Zero One Investment Research Daily Intelligence Brief, August 24, 2026

Even NVIDIA cannot absorb memory costs as Microsoft installs the first Vera Rubin racks

Top Market Signals

NVIDIA passes memory costs into server prices

NVIDIA (NVDA US) has told several of its largest customers that servers built around its AI chips will cost more than 15% more, with the increases landing on Grace Blackwell and Vera Rubin systems that ship early next year. Bloomberg reported the notices on Saturday, citing people familiar with communications that have not been made public. NVIDIA has not confirmed them.

The size of each increase depends on the chip generation and the memory configuration involved. Contract manufacturers that build servers for Microsoft, Google and Oracle have already passed the notices to their own customers. The Information puts the rise at about 17%, citing server makers.

Memory is the reason. Analysts quoted by Tom's Hardware project conventional DRAM contract prices rising 58% to 63% QoQ in 2Q26, after a 90% to 95% jump in 1Q26, as suppliers move advanced nodes toward HBM and server products. A Rubin GPU ships with up to 288GB of HBM4 per package, and an NVL72 rack combines 72 of them, more than 20TB of HBM before counting the LPDDR attached to the Vera CPUs. HBM consumes roughly four times the wafer area of the same capacity in conventional DRAM.

Taiwan's assemblers gain average selling price and carry the volume risk. Economic Daily News reports that Foxconn (2317 TT), Quanta (2382 TT) and Wistron (3231 TT) all see rack ASP rise if the increase holds, and that the open question is whether higher system prices thin out cloud operators' order volumes. Supply chain sources cited by Commercial Times say the tightness has spread from HBM into server DRAM, LPDDR and SOCAMM, lifting the bill of materials for a whole rack rather than one component. TrendForce expects DRAM supply to stay tight into 2027 and says NVIDIA is evaluating a change to the HBM capacity configuration on Rubin Ultra to hold cost down.

Linked stocks: NVDA US, 2317 TT, 2382 TT, 000660 KS

Microsoft turns on the first Vera Rubin racks

NVIDIA and Microsoft (MSFT US) said the Vera Rubin platform has begun installation, making Microsoft the first cloud operator to deploy it at data center scale. NVIDIA announced volume production for the platform in May, and the first racks are now going in.

Vera Rubin is the largest system NVIDIA has built at the POD level, assembled from five dedicated rack types and seven purpose-built chips, among them the Rubin GPU, the Vera CPU, the NVLink 6 switch, the ConnectX-9 SuperNIC, the BlueField-4 DPU and the Spectrum-6 Ethernet switch. Microsoft will run Vera Rubin NVL72 racks in its next generation AI data centers, including the new Fairwater sites, supplying Azure customers and OpenAI model training and inference workloads.

The supply chain behind it is twice the size of Grace Blackwell by NVIDIA's own count, spanning 350 factories with more than 150 Taiwanese partners. Dell, HPE, Lenovo, Supermicro, ASUS, Gigabyte, Pegatron, Quanta's QCT unit and Wistron are all using NVIDIA's DSX platform to speed the ramp.

Foxconn is expected to supply first. The company has kept its 3Q26 volume production target for Vera Rubin racks, with shipments starting in 4Q26 and the platform becoming its main product next year. Foxconn also said Grace Blackwell demand runs at least into next year, that AI rack shipments grow each quarter, and that 3Q26 shipments should rise by a high double-digit percentage QoQ with full-year volume more than doubling. The window from 2H26 into 2027 is when the installed base shifts from Blackwell to Rubin.

Linked stocks: NVDA US, MSFT US, 2317 TT, 2382 TT

Largan becomes a direct NVIDIA optical supplier

Largan Precision (3008 TT), the smartphone camera lens maker, has extended from optical components into fiber array units, the assembled module that carries light into and out of a co-packaged optics link. Supply chain sources cited by Economic Daily News place Largan alongside Taiwan's Browave (3363 TT) and China's T&S Communications in NVIDIA's fiber array unit supply chain, with volume production as early as 2H27 and market talk pointing at the Rubin Ultra platform. Largan declined to comment on 23 August and said its current focus is the fiber array itself.

The step up changes what Largan sells. A fiber array unit combines the fiber array with a micro lens array, carries a higher selling price and a higher gross margin than either component alone, and moves Largan from an indirect NVIDIA supplier to a direct one. TSMC has already taken Largan into its COUPE silicon photonics ecosystem as a fiber array unit supplier.

