GIS is allocating TWD4-5bn of capex to high-power laser packaging, which analysts estimate could contribute 30-50% of earnings by 2028 as co-packaged optics ramp up. The supplier has secured customer commitments from a leading US optics company for 300mW and 400mW continuous-wave laser packaging, with initial revenue expected to start in the first half of 2027. Although high-power lasers are projected to account for under 2% of total sales in 2027 to 2028 because GIS recognizes only manufacturing value-add without raw material costs, the high-margin profile delivers disproportionate operating leverage once shipments scale. The overall business is modeled to inflect from a projected net loss of TWD1,540mn in 2026 to net profits of TWD910mn in 2027 and TWD2,050mn in 2028. Capital deployment reinforces this strategic shift, with GIS committing TWD8bn in total capex this year, directing one-third of the packaging investment to Taiwan and two-thirds to Vietnam. Analysts note that this planned spending already surpasses the capital expenditure of the current largest laser packaging specialist, positioning the company to capture share as customers mandate outside-of-China manufacturing amid acute capacity shortages. This transformation aligns with broader industry momentum toward co-packaged optics, where peers like Coherent identify a $16.2bn content opportunity on advanced architectures and suppliers like FOCI prepare fiber array units. Key risks to monitor include consumer demand weakness across MacBooks and iPads, margin compression from component costs, and potential development delays in fiber array units. #Semiconductors# #Smartphones# Report date: 2026-11-10

Corning ($GLW) is delivering on its annualized Springboard revenue run-rate target of $20 bn one quarter early as third-quarter operations track toward the high end of guidance. Management indicated third-quarter revenue is tracking toward the upper end of its prior $4.9 bn to $5.0 bn guidance range, implying high-teens revenue growth that leadership expects to sustain through 2026 and 2027. Sell-side models project third-quarter revenue of $5.04 bn versus consensus of $5.02 bn, with operating margins reaching 22.3% versus consensus of 21.7%, driving an estimated EPS of $0.90 against consensus of $0.88. This operating momentum is led by Optical-Enterprise strength and a sequential rebound in Optical-Carrier shipments, offsetting smartphone softness in Glass Innovations. Delivering the run rate early reinforces progress toward an internal Springboard revenue target of $30 bn in 2028, above the high-confidence target of $27 bn. Even so, buy-side hurdles remain significantly higher than sell-side beat-and-raise consensus. Investor expectations already track ahead of official targets, likely requiring a formal raise to the $30 bn revenue goal and the operating margin target of at or above 20%. For the fourth quarter, analysts estimate revenue of $5.22 bn versus consensus of $5.2 bn and operating margins of 23.0% backed by 30%+ incremental margins. The critical risk to monitor is customer adoption in rack scale-up architectures, where any delay in the copper-to-optical transition would curb optical volume. #Corning# #OpticalNetworking#

Kioxia has restricted long-term customer agreements to three-year durations and capped committed volumes at approximately 50%, positioning the chipmaker to capture cyclical price gains, according to UBS. During an executive meeting on September 30, management indicated that market supply and demand conditions are expected to remain tightly constrained through end-2027. By deliberately limiting committed contractual volumes rather than locking in multi-year fixed quotas, Kioxia retains substantial uncommitted output to sell at escalating market rates. This commercial stance contrasts sharply with peer memory producers who committed larger production shares under extended terms, creating an operational margin advantage as spot and short-term contract prices climb. UBS highlights that this flexible contracting framework allows Kioxia to maximize upside exposure while competitors remain constrained by rigid low-price agreements negotiated during prior cycle troughs. Although management deferred formal guidance on 2028 market dynamics due to the extended forecast horizon, internal indicators display zero evidence of demand deterioration. The strategy delivers higher operating leverage during inflationary price waves, though it exposes the business to immediate downside pressure if macroeconomic conditions soften before end-2027. The critical variable to monitor is the quarterly trajectory of uncommitted contract pricing versus fixed long-term realizations, which will confirm whether margin expansion accelerates as management envisions. #Kioxia# #Apple# #Memory# Report date: 2026-10-15

