Producten
14 ankerproducten van de hardwarecrisis. Per product: trendlijn uit de tijdseries (alleen waar ≥2 datapunnen), events en leverancierssignalen uit 705 events.
NVIDIA B200 GPU
25 events in dataset
Nog onvoldoende datapunten voor een trendlijn (serie heeft minder dan 2 punten). Het signaal staat hieronder in de events.
Signalen: 20 MW -> 84 MW planned with GB300 NVL72 liquid-cooled racks; on-site fuel cells bypass incremental utility interconnect, ABF substrate, ASML EUV, Advanced packaging, Board design, CoWoS capacity, CoWoS packaging, Copper interconnect
- B200 Blackwell in mass production through Q1–Q2 2026; GB200 NVL72 rack-scale shipping. $30K–40K per unit. AMD vs NVIDIA AI GPU Market Share 2026: MI350X vs B200 Competitive Analysis
- MI350X CDNA 4 with 288 GB HBM3e, 8 TB/s bandwidth, ~4,600 TFLOPS FP8 matches B200 specs. Growing deployment at Microsoft/Meta/OpenAI. AMD vs NVIDIA AI GPU Market Share 2026: MI350X vs B200 Competitive Analysis
- NVLink 5.0 providing 1.8 TB/s bidirectional bandwidth between GPUs in GB200 NVL72. NVIDIA GB200 Supply Chain: The Global Ecosystem Explained
- TSMC N3P (3nm) process node used for B200/GB300; capacity constrained by equipment and wafer starts. NVIDIA GB200 Supply Chain: The Global Ecosystem Explained
- B200 priced at $30K–40K per unit; GB200 NVL72 rack-scale systems at $2M–3M. AMD vs NVIDIA AI GPU Market Share 2026: MI350X vs B200 Competitive Analysis
- AMD MI350X priced 15-20% below NVIDIA B200 to gain market share. AMD vs NVIDIA AI GPU Market Share 2026: MI350X vs B200 Competitive Analysis
GB300 / DGX Station
11 events in dataset
Nog onvoldoende datapunten voor een trendlijn (serie heeft minder dan 2 punten). Het signaal staat hieronder in de events.
Signalen: 20 MW -> 84 MW planned with GB300 NVL72 liquid-cooled racks; on-site fuel cells bypass incremental utility interconnect, 9-CSP CapEx: $957B 2026 (+106%) → >$1.5T 2027 (+60%, revised up from 50%), 9-CSP capex: $957B 2026 (+106%) -> >$1.5T 2027 (+60%); AI server shipments +32.2% 2026, +~36% 2027, ASML EUV, Advanced packaging, CoWoS packaging, Dedicated Ethernet training back-end spanning buildings on a ~1M-GPU, gigawatt campus — direct pull on 800G/1.6T optics and switch ports, Direct demand-side datapoint for nearline HDD: the L4 'source of truth' for a ~1M-GPU cluster is durable HDD-backed BLOB/Tectonic storage, regionally colocated with GPUs
- GB300 transitioning into rack-scale shipment late 2026; 288 GB HBM3e per GPU. Power-Bound, Not GPU-Bound: AI Data Center Power Constraints Are the Real 2026 Bottleneck
- TSMC N3P (3nm) process node used for B200/GB300; capacity constrained by equipment and wafer starts. NVIDIA GB200 Supply Chain: The Global Ecosystem Explained
- GB300 Blackwell Ultra unveiled October 2025: 160 streaming multiprocessors vs 144 on GB200, 20,480 total CUDA cores. Targeted for late-2026 rack-scale shipping with 288 GB HBM3e per GPU. DGX Station version priced at $92,900 with 748 GB unified memory, 20 PFLOPS FP4 performance, trillion-parameter model support. Nvidia GB300: 288 ГБ, PCIe 6, 1400 Вт — specs detail — Habr
- NVIDIA has reserved the majority of TSMC's CoWoS advanced-packaging capacity through at least 2027. Vera Rubin's dual-chiplet GR100 design (two dies per GPU package) increases the number of large-die assemblies per rack compared to monolithic Blackwell B200. While TSMC's CoWoS yield has improved from earlier quarters, the combination of increased die count per GPU plus continued Blackwell GB300 demand keeps allocation extremely tight. Foxconn and Wistron provide ODM assembly for NVL72 rack-scale systems under NVIDIA direction. Yield improvements alone cannot offset the geometric increase in required CoWoS operations. AI Server Demand Puts Supply Chains Under 2027 Pressure — WinBuzzer
- TrendForce AI-server industry research (Press Center, 14 Oct 2026): combined CapEx of the nine leading CSPs (Google, Amazon, Meta, Microsoft, Oracle, ByteDance, Tencent, Alibaba, Baidu) projected to surge 106% YoY to US$957B in 2026 and surpass US$1.5 trillion in 2027 (+60% YoY, revised up from 50%) because of rising complete-server-system prices, higher component quotations amid supply shortages, and advance payments some component suppliers demand to secure supply. AI-server shipments: +32.2% YoY in 2026, ~+36% in 2027. Total server market revenue: US$639.2B expected in 2026 (AI servers ~US$479B, ~75% share); first-ever US$1T+ server market in 2027 with AI-server revenue up >90% YoY to nearly US$920B (~81% share). End-customer orders in 2027 expected to center on rack-scale solutions: NVIDIA GB300 and VR200, plus AMD Helios series; memory shortages and rising component prices push next-gen rack ASPs up. Tier-2 operators incl. Tesla and CoreWeave also keep buying AI servers. CSP CapEx Projected to Exceed US$1.5 Trillion in 2027, with AI Server Revenue Growing Over 90% YoY on Higher Shipments and Prices, Says TrendForce
- TrendForce (14 Oct 2026 press release): combined CapEx of the nine leading CSPs (Google, Amazon, Meta, Microsoft, Oracle, ByteDance, Tencent, Alibaba, Baidu) is projected to surge 106% YoY to US$957B in 2026 and to surpass US$1.5T in 2027 (+60% YoY) — the 2027 growth forecast revised up from 50%. The upward revision is attributed to rising complete-server-system prices, higher component quotations amid supply shortages, and advance payments some component suppliers demand to secure supply, prompting CSPs to bring cash outlays forward. AI server shipments are forecast +32.2% YoY in 2026 and nearly +36% in 2027; total server market revenue passes $1T for the first time in 2027, with AI server revenue +90% YoY to ~$920B (~81% of the server market). Shortages and rising memory prices are explicitly cited as drivers of next-gen rack ASPs (GB300/VR200, AMD Helios). CSP CapEx Projected to Exceed US$1.5 Trillion in 2027, with AI Server Revenue Growing Over 90% YoY on Higher Shipments and Prices, Says TrendForce
HBM3e geheugen
40 events in dataset
Nog onvoldoende datapunten voor een trendlijn (serie heeft minder dan 2 punten). Het signaal staat hieronder in de events.
