufacturing in Albany, New York, and said that the first quantum chips are currently passing through a 300mm facility, with a separate investment of $1 billion.
In India, Applied Materials announced an investment plan of up to $5 billion through 2035, including a 140-acre chip research complex and significant expansion of the supply chain, development, and workforce. Fujifilm is investing approximately $83 million in a semiconductor materials plant in Gujarat, while Tata Electronics is working with Besi to develop advanced-packaging capabilities in Assam. Nexperia and Tata Electronics also announced a partnership covering manufacturing, assembly, and testing.
Why Does This News Matter?
This picture shows that AI expansion depends on an interconnected ecosystem, not merely on the availability of faster processors. Memory, packaging, interconnects, power, materials, and manufacturing equipment have become potential bottlenecks, while government and private investments are geographically redistributing production capacity.
The European gap is particularly clear; a report by the Global Electronics Association and DECISION Études & Conseil found that companies headquartered in the European Union account for only 6% of the EU data-center chip market, 7% of server manufacturing and assembly, and 8% of cloud infrastructure. Since 2021, European demand for server systems has grown by 36% annually, compared with only 20% growth in local manufacturing.
But the figures do not mean that all announcements have translated into actual production capacity. The discussions between SK hynix and Intel remain exploratory, and some projects have been announced as plans extending through 2035, while revenue forecasts and funding reflect market trends more than they prove product success or factory readiness. Therefore, the practical question is not only the size of the announced investment, but also these projects’ ability to provide large-scale production, secure materials and power, and achieve an economic return amid rising design costs.