Nieuws — events, nieuwste eerst705 events
Nieuws
Alle 705 events, nieuwste eerst. Filter op tag.
KT Cloud and Vertiv announced (28 Oct 2026) a 'Strategic Technology Partnership for Next-Generation AI Data Centers': KT Cloud will deploy Vertiv modular infrastructure solutions - including the SmartRun prefab power/cooling system - at its Gunsan AI Data Center in South Korea, expanding cooperation on core AIDC infrastructure technologies.
At Oracle AI World 2026 (Las Vegas), Oracle unveiled new accelerator choices for OCI (The Next Platform, 27 Oct 2026): (1) Google's Tensor Processing Units will be available on OCI next year — the first time Google's specialized AI chips are offered in a hyperscale environment other than Google's own; TPU targets LLM training/post-training and reasoning/agentic inference. Google's eighth-generation TPU line (April 2026) split into TPU 8t (training) and TPU 8i (inference). (2) Cerebras fast-inference hardware: CS-4 AI accelerator (introduced Aug 2026; rack-scale with three Wafer Scale Engine 3 Turbo chips; up to 30x tokens-per-second-per-user vs GPUs per Cerebras) arrives in OCI in 2027. (3) SambaNova fifth-generation SN50 Reconfigurable Dataflow Units (released Feb 2026), designed to pair with GPUs — GPUs handle prefill, RDUs handle decode. Context: OCI revenue $7.4B in Q1 FY2027 (+121% YoY), remaining performance obligations $664B (+$209B YoY), FY26 capex $55.7B with FY27 possibly beyond $90B; OCI SVP Nathan Thomas: 'we partner very deeply' with Nvidia and AMD and want a wide range of options.
Oracle is Expanding the Accelerator Options in OCI for Enterprises (The Next Platform)
TrendForce (27 Oct 2026) on two reported NVIDIA memory-strategy shifts: (1) Samsung and SK hynix are preparing to ship 8-Hi HBM4 for Rubin Ultra in 2Q27, departing from the original 12-Hi HBM4E plan (ZDNet); talks gained traction in 3Q26; a mix of 8-Hi HBM4 and 8-Hi HBM4E is expected across Rubin Ultra NVL rack configurations by rack type. Both vendors completed NVIDIA qualification for 12-Hi HBM4 but 8-Hi products need separate certification. TrendForce: 8-Hi uses fewer DRAM dies, cutting cost per stack and GPU BOM and letting the same HBM bit output support more GPUs, so 8-Hi evaluation is expected to be a priority for many GPU/ASIC vendors in 2027; 12-Hi HBM4 remains mainstream in 2026 with 8-Hi adoption rising sharply from mid-2027. Economy Tribune (citing Daishin Securities) suggests Anthropic's in-house chip could also use 8-Hi HBM4, implying Samsung supplies both foundry and memory. (2) Wccftech: NVIDIA quietly trimmed Vera Rubin HBM4 bandwidth specs by ~13% per Rubin GPU and up to 26% for HGX Rubin NVL8 systems (changes dating to September). An archived Sept 1 spec listed Vera Rubin NVL72 at 20.7 TB HBM4 / 1,580 TB/s, the Superchip at 576 GB / 44 TB/s and each Rubin GPU at 288 GB / 22 TB/s before the cuts.
Supply-chain reporting relayed by TrendForce (27 Oct 2026): TSMC is reportedly weighing CoWoS price hikes of up to 20%, with foundries and OSATs considering a further 10-20% round across CoWoS-type packaging (Commercial Times); some packaging lead times already stretching. Test time is an emerging bottleneck: Nomura (via Seoul Economic Daily) estimates NVIDIA Blackwell final testing takes ~4x Hopper's and Rubin could require ~7x. Packaging-equipment lead times have lengthened to six to eight months from three to six, with customers prioritizing delivery over price; TSMC itself reportedly placed equipment orders at existing prices this year without its customary discount/feature demands. Separately (Liberty Times), TSMC has finalized wafer-out price increases of 3-6% from January 2027 by node, steepest at 2nm/3nm. Capacity: analysts cited by Commercial Times estimate monthly CoWoS capacity exceeding 130,000 wafers in 2026 with a 200,000 target for 2027; Goldman Sachs projects 1.275M CoWoS wafers in 2026 (~106K/month), +114% YoY to 2.73M in 2027 (227.5K/month) and 3.48M (~290K/month) in 2028.
Applied Materials plans to more than double its stake in Dutch packaging-equipment maker BE Semiconductor Industries from 9% to ~20% via an accelerated share purchase, subject to U.S. antitrust review; AMAT (largest shareholder since April 2025) says it will not go beyond ~20% or seek board seats (Reuters via TrendForce, 27 Oct 2026). The partnership targets next-gen interconnect/packaging for AI semiconductors — hybrid bonding joins copper interconnects and dielectrics bumplessly for denser die stacking; the two have jointly developed the Kinex die-to-wafer hybrid-bonding system. BESI holds ~70% of the global hybrid-bonding equipment market (ijiwei); its Datacon 8800 CHAMEO ultra plus AC runs 100 nm alignment at up to 2,000 cycles/hour. Demand momentum: BESI Q3 2026 orders rose 96% YoY to EUR 342.6M (Reuters, 22 Oct), hybrid-bonding orders in the first nine months of 2026 more than doubled the full-year 2025 level, customer base reached 23 and cumulative equipment demand exceeded 215 systems (BESI earnings). Chinese challenger Piotech booked RMB 136M hybrid-bonding revenue in 2025 (+41.9% YoY) with W2W repeat orders.
DOE offered Constellation Energy a conditional loan commitment of up to $5B (announced Oct 27) to finance nuclear uprates and modernization at eight plants (15 reactors, ~15.5 GW attributable) in Illinois, Maryland and Pennsylvania, adding nearly 1.1 GW of generating capacity; the second major nuclear uprate financing offer in a month, bringing DOE's UPRISE program to ~4.7 GW of identified potential capacity additions against a 5-GW target.
HD Hyundai Electric's Q3 2026 regulatory filing (Oct 27) reports record quarterly revenue of KRW 1.1799T (~$888.8M, +18.5% y/y) and operating profit of KRW 284.5B (+15.1%, 24.1% margin); distribution equipment & new-business revenue surged +85.8% y/y on AI-datacenter/semiconductor distribution-transformer demand and North American sales; order backlog reached $8.843B (+26.6% y/y, +4.2% q/q) despite Q3 new orders of $980M (-19.1% on H1 base effect from $3.231B first-half bookings); alongside results the company disclosed a KRW 198B (~$149M) phase-2 Cheongju distribution campus expansion over 26 months targeting +KRW 250B (~$188M) annual revenue.
Antares Nuclear signed a multi-year agreement (announced Oct 27) with BWX Technologies' Advanced Fuels to fabricate TRISO fuel for its Mark-1 electricity-producing microreactor through the end of the decade, with an initial allocation of one metric ton uranium and scale-up provisions; this follows DOE approval of the Mark-1 Nuclear Safety Design Agreement on Oct 23, clearing the path to operation at Idaho National Laboratory in 2027 and first US military-site deployments from 2028.
CenterPoint Energy added $4B to its 10-year capital plan on its Q3 2026 earnings call (27 Oct 2026), to $70.7B total through 2035 ($51.5B electric, $19B gas), and accelerated ~$1.5B of capex into the first half of the plan — a $4.5B increase to planned investment through 2030 — to serve ~14 GW of new demand expected online through 2030.
CenterPoint expands 10-year capital plan by $4B to handle load growth (Utility Dive)
BloombergNEF raised its 10-year US energy-storage deployment forecast by ~30% versus six months ago despite trade barriers and project cost inflation: cumulative US deployments of 213.9 GW by 2030 and 434.4 GW by 2036; global deployments of 3,098 GW / 11,106 GWh by 2036, roughly 10x today's levels. Near-term US drivers: colocated storage economics with renewables and data-center appetite for behind- and front-of-meter capacity.
BNEF ups US energy storage deployment forecast by 30% (Utility Dive)
PJM urged FERC on 26 Oct 2026 to dismiss Monongahela Power's 6 Oct complaint seeking to recoup $145M of losses incurred during Winter Storm Fern, calling it 'a fishing expedition without substance' and attributing the utility's LMP differences mainly to downed lines on Mon Power's own system; PJM Power Providers Group and EPSC said the requested retroactive rerun/resettlement of cleared market outcomes 'violates core principles of market finality' and should be rejected.
PJM urges FERC to dismiss Mon Power 'fishing expedition' seeking to recoup $145M (Utility Dive)
Vertiv completed (PR dated 27 Oct 2026, filed same day on 8-K) its ~$1.45 billion upfront-cash acquisition of Utility Innovation Holdings (UtilityInnovation Group), plus up to $1.15 billion contingent EBITDA earn-out over 12/24 months. Price is ~13x expected UIG 2027 EBITDA; Vertiv expects first-year adjusted-EPS accretion. Strategically, Vertiv extends its power/cooling portfolio 'from grid interconnect to chip', adding microgrid controls and behind-the-meter generation orchestration to help data-center operators 'secure power faster as grid constraints increasingly limit AI infrastructure deployment'.
SpaceX has secured a binding commitment for an approximately $40 billion master finance lease covering NVIDIA Vera Rubin GPUs and related equipment, financed through an entity managed by Apollo Global Management with Goldman Sachs as sole structuring agent and anchor commitments from BlackRock-managed funds/accounts, Goldman Sachs Asset Management and KKR-advised capital accounts; SpaceX calls it the largest committed GPU financing to date. Closing is expected before end-2026 subject to conditions; lease schedules commence as GPU equipment is delivered in the first half of 2027. The deployment supports SpaceX's terrestrial AI compute expansion including infrastructure for Grok and cloud services. GPU count, data-center locations and power requirements were not disclosed; SpaceX's CFO cited sub-one-year payback on superintelligence infrastructure without supporting calculations.
SpaceX Secures $40 Billion Financing for NVIDIA Vera Rubin (Converge Digest)
Anthropic, PBC filed Form S-1 with the SEC on 26 Oct 2026 (accession 0001628280-26-067168). Verbatim from the primary document: Anthropic holds 'master equipment leases with Broadcom, pursuant to which we receive such hardware for deployment in dedicated data center facilities operated by others on our behalf'; the arrangements are accounted for as finance leases and 'our rent payments over the life of these leases are equal to an aggregate of approximately $161.2 billion excluding interest, with rent payable in quarterly installments and payment obligations typically commencing upon AI infrastructure equipment becoming ready for service.' The S-1 further states Anthropic trains and serves Claude 'natively across four major chip families (AWS Trainium, Google TPUs, NVIDIA GPUs, and AMD GPUs) using a proprietary orchestration layer,' has 'consistently operated in a compute-constrained environment where customer demand exceeds available compute,' and committed to procure additional TPU compute from TPU providers including Google for delivery between 2028 and 2031. Trade press (Converge Digest, 26/27 Oct) characterizes the Broadcom-leased hardware as Google TPU infrastructure.