Signalen: 1-gamma process would match or exceed Samsung/SK Hynix 1X-class nodes for HBM4E performance targets, 71% market share = extreme concentration risk. If Hanmi experiences production disruption (fire, labor dispute, export restriction), HBM4 assembly at all three memory vendors is constrained simultaneously., ABF substrate, AI demand growth, ARM ecosystem maturity, Advanced packaging, Both HBM3E suppliers raising prices simultaneously indicates collective capacity constraint rather than competitive undercutting, CoWoS capacity
- B200 Blackwell in mass production through Q1–Q2 2026; GB200 NVL72 rack-scale shipping. $30K–40K per unit. AMD vs NVIDIA AI GPU Market Share 2026: MI350X vs B200 Competitive Analysis
- GB300 transitioning into rack-scale shipment late 2026; 288 GB HBM3e per GPU. Power-Bound, Not GPU-Bound: AI Data Center Power Constraints Are the Real 2026 Bottleneck
- MI350X CDNA 4 with 288 GB HBM3e, 8 TB/s bandwidth, ~4,600 TFLOPS FP8 matches B200 specs. Growing deployment at Microsoft/Meta/OpenAI. AMD vs NVIDIA AI GPU Market Share 2026: MI350X vs B200 Competitive Analysis
- Grace Hopper GH200 superchip shipping; 141GB HBM3e unified memory. Used in HPE Cray EX254n for exascale systems. 8x NVIDIA Grace Hopper Superchips in a Blade — HPE Cray EX254n at GTC 2024
- HBM3e supply constrained by wafer allocation from Samsung/Micron/SK Hynix. 3:1 conversion ratio between HBM and DDR5 wafers. High Bandwidth Memory (Wikipedia) - Micron quote
- SK Hynix primary HBM3e supplier to NVIDIA; estimated 50%+ market share. High Bandwidth Memory (Wikipedia)
HBM4 geheugen
54 events in dataset
Nog onvoldoende datapunten voor een trendlijn (serie heeft minder dan 2 punten). Het signaal staat hieronder in de events.
Signalen: 1-gamma process would match or exceed Samsung/SK Hynix 1X-class nodes for HBM4E performance targets, 1.8 MW/CDU in-row cooling at N+1; NVL72 rack = 20.7 TB HBM4; scalable-unit blueprint to GW scale, 70% allocation = structural moat for SK Hynix; base die yield respin a known risk factor, 71% market share = extreme concentration risk. If Hanmi experiences production disruption (fire, labor dispute, export restriction), HBM4 assembly at all three memory vendors is constrained simultaneously., 8-Hi HBM4 shift raises the number of Rubin Ultra GPUs supportable per unit of HBM bit output from mid-2027, >1M Maia 300 accelerators planned for 2027 at 96 chips/~167 kW per rack — incremental liquid-cooled rack demand outside the NVIDIA ecosystem, AATT2 HBM assembly/test capacity online Oct 2026; 100% of planned 2027 output committed, AMD; Microsoft Azure; HBM4 supply for MI455X
- Vera Rubin platform targets H2 2027; 336B transistors, HBM4, NVLink 6.0 (3.6 TB/s per GPU), NVL72 system at 3.6 ExaFLOPS FP4. NVIDIA GB200 Supply Chain: The Global Ecosystem Explained
- MI355X successor to MI350X targeting 2027; expected 30-40% performance uplift. AMD MI350X vs NVIDIA B200: Which AI GPU Wins?
- Samsung targeting HBM4 production H2 2027; 12-stack architecture with 24GB per die. High Bandwidth Memory (Wikipedia)
- Micron HBM4 targeting 2027 production; competing with Samsung and SK Hynix. High Bandwidth Memory (Wikipedia)
- Samsung expanded 1c DRAM capacity to target 60,000 wafers/month specifically for HBM4 production. 2025–present global memory supply shortage (Wikipedia)
- AMD secured HBM4 supply from Samsung for its next-gen Instinct MI455X GPU. Deal includes condition benefiting Samsung foundry business — Samsung supplies both HBM4 for MI455X accelerators and DDR5 for EPYC processors. May signal partial AI chip shift toward Samsung Foundry beyond just memory. [News] AMD Secures Samsung HBM4 for MI455X — TrendForce
CoWoS advanced packaging
35 events in dataset
Signalen: $1.3T hyperscaler capex in 2027 implies continued GPU procurement at or above current pace — supports long-term CoWoS/HBM demand projections., $161.2B aggregate rent under Broadcom master equipment leases + additional TPU compute procurement committed for 2028-2031 delivery (S-1), 5.5x-reticle CoWoS in HVM at 98–99% yield; 14x target 2029, 6 GW cumulative Anthropic TPU deployment planned through 2027, ABF substrate, AI GPU (Blackwell/Rubin), server-CPU and custom-ASIC packaging demand; test-time growth consuming assembly/test capacity; upstream equipment delivery (6-8 months), Advanced packaging, Advanced packaging equipment
- B200 Blackwell in mass production through Q1–Q2 2026; GB200 NVL72 rack-scale shipping. $30K–40K per unit. AMD vs NVIDIA AI GPU Market Share 2026: MI350X vs B200 Competitive Analysis
- GB300 transitioning into rack-scale shipment late 2026; 288 GB HBM3e per GPU. Power-Bound, Not GPU-Bound: AI Data Center Power Constraints Are the Real 2026 Bottleneck
- Vera Rubin platform targets H2 2027; 336B transistors, HBM4, NVLink 6.0 (3.6 TB/s per GPU), NVL72 system at 3.6 ExaFLOPS FP4. NVIDIA GB200 Supply Chain: The Global Ecosystem Explained
- MI350X CDNA 4 with 288 GB HBM3e, 8 TB/s bandwidth, ~4,600 TFLOPS FP8 matches B200 specs. Growing deployment at Microsoft/Meta/OpenAI. AMD vs NVIDIA AI GPU Market Share 2026: MI350X vs B200 Competitive Analysis
- MI355X successor to MI350X targeting 2027; expected 30-40% performance uplift. AMD MI350X vs NVIDIA B200: Which AI GPU Wins?
