Chipmaker introduces 400-layer BV-NAND prototype and next-generation zHBM architecture as AI drives demand for faster, denser and more energy-efficient memory

Samsung Electronics on Tuesday unveiled a new generation of artificial intelligence memory technologies, including a prototype flash memory chip with more than 400 layers and a new three-dimensional memory architecture, as the world’s largest memory chipmaker seeks to consolidate its leadership in one of the fastest-growing segments of the semiconductor industry.

The announcements, made at the Future of Memory and Storage (FMS) conference in Santa Clara, California, underscore Samsung’s strategy of staying at the forefront of AI infrastructure by developing memory technologies capable of handling the explosive growth in data generated by advanced AI systems.

As artificial intelligence models become larger and increasingly complex, the need for faster, denser, and more energy-efficient memory has become one of the defining trends in the semiconductor industry. While much attention has focused on AI processors from companies such as Nvidia and AMD, memory has emerged as an equally critical component because modern AI systems must rapidly move and store vast amounts of data during both training and inference.

Samsung’s flagship announcement was its V10 Bonding V-NAND (BV-NAND) prototype, a next-generation NAND flash memory chip built with more than 400 stacked memory layers and a newly developed wafer-bonding architecture designed to increase storage density while improving performance.

The company said the technology delivers approximately 58% higher memory density than its previous V9 NAND generation while improving read speeds, write performance and input/output throughput. Those gains are intended to meet rising demand from AI data centers, cloud providers and enterprise systems that require high-capacity storage capable of processing enormous datasets efficiently.

NAND flash memory stores data even when power is removed and is widely used in smartphones, personal computers, solid-state drives and enterprise storage systems. As generative AI applications expand, demand for higher-capacity NAND has accelerated because AI inference, the process of generating responses to user prompts after a model has been trained, requires rapid access to massive volumes of stored data.

Industry analysts now see inference as the next major growth engine for AI infrastructure. Unlike model training, which occurs periodically, inference workloads are expected to grow continuously as AI assistants, enterprise software and consumer applications handle billions of daily user interactions.

Samsung also unveiled concept designs for what it described as the industry’s first zHBM and zNAND-O architectures, outlining its long-term vision for three-dimensional memory systems that could significantly improve the performance and efficiency of future AI servers. The proposed zHBM architecture represents a departure from today’s conventional high-bandwidth memory (HBM) designs. Current HBM chips are positioned adjacent to AI processors and connected through advanced packaging technologies.

Samsung’s concept instead stacks memory vertically above AI accelerators, shortening the physical distance data must travel between the processor and memory. The company said this design could substantially increase bandwidth while lowering latency and reducing power consumption, addressing key bottlenecks facing next-generation AI computing systems.

The company said its wafer-bonding technology could ultimately enable memory densities more than 10 times greater than conventional HBM5 while tripling energy efficiency and reducing thermal resistance by more than half. Improving thermal performance has become an important objective across the AI semiconductor industry as more powerful processors generate greater heat, making cooling systems a critical factor in data center design and operating costs.

Samsung’s latest announcements come amid intense competition among global memory manufacturers to capitalize on the AI boom.

The company is competing aggressively with SK Hynix and Micron Technology, both of which have benefited from soaring demand for high-bandwidth memory used in AI servers. While SK Hynix has established an early lead in supplying advanced HBM chips to Nvidia, Samsung has accelerated investment in next-generation memory technologies aimed at strengthening its long-term competitive position. The company’s emphasis on wafer-bonding and vertically integrated memory architectures points toward closer integration between processors and memory as AI computing requirements continue to expand.

Investor concerns have recently emerged over the long-term outlook for parts of the memory market, particularly if demand for smartphones in China weakens after a prolonged replacement cycle. Consumer electronics have traditionally accounted for a substantial share of NAND demand, making the sector vulnerable to fluctuations in handset sales.

However, analysts believe AI infrastructure spending is becoming the dominant driver of industry growth, offsetting weakness in more mature end markets.

That view was boosted last week when Samsung disclosed that long-term supply agreements with customers could eventually account for between 60% and 70% of its total memory sales, providing greater revenue visibility and reducing exposure to the semiconductor industry’s historically volatile pricing cycles.

Analysts at Citi said in a research note last month that inventories across both DRAM and NAND supply chains remained well below historical averages, indicating that the market continues to be supported by healthy demand despite concerns about slowing consumer electronics sales.

The combination of low inventories, sustained investment by hyperscale cloud providers and accelerating deployment of generative AI applications has strengthened expectations that memory manufacturers could continue benefiting from favorable supply-demand dynamics over the medium term.

Microsoft, Meta Platforms, Oracle, Amazon and Alphabet have collectively committed more than $1 trillion in future lease payments for data centers and other infrastructure that have not yet entered service, revealing the extraordinary scale of the artificial intelligence investment cycle and the long-term financial commitments underpinning it.

