On October 14, 2026, TrendForce reported from the OCP Global Summit 2026 that Arm said the CPU market is expected to surge from $52 billion in 2026 to $300 billion by 2030 due to agentic AI redefining data center architectures; Arm’s AGI CPU debuted in March 2026.
AP Memory Technology raised $415.5 million via a GDR offering to bolster working capital and foreign-currency raw material procurement. This financial maneuver supports operational stability but does not immediately alter supply or pricing dynamics.
A repost by TrendForce states that Elon Musk suggested China is making rapid progress on RAM, noting China is rapidly advancing in LPDDR5/LPDDR6 and highlighting its mass manufacturing scale, citing lithium batteries and solar as examples. The post raises the question of whether conventional memory could follow a similar path. This is commentary only, with no price, capacity or supply-demand data; the source does not state the cause.
In an SK hynix Newsroom article series [Inside the Research] Part 2 dated October 14, 2026, the company detailed its research paper "Co-Optimizing Cell, Non-Cell, and Page Schemes for Energy Efficient Analog Computing in 3D FeNAND," published in the 2025 IEEE International Electron Devices Meeting (IEDM), authored by Won-Tae Koo and Jihun Kim of SK hynix's Global RTC. The research proposes a 3D FeNAND-based analog compute-in-memory (A-CiM) architecture that merges storage and computation, and by co-optimizing cell, non-cell (interconnect/word line), and page schemes, achieves approximately 10 to 20 times higher throughput and around 4,000 to 5,000 times greater energy efficiency than conventional 3D arrays under the same chip-area constraints; the article notes these figures reflect 3D FeNAND's high-density advantage over conventional 2D nonvolatile memory arrays widely studied as A-CiM candidates. Reported optimization results include: using a low-resistance word line metal improved throughput by up to 16% and energy efficiency by up to 3%; air gaps reduced word-line parasitic capacitance by approximately 60%, improving throughput by 1.2 times and energy efficiency by two times; jointly optimizing word line resistance and capacitance improved throughput by up to 1.4 times and energy efficiency by up to 2.2 times; in-page pipelining (IPP) doubled energy efficiency and improved throughput by nine times versus the baseline; and splitting a conventional 16KB page into four 4KB subpages improved energy efficiency by 1.2 times and throughput by 1.7 times. The team also demonstrated four-level-per-cell operation in the low-current region with stable data retention at the 300mm wafer scale. The article states the research shows NAND's potential to expand beyond a storage medium into a computing medium; the source contains no price, shipment, capacity, or market-condition information.
SEMI expects Americas 300mm fab equipment spending to more than double from US$18.1 billion in 2026 to US$40.6 billion in 2029, driven by U.S. investment and expansions by TSMC, Intel, Micron, and Samsung.
The second-generation Lenovo Legion Go handheld devices have already received a substantial price increase after their initial launch. The AMD Ryzen Z2 Extreme-powered Legion Go 2 was previously increased from $1,349.99 to $1,999.99; now the same Legion Go 2 with 32 GB of RAM and 2 TB of storage, which had previously jumped to $2,849, rose by another $233 (around 8% of the previous price) and is listed at $3,081.99 on the Lenovo US online store, while the 1 TB version can still be found at Best Buy for $1,999.99. The entry-level Legion Go S jumped from $899.99 and $649.99 to $1,579.99 and $989.99 for the Z1 Extreme and Z2 Go models, respectively. The latest price increase comes as the gaming handheld continues to be out of stock at Lenovo, although partner stores like Best Buy still have stock. Lenovo has not yet announced the price hike for the Legion Go; the source does not state the cause (the article's suggestion of a link to the 'RAMpocalypse' affecting RAM and DRAM pricing across PC hardware segments is presented as an assumption, not a confirmed cause).
CanSemi debuted on the ChiNext board on October 13 with an IPO price of CNY 12.01 per share; shares opened at CNY 45, up 274.69%, raising approximately CNY 7.5 billion.
CXMT detailed its 4F² DRAM architecture roadmap, targeting DDR5 RDIMMs by end-2026, while its global DRAM market share rose to 9.5% in 2Q26. The 4F² tech reduces cell area by 33%, placing CXMT roughly two years ahead of SK hynix and Samsung. Bearish for global DRAM prices due to intensified competition and potential supply expansion.
Fabless semiconductor startup Bolt Graphics announced the bring-up of its Zeus test chip and the opening of a Taipei office to manage manufacturing partnerships as it prepares for production of its first fully commercial GPU, which is planned for general availability in 2028 (a plan, not an achieved result). Built on Bolt's rendering technology architecture, the 12 nm TSMC test chip accelerates electromagnetic simulation and path tracing in software and hardware, enabling full-product simulation without physical prototypes. Bolt stated its GPU exceeds performance expectations, validating the chip architecture, and that it has proven complex physics algorithms can be executed on fixed-function hardware, accelerating training of advanced robotics and world models. The source does not state the cause (this is a company progress announcement, not a price or supply change).
Marvell launched the second-generation Structera X 3406/3506 controllers supporting CXL 3.2 and PCIe Gen6, offering up to 9TB DDR4 or 6TB DDR5 per unit. Designed to address AI KV caching and large language model inference bottlenecks, the new chips provide 5x LZ4 compression throughput. Bullish for CXL controller and AI server memory expansion markets.