Anthropic, PBC — Form S-1 registration statement (filed 2026-10-26)
TrendForce preview (26 Oct 2026), citing Newsis/Chosun Biz: market consensus projects SK hynix 3Q26 consolidated revenue KRW 98.23T and operating profit KRW 76.96T (+301.76% / +576.11% YoY), a ~78.4% operating margin and an all-time high, with Q4 operating profit projected at KRW 87.27T (FnGuide consensus); earnings report 28 Oct. Brokerage estimates split by ~KRW 10T (SK Securities KRW 79T vs DS Investment KRW 70.1T vs Meritz KRW 80.1T). Samsung has already guided 3Q26 operating profit ~KRW 107.4T (+782.5% YoY). Separately, SK hynix disclosed it has shelved its US subsidiary Solidigm's IPO, raising questions over financing of future US NAND capacity expansion; reported options include internal cash, borrowing, joint ventures and customer-backed financing tied to long-term supply agreements (Hankook Ilbo, The Bell via TrendForce); a memory fab can cost tens of trillions of won, making JV/CSP-partner financing a live path.
Chosun Biz, citing industry sources (via TrendForce, 26 Oct 2026): SK hynix's HBM4 shipments may have fallen short of expectations for a second consecutive quarter in 3Q26, amid delays in NVIDIA's Rubin rollout and longer-than-expected performance tuning for mass-produced chips. Offsetting factor: pricing — JPMorgan (via Money Today) estimates HBM4 ASPs rising 121% by 2027 vs 2025 levels. Broader ASP momentum per DealSite citing Hana Securities: after 2Q26 jumps (DRAM +~30%, NAND +mid-50s% QoQ), 3Q26 DRAM ASPs are expected up another ~21% QoQ (potentially lifting the DRAM segment operating margin to ~83%) with NAND ASPs +~17% QoQ.
Clients are locking in multi-year memory supply amid tightening availability (via TrendForce, 26 Oct 2026): Samsung reportedly plans to allocate 60-70% of its memory production capacity to long-term contracts; SK hynix disclosed at its July earnings call that it had signed multi-year deals with ~10 customers including major clients. Business Post, citing KeyBanc: HP and Dell have extended memory supply agreements with Micron through 2031, while Alphabet is seeking contracts with memory suppliers lasting until 2035. These moves strengthen demand visibility for Samsung and SK hynix as hyperscalers/OEMs secure supply years in advance.
TrendForce News preview (26 Oct) ahead of SK hynix 3Q26 results (28 Oct): Korean market consensus sees revenue KRW 98.23T and record operating profit KRW 76.96T (+301.8% / +576.1% Y/Y; ~78.4% margin), brokerage spread KRW 70.1-80.1T. On NAND specifically: after 2Q26 NAND ASPs surged mid-50% QoQ, 3Q26 NAND ASPs are projected +17% QoQ (Hana Securities via DealSite); DRAM ASP +21% QoQ forecast. LTA wave: Samsung 'reportedly plans to allocate 60-70% of its memory production capacity to long-term contracts'; SK hynix already has multi-year deals with ~10 customers; KeyBanc (via Business Post) reports HP and Dell extended memory commitments; Solidigm's shelved IPO adds capex-allocation questions for the call.
11 of the 15 states in MISO's footprint, plus clean-energy trade groups and ratepayer advocates, urged FERC (Friday 23 Oct 2026 filings, reported 26 Oct) to reject MISO's LARS proposal to fast-track combined large-load and generation interconnection studies, arguing it lacks sufficient customer protections and favors vertically integrated utilities over independent power producers; Duke Energy Indiana, Entergy and other utilities backed the filing.
Clean energy groups, states oppose MISO's large load-generation study plan (Utility Dive)
US orders for wind turbines reached 2.8 GW in Q2 2026 as developers raced to commence construction before the 4 July 2026 tax-credit deadlines moved up by the One Big Beautiful Bill Act (45Y PTC / 48E ITC eligibility); the US installed ~5 GW of onshore capacity in Q2 2026, most of it the 3.65 GW SunZia project in New Mexico — the largest US onshore project to date.
Wind turbine orders hit 2.8 GW in Q2 ahead of tax credit deadline: WoodMac (Utility Dive)
Celestica Q3 2026 (quarter ended 30 Sep, PR 26 Oct 2026): revenue $5.83B, up 83% YoY vs $3.19B; adjusted EPS (non-GAAP) $3.30, up 109% YoY; both above the high end of guidance; adjusted operating margin 8.5% vs 7.6% a year ago. Company also completed a public offering for ~$3.45B aggregate gross proceeds and raised its FY2026 outlook to $21.0B revenue (from $20.5B) and $11.55 adjusted EPS (from $11.30), with Q4'26 revenue guidance of $6.275-6.575B.
Celestica Announces Third Quarter 2026 Financial Results and Will Hold 2026 Investor and Analyst Day
Celestica's initial 2027 annual outlook (given 26 Oct 2026 with Q3 results): revenue of $40.0 billion (+90% YoY), record adjusted operating margin (non-GAAP) of 9.0%, adjusted EPS $21.50, and capital expenditures of $1.5-2.0 billion. At the same time FY2026 free-cash-flow guidance was CUT from $600M to $460M, attributed to 'targeted working capital investments to support the accelerating growth in 2027' - i.e. component inventory build ahead of the 2027 ramp.
Celestica Announces Third Quarter 2026 Financial Results and Will Hold 2026 Investor and Analyst Day
Vertiv announced (8-K + EX-99.1, event 26 Oct 2026) it has COMPLETED its acquisition of King Environmental Services Ltd (KES), a Europe-based provider of fluid management, commissioning and load-testing services for high-density liquid-cooled data centers. The close (announced as pending since 24 Sep 2026) extends Vertiv's fluid-management services platform into EMEA, building on the 2025 PurgeRite acquisition in North America.
Sankyung Today, 24 Oct 2026 (via VLSI Korea Daily Silicon digest, 25 Oct): Samsung Electronics is developing zHBM, which stacks memory vertically with AI compute, targeting up to 8x performance per GPU and 3x performance per watt versus next-generation HBM5 — targets that still require commercial validation; a high-capacity high-speed NAND technology, zNAND-O, is in parallel development. SK hynix is countering on 3D DRAM and high-bandwidth flash (HBF). The digest frames the competitive axis as shifting from per-product speed to power efficiency and total cost of the whole AI server, with low-power server modules (SOCAMM2) and CXL memory modules as related battlegrounds; SK hynix and Samsung earnings due 28-29 Oct are the next verification points.
NotebookCheck roundup (25 Oct, citing TrendForce and manufacturer statements): 256GB TLC SSD contract price reached ~US$134.20 in 3Q26 vs $33.30 a year earlier (+303% Y/Y, ~+20% QoQ); 512GB equivalents rose ~$62.40 → $252; a 1-Tb TLC NAND component +123% in the six months to 21 Oct. TrendForce forecasts 4Q26 NAND contract prices +15-20% QoQ and enterprise SSD +23-28% QoQ on AI-inference buildout, with eSSD demand (capacity basis) growing >80% Y/Y in 2026. Allocation statements: SK hynix reportedly sold out its NAND production through end-2026; Micron says >75% of expected FY2027 output already committed; SanDisk capacity remains ~30% below 2022 peak after retiring ~560K wafers/month; meaningful new supply from SK hynix/Micron not before mid-to-late 2027.
NotebookCheck — SSD prices explode by 303% in a year as AI fuels NAND flash shortage
Samsung has placed a new equipment order that will enable 6,000 additional 2nm wafers per month at its Taylor, Texas fab, building on earlier procurement aimed at 10,000 WSPM; Taylor Fab 1 (EUV-equipped, 2nm) is expected to become operational by end-2026, and ZDNet Korea sources say the site should be capable of producing up to 20,000 wafers per month — pilot capacity included — by the second half of 2027 (wccftech relay, 23 Oct 2026). Contrast: TSMC's Arizona sites target 3nm production by end-2027 and 2nm by decade end, so Taylor could give Samsung the first US-produced 2nm wafers at volume.
Western Digital FQ1 FY2027 (reported 22 Oct after close, B&F write-up 23 Oct): revenue $4.38B, +55% Y/Y, above the $4.1B outlook; exabytes shipped 248 EB, +21.6% Y/Y and up from 231 EB the prior quarter; revenue per exabyte shipped $17.7M, +28% Y/Y and +8.8% QoQ — explicit price re-rating. Cloud = 88% of revenue; 90% of shipped drives were mass-capacity nearline. Non-GAAP gross margin 57.8% vs 50.2% prior quarter. Tech transitions: next-gen 40TB ePMR platform ramping to ~30% of nearline shipments by FQ2 FY27 and ~50% by FQ3 FY27; first-gen 44TB HAMR finishing qualification with 2 hyperscalers in CY1 2027 with volume shipments planned CY1H 2027, 2 more customers qualifying in CY1H27; second-gen HAMR (5TB/platter, 11-disk, ~55TB) qualification starts CY2H 2027; high-bandwidth HDD sampling with multiple customers.
SK hynix confirmed 23 Oct 2026 it 'will halt its review of potential pre-IPO and IPO options' for Solidigm, which had been explored to finance a new US NAND fab (estimated $15-30B by B&F). Statement: the review was 'discontinued in light of market conditions and share price trends' while 'the need to invest in Solidigm's growth remains unchanged'. Contributing factors: SK hynix shareholder dilution concerns, semiconductor-stock correction after the 20-Oct rout, and Korean regulatory debate over listing a subsidiary of a listed parent; Goldman Sachs / Morgan Stanley underwriter work reported earlier has come to naught. Solidigm NAND is made at Dalian (Fab1 192L; Fab2 equipping limited to ~240-layer tools under US export rules), 2+ generations behind the 270-300L frontier.
The US Senate is working toward a 9 Nov 2026 floor vote on the bipartisan permitting deal (BAAJA); NRDC and the Sierra Club are withholding endorsement — praising the bill's transmission planning/permitting changes as 'a clear win for clean energy' but objecting to 'project certainty' provisions and limits on state/court blocking — while developers and abundance movement groups (Third Way, Searchlight Institute) rally behind it; a bipartisan Group of Four is weighing technical fixes via a managers amendment.
