All published memory-market news.

Kioxia announced the GP1 Series PCIe 6.0 NVMe SSD for GPU-direct AI workloads. The drive uses second-generation XL-FLASH and is designed to provide up to 10 million random-read IOPS, with evaluation samples planned for selected customers by the end of 2026. The product is aimed at architectures that use fast flash as an extension tier alongside HBM rather than relying only on additional HBM4. For the memory market, this reinforces a broader trend: AI infrastructure buyers are looking for storage-class memory tiers to stretch expensive HBM capacity. The approach could increase demand for high-end enterprise SSDs and specialty NAND while moderating some pressure to add HBM in every configuration.

Semiconductor Engineering reports that package warpage is becoming harder to manage as advanced packages grow in size and complexity. The issue is especially relevant to HBM integration, hybrid bonding, glass carriers, and panel-level packaging because stress can build across multiple materials and process steps. For memory suppliers and packaging partners, this is a yield and reliability concern rather than a simple mechanical specification. More metrology, simulation, and material control may be needed earlier in the process flow. The market impact is indirect but important: if warpage challenges limit packaging yield or slow qualification, advanced HBM supply can remain constrained even when wafer output improves.

StorageNewsletter reports that Samsung and SK hynix signed large U.S. AI chip supply agreements extending through 2030. The same item says HBM capacity from Samsung, SK hynix, and Micron was fully sold out at the time. Even without detailed contract terms in the compact payload, the implication is clear for memory buyers: large customers are trying to secure advanced AI-memory supply years ahead of delivery. Long-term commitments can reduce spot availability, support supplier pricing discipline, and make late entrants compete for future capacity rather than current inventory. For HBM, this reinforces the view that allocation pressure is becoming a strategic planning issue rather than a short procurement cycle.

Sony has secured enough DRAM-related supply to support planned PlayStation 5 and PlayStation 5 Pro production, according to TechPowerUp's summary of the company's earnings discussion. The timing matters because Grand Theft Auto VI is expected to lift console demand around its launch window. Sony is reported to have worked with major DRAM suppliers including Samsung, SK hynix, and Micron to protect availability. The company did not announce an immediate PS5 price increase tied to memory volatility. For the memory market, this shows large consumer-electronics buyers locking in supply ahead of demand spikes. Those commitments can stabilize one customer's production plan while reducing flexible supply available to other buyers.

TechPowerUp reports that AMD has notified add-in-board partners of an August GPU kit price increase of at least 10%. The cited driver is the cost and availability of GDDR6 memory, which AMD sources from suppliers such as SK hynix, Micron, and Samsung before bundling memory with GPU dies for board partners. Cards with more VRAM are expected to feel the largest increase because they require more GDDR6 devices. This is a direct pass-through signal from graphics memory into finished hardware pricing. It also shows that the memory shortage is not limited to HBM or server DRAM; older graphics-memory families can still pressure consumer GPU pricing when allocation tightens.

Samsung has shut down and moved equipment from an older NAND line at its Hwaseong campus, according to DigiTimes, using the site to expand DRAM production. The report says general-purpose DRAM capacity at the location is expected to rise about 15% by the end of 2026. The shift highlights how suppliers are reallocating assets toward DRAM while AI demand keeps memory markets tight. It also points to a tradeoff inside Samsung's own capacity plan: more DRAM output can help address shortages, but converting NAND space may reduce flexibility in flash supply. For buyers, the added DRAM capacity is constructive, though the scale is unlikely to fully offset broader demand from AI servers, PCs, and mobile devices.

Samsung told investors it expects HBM4 sales to rise sharply in the third quarter of 2026. DigiTimes reports that HBM4 is projected to make up more than 60% of Samsung's HBM revenue during the second half of the year. The ramp matters because Samsung has been trying to close the gap with the strongest HBM suppliers during the AI buildout. If Samsung executes, buyers could see another major source of advanced HBM scale, even while overall AI memory demand remains tight. For pricing, the near-term signal is still supportive: a stronger Samsung ramp reflects heavy demand rather than excess supply, and it may take time before additional output materially softens HBM allocation.