- TSMC CoWoS capacity expanding but remains the primary bottleneck for all advanced GPU production. 2026 capacity estimated at 100K wafers/month. NVIDIA GB200 Supply Chain: The Global Ecosystem Explained
ABF-substraat
9 events in dataset
Signalen: ABF substrate, CoWoS packaging, Grid power, HBM3e memory, Ibiden, Nippon Steel, Planning horizon extended to 3–5 years across fab + packaging + substrates, Samsung Electro-Mechanics + LG Innotek AI substrate utilization near 90% (H1 2026). Ibiden constrains Nvidia shipment volumes; Ajinomoto constrains Ibiden throughput.
- B200 Blackwell in mass production through Q1–Q2 2026; GB200 NVL72 rack-scale shipping. $30K–40K per unit. AMD vs NVIDIA AI GPU Market Share 2026: MI350X vs B200 Competitive Analysis
- ABF substrates for advanced GPUs lead time 30-40 weeks, constrained by Nippon Steel and Ibiden capacity. AMD vs NVIDIA AI GPU Market Share 2026: MI350X vs B200 Competitive Analysis
- ABF substrates for advanced GPUs face lead time of 30–40 weeks, constrained by Ibiden (Japan), Ajinomoto (substrate supplier to Ibiden), and Unimicron expansion. South Korean suppliers Samsung Electro-Mechanics and LG Innotek ramping AI substrate capacity with 1H utilization near 90%. Unimicron raised 2026 capex from NT$34B to >NT$53.7B, 80% directed at AI-related substrate. The ABF Substrate Bottleneck — DataGravity.dev
- AMD CEO Lisa Su told Taiwan press during a 6 Oct 2026 visit (Liberty Times/Wealth Magazine/Economic Daily News, via TrendForce) that AMD's capacity planning horizon has stretched from 1–2 years to three, four, or even five years ahead — locking in wafer fab, packaging (CoWoS) and ABF substrate capacity simultaneously so the chain expands in lockstep. AMD is increasing supply in 2026 and plans to 'ramp significantly further in 2027' as AI CPU/GPU demand exceeds expectations; the MI450-series AI platform (TSMC 2nm + CoWoS) entered production at TSMC in Q3 2026 with capacity ramping gradually to end-2026 before higher-volume shipments begin in 2027; the succeeding MI455 series is scheduled to enter production in 1H 2027. Su said AMD is exploring Fan-Out panel-level packaging with the Taiwan supply chain but does not view it as a replacement for CoWoS, and confirmed the >US$10B Taiwan ecosystem investment plan remains fully on track. Foxconn is among the AMD AI-server suppliers scaling with the ramp. TrendForce — AMD's Lisa Su Visits TSMC, AI Supply Chain Partners as Company Locks In Capacity 3–5 Years Ahead
- The ABF substrate bottleneck moved from analyst claim to industry self-admission in late September 2026: at the Semicon Taiwan show Unimicron's chairman publicly said substrates are in short supply, and Broadcom named substrates as an AI-chip bottleneck (Growth Papa research letter dated Sep 10, published Oct 9). The letter maps the choke points to Japan — T-glass low-CTE glass fabric ~90% Nittobo, ABF film >95% Ajinomoto, ultra-thin package copper foil ~99% Mitsui Mining & Smelting — with finished-substrate leadership at Unimicron (TW) and Ibiden (JP); Korean players (Samsung Electro-Mechanics FC-BGA, Doosan CCL, Isu Petasys MLB, Lotte copper foil) hold seats but not the bottlenecks. As packages grow toward 5.5x-reticle CoWoS, substrate area per GPU expands super-linearly, tightening effective supply. Corroborating move: Samsung's announced ~$9.15B chip-segment investment (Oct 13) includes W7.67Tn from Samsung Electro-Mechanics for FC-BGA substrate capacity split between Sejong (Korea) and Vietnam. Growth Papa — The Bigger the Chip, the Fewer the Substrates: Three Layers of Materials, Two Seeds
- Unimicron Chairman S.C. Chien at the Taiwan Printed Circuit Association (TPCA) Show on 20 Oct 2026 (per Wealth Magazine, Economic Daily News, CNA, MoneyDJ — via TrendForce): lead times for some key substrate equipment have stretched to 30 months; lead times have lengthened for more than 80% of equipment, with over 20% delayed by more than double their original lead times. Low-CTE glass fiber cloth, copper-clad laminates (CCL) and ABF are particularly tight; critical materials could face a 20% supply shortfall in 2027. Capacity-growth asymmetry: CoWoS projected to grow 6.8x by 2028 vs 2024 alongside 6.3x hyperscaler capex, while IC-substrate capacity merely doubles and DRAM expands 1.8x. AI substrate sizes: ~1x reticle early designs -> 3x (2025) -> 5x targeted 2026 -> 5-9x by 2027-2028, eventually 12-14x; AI chips drawing >1,000 A strain power integrity, cooling and yield. On Intel's EMIB chain, substrate yields recently improved to ~45% (industry sources), targeting 50% by 4Q26 and 60% by 1Q27. Unimicron Warns of IC Substrate Crunch: Up to 30-Month Equipment Lead Times, 20% Material Gap in 2027 (TrendForce)
DDR5 RDIMM
27 events in dataset
Signalen: 1.8 MW/CDU in-row cooling at N+1; NVL72 rack = 20.7 TB HBM4; scalable-unit blueprint to GW scale, 23% of DRAM wafers now dedicated to HBM (up from 8% in 2024); no major new DRAM fab capacity before 2028, 2Q26 DRAM revenue share table: Samsung 39.4 / SK hynix 24.9 / Micron 23.3 / CXMT 9.5 (%), >1M Maia 300 accelerators planned for 2027 at 96 chips/~167 kW per rack — incremental liquid-cooled rack demand outside the NVIDIA ecosystem, >50% more gross die per wafer at 8Gb-equivalent vs G4; LPDDR5X 24Gb products (50% more data vs prior gen), Advanced packaging, CXMT expanding toward reported ~300K wafers/month DRAM capacity per same wave of coverage, DRAM wafer capacity
- HBM3e supply constrained by wafer allocation from Samsung/Micron/SK Hynix. 3:1 conversion ratio between HBM and DDR5 wafers. High Bandwidth Memory (Wikipedia) - Micron quote
- AMD secured HBM4 supply from Samsung for its next-gen Instinct MI455X GPU. Deal includes condition benefiting Samsung foundry business — Samsung supplies both HBM4 for MI455X accelerators and DDR5 for EPYC processors. May signal partial AI chip shift toward Samsung Foundry beyond just memory. [News] AMD Secures Samsung HBM4 for MI455X — TrendForce