Why green groups won't endorse the Senate permitting deal (Latitude Media)
TrendForce AI-server industry research (Press Center, 22 Oct 2026): server CPU shipments projected +31% YoY in 2026 and over +40% YoY in 2027, driven by AI-inference/agentic workloads shifting the server CPU-to-GPU/ASIC ratio from the traditional 1:8 toward 1:2 or even 1:1. x86 (Intel+AMD) held nearly 90% of the 2025 market. Intel's next-gen product shipments are delayed by process and packaging challenges; it relies on existing Emerald Rapids and Granite Rapids lines while limited internal manufacturing capacity constrains shipments and prevents fully meeting customer orders, contributing to share decline. AMD's next-gen Venice processors entered small-scale mass production in 2026 with Genoa/Turin still the primary shipment drivers; despite tight TSMC capacity limiting AMD's ability to take additional orders, TrendForce estimates AMD's server-CPU share at ~27% in 2026, rising further on new products planned for 1H27. Arm-based CPUs are expected to capture ~30% of the server CPU market in 2027, led by a substantial NVIDIA Vera CPU ramp plus high-volume Google Axion, AWS Graviton, Microsoft Cobalt 100, Arm AGI CPU and Huawei production, and Alibaba's new Yitian 720/730.
Supply-chain checks indicate Google's MediaTek-designed TPU v9 variant 'HumuFish' remains '100%' committed to Intel's EMIB-T advanced packaging despite industry-source reports that EMIB-T substrate yields were ~35% in 3Q26 versus a 50% year-end target (Common Wealth Magazine via TrendForce, 22 Oct 2026). Yield problems raised the possibility of a CoWoS-L switch, which the checks deny. Root causes: CTE mismatch between ABF substrates and embedded silicon bridges complicating alignment, and NCF air-bubble entrapment under bridge TSVs; Intel has adjusted its substrate electroplating process while Ibiden, Unimicron and Shinko work on improvements. The TPU v9 lineup is confirmed as three variants: Broadcom-designed WhaleFish (TSMC CoWoS-L, mass production 2H27), MediaTek-designed HumuFish (exclusively Intel EMIB-T, mass production 1H28), and TriggerFish (also EMIB-T) — per analyst Ming-Chi Kuo via TechNews, TriggerFish upgrades HumuFish with 2-3x SRAM, an added simulation die, and HBM4E instead of HBM4. Google-designed Compute Die A on TSMC N2 tape-out reportedly pushed back to late October with other dies on schedule; Commercial Times reports the MediaTek-included HumuFish tape-out could slip to January 2027, delaying shipments to 1Q28.
Samsung Electro-Mechanics is reportedly set to receive a combined KRW 3.7 trillion in investment commitments from Tesla (~KRW 1.1T) and NVIDIA (~KRW 2.6T per Maeil Business Newspaper) to expand advanced FC-BGA substrate production. Tesla signed a long-term FC-BGA supply agreement — five years, 2029-2033 — for substrates for AI chips used in its autonomous EVs and Optimus humanoid robots; funding feeds SEMCO's KRW 4.27T Sejong investment plan: plant construction complete end-2027, equipment install and trial operations 1H28, mass production 2H28, deliveries to Tesla from 2029. NVIDIA separately aims to raise total IC substrate capacity 2.5x over three years: billions of dollars in capacity-reservation prepayments to Japan's Ibiden (Ono plant targeting tripled production), and SEMCO's Sejong+Busan expansion in Korea with NVIDIA expected to take a significant share of SEMCO's annual package-substrate capacity by 2028. SEMCO also disclosed a KRW 360.5B contract on 19 Oct 2026 to supply memory-package substrates to an undisclosed global customer through 2027. High-end FC-BGA remains concentrated in SEMCO, Ibiden (>70% estimated share) and Unimicron.
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.
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).
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.
PSMC (Powerchip) reported Q3 2026 revenue NT$21.3B (+23% QoQ, +80% YoY), net profit NT$6.5B (+98% QoQ, +338% YoY) and 48% gross margin (+20pts), and plans further mature-node wafer price hikes of 10-20% in 2027 while negotiating long-term agreements to lock pricing and volumes (earnings presentation via TechNews/Liberty Times, TrendForce 22 Oct 2026). Logic 8-inch/12-inch wafer-input ASPs already rose 8-10% in Q3; President Martin Chu said completed memory wafers were priced >30% below wafers entering production, signaling an in-flight ASP repricing, with Q4 revenue up double digits. Supply drivers: major DRAM makers shifting capacity to HBM and exiting specialty DRAM keeps PSMC's specialty-DRAM foundry prices firm; MLC NAND tightness pushes high-capacity SLC demand (PSMC shifting SLC platform to 24nm); AI servers require 3-5x more NOR Flash per system than conventional servers and PSMC is the only NOR-flash foundry outside China; AI-server/automotive PMIC demand has customers seeking three-year contracts, with fully-utilized 2026 capacity squeezing smartphone PMIC output. Roadmap: 1p DRAM collaboration with Micron (PDK end-2026, equipment 1Q27, trial 2Q27); Micron Post-Wafer-Finish project pilot line complete end-2026, mass production 4Q27; IPD orders visible into 2028; wafer-on-wafer hybrid-bonding trials for AI inference.
Nikkei (via TrendForce, 22 Oct 2026): Micron president & CTO Scott DeBoer said in a technology webinar that 4F2 DRAM cell architecture is 'not on the roadmap right now' after more than 20 years of evaluation, even as Chinese rival CXMT advances it — Micron will instead extend 6F2 and develop 3D DRAM longer-term, citing capacitor-pitch (not cell architecture) as the density driver plus 4F2 scaling limits, complexity and cost. Node plan: 1-gamma already in high-volume manufacturing (underpinning next 12 months of products); 1-delta volume production targeted 2H27 with mid-20% bit-density gain, supporting DDR6, LPDDR6, HBM5 and vertical HBM (vHBM); DDR6/LPDDR6 target ~2x DDR5 bandwidth and +35% power efficiency vs LPDDR5. 1-epsilon node is being developed with wafer bonding. A 1TB DDR5 RDIMM is planned following 512GB module sampling.
Via TrendForce (22 Oct 2026, citing MarketBeat from Micron's technology webinar): Micron is developing vertical HBM (vHBM) — memory stacked directly on processing chips for substantially higher bandwidth — alongside standard and custom HBM; thermal management remains the challenge; packaging work includes hybrid bonding for tighter pitches, taller HBM stacks and improved cooling. On NAND, Micron plans to ramp Generation 11 (G11) in 2H27, targeting 2x bit density vs G9, 1.5x I/O speed, lower transfer energy, wafer-on-wafer bonding and 400+ layers.
TrendForce (22 Oct 2026, citing TechNews/Liberty Times): PSMC 3Q26 revenue NT$21.3B (+23% QoQ, +80% YoY), net profit NT$6.5B (+98% QoQ, +338% YoY), gross margin 48% (+20pp), driven by rising memory prices and tight foundry capacity as major DRAM makers shift wafers to HBM and exit specialty DRAM. Q3 completed memory-wafer ASPs were >30% below wafers entering production, so recent price gains are not yet fully booked — Q4 revenue guided to grow double digits QoQ. PSMC-Micron 1p DRAM collaboration: PDK due end-2026, equipment arriving 1Q27, trial production 2Q27; Micron Post Wafer Finish (PWF) pilot line finishes end-2026, mass production 4Q27. Tight MLC NAND is pushing demand toward high-capacity SLC (PSMC shifting its SLC platform to 24nm); AI servers need 3-5x more NOR Flash than conventional servers, benefiting PSMC as the only NOR foundry outside China. 8/12-inch logic wafer-input ASPs +8-10% in Q3 with a further 10-20% hike planned for 2027 under LTA negotiation.
Maeil Business Newspaper, citing industry sources (via TrendForce, 22 Oct 2026): Tesla and NVIDIA will invest a combined ~KRW 3.7T in Samsung Electro-Mechanics advanced substrate capacity. Tesla (~KRW 1.1T) signed a long-term FC-BGA supply agreement for AI chips in its autonomous EVs and Optimus robots covering 2029-2033 and is getting dedicated lines at SEM's Sejong site (part of the KRW 4.27T Sejong plan disclosed in September): construction completes end-2027, equipment install/trials 1H28, mass production 2H28, deliveries from 2029. NVIDIA's commitment is believed worth ~KRW 2.6T and it is also securing a significant share of annual substrate capacity by 2028; NVIDIA aims to lift its total IC substrate capacity 2.5x, and is supporting Unimicron's KF expansion via a capacity-reservation deal (additional output from 2Q27, full capacity mid-2028). Separately, SEM disclosed a KRW 360.5B contract on 19 Oct to supply memory semiconductor package substrates to an undisclosed global customer through 2027. High-end FC-BGA remains concentrated: Ibiden >70% estimated share, plus SEM and Unimicron.
Common Wealth Magazine, citing industry sources (via TrendForce Breaking, 22 Oct 2026): EMIB-T substrate yields were ~35% in 3Q26 vs a 50% year-end target. Two technical blockers: CTE mismatch between ABF substrates and embedded silicon bridges complicating high-temperature alignment, and air bubbles trapped in NCF (non-conductive film) under the bridges' TSVs weakening adhesion. Google's HumuFish program remains '100%' committed to EMIB-T despite yield problems and rumored switching to TSMC CoWoS-L. Intel has adjusted its substrate electroplating process; suppliers Ibiden, Unimicron and Shinko continue yield ramp. Timing cushion: HumuFish mass production is slated 1H28, but the Google-designed Compute Die A (TSMC N2) tape-out has slipped to late October, and Commercial Times suggests the MediaTek-involved TPU v9 tape-out could slip to January 2027, potentially delaying shipments to 1Q28.
ZDNet, citing industry sources (via TrendForce, 22 Oct 2026): Groq — which licensed its inference technology to NVIDIA — is considering 3D-stacked (3DS) TSV DRAM for next-gen chips, moving away from its SRAM-centric design. 3DS differs from HBM: it prioritizes capacity using conventional DRAM interfaces rather than HBM's transfer-speed focus. Context on the SRAM limit: NVIDIA's Groq 3 LPX rack combines 256 LPUs (500MB SRAM, 150TB/s on-chip bandwidth each) for 128GB SRAM / 40PB/s aggregate, but d-Matrix CEO Sid Sheth (via 36Kr) notes SRAM capacity grows only ~10-20% per two years vs AI model sizes doubling every 6-9 months. Cerebras' future CS-6 will likewise integrate stacked DRAM with wafer-scale SRAM and compute. TSMC CEO C.C. Wei earlier confirmed work with customers on next-gen LPU development.
[News] NVIDIA-Linked Groq Reportedly Eyes Shift From SRAM to 3D-Stacked DRAM for Next-Gen AI Chips
Micron is accelerating its 3D NAND roadmap by skipping a generation (no G10): from currently-shipping G9 (276 layers, first product 2650 client SSD Jul 2024) directly to G11 — >400 layers, wafer-on-wafer bonding, 2x bit density vs G9, up to 1.5x I/O speed, 25% less energy per bit — ramping 2H CY2027, per CTO Scott DeBoer at Micron's 21 Oct technology/product investor webinar. Two AI-directed product lines on G11: high-IOPS 'AI-class NAND' (SLC-based, small blocks, 'ultra fast memory storage') and high-bandwidth AI-class NAND positioned closer to compute than the accelerator shoreline (explicitly not HBM-style stacking); DeBoer did not address the HBF path SanDisk and SK hynix promote.
