Semiconductor Market Dynamics
I. Industry Overview: AI Spending Keeps Expanding, Institutions Raise Full-Year Forecasts
Gartner projects global AI spending to reach $2.7 trillion in 2026, up 49.5% year-on-year, driven by AI-optimized IaaS, servers, networking architecture and AI processing semiconductors, largely unaffected by memory price increases. KB Securities estimates the global semiconductor market will reach approximately $1.44 trillion in 2026, with memory rising to 55% share and growing more than threefold, as the AI infrastructure bottleneck shifts from GPUs to HBM, server DRAM and NAND, signaling that the core constraint has moved to memory. Bank of America has raised its 2030 global semiconductor TAM forecast from $1.7 trillion to $3.2 trillion.
AI data center capacity buildout is being described as the largest infrastructure project in human history. Google has raised its full-year capex to as much as $205 billion, Microsoft plans to expand data center capacity from 12GW to over 38GW by 2032, Oracle's GPU utilization stands at 97.9% with a $664 billion order backlog, and CoreWeave's contract backlog has grown to $104.2 billion.
II. Memory: Structural DRAM Shortage Persists, NAND Supply-Demand to Diverge in 2027
DRAM: Supply gap extends to 2028, HBM crowding-out effect becomes core bottleneck
Goldman Sachs estimates a global DRAM supply gap of about 4.9% in 2026, while UBS expects demand to grow 36.2% in 2027 against supply growth of only 19.3%, with tightness persisting into H1 2028. A Bloomberg survey shows over 60% of industry participants believe the shortage cycle will extend beyond 2028, and in extreme cases to 2030. Citi forecasts global DRAM wafer capacity will grow only 8% next year, with shortages potentially lasting until 2031.
The HBM crowding-out effect is the core reason—HBM consumes roughly 3 to 4 times more silicon wafer than conventional DRAM for the same capacity, and suppliers are prioritizing HBM, steadily tightening general-purpose DRAM supply. Micron notes that memory has become the core bottleneck limiting AI system performance, with meaningful new capacity not arriving until 2028; Samsung and SK Hynix inventories have fallen below 10 days of consumption, a multi-year low.
On pricing, Bank of America has raised its optimistic DRAM growth scenario to 80% annually, with HBM3E 36GB spot prices at $2,100—roughly 4 to 5 times long-term contract prices—and 16-layer HBM4 spot prices reaching $3,500. Samsung has allocated about 70% of its capacity through 2031 to long-term agreements, with key customers including NVIDIA, Microsoft and Google, and some AI customers seeking supply commitments of up to 10 years.
NAND: AI inference demand reshapes the landscape, supply-demand may loosen in H2 2027
Kioxia's CEO has unusually stated that NAND prices "have risen enough" and has asked the sales team to avoid aggressive price increases. Apple has accepted Samsung's Q1 2027 quotes, with DRAM near $2 per Gb and NAND near $0.33 per Gb, up 30% to 40% from current levels. Phison's CEO warns that next year's shortage is "anyone's guess," noting that new NAND fabs take about four years from investment to mass production, making it hard for supply to catch up quickly.
TrendForce estimates the 2026 NAND supply-demand gap at -4% to -5%, with conditions potentially easing in H2 2027. SanDisk has signed multi-year agreements with eight major customers, locking in about two-thirds of its bit supply for fiscal 2028. Kioxia and SanDisk jointly announced cumulative investment of about ¥5 trillion through fiscal 2032, serving both as a capacity roadmap and as leverage in long-term contract negotiations.
III. AI Servers and CPU: CSP Capex Surges, CPU Returns to Center Stage
TrendForce data shows the world's nine largest CSPs are expected to spend over $886.7 billion in capex in 2026, up about 90% year-on-year, expanding to $1.3 trillion in 2027. AI server shipment growth has been revised up to 31%.
CPU demand is seeing a structural surge driven by agentic AI. Personal AI agents such as Meta's Muse need to run continuously in dedicated cloud virtual machines, and traditional software tasks rely heavily on CPU rather than GPU, with the CPU-to-GPU ratio expected to shift from 1:4 toward roughly 1:1. TrendForce data shows CPU lead times have nearly doubled from 16-20 weeks to 25-30 weeks. Intel CEO Lip-Bu Tan says the company "can only meet 50% of customer demand" and warns that memory shortages could worsen further. Bank of America has raised its 2030 server CPU market forecast to $210 billion, with AI-specific server CPUs accounting for about 86%.