Optical transceivers at switch front panels burden network switches with roughly 15W to 20W of digital signal processor power and $150 of component cost per module, according to packaging analysis from Bernstein and Broadcom. In traditional switch architectures, high-speed electrical data signals travel from the switch ASIC across circuit board traces and connector interfaces before reaching pluggable transceivers at the front panel. These long electrical paths produce high interface losses and degrade high-frequency signals. In addition, single-mode optical links require precise sub-assembly alignment, where coupling lasers and photodetectors to a delicate 9-micron fiber core represents roughly 30% to 35% of total optical transceiver manufacturing cost. Packaging architecture is evolving by shifting optical conversion progressively closer to switch silicon to reduce path losses and power consumption. Near-packaged optics moves the optical engine off the front panel and onto the circuit board beside the switch ASIC, shortening board-level trace lengths while running fiber to the faceplate. Co-packaged optics mounts the optical engine directly onto the ASIC substrate, bringing electrical-to-optical conversion within millimeters of silicon. This substrate-level integration minimizes electrical trace lengths and provides maximum bandwidth density, while keeping pluggable laser sources external for thermal management and replacement. Adoption depends on balancing integration density against modular repairability. #AdvancedPackaging# #NVIDIA# #TEConnectivity# #OpticalNetworking# #Networking# Report date: 2026-10-15 · Bernstein

NVIDIA ($NVDA) would incur an additional 60kW of power and $2M of cost per rack by replacing short copper cables with optical links, according to Bernstein's expert assessment with former NVIDIA leader Brad Smith. In the Vera Rubin architecture, NVLink switching sits centrally in the rack to connect compute trays above and below through roughly 5,000 short copper cables. These passive copper links operate across distances up to 18 inches at 200Gbps signaling speeds, retaining printed circuit boards with copper traces. Prior to accelerated computing workloads running across up to 1 million GPUs, copper handled roughly 90% of data center interconnects alongside 8% multi-mode optics and 2% single-mode optics, because electrical wiring provides unmatched capital and thermal efficiency. Physics dictates that achievable copper reach compresses as signaling speeds rise, falling from roughly 100m at 1Gbps to 2.5m at 100Gbps, 18 inches at 200Gbps, and 12 inches at 400Gbps. At 400Gbps, passive copper reach becomes impractical as electron radiation degrades signal integrity along the conductor. To bypass cabling congestion without abandoning electrical transmission, Feynman-era architectures approaching 1MW per rack integrate four blade assemblies directly into a midplane board with back-mounted NVLink switches. While active electrical cables with retimers can extend 200Gbps reach to 2 meters, they introduce roughly $1,200 of unit cost and 100 to 200 nanoseconds of latency. Server makers will therefore keep copper inside the rack until package-level input-output demands exceed short-reach electrical limits. #AdvancedPackaging# #NVIDIA# #TEConnectivity# #OpticalNetworking# #Networking# Report date: 2026-10-15 · Bernstein

Global memory suppliers lost $738bn in combined market capitalization following CXMT's 4F² architecture disclosure in October 2026, according to Goldman Sachs. The sell-off triggered share price declines between 11% and 26% across the sector, hitting SK Hynix down 25%, Samsung Electronics down 20%, CXMT down 15%, and Micron ($MU) alongside SanDisk ($SNDK) down 11%. Goldman Sachs considers this market reaction an overestimation of Chinese technology breakthrough potential. While equity valuations collapsed, physical market fundamentals diverged sharply, as benchmark DDR5 16Gb spot pricing continued trending upward from $54.50 past $60.00 throughout the exact same two-week sell-off period. The valuation contraction pushed Samsung Electronics to 0.9X P/B and SK Hynix to 0.9X P/B for 2028E, implying that public markets are pricing in an aggressive downcycle by 2028. Goldman Sachs views current risk-reward as highly attractive against its 'higher for longer' thesis, supported by projected 2026E/2027E ROE of 53%/54% for Samsung Electronics and 114%/68% for SK Hynix. The correction also penalized NAND producers such as SanDisk and Kioxia (down 26%) that have no technological exposure to DRAM cell architecture shifts. The primary near-term risk for equity recovery is a scarcity of immediate catalysts, alongside macroeconomic headwinds such as higher CPI prints and broader AI financing scrutiny. #Memory# #Samsung# #SKHynix# #SanDisk# #Semiconductors# Report date: 2026-10-16 · GS