TrendForce news roundup (22 Oct 2026): at SEMICON West 2026 ('Optical Integration for the Next Generation of AI') NVIDIA showed an optics-on-interposer (OOI) prototype that moves optical engines from the organic package substrate onto the silicon interposer carrying the compute die — eliminating the organic substrate from the electrical path for higher bandwidth density and lower power — plus a DWDM demo carrying eight wavelengths at up to 32 Gb/s each (256 Gb/s aggregate per fiber). Both are research-stage, no commercialization timeline. Roadmap context in the same report: NVIDIA's Kyber NVL1152 architecture will extend optical interconnects to join eight racks into one 1,152-GPU NVLink domain for the Feynman platform (expected ~2028); today's commercial CPO baseline is Spectrum-X Ethernet Photonics (full production since May 2026, 200G SerDes + CPO, early adopters CoreWeave/Lambda/OCI) on a named manufacturing ecosystem: TSMC SiPho fab, SPIL chip-scale packaging/test, TFC laser-die module packaging, Foxconn system integration, Coherent/Lumentum/Sumitomo Electric external laser sources, Corning/Senko/Browave fiber components, Fabrinet cooling/final test.
At the PECC Summit (20 Oct 2026), Nokia SVP Optical Modules Mehrdad Ziari laid out an InP manufacturing roadmap spanning the legacy Sunnyvale fab (60,000 sq ft), a new six-inch InP fab in San Jose (185,000 sq ft), and planned conversion of an existing Chandler, Arizona compound-semiconductor facility to six-inch InP production (1.1 million sq ft). Nokia also demoed at ECOC 2026 a monolithic InP PIC (4 lasers, 8 modulators, 8 SOAs + power monitors) driving 8x200G in a linear-pluggable-optics (LPO) OSFP module at 8.6 W (~5.4 pJ/bit, 1.6 Tbps aggregate). Ziari warned fab investments take 3-5 years vs 12-18-month customer pull — a structural timing gap in optics supply.
PECC Summit: Nokia Links AI Optics Scaling to InP Fab Expansion (Converge Digest)
Nokia's Q3 2026 interim report (22 Oct 2026) shows Network Infrastructure net sales +15% y-o-y constant currency, led by IP Networks +43% and Optical Networks +18%; Group net sales to AI & Cloud customers grew +123% with order intake +118%. This is a fresh first-party datapoint that AI data-center buildouts are driving a step-change in routing/switching and coherent-optical demand across vendors, not just NVIDIA/Cisco — corroborating the optics and switch-silicon shortage thesis from the demand side.
Bloom Energy's record Q3 2026 (8-K/EX-99.1 filed Oct 22): revenue $1,235.8M (+138.1% y/y), product revenue $1,073.6M (+179.4% y/y), gross margin 33.8% (+453 bps), operating income $225.4M, diluted EPS $0.78; FY2026 revenue guidance raised for the third time this year to $4.4-4.6B (+122% y/y at midpoint); company notes S&P 500 inclusion. Confirms fuel-cell on-site generation has scaled into a multi-billion annualized supplier to data centers/semiconductor fabs.
Google is adding 1,070 MW to the 100-MW virtual power plant agreement it signed with Voltus in June 2026 (~1.17 GW total), contracting residential/commercial/industrial batteries, thermostats, on-site generation and flexible process load that Voltus offers into PJM markets when the grid is strained; ramp to full capacity over 3-4 years, contracted through 2040. PJM demand-response revenue rose 240% from 2024 to 2025 to $523.5M (Monitoring Analytics), with emergency DR 88% of it; DR payments can reach $350,000/MW-year (Voltus). Voltus has ~8.5 GW under contract across North America.
Google supersizes PJM virtual power plant to more than 1 GW (Utility Dive)
At FERC's annual Reliability Technical Conference (21 Oct 2026), PJM Sr. Director Operations Planning Christopher Pilong described data-center demand as 'a rising sea level... every hour of the day,' stressing that near-constant 24/7 load destroys the traditional shoulder-season maintenance window — emergency procedures were already invoked 12-14 Oct 2026 during heat 20°F above normal ('two to three years ago, we would not have seen that'). PJM filed four reliability priorities at FERC: (1) timely, continuously updated technical information from large-load operators; (2) enforceable voltage/frequency ride-through for new AND existing large loads; (3) interconnection/energization decisions that reflect real gen+transmission build time; (4) coordinated electric-gas planning, cybersecurity and accountability.
PJM, FERC Focus on Large Load Challenges at Tech Conference (PJM Inside Lines)
Utility AI deployment for planning/operations is scaling from pilots into niche decision-support (load forecasting, vegetation management, maintenance prioritization) but remains barred from acting on physical infrastructure over auditability concerns: 'Most utilities could not reconstruct what an AI system read, what it did and why' (QualityAI's Rick Lesley via Utility Dive, 22 Oct 2026). A Guardrail Technologies review of federal regulatory filings found roughly 70% of energy and utility companies identified AI as a cybersecurity risk, yet only 23% of utilities and 9% of energy companies documented a process for managing it; FERC's October cybersecurity audit flagged 'super intelligence' tool adoption and access-control risks.
Utilities turn to AI to manage rising power demand — from AI (Utility Dive)
Carrier announced plans to invest more than $600M over the next three years in a new advanced manufacturing and engineering center in Fort Worth, Texas (AllianceTexas development). The >1,000,000 ft² (≈93,000 m²) plant begins operations in 2027, creates 1,500 jobs, and expands US production of advanced thermal-management systems including air- and water-cooled commercial chillers with engineering labs on site. Carrier states its commercial HVAC business currently has record backlog levels driven by data centers, healthcare, education and industrial demand — 'expanding the company's US manufacturing capacity for advanced thermal management systems ... and helping customers meet growing demand for intelligent climate and energy solutions across data centres' (company statement); CEO David Gitlin frames it as increasing 'capacity for critical technologies' and supply-chain resilience.
Trane Technologies is set to acquire Cancoil Thermal Corporation, a Canadian manufacturer of heat-transfer products and a current Trane supplier. Founded 1983 (Kingston, Ontario), Cancoil produces fluid coolers, air-cooled condensers and custom finned-tube coils; the deal includes its existing operations — Fin Aire Inc and Cancoil USA — with manufacturing in Ontario and 'growing US manufacturing operations' in Jacksonville, Texas. Cancoil supplies key applications across Trane's chiller and airside businesses; COO Donny Simmons: the capabilities 'will help us better serve customers in critical thermal management applications with speed, quality and reliability.'
ByteDance, Alibaba and Tencent have ordered more than 3 million NVIDIA RTX PRO 5500 GPUs worth over US$35B combined, and total Chinese orders have topped 6 million units — more than double NVIDIA's potential global production of ~3 million units in 2027 — with NVIDIA advising partners they may receive just 5-10% of requested volumes; initial samples expected December 2026, commercial shipments 1Q27 (Reuters via TrendForce, 21 Oct 2026). Context: B300 data-center GPUs remain off-limits in China and H200 shipments require case-by-case licensing; the workstation-class RTX PRO 5500 (21,760 CUDA cores, 84GB GDDR7, 1,398 GB/s, 600W, ~RMB 85,000-90,000 per The Information — comparable to Huawei Ascend 950PR) has become the viable China inference option. NVIDIA is reportedly prioritizing its next-generation Rubin platform for TSMC capacity: from 2027 Rubin is expected to absorb a substantial share of NVIDIA's TSMC manufacturing capacity, leaving less room for the RTX PRO 5500.
Elon Musk said on a SpaceX conference call that Terafab targets volume production in 3Q28 (he allowed 3Q or 4Q is possible) and called the timeline 'very aggressive' against typical multi-year timelines for advanced logic fabs in the US; a first functional SRAM test chip is targeted for around May-June 2027 at the Austin research facility (Reuters via TrendForce, 21 Oct 2026). Equipment access is a major hurdle — Musk did not confirm EUV orders from ASML, and Reuters notes ASML expects demand to outstrip supply through 2028 with discussions over future orders already extending into 2029-2030 ('ASML itself has become a bottleneck'). Tesla/SpaceX are in talks with South Korean semi equipment/materials suppliers (NDAs signed; discussions span nearly every product category; HPSP named), refusing Chinese sourcing yet seeking sub-Chinese prices (The Elec). TSMC CEO C.C. Wei confirmed ongoing Terafab discussions but stressed any fab involving TSMC must be TSMC-operated — conflicting with Musk's 7 Oct statement that his companies would build and operate it with TSMC possibly only subleasing. Intel CEO Lip-Bu Tan said Intel would remain involved (Reuters citing Bloomberg).
At the GTC Berlin keynote (21 Oct 2026), Jensen Huang headlined the DSX Pod — a containerized version of NVIDIA's DSX reference architecture: 'starts with Vera Rubin NVL72. Seven liquid-cooled compute racks connected with 17 kilometers of Spectrum X cables form a Vera Rubin DSX pod... integrated, fully cabled, and factory tested, then delivered in transportable containers' (The Next Platform, quoting the keynote video). Design uses dry cooling rather than evaporative cooling and 800V DC power delivery. Deployment-economics target: time from arrival of computing to token production cut 'from four to five months down to about a month', reducing idle capital — Huang quantified a 1GW datacenter at ~US$50B cost against ~US$40B expected revenue, so a three-month delay ≈ US$10B forgone; TNP also cites TeraWulf's Lake Mariner site where nearly half of 1,800 on-site trades are electricians (labor, not GPUs, binding first). Partners flagged: Schneider Electric, Vertiv, HPE, Lenovo, Dell — NVIDIA said in a statement it is working with them 'to deliver DSX POD AI factories worldwide in early 2027'.
Huang shrugs off protestors to kick off GTC Berlin with DSX Pods, Robots (The Next Platform)
TrendForce weekly memory spot update (21 Oct 2026): DDR5 supply remains tight — suppliers reluctant to sell and buyers raising target prices still cannot secure allocation, though intermittent inquiries have not sparked price-chasing. Branded DDR4 trading subdued as module houses strictly control capital and inventory; the mainstream DDR4 1Gx8 3200MT/s average spot price rose 0.84% to US$46.357 (as of 20 Oct). NAND: weak module-house demand and elevated inventories kept spot trading sluggish with sellers reluctant to cut; 512Gb TLC wafer average spot slipped 1.28% to US$18.592 (as of 19 Oct).