IV. Analog and Power Semiconductors: TI Raises Prices for Third Time This Year, Industry Recovery Confirmed
Texas Instruments launched its third price adjustment of the year on September 1, following earlier rounds in April and July. STMicroelectronics has also initiated a third round this year, ADI's second round took effect September 13, Renesas' new pricing takes effect January 1, 2027, and Maxscend, Nationz Technologies and Qualcomm have followed in quick succession.
The driver is no longer just cost pass-through—AI computing demand for power management ICs continues to rise, and foundries are prioritizing mature-node capacity for AI servers and automotive products, creating a "siphon effect." A traditional server needs only dozens of PMICs, while a new-generation AI server requires far more power management chips at higher specifications, significantly raising per-unit value.
In power semiconductors, onsemi will adjust prices from October 10, Allegro from October 1, and Taiwanese suppliers Panjit, Taiwan Semiconductor and Eris Technology plan increases of 10%-15% from October. The AI data center power semiconductor market could be revised up from $12 billion to $50 billion, as NVIDIA pushes 800V DC architecture and GaN and SiC adoption accelerates.
V. Passive Components: MLCC Supply-Side Contraction, Industry Enters New Price Cycle
Murata has formally begun product line optimization, phasing out selected MLCC part numbers across nine series including GRM, GRJ and GRT, with last orders due by March 31, 2028 and final shipments by March 31, 2029. The move is aimed at phasing out low-margin mature part numbers and shifting capacity toward AI servers and high-end automotive products. Murata holds about 40% global MLCC share and about 70% of high-end AI server MLCC, with utilization already at 95% and customer inquiries for high-end AI server MLCC reaching twice its current capacity.
Samsung Electro-Mechanics was first to raise Q4 2026 quotes for OEM/ODM customers, with consumer-grade X5R up 25%-30% on average and high-end AI server X6S up 10%-20%. It also signed a KRW 1.0722 trillion AI server MLCC long-term supply contract, the largest single MLCC order in company history.
AI server MLCC usage is about 8 to 12 times that of traditional servers. Murata's vice president says MLCC demand will stay strong at least through 2028, and the company is "boldly evaluating" a 3-5 year expansion plan. Sanhuan's automotive-grade MLCC is set to rise 30%-40% in Q4, with some high-capacitance part numbers up as much as 80%.
VI. HDD: Contract Prices Hit Record High for Sixth Straight Quarter, Shortage May Last to 2028
HDD benchmark contract prices rose for a sixth consecutive quarter to record highs, with 3.5-inch 1TB drives at about $67.7, up 15% quarter-on-quarter, and 2.5-inch 1TB at about $61.2, up 10%. Nearline drive prices rose 20%-30% from the prior quarter, with US hyperscalers accounting for 60% of buyers.
Seagate's CEO says nearline capacity is fully booked through the end of 2026, with many customers planning into 2029 and beyond. Morgan Stanley notes HDD demand is growing 40%-50% annually while supply grows only 30%-35%, with shortages lasting at least through 2028. Dell's COO warns that if AI agent workloads continue to grow rapidly, memory and HDD shortages could persist for more than five years.
Manufacturer Updates
Spot Market Insights
Memory
eMMC: The eMMC market stayed flat in September, with prices failing to stabilize. Samsung 4G/8G kept sliding but neared bottom by late month, as channels grew reluctant to sell lower. Sandisk sub-64G products held relatively firm on EOL expectations, though demand shrank in tandem. Kioxia supply remained ample, leaving room for negotiation. Deals continued, but panic selling by hoarders and end-user price pressure coexisted, with signs of a price floor gradually emerging.
NAND Flash: The market polarized. Winbond saw dull trading, weak prices and even inverted pricing, with ample inventory. MXIC and ISSI remained severely short, with limited OEM arrivals and suppliers holding back goods, keeping premiums elevated all month.
NOR Flash: Supply kept tightening. Winbond small-capacity held flat; Micron small-capacity quotes loosened but stayed too high to close deals. 32M/64M parts saw no OEM allocation on EOL expectations, pushing supply to the spot market. Some high-capacity and automotive-grade parts were critically short at absurdly high prices. Customers began stocking for next quarter, likely worsening the supply-demand imbalance.