- For NVIDIA Vera Rubin, SK Hynix holds approximately 70% of HBM4 allocation. Samsung targets 25–30% share after rebuilding qualification pipeline. Micron successfully qualified its HBM4 alongside Samsung and SK Hynix — first time all three major IDMs supply the same next-gen HBM product tier. HBM4 pricing carries a 30–70% premium over HBM3E; Samsung/SKHynix also implemented ~20% price hike on HBM3E for 2026. Both Korean manufacturers redirected over 50% of DRAM wafer capacity to HBM production. HBM-DDR5 wafer displacement ratio stands at 3:1, with approximately 23% of all DRAM wafers reallocated to HBM. AI Memory Bottleneck HBM: The 3-Way Race Re-Opens — ByteDive
- Microsoft unveiled the Maia 300 accelerator on 12 Oct 2026: a multi-chiplet design built on TSMC's 2nm N2P for compute chiplets (memory-controller/IO dies on 3nm), with eight HBM4 stacks per package — up from six HBM3E on Maia 200 — taking memory capacity from 216 GB to 288 GB. Microsoft claims 36% more compute FLOPs, 2.5x HBM bandwidth and 14% higher scale-up bandwidth, and ~40% more computing output per megawatt than leading GPU-based systems; it plans to roll out potentially more than one million Maia 300 accelerators next year, 96 per rack at ~167 kW. Because density doubles, HBM stacks per rack jump 2.7-fold (288 to 768), a material new draw on 2027 HBM4 supply — SK hynix is seen as leading contender (sole HBM3E supplier for Maia 200) with Samsung's mass-produced 36 GB 12-Hi HBM4 also in play (via Seoul Economic Daily / Reuters, TrendForce roundup). Diverging route: at OCP 2026 Intel officially branded its Xe3P-based 'Crescent Island' inference accelerator as the GPU D-Series, deliberately using up to 480 GB of LPDDR5X instead of cost- and supply-constrained HBM, targeting 1.3 PFLOPS FP8. TrendForce — AI Chip Memory Strategies Diverge: Microsoft's Maia 300 Fuels HBM4 Demand as Intel Bets on LPDDR5X
- NVIDIA chose a monolithic compute die for its Vera CPU over the chiplet architectures used by x86 rivals, with a custom LPDDR5X memory subsystem and second-generation Scalable Coherent Fabric, per a Wccftech interview with NVIDIA's Vera team (via TrendForce, 19 Oct 2026). NVIDIA claims: 3.4 TB/s bisection bandwidth (3x the leading x86 processor in its comparison); 1.5x the performance of Grace; up to 2x leading x86 performance on sequential agentic tasks (Python scripting, code compilation) and 1.6x on a geometric-mean basis; new 'Spatial Multithreading' that allocates dedicated resources to worker vs I/O threads. Vera's role is expanding from GPU host processor (as Grace was positioned) toward standalone Vera CPU racks for agentic AI tool calls and reinforcement learning. In the same roundup, NVIDIA SVP Jeff Fisher (per CNA) denied RTX 5090 discontinuation rumors amid acknowledged industry-wide supply constraints; TechPowerUp notes the card remains hard to find near its US$1,999 MSRP, with some Amazon/Newegg listings above US$8,000. [News] NVIDIA Chooses Monolithic Die Over Chiplets for Vera CPU, Claims 2x x86 Performance in Agentic AI (TrendForce)
- Groq — the AI-chip company that licensed its inference technology to NVIDIA — is considering a shift from its SRAM-centric design to 3D-stacked DRAM (TSV-based 3DS, capacity-focused vs HBM's bandwidth focus) for next-generation chips (ZDNet industry sources via TrendForce, 22 Oct 2026). Current NVIDIA 'Groq 3 LPX' is a rack-scale inference accelerator for the Vera Rubin platform: each LPU integrates 500MB SRAM with 150TB/s on-chip memory bandwidth; a full LPX rack combines 256 LPUs for 128GB SRAM and 40PB/s aggregate SRAM bandwidth plus 12TB DDR5 (NVIDIA specs). The capacity wall: per d-Matrix CEO Sid Sheth via 36Kr, SRAM capacity grows only ~10-20% every two years while AI model sizes may double every 6-9 months. Manufacturing: Samsung Foundry produces NVIDIA's Groq 3 LPU on its 4nm process at Pyeongtaek (Seoul Economic Daily), and Samsung executives said in June they were discussing foundry cooperation on the next-generation LPU; TSMC CEO C.C. Wei said in April TSMC was working with customers on related efforts. Cerebras is likewise pursuing 3D-stacked DRAM for its future CS-6 wafer-scale system. [News] NVIDIA-Linked Groq Reportedly Eyes Shift From SRAM to 3D-Stacked DRAM for Next-Gen AI Chips (TrendForce)
eSSD (enterprise SSD)
21 events in dataset
Signalen: 256TB in a single E3.S slot (QLC), plus PCIe Gen6 SSD shown, 321-layer NAND to ~50% of domestic capacity by end-2026; Yongin P1 cleanroom early 2027 adds new capacity; H1 2026 cumulative revenue >100T won for first time, Addresses the controller CPU scaling wall for PCIe Gen6+ high-capacity eSSDs (incl. 245TB-class QLC drives), IPO capital + Xtacking-into-DRAM signals sustained NAND capex runway for China supply, NAND flash diverted to HBM production; HDD manufacturers at full capacity. Both storage tiers face supply constraints simultaneously from different root causes., SK hynix NAND allocation through end-2026; Micron >75% of FY27 output committed; SanDisk ~30% below peak wafer capacity, Server DRAM expected to remain undersupplied in 4Q26; PC DRAM supply could decline in 2027, Supplier inventories at historic lows; production failing to keep pace with orders (per TrendForce-cited commentary)
- NAND supply also constrained by wafer reallocation to HBM; enterprise SSD prices up 100%+. 2025–present global memory supply shortage (Wikipedia)
- Some NAND flash manufacturers converted production lines to DRAM to capture HBM margins, creating a compounding effect. Phison CEO confirmed 'every NAND manufacturer told us 2026 sold out.' Western Digital confirmed completely sold out of hard drives for 2026 at hyperscale level. This created a secondary squeeze on enterprise SSDs (which use NAND), pushing the SSD/HDD price ratio to 16x. The NAND reallocation is a second-order effect of HBM demand — the same margin incentive that moves DRAM wafers to HBM also moves NAND fab capacity to DRAM. The 2026 Memory Chip Shortage: Why Server RAM Prices Have Doubled (DatacenterDisk)