According to Blocks & Files, published Tuesday 13 Oct 2026, Kioxia has made a direct liquid cooling version of its PCIe gen 5 CM9 datacenter SSD, the CM9-DLC, so customers can use the drives inside a hot rack chassis without thermal throttling. The CM9-DLC products have cold-plate, direct liquid cooling support and also support air-cooling, with a liquid-cooled metal enclosure touching the top or both sides of the device, in the E3.S form factor. Maulik Sompura, Senior Staff Director, Product Management, SSD business unit at Kioxia America, said the rise of AI infrastructure is driving a fundamental shift in data center design toward higher power densities and advanced cooling technologies. Deploying liquid-cooled CM9-DLC drives will involve the system builder matching a host server's cold-plate/manifold system and coolant loop, and undergoing qualification; Kioxia does not provide information such as mounting, thermal contact surface or specific cold-plate compatibility, saying only that the drives "support cold-plate-enabled server designs," and system builders should ensure their server chassis complies with standards like the SFF-TA-1008 rev. 3.0a specification. The liquid-cooled CM9 drive performance numbers are the same as for the air-cooled CM9. The CM9 Series drives are currently sampling to select customers and will be showcased at the 2026 OCP Global Summit, taking place October 12 – 15 in San Jose, Calif. Kioxia gave its CM10 series of SSDs liquid-cooling support earlier this year.
Microsoft plans a multi-gigawatt rollout of its next-generation Azure Maia AI accelerator, with production deployments planned for 2027. The company claims a 96-accelerator configuration will deliver up to seven times the token throughput of its Maia 200 predecessor and provide 40 percent more tokens per megawatt than leading GPU-based systems. Each Maia 300 accelerator will feature eight HBM4 memory stacks delivering up to 18 TB/s of peak memory bandwidth.
According to TrendForce, Samsung and AMD teamed up at OCP 2026 to showcase a CXL-based solution powered by Samsung's CMM-D, enabling shared memory and flexible capacity expansion for AI inference workloads. This is a product showcase/joint demonstration, posted at 12:00 PM on Oct 13, 2026; the source does not state the cause, and no price, production or shipment figures are involved.
CXMT president Cao Kanyu announced that the company's first 4F² DRAM product will be a server DDR5 RDIMM, expected to launch by the end of the year; the announcement was made at the Integrated Chips and Chiplets Conference on Saturday, and the article was published on 13 October 2026. The memory uses vertical transistors to reduce the cell footprint by about one-third on paper, but no speeds or capacities have been announced yet. CXMT expects to sell a memory cell that Samsung and SK hynix have only shown in research papers. The source does not state the cause of the announcement.
Intel proposes an open software blueprint that positions open software as connective tissue linking CPUs, GPUs, data, and business applications, emphasizing upstream contributions to frameworks like PyTorch, SGLang, and vLLM, aiming for Day 0 readiness, describing Xeon as the operational foundation for inference, extending to disaggregated inference, and presenting microservices and enterprise agent toolkits as building blocks; the article notes it is a blueprint, not proof of execution.
VCT technology, which offers even greater density advantages, is expected to gradually enter mass production in 2030 and beyond.
SK hynix introduces a memory-centric AI rack-scale system for 5x faster TTFT, while HBF expands capacity by 14x to alleviate the memory crunch.
China's ChangXin Memory Technologies (CXMT) is raising concerns over the future pricing power of Samsung Electronics, SK Hynix, and Micron as its plans for next-generation server DRAM prompt renewed scrutiny of competition in 2027 and 2028; South Korean industry observers caution, however, that CXMT has yet to demonstrate the production yields and reliability needed to challenge established suppliers at scale.
MangoBoost launched its AI inference platform Mango Inference on Oct. 12, after first unveiling it in July. The company said its pricing is more than 20% lower than competitors, and it plans to build a 1‑megawatt AI data center in Gangnam, Seoul by Q1 next year, expandable to 5MW, with GPUs from AMD and memory from Samsung Electronics and SK hynix. MangoBoost targets over 100 billion won in revenue next year, expecting around 10 billion won this year, and was founded in 2022.
At OCP 2026, per a TrendForce post, Microsoft's Maia 300 packs eight HBM4 stacks with 288GB capacity, while Intel's GPU D270 takes a different route with up to 480GB of LPDDR5X. The source describes these as two very different paths to scaling AI inference. The source does not state the cause or any price impact.
Whose chips are really inside the hottest new products? We open them up and show you, one part at a time.
iPhone DuoFirst-pass BOM: 37 parts, 12 brands and 16 public MPNs; A20 Pro, C2, N1 and the power/RF chains are located, while the NAND brand stays unconfirmed.See the full teardownAll teardowns| Analog | $0.68 | −5.9% |
|---|---|---|
| Connector | $0.75 | +0.2% |
| CPU | $386.74 | −21.3% |
| DDR4 | $88.40 | −0.2% |
| DDR52026-W13 | $14.89 | −11.2% |
| GPU | $478.20 | +8.2% |
| MCU | $0.68 | −61.7%Unusual |
| MLCC | $0.01 | +54.0%Unusual |
| MOSFET | $0.12 | −82.8%Unusual |
| PMIC2026-W13 | $0.44 | −12.5% |
| SSD/HDD | $288.43 | n/a |