Reuters (via TrendForce, 21 Oct 2026): Samsung is planning to shift its Xi'an plant from roughly 200-layer to 286-layer V9 NAND under an internal project dubbed 'Monami', while SK hynix — already authorized to produce up to 240-layer NAND in Dalian — may seek US approval for R&D on 300-plus-layer technology; both confirmed upgrades are underway 'in close coordination with Washington' but declined details. TrendForce estimates China accounts for ~30-35% of Samsung's total NAND output and ~40-45% of SK hynix's. YMTC (160/232-layer dominated lineup) is already producing 294-layer chips in limited volumes and plans beyond 300 layers. Both Korean makers must obtain annual licenses to install US equipment in China, and Washington indicated in 2025 it would not approve technology upgrades/expansions there. Beyond China (via Seoul Economic Daily / Bloter, cited by TrendForce): Samsung has shifted ~60% of V-NAND capacity to V9 (286L), ramping NVIDIA shipments, begun V10 mass production, and is accelerating ~500-layer V11 development; SK hynix's 321-layer NAND became its largest production segment in 2Q26 with ~half of domestic NAND capacity expected on 321L by year-end.
Bank of America note ahead of WD earnings (21 Oct 2026): expects WD FQ1 exabyte shipments +21% Y/Y — BELOW management's ~25% longer-term exabyte-growth target — while average price per terabyte rises +24% Y/Y; 'Demand for hard-drive storage continues to exceed available supply, giving manufacturers room to charge more for each unit of storage capacity'; WDC price target raised to $800 from $720 on tighter HDD supply supporting prices and margins (forecast FQ1 gross margin 56.4%, above WD's own 55-56% guidance).
TheStreet via AOL — BofA lifts Western Digital target despite lower capacity outlook
TrendForce News (21 Oct, aggregating Reuters, Seoul Economic Daily, Bloter): Samsung is upgrading its Xi'an plant (~30-35% of Samsung's NAND output; China share for SK hynix is 40-45%) from ~200-layer to 286-layer V9 NAND under internal project 'Monami'; SK hynix — already authorized for up to 240-layer production at Dalian — may seek US approval for 300+ layer R&D there. Both must obtain annual US licenses for equipment installs and Washington said in 2025 it would not approve tech upgrades/capacity expansions in China. YMTC (160/232L-dominant lineup) is producing 294-layer chips in limited volumes targeting >300. Separately, Samsung has shifted ~60% of V-NAND capacity to 286L V9 (ramping to NVIDIA) and begun V10 mass production, with ~500L V11 accelerating; SK hynix's 321-layer became its largest NAND segment in 2Q26 with ~half of domestic NAND capacity on 321L by year-end.
TrendForce weekly memory spot report (21 Oct, prices as of 19 Oct): 512Gb TLC NAND wafer average spot price slipped 1.28% this week to US$18.592. Module houses (the largest spot buyer) show weak demand against persistently high inventories, keeping trading sluggish; sellers believe further cuts won't stimulate demand and show limited willingness to lower prices — spot stuck in a narrow range. This contrasts with contract-side momentum (eSSD/NAND contract prices still rising per the same week's reporting).
TrendForce Insights — Memory Spot Price Update (21 Oct 2026)
VAST Data announced (GTC Berlin, reported 21-22 Oct 2026) its Foundation Stack for Governed Agent Sandboxes: persistent cross-session AI-agent state (prompts, tool calls, tokens, costs, decisions) stored on the VAST AI OS under NVIDIA's OpenShell secure agent runtime and the open-source NemoClaw reference stack, with BlueField-4-based hardware security enforcement. Same-day StorageReview relay: DeepL deployed VAST's AI OS for voice-translation training and expects 3x usable capacity from its existing NVMe estate. Signals that inference/agent workloads are converting ephemeral GPU sandboxes into durable, audited NVMe-tier capacity demand.
Blocks & Files — VAST agent sandbox stores cross-session persistent agent state
At GTC Berlin (21 Oct 2026), NVIDIA announced — alongside substrate/packaging/test expansion — that it and Siliconware Precision Industries (SPIL) plan to expand Taiwanese optical-engine (CPO OE) manufacturing capacity by 15x by mid-2027, and to double that again by the end of 2028. StorageReview explicitly notes every figure is a planned capacity target with no current baseline disclosed. Related context in the same announcement: NVIDIA is raising overall substrate supply 2.5x over three years (IBIDEN Ono plant 3x with multibillion-dollar reservation, Unimicron KF expansion from Q2 2027), Amkor Korea interposer capacity +300% by end-2028, and total semiconductor test capacity doubling 2026->2027 (incl. $1.1B KYEC reservation). The 15x OE step is the clearest signal yet that NVIDIA is provisioning CPO for volume Vera Rubin / Spectrum-X Photonics deployment rather than niche builds.
Zacks analysis relayed on Yahoo Finance (21 Oct 2026), from Ciena's fiscal Q4-26 earnings (CEO Matt Graff): latest quarter revenue $1.67B, +37% YoY, backlog $8.5B expected to exceed $10B by end of fiscal 2026 (into fiscal 2028 shipments), fiscal-2026 revenue guidance raised — but the binding constraint is parts, not demand: component shortages across the industry are 'unlikely to ease before fiscal 2028', so Ciena has locked in key components through fiscal 2029, paying suppliers up front (cash from operations expected to decline in fiscal Q4 under those component agreements). Optical networking was ~71% of last-quarter sales; two customers were each >10% of revenue (41.7% combined). Confirms the optics-supplier queue persists at the equipment layer into 2028 even for the non-Chinese coherent/DCI segment.
Ciena (CIEN) Has $8.5 Billion in Backlog, but Can It Justify Its Premium?
TrendForce news (21 Oct 2026): Nexilumen, a Fudan University spin-off (with Peking University collaboration), advanced Micro-LED optical interconnects: a 4x4 green Micro-LED array with 16 independently driven pixels at ~2.3 Gb/s each achieves 37 Gbps aggregate via space-division multiplexing, and a five-wavelength InGaN/hybrid-QD WDM stack on a 320x290 um area reaches 18.66 Gbps aggregate (~201 Gbps/mm2 data-rate density); earlier Nature Communications work showed 0.056 pJ/bit at 1.5 Gb/s. Demonstrated at CIOE 2026 with long-wavelength InGaN devices (2.85 GHz modulation bandwidth) and >3 GHz 405 nm sources; seed funding (tens of millions RMB) closed July 2026. Micro-LED links target short-reach interconnect as a lower-power alternative lane for in-cabinet optics — same 'optical engine relocation' theme as CPO/NPO, but from the light-source side; still R&D-to-early-engineering stage.
At Optica's PECC Summit (20 Oct 2026), Coherent EVP/CTO Dr. Julie Eng said Coherent's indium-phosphide (InP) capacity was essentially sold out through the end of 2027, with fab expansion underway to address demand. Corroborates the Ciena 'component scarcity into FY28' and Lumentum 'sold out through 2029' signals from another top-3 Western optics supplier — the InP laser bottleneck (EML/CW DFB feeding 800G/1.6T pluggables) is not clearing near-term.
PECC Summit: Coherent Sees Chip-to-Chip Optics as the Next Frontier (Converge Digest)
The Nuclear Company (venture-backed developer) announced (21 Oct 2026) plans for 'Prosperity One', a 6.6-GW nuclear campus in Lewis County, Kentucky on the Ohio River, accommodating up to six Westinghouse AP1000 reactors — which would make it the largest US nuclear site (vs Plant Vogtle's ~4.5 GW net summer across four units). TNC says it has secured site control and intends to pursue federal approvals to begin construction in 2028; initial work covers site characterization, geotechnical, environmental, and transmission interconnection planning.
The Nuclear Company Proposes 6.6-GW AP1000 Nuclear Campus in Kentucky (POWER Magazine)
Avangrid (Iberdrola subsidiary) is investing in the modernization of more than 40 US substation facilities in 2026 across Maine, New York and Connecticut, unveiled this month (T&D World, 21 Oct 2026): major substation rebuilds and equipment replacements including flood-mitigation works, automation upgrades and capacity enhancements, explicitly framed by CEO Jose Antonio Miranda as preparation for meeting growing energy demand (data-center-driven load growth in the service territories).
Avangrid Advances Grid Modernization Investment for 40+ US Substation Facilities (T&D World)
LevelTen Energy index (via Utility Dive, 21 Oct 2026): the average US solar PPA price rose 4.9% in Q3 2026 to $64.23/MWh (+6.0% y/y, +12.7% vs Q3 2024); wind PPAs were roughly flat on the quarter but +13.8% y/y and +27.5% since 2024. LevelTen attributes most of the recent increase to new trade tariffs (Section 232 measures announced in early August, which could add $3-7/MWh once effective later this year), with 'more than half of solar modules repriced' and buyers contracting ahead of tariff effect; hyperscaler demand is reshaping the PPA landscape.
Solar PPA prices jump 5% in Q3, mainly on new trade tariffs: LevelTen (Utility Dive)
At NVIDIA GTC Berlin (21 Oct 2026), Vertiv and NVIDIA introduced an 'AI factory' approach combining NVIDIA's Vera Rubin DSX AI Factory reference design and Omniverse DSX Blueprint with Vertiv SimReady power/cooling assets on Dassault Systemes' 3DEXPERIENCE platform: pre-engineered designs are configured and virtually validated (requirements-to-validated-solution time cut from ~30 days to ~3 days, a claimed 10x), then translated into factory-built deployment. Under defined assumptions the approach is designed to generate up to 30% more cumulative token output in the first 12 months at the same installed capacity. Polar Data Centers is applying it to the first 41 MW phase of its multiphase AI data center at Heroya, Norway. Vertiv is contributing SimReady OpenUSD assets of its power/cooling equipment to the SimReady Central Library on Hugging Face.
The White House investment announcement for South Carolina (21 Oct 2026) lists Eaton investing $340 million to expand US transformer production at Jonesville, SC. It is the second Eaton power-equipment capacity move logged in the DB inside eight weeks (after the >$242M North Little Rock, AR modular electrical enclosure plant announced mid-Sep) and lands while large-power-transformer lead times remain ~128 weeks and US transformer producer prices are +86% since Jan 2020.
The same White House South Carolina release (21 Oct 2026) confirms a cluster of grid-equipment manufacturing investments: Okonite $400M for a new high-voltage electric cable plant in Orangeburg (105 jobs) — re-confirming a project first logged in the DB from the Jul-Sep build-out wave; TS Conductor $134M for an advanced manufacturing facility in Hardeeville (composite-core/high-capacity conductor); and Red Metals $70M for a copper manufacturing operation in Charleston County — upstream input for wire/cable. Context: US power wire & cable producer prices are easing (-13.8% y/y per BLS via StackPath) even as new domestic cable capacity is being funded.
Virginia's major investor-owned utilities filed the first publicly quantified distribution-grid utilization metrics in the US (Latitude Media, 21 Oct 2026): Dominion Energy, serving northern Virginia's Data Center Alley, reported 41% annual utilization (Jul 2025-2026) rising to ~80% average during infrequent peak hours; rural Appalachian Power reported 14% annual / 54% peak utilization. The filings propose standardized measurement methods, starting a potentially years-long utility/regulator process to decide between building new capacity and shifting/expanding load on existing wires (storage, advanced reconductoring, flexible data centers).