- TrendForce 4Q26 memory research (30 Sep 2026): conventional DRAM contract prices projected +10-15% QoQ in 4Q26 and NAND Flash contract prices +15-20% QoQ, with the pace of DRAM increases moderating vs prior quarters. Suppliers keep prioritizing advanced-process capacity for high-performance server products, so server DRAM stays undersupplied in 4Q26; enterprise SSDs are the only category with accelerating price growth (CSP demand). PC OEMs keep procuring aggressively on expectations DRAM supply remains tight in 2027, while PC DRAM supply itself could decline in 2027 as capacity shifts to server. Some server suppliers' price increases lag the market average because of pricing mechanisms stipulated in long-term agreements (LTAs). AI Server Demand Sustains Memory Contract Price Increases in 4Q26, While Consumer-Side Pressure Persists, Says TrendForce
- Counterpoint Research (13 Oct 2026): Yangtze Memory Technologies Co., China's only NAND maker at commercial scale, has filed for a Shanghai STAR Market IPO, ~a month after CXMT's July debut. YMTC rose to No. 3 globally by Q2 2026 NAND shipment volume — trailing only Samsung and SK hynix and ahead of Kioxia and Micron (No. 2 if SK hynix's Solidigm volumes are unbundled). Its NAND revenue share climbed from 9% to 14% between 1Q25 and 1Q26 with YoY revenue growth peaking at 478%; enterprise SSD revenue, near zero in 2024, is projected to approach 39% of YMTC's total revenue by 2028. YMTC is extending its proprietary Xtacking wafer-bonding technology beyond NAND into DRAM, targeting AI memory rather than only domestic commodity supply. YMTC Files for Landmark IPO After Climbing to No. 3 in Global NAND Race, With DRAM Next (Counterpoint Research)
- ZDNet Korea (via TrendForce, 16 Oct 2026): SK hynix VP Koo In-jae (SoC design lead) unveiled 'H-TPU' (Hynix Transaction Processing Unit) — an ultra-lightweight programmable processor dedicated to SSD data-plane control, deployed in clusters of dozens instead of a large CPU (superscalar/out-of-order features removed, 4-stage pipeline). H-TPU cuts processor area to ~4% and power to ~2.5% of a conventional CPU core, with per-operation module activation for further efficiency; SK hynix plans hardware automation (HWA) on top, moving instruction processing, address lookup and buffer copies into dedicated circuits. Framing: AI inference is read-intensive (write-once-read-many, read:write ratios ~300,000:1), and H-TPU-based controllers are positioned to complement high-bandwidth flash (HBF) by prefetching NAND data in parallel with compute. Same article (via Yonhap): Samsung began mass production of the PCIe 6.0 PM1763 enterprise SSD in July 2026 with a newly developed 4nm controller and 9th-generation V-NAND — >2x performance and >1.8x power efficiency vs predecessor. [News] SK hynix Unveils H-TPU SSD Processor That Slashes Power Consumption, Eyes HBF Applications (TrendForce, citing ZDNet Korea / Yonhap)
- Enterprise SSDs accounted for 48% of global NAND bits shipped in Q2 2026, up from 26% a year earlier. Enterprise SSD Prices 2026: How to Buy Server Storage Now
Nearline HDD
28 events in dataset
Nog onvoldoende datapunten voor een trendlijn (serie heeft minder dan 2 punten). Het signaal staat hieronder in de events.
Signalen: +100% Toshiba HDD capacity by FY2027 on ~$380M capex (Philippines), 195 EB/quarter nearline to data centers (+43% YoY); nearline = 89.4% of shipped capacity, 209 EB nearline shipped in single quarter (Q2 FY2026). Hyper-concentration on cloud (>89% revenue) creates hyperscaler dependency risk but secures committed volumes., 248 EB shipped, +21.6% Y/Y; 40TB ePMR at 50% of nearline by FQ3 FY27; 44TB HAMR volume shipments CY1H 2027, 40TB ePMR volume at customer. 44TB HAMR H1 CY27 on track. LTA negotiations covering 2029 to 2031 in active progress. Exabyte growth greater than 25% guided forward., All 2026 nearline capacity allocated. NVMe $/TB ratio 16x from 6x YoY. Consumer HDD approximately 5% of WD output. HDD price rise 35-46% YoY., Both Seagate and WD report utilization at or near full capacity across all product lines. China PC market also driving demand rebound after years of SSD substitution., Both WD CEO Irving Tan Feb 2026: pretty much sold out for calendar 2026 and Seagate CEO Dave Mosley confirmed identical capacity allocation position. Consumer HDD represents approximately 5% of WD output with rest absorbed by hyperscale CSPs.
- Some NAND flash manufacturers converted production lines to DRAM to capture HBM margins, creating a compounding effect. Phison CEO confirmed 'every NAND manufacturer told us 2026 sold out.' Western Digital confirmed completely sold out of hard drives for 2026 at hyperscale level. This created a secondary squeeze on enterprise SSDs (which use NAND), pushing the SSD/HDD price ratio to 16x. The NAND reallocation is a second-order effect of HBM demand — the same margin incentive that moves DRAM wafers to HBM also moves NAND fab capacity to DRAM. The 2026 Memory Chip Shortage: Why Server RAM Prices Have Doubled (DatacenterDisk)
- HAMR (Heat-Assisted Magnetic Recording) HDDs targeting 30TB+ capacity; first products expected 2027. Enterprise SSD Prices 2026: How to Buy Server Storage Now
- Nearline HDD supply constrained; Seagate and WD prioritizing enterprise over consumer. Enterprise SSD Prices 2026: How to Buy Server Storage Now
- Organizations adopting multi-tier storage: NVMe for training, HDD for cold data. Enterprise SSD Prices 2026: How to Buy Server Storage Now
- Seagate shipped 218 EB of total capacity in fiscal Q4 2026, up 34% YoY, of which 195 EB went to hyperscaler/large-enterprise data center nearline drives, up 43% YoY. Nearline drives accounted for 89.4% of shipped capacity; non-nearline fell 10% YoY to 23 EB. Seagate HAMRing customers - generating more and more $/TB (Blocks & Files)
- Seagate: 'Based on the long-term supply agreements in place today, the vast majority of our nearline exabytes are now allocated into calendar 2028. Importantly, we are not seeing customers pull back on planning horizons. As our strategic relationships deepen, many are actively seeking to extend planning horizons through 2029 and beyond.' Seagate HAMRing customers - generating more and more $/TB (Blocks & Files)
800G / 1.6T transceivers
61 events in dataset
Nog onvoldoende datapunten voor een trendlijn (serie heeft minder dan 2 punten). Het signaal staat hieronder in de events.