A new analysis highlighted by Tom's Hardware argues that memory pricing has moved back toward levels last seen nearly two decades ago when measured per gigabyte. The article attributes the reversal to AI demand, especially the pull from HBM and related DRAM capacity. That matters for RamTrend because it turns a familiar cyclical shortage into a broader historical signal: memory no longer looks like a component following its old deflationary curve. Higher-value AI products are drawing wafer capacity and supplier attention away from conventional modules, lifting costs that reach PC builders and other buyers. The exact duration is uncertain, but the current pricing environment shows that AI infrastructure demand can overwhelm years of normal cost declines in a short period.

A report cited by TechPowerUp says South Korean distributors have been warned of higher GeForce RTX 50-series pricing, with some models potentially rising by about 30%. The cited driver is GDDR7 memory, which Nvidia supplies to add-in-board partners as part of the GPU kit. The article notes that 2GB GDDR7 devices are priced in the low $20 range, meaning a flagship card using 16 packages carries a memory bill of at least $320 before the GPU die and board costs are included. The market signal is straightforward: graphics memory inflation is now large enough to flow through to finished GPU prices, even in a region close to major DRAM suppliers. That keeps GDDR7 availability and pricing important for gaming GPU demand and AIB margins.

Microsoft is reportedly working on Windows 11 changes aimed at making lower-memory PCs feel faster. The context is a market where many builders still view 16GB as the practical baseline, yet shortages and cost pressure have made 8GB laptops more visible again. This is not a direct supply announcement, but it is a useful demand-side signal. If operating systems and PC vendors adapt around smaller RAM configurations, some consumer systems may ship with less memory than buyers would otherwise choose. That could slightly temper unit demand for client DRAM, but it also confirms that memory pricing has become noticeable enough to influence product planning and user-experience tradeoffs.

TrendForce's latest memory outlook points to a split market in 2027. DRAM is expected to stay tight as AI servers absorb HBM capacity and increase demand for both accelerator memory and CPU-attached memory. That would keep upward pressure on DRAM pricing even as suppliers work to expand output. NAND Flash is described differently: new production capacity is expected to arrive while consumer electronics demand remains soft, creating a risk of looser supply in the second half of 2027. For RamTrend, the key signal is that AI is not lifting every memory segment equally. DRAM appears positioned for continued pricing strength, while NAND may need stronger end-device demand to absorb additional supply.

Marvell introduced technologies at FMS 2026 that address memory pressure in large AI inference deployments. The portfolio covers SSD-side infrastructure, CXL-based expansion at the rack level, and optical links for shared memory across larger systems. The focus is the growth of key-value cache requirements as models use longer context windows and produce more memory traffic. For the memory market, Marvell's announcement is less about selling DRAM or NAND directly and more about enabling architectures that can consume those resources more efficiently. If hyperscalers adopt more disaggregated designs, demand could shift toward higher-capacity SSDs, CXL memory modules, and networking components that keep accelerators supplied with data.

A Counterpoint Research update cited by TechPowerUp shows Samsung retaining the largest share of global DRAM revenue in the second quarter of 2026. Samsung is reported at 39%, while SK hynix holds about 26% after strong year-over-year revenue growth but a lower share than a year earlier. The same report says South Korean suppliers together account for 65% of global DRAM revenue. The data points to a market where higher prices for mainstream DRAM and HBM are lifting supplier revenue, while share shifts depend on product mix, execution, and contract strategy. SK hynix is described as leaning into longer customer agreements with preset pricing terms, while Micron is characterized as gaining ground. For buyers, the key signal is that pricing strength is broad enough to expand the revenue pool even when vendor rankings move.

Montage Technology is beginning trial production of a CXL 3.2 controller designed for memory expansion in AI infrastructure. The product targets systems that need to attach and share more memory without tying every capacity increase directly to a new processor or accelerator configuration. For the server market, this points to continued investment in CXL as a way to address memory capacity pressure created by larger AI workloads. The near-term price effect is limited because controller qualification does not immediately create new DRAM supply. Still, broader CXL adoption could change how data centers consume memory by making pooled capacity easier to deploy across servers.