What to know about the nation's first grid utilization metrics (Latitude Media)
At NERC's standards progress reviews reported in the Oct 15, 2026 ERCOT Large Load Working Group (Zero-Emission Grid summary, 21 Oct): a Computational Load definition now exists and proposed registration criteria apply to qualifying sites of at least 50 MW connected at 100 kV or above; NERC's Standards Committee approved a new standards project on ride-through and active-power recovery for computational loads (RTO Insider, 22 Oct), with criteria/definitions and Rules of Procedure revisions due 31 Dec 2026 and a further-standards workplan due 1 Mar 2027; ballot results exceeded the 66.67% approval threshold. Dominion's event analysis showed 1,500 MW initial load response in both the July-2024 and April-2026 disturbance events, but sustained load transfer dropped from 1,500 MW (Jul 2024) to 500 MW (Apr 2026).
NERC Computational Load Requirements: ERCOT LLWG October 2026 Update (Zero-Emission Grid)
PJM's latest capacity auction cleared at the FERC-approved price cap for the third consecutive time, at more than eleven times the price four years ago, and fell short of PJM's reliability requirement (Atlantic Council analysis, Utility Dive 21 Oct 2026). The authors argue the outcome evidences an underinvestment crisis produced by over-reliance on RTO spot/capacity markets, not a failure of competition itself, and press FERC/states/RTOs to expand long-term offtake — bilateral IPP-LSE PPAs under regulatory hedging mandates, or IPP-RTO contracts funded by a reliability charge — citing PJM's recent white paper on such constructs.
Supermicro's 20-Oct preliminary raise (revenue $18.5-19.5B vs $14.5-15.5B guided) triggered a sector-wide repricing of AI-hardware supply chains: SMCI +~12% after-hours 20 Oct (+16% intraday per Stocktwits); on 21 Oct KOSPI jumped 4.6% to 5,805.63, Samsung Electronics +6.2%, SK Hynix +6.5%, Kioxia +4%, TDK +3.5%, Murata +3.7%, TSMC +0.8%, Foxconn Industrial Internet +3.1%, Luxshare +2.5%; Nikkei +1.7% to 65,162. Dell and HPE shares also jumped (IBD headline, 20 Oct). Coverage frames the move as relief that AI infrastructure spending is sustainable after last week's selloff on financing-cost and CoreWeave-debt concerns; SMCI server assembly (Nvidia/AMD-based) cited as a demand indicator for data-centre hardware. Investors now await Alphabet/Microsoft prints and SMCI's full 3-Nov result for confirmation.
At NVIDIA GTC Berlin (21 Oct 2026) Vertiv and NVIDIA introduced an AI-factory approach combining pre-engineered Vera Rubin DSX reference architectures virtually validated on NVIDIA Omniverse DSX Blueprint with Vertiv SimReady power/cooling assets (Dassault Systèmes 3DEXPERIENCE integration), then factory-built deployment via Vertiv OneCore and SmartRun (overhead infrastructure for high-density power distribution, liquid cooling, networking and containment). Modeled scenario: requirements-to-validated-design time cut ~30 days to 3 (10x), infrastructure deployment ~90 to 45 days (up to 50%), productive capacity ~10 weeks sooner, 'up to 30% more cumulative token output within the first 12 months with the same installed capacity'. Polar Data Centers is applying the model at its Herøya campus (Norway) for the first 41 MW phase of a multiphase AI data center scalable to 100 MW, with Vertiv supplying prefabricated power and thermal modules; the deployment model is designed to bring high-density, liquid-cooled capacity online sooner.
Chemours confirmed that injunctions issued against four exhibitors at Chillventa 2026 (Nuremberg, 13-15 Oct) relate to its patents on blends containing R1234yf — the family it holds covering R454B, R452A, R454A, R454C and R513A. In the same show week rival Solstice claimed five injunctions over an unnamed HFO manufacturing process and proprietary blend. Context: Chemours launched R1234ze(E) and the R1234ze(E) blend R515B under Opteon in August 2026; Solstice then stated its patent portfolio covers key HFO-1234ze manufacturing processes and that 'Solstice has not licensed any manufacturing-related IP to others.' Chemours and Solstice share a long history of HFO cross-licensing and joint patent protection (ex-Honeywell/duPont lineage).
Cooling Post — Chemours' injunctions relate to R1234yf blends
Chemours faces a securities class action (filed via Levi & Korsinsky) alleging misrepresentation of aftermarket demand for Opteon refrigerants and of its 2026 guidance: ~$65M of Opteon aftermarket sales booked in 2025 that management later estimated should have been allocated to 2026, following excess distributor inventory built during the US residential AC equipment transition driven by the EPA AIM Act. The complaint says Chemours described aftermarket demand as 'strong' between 20-Feb and 4-Aug-2026 without disclosing the buildup; Q2 2026 net sales were $1.6B (~1% below prior year) with a 4% volume decline attributed to lower Opteon blends aftermarket; shares fell 18.63% ($3.34) after the 4-Aug reduced full-year guidance; lead-plaintiff deadline 7-Dec-2026.
Refindustry — Chemours Faces Investor Lawsuit Over Opteon Refrigerant Sales in the US
Hitachi, Brookfield, Hartree and Tokyo Gas signed an MoU (30 Sep 2026, reported 21 Oct) to develop 'green data centers' in Japan on a standardized model: on-site generation as primary power with battery storage and grid backup, and designed redundancy across cooling equipment and power-distribution systems to maintain operation during equipment failure. Hitachi contributes data-center cooling and power-generation facility design, equipment supply and maintenance, plus its HMAX Data Center Physical-AI suite for autonomous integrated operation of power/IT/cooling; Tokyo Gas brings gas-cogeneration 'Power Hybrid DC Solution'. First projects target Kanto and Kyushu; the model builds on the July 2025 Hitachi-Brookfield-Hartree collaboration and will involve IT companies, manufacturers and financial institutions.
Equinix launched a network of Enhanced Performance Innovation Centres (EPIC), live in Sydney, Frankfurt, London, Osaka and Paris, letting enterprises test liquid-cooled AI infrastructure inside operating data centers before committing to deployment — GPU/CPU systems, training/inference workloads, and cooling-power-networking-facility integration with hardware partners. SVP Brian Stein: 'Liquid cooling is an essential component of modern enterprise infrastructure and a hard requirement for customer's deploying AI'; the centres target the skills/integration gap (maintenance, consistency, integration) that blocks distributed enterprise DLC adoption. More locations expected in 2027; extends Equinix's 2022 Co-Innovation Facility and 2024 Solution Validation Centre model; Criteo quoted on stepwise liquid-cooling confidence-building.
DataCenter.news (DCD syndication) — Equinix opens Sydney centre for liquid-cooled AI tests
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.
Commercial Times (via TrendForce, 20 Oct 2026): institutional investors expect 4Q26 DRAM contract prices to rise ~20% QoQ — a slowdown from ~80% in 2Q and ~40% in 3Q reflecting the higher base, weakening end-market purchasing power, and price ceilings under long-term agreements gradually taking effect — while remaining optimistic on the cycle given AI inference demand and limited new supply of mature memory products. Separately (via MoneyToday): CXMT's 4F² announcement has yet to affect DRAM spot prices, which keep rising; mass production still faces yield/leakage/data-retention challenges and full commercialization is generally expected after 2028. On HBM: with memory supply still tight, GPU makers are weighing lower HBM capacities to lift GPU shipments — if NVIDIA switched Rubin Ultra from 12-Hi to 8-Hi HBM (same stacks/GPU, same capacity per layer), HBM capacity per GPU would fall ~33% while the same DRAM die supply could theoretically support 50% more GPUs; TrendForce notes both stack heights use one base die, so 8-Hi HBM is expected to cost 10-20% MORE per Gb than 12-Hi in 2027.
CXMT (via TrendForce, 20 Oct 2026, citing Star Market Daily / Eastmoney / SCMP): at its 19 Oct extraordinary shareholders' meeting the company projected 9M26 revenue of RMB 265-270 billion (+>7x YoY) and net profit attributable to shareholders of RMB 131-132.5 billion (vs a RMB 5.3 billion loss a year earlier) ahead of its 31 Oct earnings release. Implied 3Q26: revenue RMB 114.69-119.69 billion (+15.25-20.28% QoQ) but net profit of only RMB 53.40-54.90 billion — +1-4% QoQ — after 2Q26 profit more than doubled to RMB 52.84 billion; Eastmoney attributes the stall to moderating DRAM price gains. Fabs in Hefei (x2) and Beijing run at high utilization with full order books; each fab has up to 100,000 wafers/month capacity (36Kr); Hefei test facility expanding and new DRAM + HBM production lines being developed in Shanghai. Depreciation expense rose 27.85% YoY to RMB 14.51 billion in 1H26. Chinese module maker Longsys guided 3Q26 profit of RMB 3.42-4.32 billion, a sequential drop of up to 49% from 2Q's RMB 6.72 billion.
Meta detailed (mid-Oct 2026) the storage system behind Prometheus, its New Albany, Ohio gigawatt GPU supercluster scaling toward ~1 million GPUs (GB200 'Catalina'/'Ariel' and GB300 'Clemente' platforms, Vera Rubin 'San Miguel' to follow): it has migrated from a file-based (NFS-like on Tectonic) Llama-training design to multi-tiered cached BLOB/object storage to cut GPU idle time. The 4-layer hardware stack is L1/L2 memory + local flash on GPU hosts, L3 regional flash-backed BLOB storage, and L4 a durable HDD-backed BLOB/Tectonic 'source of truth' tier, with distributed caching reusing spare GPU-host memory (Owl components), metadata caching, prefetch, hedged reads and dynamic concurrency control; Tectonic itself erasure-codes across HDD+flash media tiers with hot/warm/cold placement.
Blocks & Files — Meta's Prometheus using object storage for massive AI training jobs
SK hynix shares fell as much as 6% on 20 October 2026 to the ₩1.18M range (~$890), breaking below ₩1.2M intraday for the first time in about six months (first time below since April), extending a 24.3% weekly rout over Oct 12-16 (its worst week since 2008; Samsung Electronics -17.8% the same week). BigGo Finance attributes the renewed selling to three pressures: (1) the Solidigm Nasdaq-listing controversy — critics call it a duplicate listing and the head of South Korea's Financial Supervisory Service told lawmakers the plan may violate listing rules; (2) the imminent end of a ₩40T (~$30.1B) share buyback, removing a key demand source; and (3) CXMT DRAM-advance and AI-profitability worries. Foreign investors sold a net ₩9.612T (~$7.2B) of SK hynix shares in October to date, pushing foreign ownership to 48.85% (lowest since Dec 2021). The article recaps Solidigm's pre-IPO state: Goldman Sachs and Morgan Stanley selected as lead managers, pre-IPO procedures entered, Reuters reporting a US listing as early as 2027 at a valuation of up to $150B and raise of up to $15B, with the ownership chain SK Inc. → SK Square → SK hynix → 'AI Company' → Solidigm.