Signalen: $115B by 2027 implies ~$57B compound growth from FY2025 baseline; networking and interconnect represent an increasing share as GPU-count-per-cluster grows non-linearly, $16.7B AI semiconductor revenue in single quarter = ~$67B annualized rate; bandwidth of demand far exceeds historical networking TAM, $600M+ optical revenue target for FY2027 across AEC, DSP, and optical engine product lines; suggests optical segment alone approaching $300M+ annually, 12.8 Tb/s per OSFP-sized module; 4x front-panel density vs 200G/lane pluggables, 16-lane 224G PAM4 retimer + 3nm 1.6T AEC DSP extend 200G/lane scale-up silicon, 1RU Davisson CPO 102.4T box previewed for Q1 2027 production — earliest merchant CPO switch timeline observed, 300mm InP-on-SiPh platform would multiply laser/PIC die-per-wafer vs 150mm InP mainstream — unconfirmed volume, 300mm InP-on-silicon platform could raise InP laser/PIC wafer output vs industry-standard 150mm and below
- At Huawei Connect 2026 (Sept 17-18, Shanghai) Huawei published its accelerated Ascend roadmap: Ascend 960DT in Q1 2027 (pulled 3 quarters earlier than planned; 2 PFLOPS FP8 / 4 PFLOPS FP4, 288 GB HBM @ 9.6 TB/s = 2.4x the Ascend 950 series), Ascend 960PR in Q3 2027 (8 PFLOPS FP4, 192 GB HBM @ 2.4 TB/s), Ascend 970 in 2028 (3.6 PFLOPS FP8, 288 GB @ 14.4 TB/s), Ascend 980 in 2029 (preliminary: 7.2 PFLOPS FP8, 384 GB HBM @ 38.4 TB/s). Huawei also unveiled the Atlas 960E SuperPoD (industry-first Near-Package Optics, 7.2 Tbps links), upgraded TaiShan 950 SuperPoD and OceanStor M900 memory-context storage; its UnifiedBus agentic SuperCluster targets scaling to one million NPUs. Wccftech: Huawei Brings Huge Inference Boost With Its Next-Gen Ascend 960 Chips In 2027...
- At SEMICON West 2026 ('Optical Integration for the Next Generation of AI'), NVIDIA presented a next-generation CPO research prototype that moves optical engines from the organic package substrate onto the silicon interposer carrying the compute dies (optics-on-interposer, OOI) — eliminating the organic substrate from the electrical signal path for higher bandwidth density and lower power (Wccftech/The Elec via TrendForce, 22 Oct 2026); it also demonstrated DWDM at eight wavelengths up to 32 Gb/s each = 256 Gb/s aggregate per fiber. Both are research-stage with no commercialization timeline; NVIDIA says it is evaluating them for future products. Ahead, NVIDIA plans to extend optical interconnects with the Kyber NVL1152 architecture — eight racks in one 1,152-GPU NVLink domain supporting the Feynman platform, expected 2028 (NVIDIA/Wccftech). Current-generation CPO is already in production: Spectrum-X Ethernet Photonics switches entered full production May 2026 (200Gb/s SerDes + CPO, Vera Rubin platform; early adopters CoreWeave, Lambda, Oracle Cloud Infrastructure), with the manufacturing ecosystem named as TSMC (silicon-photonics fab), SPIL (chip-scale packaging/assembly/test), TFC (laser-die modules), Foxconn (system integration), Coherent/Lumentum/Sumitomo Electric (external laser sources), Corning/Senko/Browave (fiber), Fabrinet (cooling/final test). [News] NVIDIA Unveils Next-Gen CPO Prototype With Optics on Silicon Interposer, Targets Higher Bandwidth Density (TrendForce)
- Frank Rohmund, head of semiconductors at Carl Zeiss (the sole supplier of EUV optical systems to ASML), stated in an interview that China would need approximately 15 years to develop its own EUV lithography machine. He characterized this as a reasonable estimate while acknowledging China's actual progress is difficult to quantify. Critically, Rohmund warned that export restrictions may accelerate Chinese innovation rather than suppress it, paralleling ASML CEO Peter Wennink's prior assessment that tighter controls push Chinese R&D harder. A Shanghai-based company reportedly began producing immersion DUV lithography equipment as of July 2026, indicating active parallel development tracks. Carl Zeiss executive: China needs ~15 years to build EUV; sanctions backfire by accelerating innovation
- Tomahawk 6: 5nm, 64×800G ports = ~51.2 Tbps switch fabric. Adds TSN support for AI cluster scheduling. Volume expected 2027–2028 only. Enterprise Network Equipment Market Size & Share Analysis
- Pluggable-optics architecture. 51.2 Tb/s per-chip. Used in Blackwell-era NVIDIA networking builds. Legacy product now, succeeded by Spectrum-6 CPO line. Nvidia's CPO Switches Enter Full Mass Production; TSMC, Foxconn Lead Supply Chain Positioning
- World's first volume-produced 200G/lane CPO Ethernet switch system. 102.4 Tb/s per-chip (2× Spectrum-4). SN6800: 512×800G or 2,048×200G, 409.6 Tb/s total fabric. Mass production Aug 2026. Customers: CoreWeave, Lambda, Oracle. Nvidia's CPO Switches Enter Full Mass Production; TSMC, Foxconn Lead Supply Chain Positioning
Transformatoren
33 events in dataset
Nog onvoldoende datapunten voor een trendlijn (serie heeft minder dan 2 punten). Het signaal staat hieronder in de events.