DDR3 & DDR4: DDR3 diverged — 4G fell early then saw trader stockpiling late month, ending earlier dump pricing, while 8G/16G held steady. In DDR4 4G, Winbond bulk demand and Hynix 4G key-account demand drove prices up fast, while Samsung 4G/8G stayed soft and declined. DDR4 16G and DDR5 16G traded actively, with 8G/16G chips priced high on thin supply, and DDR5 16G firming upward. Overall: chips strong, modules weak; large-capacity strong, small-capacity mixed.
LPDDR: Early September saw weak demand and flat prices; mid-to-late month brought renewed demand for low-capacity LPDDR5/5X, with more sourcing activity and livelier trading. LPDDR4/4X stayed flat on low inventory but equally dull demand after OEM EOL and shipment cuts; some Micron parts that rose earlier saw slight pullbacks.
Server Memory (DIMM): Early-to-mid month saw slight price gains on overseas demand, with DDR5 32G/64G spot tightening. But domestic customers balked at high prices, turning to domestic or used supply. Late month demand turned down as end-user buying paused, prices loosened, and quotes struggled to close. Overall: overseas holds prices up, domestic stays on the sidelines.
Storage
SSD: September showed clear divergence. Samsung 15.36T stayed tight at high prices, though quotes eased late month; other capacities (SATA and PCIE) fell broadly as weak demand pushed hoarders to cut prices. Interest centered on PCIE 5.0 (Samsung PM9D3A, Solidigm PS1010) and SAS drives. Solidigm PS1010 tightened late month on strong demand and thin supply, mostly from overseas customers.
HDD: Overall demand was flat in September, but 24T remained the main enterprise driver, with tight supply and high prices; channels expect post-holiday price hikes. In 20T, Seagate/WD held steady while Toshiba rose then fell, with branded supply tight late month and O-goods filling the gap. 16T was well supplied, with more room to negotiate and prices drifting down as some channels moved volume at low quotes. Overall: large-capacity tight, mid/small-capacity loose.
CPU
Server CPU: Intel fell slightly across the board early month, with 6th-gen well supplied and clearly weaker, and old 3rd/4th-gen under heavy inventory pressure. Mid-month brought a demand rebound in 3rd/5th-gen, with active inquiries on 6338/6530/6526Y pulling 4th-gen along. Late-month demand kept recovering, channel pricing stayed competitive, and a modest Q4 rebound is expected. AMD saw 5th-gen shortages ease and prices stabilize, but 9004 series kept falling on heavy overseas arrivals and OEM inventory buildup, 9005 held flat, and mid/low-end slipped short-term amid serious imbalance; 9135 stayed short all month.
Mobile Platform CPU: Low-end industrial models (J6412 etc.) moved well, with J6412 falling first then rising on end-user stockpiling. Q370 and H310 demand persisted, trading actively mid-month; late month saw Q370 briefly short and H310 appearing only in small full-pack lots, with end budgets unchanged but market prices high. X-series pure industrial offers stayed scarce. N305 remained short with no relief in sight.
RAID/NIC
Broadcom RAID demand was scattered, with 9670-24i and 9361-8i 1G the steady hot spots and prices moving little. NIC demand centered on hot parts like N425G/P425G, which stayed short with rising prices; P2200GQF00 was short and bid up. Overall, Broadcom cards stayed tight, with hot models increasingly hard to source.
GPU
Demand was light early-to-mid month, with PRO6000 Server sharply up on shortage and no spot available; some Pro6000SE arrivals rose in price and moved fast. Late month brought hot demand for H200 141G NVL PCIe, making it the new demand focus.
October Market Outlook
Memory Chips (DDR/LPDDR): High-end large-capacity DDR5, HBM and enterprise SSD should stay strong; consumer-grade DRAM and NAND remain pressured by high prices and may keep softening.
NOR Flash: High-capacity automotive/industrial parts still have upside; small/mid-capacity gains narrow and hold high.
HDD: 24T shortage persists with post-holiday hike expectations; 16T well supplied and under pressure.
CPU: Intel may see a modest Q4 rebound; AMD 9004 series keeps facing downside, with mid/low-end shortages yet to be filled.
Server market: Enterprise high-end memory and SSD stay robust; consumer PC modules may return to range-bound trading after a brief recovery.
GPU: High-end professional cards stay tight, H200 demand heats up, and upward price pressure persists.
Broadcom cards: Lead times stretched to one year, spot tight and prices rising — early ordering is advised.