At OCP 2026, Intel claimed to be the only high-volume foundry able to supply 300mm indium-phosphide silicon photonics at scale, integrating InP laser technology into its 300mm SiPh platform to feed NPO/CPO demand; conventional InP substrates remain largely 100-150mm (6-inch) wafers, which caps laser chip output per wafer. TrendForce's companion recap: NVIDIA has begun shipping Spectrum-X CPO switches to select partners (built on TSMC COUPE); Broadcom continues limited shipments of 51.2T Bailly CPO and showcased 3rd-gen 102.4T Tomahawk 6-Davisson (200G/channel); TSMC plans to start COUPE-on-substrate production in 2026 (2x power efficiency, ~10x latency reduction vs board-level pluggables); Zhongji Innolight has a dedicated CPO line in small-volume 800G CPO module production with customer orders.
TrendForce news (20 Oct 2026): at OCP Global Summit 2026 Intel claimed to be the only high-volume foundry able to supply 300mm InP silicon photonics at scale to the broader industry, aimed at next-gen NPO/CPO for AI and HPC. Context on why this matters: conventional InP substrates are largely limited to 150mm (6-inch) wafers, with 3-inch and 4-inch formats still widely used; InP crystals are brittle, grow slowly and are defect-sensitive, limiting laser chips per wafer and slowing capacity expansion. Intel integrates InP-based laser technology into its 300mm silicon-photonics manufacturing platform to raise wafer throughput and scalability. This is the upstream laser/photonic-IC capacity constraint flagged repeatedly by optics vendors (Lumentum 'sold out through 2028', EML shortfalls — already in DB).
TrendForce news (20 Oct 2026) status round-up of CPO supply: NVIDIA has begun shipping Spectrum-X CPO switches to select partners (built on TSMC COUPE); Broadcom continues limited shipments of its 51.2 Tb/s Bailly CPO switches and showcased the third-generation Tomahawk 6-Davisson CPO switch at OCP 2026 (102.4 Tb/s, 200 Gb/s per channel); TSMC plans to start production of its COUPE-on-substrate solution in 2026, targeting ~2x power efficiency and ~10x latency reduction vs board-level pluggables; Zhongji Innolight has a dedicated CPO line in small-volume 800G CPO module production with customer orders, volume shipments potentially from 2027 and broader adoption expected 2028; Huawei introduced its NPO-based Ascend 960 supernode in Sep 2026. Bottleneck has shifted to test: optical alignment now takes seconds or less, but full testing of a single PIC can exceed 100 seconds — TrendForce expects 2027 to be pivotal for CPO test-system integration and 2028 the readiness checkpoint for large-scale CPO production.
ABB Q3 2026 (reported 20 Oct 2026): group orders hit a record $12.0B, +40% y/y; Americas orders +~70% led by US +82% y/y; Electrification segment booked record $7.67B orders, +70% y/y; CFO Christian Nilsson called data centers the 'standout' customers with segment orders up triple digits (Americas +117% comparable, Europe ~+20%, China ~+40%). Revenue +14% to $9.7B; net income $1.9B (-2%, incl. $115M non-cash exchange-hedging charges tied to the pending Rotork acquisition). ABB also announced a 51% stake in a JV with Mitsubishi Electric's carved-out low-voltage power distribution business.
Grid equipment manufacturer ABB sees US orders rise 82% (Utility Dive)
Cross-manufacturer tally (StackPath 'Power and the grid' data build updated 20 Oct 2026): the three big heavy-frame gas-turbine makers hold ~157 GW of firm turbine backlog at their latest reported quarters — GE Vernova 53 GW firm orders (2026 Q2; +63 GW slot reservations excluded), Siemens Energy 69 GW (Q3 FY2026), Mitsubishi Heavy 35 GW (Q1 FY2026; EUCI reports 32 GW). Years of firm backlog vs stated annual output: GEV 2.6 yrs (20 GW/yr target from Q3 2026), Siemens Energy 4.5 yrs (CEO guide 15-16 GW deliveries/yr), MHI 2.2 yrs (~16 GW/yr basis from FY2025 large-frame sales).
StackPath — Power equipment: gas turbines and transformers (data build)
US producer prices have decoupled across power equipment (BLS series via StackPath build updated 20 Oct 2026, data through Aug 2026): switchgear PCU335313335313 +89% since Jan 2020 and +13.3% y/y at an all-time peak; transformers PCU335311335311 +86% since Jan 2020 (+7.8% y/y, at peak); turbines PCU333611333611 only +28% since 2020 (+5.3% y/y, -0.2% in Jul 2026); power wire and cable falling (-13.8% y/y, -20.3% from Jul 2025 peak) while copper wire & cable +27.2% y/y — evidence distribution-transformer supply is easing (per Wood Mackenzie) even as large transformer lead time stays ~128 weeks.
LS Power closed a $6B private-equity funding round (up from an initial $4B target) earlier in October 2026 and has earmarked about $1.7B of it for its planned purchase of roughly 5 GW of generation from Constellation Energy, COO Darpan Kapadia told Utility Dive (interview published 20 Oct 2026). Kapadia framed the raise around 'real' demand growth (incl. data centers) with supply taking time to catch up; the firm (>$19B PE raised to date) spans generation (REV Renewables), storage (Endurant), transmission (LS Power Grid) and EV charging (EVgo).
Entergy Arkansas and Google announced (20 Oct 2026) phase two of the West Memphis data-center campus partnership: ~$600M of Google-funded infrastructure investment plus a Google-funded power purchase agreement covering new solar, battery storage and transmission capacity. Entergy projects at least $500M in additional customer savings beginning 2028 over 20 years and $2.1B total net benefits across both phases under Entergy's 'Fair Share Plus' large-load tariff pledge (large loads pay 100% of cost of service); Google also runs a $25M Energy Impact Fund for the community.
Entergy Arkansas, Google Expansion to Bring Additional Energy Resources to Arkansas Grid (T&D World)
Graz-based Emerald Horizon AG (Vienna-listed June 2026) is pitching SMRX, a container-sized, accelerator-driven subcritical thorium 'generator' paired with storage, sold storage-first: its DUALstorePLUS product (LFP battery ~5 MWh/module at ~94% roundtrip, plus proprietary CALstore molten-salt thermal storage 5 MWh(th)/module, 200-550C, configurable 5-100 MWh) ships battery units first 'as soon as this year' with molten-salt units in 2028, while the ADES thorium generator is at TRL 4.5 with prototype completion targeted 2029 and a demonstration plant 2031 before serial production (POWER Magazine, 20 Oct 2026).
Virginia Gov. Abigail Spanberger (D) proposed (20 Oct 2026) that the SCC impose a six-year customer rate freeze and raise Dominion's allowed debt share from 48% to 60% as conditions of approving the $67B Dominion-NextEra merger, which would create the largest regulated utility in the US. The merger — under federal (FERC) and state review — is a contested data-center load exposure story, with the governor's package explicitly aimed at protecting ratepayers from grid buildout costs.
Virginia governor seeks 6-year rate freeze if Dominion-NextEra merger is approved (Utility Dive)
Georgia electric cooperative Central Georgia Electric Membership Corp. plans a 50-MW home battery deployment with vendor Base Power (750+ MWh fleet across 30,000+ installs) to shave peaks and cut costs as hyperscale data centers seek to connect to its grid — a utility-led VPP hedge against large-load-driven capacity stress in a non-RTO (Georgia) market.
Georgia electric cooperative plans 50-MW home battery deployment (Utility Dive)
Grid Forward executive director Bryce Yonker (Utility Dive, 20 Oct 2026): average US retail electric rates rose from 10.59 cents/kWh (2020) to 13.63 cents/kWh (2025), +29%, while outage metrics deteriorated — the industry group's '12 actions' package argues affordability and grid modernization must be pursued jointly, reflecting the shift of grid discourse from pure buildout to cost recovery politics amid data-center-driven capex.
12 actions to tackle energy affordability (Utility Dive, Grid Forward opinion)
The Department of Energy is redirecting Defense Production Act funds to squeeze more capacity from existing lines as Biden-era grid-resilience grant money runs dry (Latitude Media, 20 Oct 2026): up to $50M for advanced reconductoring and dynamic line rating projects across Appalachian Power and Potomac Edison service territories (Virginia/West Virginia), and up to $150M awarded earlier in October to the Alaska Energy Authority for a second transmission line serving expanding military-base load with potential data-center and mining growth.
DOE turns to defense funds to squeeze more from power grid (Latitude Media)
Supermicro (8-K filed 20 Oct 2026, Exhibit 99.1): preliminary revenue for FQ1'27 (quarter ended 30 Sep 2026) estimated at $18.5-19.5B vs prior guidance of $14.5-15.5B — a ~$4B raise 'primarily due to increased demand for AI platforms'. Preliminary gross margin estimated at 12.0-12.5% vs guidance of 10.4-10.8%, attributed to customer and product mix. Earnings call scheduled 3 Nov 2026 5pm ET. Figures are unaudited management estimates, subject to revision in closing procedures. Syndicated coverage (Yahoo/Investing.com, 21 Oct) notes the midpoint beat ~$15.0B consensus and follows July's low-end pre-warning; the raise is read as evidence data-centre hardware demand remains strong.
Goldman Sachs (report published 18 Oct 2026, per BigGo Finance coverage dated 20 Oct): raised forecasts for the global server market size in 2027/2028 by +31%/+45%, upgrading both AI and general-purpose servers. Rack-level AI server TAM projected at US$1 trillion by 2028 (163% CAGR), equal to 63% of the total server market; general-purpose server TAM ~$271B by 2028. Second major upward server-outlook revision by Goldman in recent weeks. Basis cited: demand-side capital flows, chip shipments and upstream capacity all moving higher simultaneously. Memory price increases read as a demand signal — CSPs are folding higher memory costs into capex rather than cutting procurement. Goldman's value-chain buy list names Hon Hai, Wiwynn, Wistron, TSMC, Asia Vital Components, King Slide Works and Aspeed; global ASIC penetration ~58-60% by 2028 (~49% among leading cloud providers).
HPE announced (Business Wire, 20 Oct 2026) HPE Spaceborne Computer-4, a lightweight, self-contained, fanless compute module built for space environments, planned to power AI and the first HPC benchmarks on the lunar surface. It deploys on Astrolab's FLIP rover for the Moon Base program via NASA's CLPS initiative, launch planned late 2026 (Griffin-1 lander, lunar South Pole). Runs LLM RAG, vision inference and LiDAR processing; equipped with low-power CPU + GPU and Kioxia enterprise flash/SSDs; HPE supports mission control with HPE Private Cloud and CloudOps software. Workload partners: NASA, Deloitte, HPE.