Signalen: $1.84B+ announced across 7 US plants in 8 weeks; earliest production 2028 (Southwire) / 2029 (Hitachi Gallman), $340M transformer production expansion (Jonesville SC), in addition to Eaton's September $242M Arkansas enclosure plant, $800M H1 2026 US transformer orders, +1.1 GW near-term firm capacity from existing fleet; UPRISE program now ~4.7 GW identified vs 5-GW target, +14 GW new demand expected online through 2030; electric capex $51.5B of $70.7B plan, 40+ substation facilities in one utility group's 2026 rebuild slate — replacement-market demand layered on the AI-driven new-build queue, ABB electrification Americas DC orders +117% comparable y/y (Utility Dive/ABB release), Backlog $8.49B = ~3.7x annualized Q2 revenue; H1 orders $3.237B vs raised FY target $5.185B
- AI data centers requiring 50-100MVA transformers per facility; lead times 12-28 months. Switchgear, Transformer & Generator Lead Times 2026
- Solid-state transformers and wide-bandgap semiconductors (SiC, GaN) targeting 2028 for power distribution. Switchgear, Transformer & Generator Lead Times 2026
- GE Vernova booked $800M in US transformer orders in H1 2026, with significant growth in US substations, switchgear and transformers to support data center development. GE Vernova 2Q 2026 Earnings Call Webcast Transcript
- Siemens Energy Grid Technologies orders rose 27.6% comparable to €5.4B in Q3 FY2026, with the major contribution from substantial transformer business growth including data center projects; order backlog reached €51bn and book-to-bill was 1.48. Pre-close call: ~€2bn of US data-center-related orders booked in H1 FY2026. Siemens Energy Earnings Release Q3 FY 2026
- Siemens Energy states growth remains limited by execution capacity rather than demand, and will increase capacity for Large Power Transformers and Switchgears by 50% between 2026 and 2030; a further big step-up in capacity additions is planned for 2029-2030. Siemens Energy Q3 FY2026 Pre-Close Call Transcript
- Hitachi Energy announced (15 Sep 2026) a new $528 million transformer factory in Gallman, Mississippi — its largest single US investment to date — as the cornerstone of an ~$1.5 billion US manufacturing expansion program (with other major projects in Pennsylvania, Virginia, Tennessee). The new plant is >2x the size of the Crystal Springs facility and will more than double production capacity; construction starts late 2026, transformer production scheduled to start 2029. Hitachi deepens commitment to U.S. manufacturing with $528 million Mississippi transformer factory
Grid-interconnect
147 events in dataset
Nog onvoldoende datapunten voor een trendlijn (serie heeft minder dan 2 punten). Het signaal staat hieronder in de events.
Signalen: $1.5B annual liquid cooling revenue base from Boyd Thermal adds ~5K employees across NA/EU/Asia; Eaton's combined power + cooling scale creates #2 position after Vertiv, $1.84B+ announced across 7 US plants in 8 weeks; earliest production 2028 (Southwire) / 2029 (Hitachi Gallman), $115B by 2027 implies ~$57B compound growth from FY2025 baseline; networking and interconnect represent an increasing share as GPU-count-per-cluster grows non-linearly, $2.6B total potential consideration signals how expensive 'time to power' has become; complements Eaton↔Siemens Energy modular onsite-power partnership, $200M combined announced DOE uplift awards in October 2026 (reconductoring/DLR + AK line), $600M+ optical revenue target for FY2027 across AEC, DSP, and optical engine product lines; suggests optical segment alone approaching $300M+ annually, $720B Goldman Sachs grid-upgrade estimate; 200 GW global data center capacity by 2030 doubles current ~100 GW, +1,070 MW contracted distributed capacity; PJM DR revenue $523.5M 2025 (+240% YoY)
- B200 Blackwell in mass production through Q1–Q2 2026; GB200 NVL72 rack-scale shipping. $30K–40K per unit. AMD vs NVIDIA AI GPU Market Share 2026: MI350X vs B200 Competitive Analysis
- NVLink 5.0 providing 1.8 TB/s bidirectional bandwidth between GPUs in GB200 NVL72. NVIDIA GB200 Supply Chain: The Global Ecosystem Explained
- Copper cable assemblies for GPU-rack interconnect lead time 20-30 weeks; constrained by TE Connectivity and Amphenol capacity. NVIDIA GB200 Supply Chain: The Global Ecosystem Explained
- Huawei shipped its Ascend 910C AI chip to select Chinese customers in Q2 2026, positioned as its highest-volume domestic AI training chip. The 910C features improved interconnect (Huawei HCCS) compared to the 910B series, enabling multi-chip training clusters that can compete with Nvidia H100 cluster workloads. ByteDance and several large model firms reportedly integrated the 910C into their training pipelines alongside existing Nvidia H20 inventory. Huawei also announced plans for a subsequent generation 'Ascend 920' targeting 2026–2027. Huawei Ascend 910C alleged specs suggest it a tough rival to Nvidia H100 — Huawei Central (via Google News)
- Production shipments of NVIDIA's Vera Rubin platform are scheduled to begin fall 2026 at all eight confirmed cloud partners - AWS, Google Cloud, Microsoft Azure, Oracle Cloud, CoreWeave, Lambda, Nebius, and Nscale. NVIDIA confirmed the platform entered full production on June 1, 2026 (GTC Taipei keynote). The 10x inference token-cost reduction rests on tripling per-GPU memory bandwidth to 22 TB/s with HBM4 and doubling rack interconnect to 260 TB/s with NVLink 6; Vera Rubin also powers upcoming supercomputers at LRZ/NERSC/Los Alamos (announced June 22 at ISC Hamburg). TechTimes: NVIDIA Vera Rubin Ships This Fall: 8 Cloud Partners, 10x Lower Token Cost, HBM4 Triples Bandwidth
- Intel announced Wednesday (21 Oct 2026) that production-equipment installation has begun at Arizona Fab 62, with mass production targeted for late 2028 (Tom's Hardware via TrendForce, 22 Oct 2026). Fab 62 could support 18A, 18A-P and 14A nodes — either expanding 18A capacity (Fab 52 already makes 18A chips) or hosting next-gen 14A; Intel has not confirmed node assignment. At Ohio One (New Albany) the first major structural truss for the site's first fab was recently installed; opening was pushed to 2030-2031 and CEO Lip-Bu Tan ties projects to secured customer demand. CNBC reports Intel is in talks with SK hynix about leasing part of the Ohio facility for US memory production (unconfirmed by either company). Intel has also integrated High-NA EUV into high-volume manufacturing (first commercial High-NA scanner at D1X, used for Panther Lake). Advanced packaging spans Oregon (R&D, 3D interconnects), Arizona (packaging R&D, assembly, test) and New Mexico (multi-chiplet production using EMIB, next-gen EMIB-T and Foveros); a new photomask fab broke ground in Santa Clara. Intel cites 4,500+ US suppliers in 49 states and >US$29B annual domestic spending. [News] Intel Begins Equipment Installation at Arizona Fab 62, Targeting 2028 Operations; Ohio One Advances (TrendForce)
CDU / vloeibare koeling
46 events in dataset
Nog onvoldoende datapunten voor een trendlijn (serie heeft minder dan 2 punten). Het signaal staat hieronder in de events.