According to a Reuters analysis of company filings, the five technology giants had committed approximately $1.09 trillion in lease payments for facilities that are still under development or not yet operational. After Meta signed an additional $68 billion of data center leases in July, the disclosed pipeline rose to roughly $1.16 trillion.

The figures show that a significant portion of Big Tech’s AI infrastructure expansion has already been contractually committed, even though much of it has yet to appear on balance sheets as lease liabilities under current accounting rules.

The commitments are nearly four times the roughly $285 billion of lease liabilities currently recognized across the five companies’ balance sheets. The discrepancy stems from accounting standards rather than undisclosed obligations.

Companies typically recognize lease liabilities only when a facility becomes available for use. Until then, future lease payments are disclosed in the notes to financial statements rather than recorded as liabilities on the balance sheet. As a result, the commitments are visible to investors but do not yet affect reported leverage ratios, lease liabilities or other debt-related metrics.

The figures therefore provide a clearer picture of the financial obligations technology companies have already undertaken to support future AI capacity.

The unprecedented leasing commitments reflect executives’ confidence that demand for cloud computing and AI services will continue expanding over the coming decade. Most of the facilities are expected to house advanced graphics processing units (GPUs) and networking equipment needed to train and deploy increasingly sophisticated AI models.

If enterprise and consumer demand for AI computing continues to accelerate, the new capacity will support the next phase of cloud revenue growth while helping providers meet rapidly rising computing requirements. However, the commitments also expose companies to substantial long-term financial obligations if AI adoption grows more slowly than anticipated.

Because data center leases typically run for well over a decade, companies could be left paying for excess computing capacity that cannot easily be repurposed or terminated without significant cost.

Microsoft Leads in Total Future Commitments

Among the five companies, Microsoft disclosed the largest pipeline of uncommenced leases. The software giant reported $329.1 billion in future lease commitments compared with $88.52 billion in recognized lease liabilities already on its balance sheet.

The figures underscore Microsoft’s aggressive expansion of Azure cloud infrastructure as it competes to meet surging demand for AI services powered by its partnership with OpenAI and its growing portfolio of enterprise AI products.

Oracle Faces The Greatest Concentration Risk

Oracle reported $260 billion in lease commitments that have not yet commenced, compared with only $37.89 billion in recognized lease liabilities. The company said the commitments primarily relate to new AI data centers expected to enter service between fiscal 2027 and 2029, with lease terms generally extending 15 to 19 years.

Oracle has acknowledged that these long-duration lease agreements create financial risks because customer contracts may not match the timing, pricing or duration of its infrastructure commitments. If customers reduce demand, fail to renew contracts or are unable to meet their obligations, Oracle could remain responsible for substantial lease payments tied to underutilized facilities.

According to Reuters calculations using company filings and LSEG data, Oracle’s borrowings represented approximately 4.4 times trailing EBITDA at the end of May. Including recognized operating and finance lease liabilities increased that leverage ratio to approximately 5.7 times.

However, S&P Global Ratings said it already incorporates Oracle’s uncommenced lease commitments into its credit analysis and expects adjusted leverage to remain around 4.4 times during fiscal 2027.

Meta Accelerates AI Infrastructure Buildout

Meta disclosed $278.99 billion in future operating and finance lease commitments before announcing an additional $68 billion in data center agreements during July. The latest contracts are part of Chief Executive Officer Mark Zuckerberg’s aggressive push to expand AI infrastructure as Meta races to strengthen its large language models, AI assistants and supercomputing capabilities.

The additional agreements significantly increase the company’s long-term infrastructure commitments and reinforce AI as Meta’s largest strategic investment priority.

Amazon and Alphabet Continue Expanding

Alphabet disclosed $85.2 billion in uncommenced lease obligations, underlining continued investment in cloud infrastructure supporting Google Cloud and its AI services.

Amazon reported $137.21 billion in future lease commitments, although its portfolio differs from peers because it includes not only data centers but also warehouses, logistics facilities, aircraft, offices and delivery vehicles supporting its broader e-commerce and cloud businesses.

Consequently, Amazon’s disclosed commitments are not directly comparable with those of companies whose leasing activity is more heavily concentrated in AI infrastructure.

Together, the scale of these commitments demonstrates that the AI infrastructure race has progressed well beyond announcements of capital expenditure plans. Technology companies have already contractually committed more than $1 trillion to facilities that will support AI computing over the next decade or longer.

The disclosures also reveal that traditional balance sheet metrics may understate the long-term financial obligations associated with the AI buildout. While these future lease commitments are fully disclosed in financial statements and considered by many credit rating agencies, they have yet to appear as recognized liabilities under accounting rules.