Schneider Electric Q3 2026 organic revenue jumped 17% to EUR 11.5B (USD 12.9B), a record quarter; data-center business 'continues to be exceptionally strong with another quarter of triple digit demand growth in Q3, further improving our visibility into 2027 and 2028' (CFO Nathan Fast, Oct 16 earnings call/release). Energy Management segment grew 19% organic; North America largest revenue source, China/East Asia higher organic growth.
Schneider Electric sees record-high Q3 revenue on global data center boom (Facilities Dive)
OCP's official post-summit summary (Oct 20): the 2026 Global Summit (San Jose, Oct 12-15, theme 'Scaling Innovation for the AI Era') drew 15,000 attendees — a sold-out event, +38% vs 10,835 in 2025 and the largest in the Foundation's 15-year history — from 3,400+ organizations in 64 countries; 53% first-time attendees; 200+ media/analysts. Starting 2027 the Global Summit moves to Moscone Center, San Francisco.
2026 OCP Global Summit By the Numbers (Open Compute Project)
MHI has obtained ETL Listing (Intertek safety) and AHRI certification (performance) for its JHT-Z270I centrifugal chiller — a 10 MW-class unit designed for AI data center cooling — and has begun North American market-introduction activities. The chiller anchors MHI's Modular Chiller Plant (MCP) pre-engineered platform integrating centrifugal chillers, pumps, heat exchangers and controls, with hybrid free-cooling modes (outside-air priority, chiller on demand) and closed-loop cooling-water configurations to improve PUE and WUE. MHI cites ~60% share of Japan's centrifugal chiller market (FY2025 JRAIA domestic shipment data) and is developing integrated power+cooling+energy-management for AI infrastructure.
Refindustry — Mitsubishi Heavy Industries Certifies 10 MW Chiller for US Data Centers
NVIDIA has asked Samsung, SK hynix and Micron to adopt TSMC's 3nm process for the base dies of its NVHBM (unveiled August 2026), per Seoul Economic Daily via TrendForce (19 Oct 2026); Samsung's DS division reportedly agreed after internal review. Rationale: consistent HBM performance and simplified validation, plus yield economics — industry estimates put TSMC 3nm yields above 90% vs ~60% for Samsung Foundry 2nm (SF2). NVHBM integrates the memory controller into the base die; NVIDIA claims up to 30% more memory bandwidth, 15% lower HBM power and 25% more XPU compute-die area vs standard HBM4E. Samsung is also validating NVHBM base-die production on its own 2nm as a hedge 'should TSMC face capacity constraints'. Context: TSMC flagged tight 3nm capacity on its 15 Oct earnings call (3nm = 33% of 3Q26 wafer sales, its largest node share) and announced another 3nm fab in Tainan (mass production H1 2028) on top of Arizona, Kumamoto and Southern Taiwan Science Park expansions. Samsung keeps in-house 4nm base dies for standard HBM4/HBM4E and a turnkey approach for other customers incl. AMD.
Google's TPU v9 project with MediaTek could see its initial tape-out delayed to January 2027 and shipments pushed to 1Q 2028 following design revisions, per Commercial Times citing institutional investors (via TrendForce, 19 Oct 2026). Revisions to TPU v9 Core A (codename 'Humufish') have delayed development ~3 weeks, affecting Core B and I/O designs handled by MediaTek; the November 2026 sampling target remains unchanged, with the production-version tape-out in 2Q27. TPU v9 is expected to adopt Intel's EMIB-T advanced packaging, with substrate yields a key challenge to production scalability and cost. Same roundup: Bloomberg reports ByteDance is nearing an agreement to buy MediaTek data-center chips, potentially its second major ASIC customer after Google; TrendForce projects global AI-ASIC accelerator shipments reach 14.1M units in 2027, up ~69% YoY. ASIC design-house momentum: GUC record Sept revenue NT$7.145B (+97.8% YoY); Alchip NT$9.979B (+496% YoY on 3nm AI-accelerator shipments). Faraday has obtained Samsung's 2nm PDK and completed Intel 18A-P design-environment setup as Samsung 4nm reportedly runs near full capacity.
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.
NVIDIA is reportedly pushing all three memory makers (Samsung, SK hynix, Micron) to manufacture NVHBM custom base dies on TSMC's 3nm node — uniting foundry rivals on a single platform for consistent performance and simpler validation; Samsung's DS division reportedly agreed after internal review. Industry estimates place TSMC 3nm yields above 90% vs ~60% for Samsung Foundry 2nm (SF2). NVIDIA specs NVHBM at up to +30% memory bandwidth, -15% HBM power and +25% XPU compute-die area vs standard HBM4E. Samsung is simultaneously validating NVHBM base dies on its own 2nm as a fallback against TSMC 3nm capacity constraints — TSMC flagged tight 3nm capacity on its Oct 15 2026 earnings call (3nm = 33% of 3Q wafer sales, largest node share) and announced an additional 3nm fab in Tainan for mass production in 1H 2028. Standard (non-NVHBM) HBM4/HBM4E base dies stay in-house at Samsung.
TrendForce (19 Oct 2026, citing TechNews): some Chinese smartphone brands are sourcing memory chips recovered from discarded phones via Shenzhen Huaqiangbei traders (potentially including used iPhones) as memory procurement costs rise by up to RMB 500 per unit; the Shenzhen government is rumored to be subsidizing brands ~RMB 300/unit, leaving ~RMB 200 absorbed by manufacturers, and some brands are reportedly cutting production and sales volumes rather than passing costs on. Salvaged chips are largely limited to ultra-low-cost unbranded phones, not flagships. TrendForce frames it against continued AI-server DRAM displacement with shortages and price increases expected to persist into 2027. The price rises, subsidies and salvaged-chip use were unconfirmed by Chinese authorities or the companies as of publication.
The Elec (via TrendForce, 19 Oct 2026): Micron VP Lee Cheol said at SEDEX 2026 (Korea) that Micron is developing edge inference memory (EIM) for physical-AI applications — stacking LPDDR atop a logic die to cut processor-memory latency and raise bandwidth — with industry observers expecting mass production in 2H2028; unlike PIM, EIM's computing architecture has not been disclosed, and Micron is also exploring incorporating EIM into LPDDR6. Context (via BigGo/The Elec): on its September earnings call Micron estimated future humanoids could need >200GB DRAM and several TB NAND (vs ~12GB/256GB in today's service robots), with physical-AI customers already sampling; Lee said LPDDR suits current battery-powered humanoids (models of ~1-2B parameters) but HBM could enter high-performance humanoids as VLM bandwidth demands grow. Competitors: Samsung is adapting its automotive removable SSD for robotics and SK hynix works with NVIDIA on low-power memory for Jetson Thor (128GB LPDDR5X) (Korea JoongAng Daily); Morgan Stanley sees 2030 humanoid deliveries of 250K-1.2M units.
SK hynix's potential U.S. listing of NAND subsidiary Solidigm is fueling a debate in South Korea over whether the move would unlock value or come at the expense of SK hynix minority shareholders (KED Global, 19 Oct 2026). Critics argue the IPO would run counter to South Korea's restrictions on multiple listings within the same corporate group — Seoul has moved to prohibit listed companies from taking spun-off units public (dual-listing ban announced around the March 2026 cabinet meeting, per KED's related coverage). The listing debate is a new regulatory risk vector on top of the reported Goldman Sachs/Morgan Stanley-led, up-to-$10B IPO.
KED Global — SK Hynix's potential Solidigm IPO reignites debate over multiple listings
Rising memory prices and tightening supply are pushing some Chinese smartphone makers to source salvaged memory chips recovered from discarded phones (incl. used iPhones) via Shenzhen Huaqiangbei traders, as some chips reportedly cost up to RMB 500 more per unit (TrendForce News citing TechNews, 19 Oct 2026). Shenzhen local government is rumored to offer brands ~RMB 300/unit subsidies, leaving ~RMB 200 net extra cost; unable to fully pass costs through, some brands are reportedly cutting production and sales volumes. TrendForce framing: AI-server demand continues to squeeze memory supply, with shortages and price increases expected to persist into 2027. Salvaged chips are largely confined to ultra-low-cost unbranded phones, unlikely to enter flagship supply chains.
ExaGrid reported (19 Oct 2026) its best-ever third quarter for bookings and revenue (quarter ending 30 Sep 2026), with double-digit revenue growth over Q3 2025 and over the first nine months of 2025, exceeding plan in all three quarters of 2026; 23rd consecutive quarter of positive P&L, EBITDA and free cash flow; 100% debt-free; highest-ever share of bookings from deals over $1M; 4-quarter average win rate above 70%; 5,000+ organizations, installations in 108 countries; over 50% of bookings outside the US.
StorageNewsletter — ExaGrid Announces Record 3rd Quarter Bookings and Revenue
TrendForce telecom research (19 Oct 2026): AI data centers, DCIs and backbone builds drive a projected global optical-fiber supply shortfall of ~180 million fiber-km in 2026; finished-fiber spot prices at 2.8–5x 2025 levels or higher (BIF >5x, AI-backbone ~2.8x, mainstream telecom ~3.7x). Capacity utilization reached 92% with inventory down from ~4 months (2024) to ~1 month; some leading suppliers booked into Q1 2027; shortages expected to persist through at least 2028 since preform capacity adds need 18–24 months. Demand 577M fiber-km in 2026 with datacenter share rising from 18–20% to 30–35% in 2027; operators shifting to LTAs — e.g. Verizon–Corning multi-year agreement (Sept 2026) covering 129M fiber-km for 2027–2032 delivery.
StorageReview (19 Oct 2026) on Meta's Prometheus deep-dive published that day: gigawatt-scale, New Albany OH cluster scaling toward ~1M GPUs on two NVL72 generations. Catalina racks = GB200 NVL72 'Ariel' variant with Grace+2x ConnectX-7 400G NICs engineered to 800G of scale-out bandwidth per GPU (vs the standard 400G); Clemente racks = GB300 NVL72 with ConnectX-8 SuperNICs making 800G/GPU the native baseline; San Miguel will add Vera Rubin NVL72 (288GB HBM4/GPU, 216 TB/s NVLink, ~1,500 kW-class racks at >170 kW) to the same fabric. Training runs on a dedicated Ethernet back-end network separate from the front end, spanning buildings; reliability disclosed at >85% pre-training goodput and MTBI ~47,000 GPU-hours (post-firmware-fix >54,000). The post notes CoreWeave put Vera Rubin NVL72 into production 'earlier this month' (~mid-Oct 2026).
Meta Prometheus: Inside the Gigawatt Supercluster Scaling to 1 Million GPUs on GB200 and GB300 NVL72