Spot prices for 4-inch indium phosphide substrates climbed from around RMB4,400 to RMB7,000 per unit in September 2026, according to SMM data cited by UBS, as acute supply tightness bolstered producer pricing power. Three-inch substrate prices similarly rallied from RMB2,500 to RMB4,000 per unit, prompting UBS to model near-term contract price hikes of 20% to 40% for JX Advanced Metals. To satisfy high-speed 1.6T transceiver roadmaps, optical component manufacturers are demanding large-diameter 6-inch substrates. Optical leader Coherent highlighted that migrating from 3-inch to 6-inch wafers more than quadruples device yield per wafer while roughly halving unit production costs. However, supplying 6-inch wafers requires compound semiconductor manufacturers to master entirely new crystal growth physics under severe thermal stress. While JX Advanced Metals historically utilized liquid encapsulated Czochralski (LEC) pulling for smaller wafers, 6-inch fabrication forces a transition to vertical gradient freeze (VGF) solidification, where the firm lacks mass-production experience. Controlling phosphorus volatilization and thermal stress across larger diameters triggers steep yield penalties during initial scaling. Furthermore, commissioning ¥120bn in new capacity will expand depreciation charges, while depleting low-cost legacy indium stockpiles will elevate raw material costs. Consequently, UBS forecasts JX Advanced Metals' InP operating profit will rise from ¥3bn in FY3/27 to ¥10bn in FY3/28 and ¥16bn in FY3/29, but expects operating margins to peak at 42% in FY3/28 and erode thereafter as rival supply curbs pricing power. #Commodities# #Materials# Report date: 2026-10-15

CoreWeave ($CRWV) needs its active data center power to reach 4x to 5x annual additions to generate positive free cash flow, an inflection J.P. Morgan models only in 2030E at a 4.5x ratio. The current cash deficit stems from the sheer pace of building capacity rather than poor economics on individual facilities. Each gigawatt demands roughly $35.0B in upfront capex during its build year against $10B to $15B in stabilized annual revenue and $9.4B in annual cash EBITDA at a 75% margin. Because CoreWeave entered 2026 with 0.85GW of active power while adding 1.0GW during the year, its ratio of active to added capacity sits at just 0.9x. The company cannot self-fund while expanding its footprint fourfold from 1.85GW exiting 2026 to 8GW by 2030. J.P. Morgan projects annual additions will accelerate to 1.5GW in 2027E, keeping the ratio depressed at 1.2x before rising to 2.1x in 2028E on 1.6GW added, 3.1x in 2029E on 1.6GW added, and finally 4.5x in 2030E when adding 1.5GW. As stabilized gigawatts accumulate, cash EBITDA from the mature base expands rapidly enough to absorb the $35.0B per gigawatt capex required for new facilities. The primary downside risk centers on power availability or facility delays that could stretch build cycles beyond planned timelines, which would prolong debt reliance before the installed base reaches critical mass. Investors should track annual megawatt commissioning versus guidance and debt service costs as new sites come online. #CoreWeave# #DataCenters# Report date: 2026-10-16 · JPM

MediaTek is set to generate US$18bn in Google TPU sales in 2027, according to UBS forecasts. TPU contribution will surge from US$2,113mn in 2026 to US$43.5bn in 2028, representing 65% of company sales. Operating margins on custom silicon should widen from 22% in 2026 to 37% in 2027 and 39% in 2028, lifting consolidated EPS from NT$67.6 to NT$363.4. Volume deliveries of v8t chips remain on track for over US$2bn in Q4 2026. On next-generation v9 silicon built on TSMC N2 and EMIB-T packaging, Ibiden packaging substrate yields have reached the 50% target for mass production ahead of schedule. The key near-term indicator is Google tape-out delivery by early November, with any push-out requiring customer notice within three weeks. #Semiconductors# #TSMC# Report date: 2026-10-12 · UBS

Samsung Electronics HBM revenue is projected to surge 321% YoY to US$81bn in 2027, according to Goldman Sachs. Blended HBM average selling prices are estimated to exceed US$4.0 per Gb, representing a 159% annual increase driven by severe industry supply shortages. High-bandwidth memory will expand from 8% of total DRAM sales in 2026 to 20% in 2027. Foundry fundamentals are improving alongside memory demand, led by 4nm base die production for next-generation HBM4 architecture. Sustaining this revenue expansion requires successful HBM4 qualification and execution on long-term customer agreements. Key downside risks highlighted by the analysts include memory supply-demand deterioration and potential customer despeccing. #Samsung# #HBM# Report date: 2026-10-08 · GS