Nanya said its board approved a multi-year investment program for Fab5A running from 2026 through 2029, with spending of up to NT$346.6 billion. The project spans 1b through 1e 10 nm-class DRAM nodes and introduces EUV lithography. The company also raised its 2026 capital expenditure budget to NT$69.7 billion, with the additional allocation aimed at equipment prepayments to keep the project on schedule. According to the roadmap described in the source, wafer output is expected to begin in the second half of 2027, rise to 30,000 wafers per month in 2028, and reach 35,900 wafers per month in 2029, with longer-term capacity planned at about 45,000 wafers per month. Nanya also said 1c pilot production has started and that 1d is moving toward pilot production. For the memory market, the announcement points to a larger long-range DRAM supply base rather than an immediate shift in available output. Near-term pricing remains supported by the recent strength in DRAM, especially DDR4, but the scale of this expansion indicates Nanya is positioning for more advanced process competition and higher future supply flexibility.

CoreWeave has arranged several years of priority enterprise SSD supply from Solidigm. The agreement is notable because AI cloud operators are starting to reserve storage capacity in advance, much as they already plan around accelerator availability. For NAND and enterprise SSD vendors, that points to steadier demand from AI infrastructure and less uncommitted capacity for high-capacity drives. It also moves flash procurement earlier in data-center planning, especially for large training and inference environments that need consistent storage throughput. If similar agreements spread across the cloud sector, enterprise SSD allocation could tighten and support NAND pricing.

SK hynix set a 54.3 trillion won investment plan across two Korean production projects aimed at future AI memory demand. The budget is divided between 35.2 trillion won for Yongin Y2 and 19.1 trillion won for Cheongju M17. Cleanroom openings are planned for June 2029 at Y2 and December 2028 at M17. The announcement is a long-horizon capacity signal rather than immediate market relief. It shows that SK hynix is preparing for sustained demand from AI servers across HBM, DRAM, and NAND, while also reminding buyers that advanced memory supply takes years to bring online. Capacity decisions being made now are more likely to shape late-decade pricing than the current cycle.

ChangXin Memory Technologies is evaluating another DRAM manufacturing site in China while pursuing a far larger role in global memory supply by 2030. If the current project pipeline is built out, CXMT's output could rise sharply over the next several years and give buyers a broader source of mainstream DRAM. The main constraint is execution. Limited access to advanced chipmaking equipment could slow the company's move into higher-end nodes and keep Samsung, SK hynix, and Micron ahead in premium segments. For pricing, this is not near-term supply relief because fabs take years to qualify and ramp. Over a longer horizon, a successful CXMT buildout would add competitive pressure in commodity DRAM, while tooling limits would keep that pressure uneven.

TSMC reportedly has a large inventory of Apple processors awaiting final packaging because the required DRAM has not arrived from Apple's memory suppliers. The issue centers on package-on-package integration, where LPDDR5X is placed with the application processor rather than installed elsewhere on the board. That makes memory availability a direct gating item for finished chip supply. The report highlights how tight mobile DRAM availability can ripple beyond memory buyers and affect semiconductor packaging schedules. Apple is said to rely mainly on Micron for LPDDR5X while also sourcing from SK hynix and Samsung. For the memory market, the signal is that leading smartphone demand is competing for constrained high-end DRAM at a sensitive point in the product cycle. That supports the broader view that mobile DRAM pricing power remains with suppliers until allocations loosen.

This is a stronger signal than earlier validation-only reports because it places CXMT closer to a manufacturing ramp. The reported LPDDR6 parts are 16Gb chips in package-on-package form, aimed at mobile memory use cases where Samsung, SK hynix, and Micron have historically held technology leadership. Risk production does not guarantee high-volume yield or major customer wins, but it suggests Chinese mobile DRAM competition is advancing faster than many buyers expected.