Signalen: $1.5B annual liquid cooling revenue base from Boyd Thermal adds ~5K employees across NA/EU/Asia; Eaton's combined power + cooling scale creates #2 position after Vertiv, +150% modular hydronic/electric plant production capacity, Phoenix facility 17,750 m2, 1.8 MW/CDU in-row cooling at N+1; NVL72 rack = 20.7 TB HBM4; scalable-unit blueprint to GW scale, 100% liquid-cooled reference architecture shifts all new Rubin builds to DLC; Motivair (Schneider) cited as a long-standing design partner on the roadmap, 144 GPU die assemblies per rack + 36 CPU + full liquid cooling = one of the most complex manufactured systems ever produced at scale. Initial deployment limited by cooling infrastructure readiness at customer sites., 2.3 MW single-CDU class supports multi-megawatt AI halls in a small footprint — reduces CDU count and floor space, a deployment-speed lever for operators, 2.3 MW single-unit liquid-to-liquid CDU capacity point now qualified for DSX AI-factory reference designs, 20 MW -> 84 MW planned with GB300 NVL72 liquid-cooled racks; on-site fuel cells bypass incremental utility interconnect
- HPE Cray EX254n: 8U supercomputing blade with 8× NVIDIA Grace Hopper GH200 liquid-cooled GPUs. 8x NVIDIA Grace Hopper Superchips in a Blade — HPE Cray EX254n at GTC 2024
- Dell PowerEdge XE9680: 8× NVIDIA B200 GPU server; liquid-cooled variant available. Dell PowerEdge XE9680 Specifications
- Lenovo ThinkSystem SR670 V3: 8× NVIDIA B200 GPU server; liquid-cooled option. Lenovo ThinkSystem SR670 V3 Specifications
- The Vera Rubin NVL72 configuration integrates 72 Rubin GPU packages (each containing 2 GR100 chiplets = 144 GPU dies) and 36 Arm-based Vera CPUs connected via NVLink-C2C bridge. Per-rack performance targets: 3.6 EFLOPS NVFP4 inference, approximately 1.2 EFLOPS FP8 training. Full liquid cooling required — air cooling limit is 15-20 kW while NVL72 draw reaches 130-140 kW. Single-rack inference cost claimed at 10x lower token cost vs Blackwell. First ODM pilot production at Foxconn/Wistron mid-2026, initial rack deliveries to North American hyperscalers began July 2026, with broader commercial availability targeting September 2026 through providers like Axe Compute. NVIDIA's Vera Rubin Makes Your Current AI Infrastructure Obsolete — TechFastForward
- Production shipments of NVIDIA's Vera Rubin platform are scheduled to begin fall 2026 at all eight confirmed cloud partners - AWS, Google Cloud, Microsoft Azure, Oracle Cloud, CoreWeave, Lambda, Nebius, and Nscale. NVIDIA confirmed the platform entered full production on June 1, 2026 (GTC Taipei keynote). The 10x inference token-cost reduction rests on tripling per-GPU memory bandwidth to 22 TB/s with HBM4 and doubling rack interconnect to 260 TB/s with NVLink 6; Vera Rubin also powers upcoming supercomputers at LRZ/NERSC/Los Alamos (announced June 22 at ISC Hamburg). TechTimes: NVIDIA Vera Rubin Ships This Fall: 8 Cloud Partners, 10x Lower Token Cost, HBM4 Triples Bandwidth
- 50kW per rack becoming standard for AI workloads; requires liquid cooling. Switchgear, Transformer & Generator Lead Times 2026
Server CPU / OEM
119 events in dataset
Signalen: $1.2B named rack order; >$3B FY26 AI-networking orders guided; Helios ramp starts FY27, $3.45B equity raise + FY26 outlook raise to $21.0B revenue: equity-funded AI-rack build capacity continues, $54B AI server backlog vs ~$8.5B/quarter run-rate implies >6 quarters of visibility for Lenovo's server lines, $942M H2-26 capex + NT$15B converts + $1.5B loan = multi-billion USD committed expansion pipeline, $95B AI server backlog signals multi-year delivery pipeline; Dell entering Q3 FY2027 with record backlog, suggesting capacity utilization at or near maximum, $95B backlog + $192B guided revenue implies Dell is scaling production to fill ~$74B AI server revenue run-rate; COO warned investors of "unprecedented" memory shortage affecting cost basis, $95B backlog = 1.28x the raised $74B annual AI-server sales forecast — over a year of forward demand visibility, +1.1 GW near-term firm capacity from existing fleet; UPRISE program now ~4.7 GW identified vs 5-GW target
- Grace Hopper GH200 superchip shipping; 141GB HBM3e unified memory. Used in HPE Cray EX254n for exascale systems. 8x NVIDIA Grace Hopper Superchips in a Blade — HPE Cray EX254n at GTC 2024
- HPE Cray EX254n: 8U supercomputing blade with 8× NVIDIA Grace Hopper GH200 liquid-cooled GPUs. 8x NVIDIA Grace Hopper Superchips in a Blade — HPE Cray EX254n at GTC 2024
- Dell PowerEdge XE9680: 8× NVIDIA B200 GPU server; liquid-cooled variant available. Dell PowerEdge XE9680 Specifications
- Lenovo ThinkSystem SR670 V3: 8× NVIDIA B200 GPU server; liquid-cooled option. Lenovo ThinkSystem SR670 V3 Specifications
- Supermicro AS-4124GO-NAT: 4× NVIDIA B200 GPU server; air-cooled variant. Supermicro AS-4124GO-NAT Specifications
- GB300 Blackwell Ultra unveiled October 2025: 160 streaming multiprocessors vs 144 on GB200, 20,480 total CUDA cores. Targeted for late-2026 rack-scale shipping with 288 GB HBM3e per GPU. DGX Station version priced at $92,900 with 748 GB unified memory, 20 PFLOPS FP4 performance, trillion-parameter model support. Nvidia GB300: 288 ГБ, PCIe 6, 1400 Вт — specs detail — Habr