OpenRouter LLM token volume accelerated 71% MoM and expanded 30x YoY in September 2026, driven by surging open-source developer activity, according to J.P. Morgan. Open-weight model volumes surged 83% MoM and 101x YoY, far outpacing closed-weight volume growth of 39% MoM and 10x YoY. The rapid shift highlights accelerating enterprise migration toward cost-effective open-source architectures. Closed-weight models saw their share of platform volume slip to 27% in September, down from 33% in August and 29% in July. While average catalog pricing rose 4% MoM due to new image models, total spend growth reached 28% MoM. A key operational constraint is that OpenRouter skews toward developer and agentic-coding traffic, excluding direct first-party hyperscaler API endpoints. #AIChips# #Tencent# Report date: 2026-09-29 · JPM

Nvidia GPU rental prices diverged across hardware tiers in September 2026 as non-hyperscaler B200 rates rebounded 1.3% MoM to $5.70 per GPU-hour, according to J.P. Morgan. The recovery widened the B200-to-H100 pricing multiple to 2.16x from 2.08x in August, reflecting resilient developer demand for cutting-edge architecture even as broader compute supply loosened across older accelerator tiers. In contrast, previous-generation H100 rental rates fell 2.6% MoM to $2.64 per GPU-hour, while A100 pricing dropped 2.8% MoM to $1.59 per GPU-hour. The H100-to-A100 price ratio ticked up slightly to 1.66x. If enterprise workloads continue migrating toward next-generation silicon, lingering rental price erosion on H100 and A100 clusters could compress operating margins for independent cloud hosters. #AIChips# #Tencent# Report date: 2026-09-29 · JPM

Amazon AWS Trainium 3 liquid-cooled racks will reach ~$86,000 in cooling component content, exceeding NVIDIA ($NVDA) Vera Rubin racks at ~$66,250, according to J.P. Morgan. Despite a lower 700W ASIC TDP compared to 2,300W on GPUs, aggressive thermal layouts drive compute tray value to $3,000-$3,500 across 18 trays alongside 10 switch trays contributing up to $2,000 each. Each compute tray integrates a double-sided configuration of 8 cold plates for four ASICs plus an extra tray-sized baseboard plate. The system deploys 512 total quick disconnects per rack valued at ~$10,200. Liquid cooling suppliers AVC and Fositek face scheduled 3Q/4Q26 volume ramps, with multi-rack cluster expansion serving as the decisive operational catalyst. #AIChips# #NVIDIA# Report date: 2026-09-30 · JPM

NVIDIA ($NVDA) Rubin Ultra GPUs will increase heat spreader content by 4x to 5x over standard Rubin lids by shifting to a three-piece removable design, J.P. Morgan projects. The architecture combines a stiffener, temporary lid, and multiple screws during advanced packaging and testing. ODMs then remove the lid to place liquid cold plates directly over TIM1.5 interfaces. Trial production begins in small volumes in 4Q26 for Rubin GPUs before entering full mass production in 1Q-2Q27 for Rubin Ultra. Jentech leads supply but faces valuation risk at ~50x forward P/E if initial adoption remains restricted to trial volumes rather than immediate fleetwide conversion. Investors should monitor trial yield rates and packaging pass rates throughout 2026. #AIChips# #NVIDIA# Report date: 2026-09-30 · JPM

MediaTek is transforming its business model as UBS forecasts Google TPU sales to surge from US$2,113mn in 2026E to US$43.5bn in 2028E, reaching 65% of total corporate revenue. Improving EMIB-T substrate yields at Ibiden, Unimicron, and Shinko underpin projected volume exceeding 2mn TPU v9 units in 2028. UBS models TPU operations generating a 42% gross margin and contributing NT$291.4 to MediaTek's NT$363.4 consolidated EPS in 2028E, alongside potential US$10-15bn pipeline projects across SpaceX Dojo 3 and Tesla AI6. Key variables to track include substrate yield learning curves and ASIC delivery schedules. #Semiconductors# #Technology# Report date: 2026-09-28 · UBS