Alignment precision is what has kept the rest of the field out. A fiber array holds fibers in V-grooves, and at 64 and 128 channels the positional error on every fiber has to be tiny before the assembled unit can be aligned well enough to mass produce. Largan chairman Lin En-ping has said that the company pairs an imperfect V-groove with an imperfect fiber to produce a good fiber array, at a precision below 0.3 micron against roughly 0.5 to 0.8 micron for the best of the rest.

Our 18 Aug 2026 Blade reported that NVIDIA's Spectrum-X Ethernet Photonics switch had entered full volume production and named Browave among four Taiwanese partners on it. Today's reporting sizes what those switches consume: at roughly 20,000 Spectrum-X units this year, the optical engines inside them need 600,000 to 700,000 fiber array units, rising to several million in 2027, against a supply chain that cannot currently meet it. Largan is separately competing in micro lens arrays, using molded glass against the wafer level optics that Himax Technologies (HIMX US) builds with semiconductor processes.

Linked stocks: 3008 TT, 3363 TT, NVDA US, HIMX US

TSMC reportedly buys two AUO panel fabs

TSMC (2330 TT) is reported to be paying more than NT$30bn for two older AU Optronics (2409 TT) panel plants, L7 and L5C, at the Central Taiwan Science Park next to its own site there. Commercial Times reports that TSMC has completed on-site due diligence and that AUO plans to dispose of the two plants first, aiming to take the transaction to its board in October. Neither company has confirmed it.

What a foundry wants from a display fab is the building, not the equipment. Panel plants come with large-area cleanrooms, ample power and the handling systems for moving large sheets of glass, which shortens construction time and fills the gap in square-format packaging: CoPoS, chip-on-panel-on-substrate, and FOPLP, fan-out panel-level packaging. Both are the formats that large optical and electronic integration will need. Line relocation and vacating of the buildings are expected before the end of 1Q27.

The park is already turning from a wafer site into a process and packaging cluster. The existing Fab 15A there runs mainly 28nm, which can support silicon interposer manufacturing, and the second phase Fab 25 plans four A14 fabs at the 1.4nm node. The first two are in steel construction, with the first building targeted for completion before April 2027 and trial production as early as 3Q27 against official A14 volume production in 2028. The advanced packaging site AP5 already runs CoWoS, chip-on-wafer-on-substrate.

Optics is filling in around it. Largan has bought land and plants in Taichung's Nantun district, the Taichung Industrial Park and near the precision machinery cluster, three transactions totaling about NT$2.568bn, and its first automated fiber array pilot line is targeted for completion by the end of 3Q26. The AUO board decision in October is the next thing that would confirm or kill the fab transaction.

Linked stocks: 2330 TT, 2409 TT, 3008 TT

NVIDIA pays US$6bn for Poolside's technology

NVIDIA is paying US$6bn to license technology from the AI coding startup Poolside and plans to use it to build one of the world's most powerful open-source AI models, according to Wall Street Journal reporting carried by Liberty Times. The stated target is Chinese open models such as DeepSeek and Kimi K3.

Poolside wrote to shareholders on 20 August announcing a reorganization of the business. Per the letter, NVIDIA will invest US$1bn at a US$12bn pre-money valuation and pay the US$6bn separately for the technology license, and will hire most of Poolside's engineers. More than 100 Poolside employees join NVIDIA and work on Nemotron, the open model program NVIDIA first disclosed in 2023, contributing to its largest and most complex model. Founders Eiso Kant and Jason Warner, formerly GitHub's chief technology officer, stay out and continue on research projects they have not described.

The licensing deal also points at the closed model developers. Open models generally cost far less to run and are easier to customize, so a well-funded open alternative competes directly with OpenAI and Anthropic as well as with the Chinese labs. American AI companies have put most of their resources into closed models, which raised the prospect of companies and governments worldwide standardizing on cheaper Chinese open weights instead. NVIDIA has widened its own open model work over the past six months, partly in response.

Poolside's letter framed the deal as securing a future in which general AI is built in the open by many participants rather than held as closed technology by a few, translated from the Chinese-language report. What NVIDIA actually releases from Nemotron, and under which license, is what would test that framing.

Linked stocks: NVDA US, MSFT US, GOOGL US

Compal reportedly wins Google TPU rack orders

Compal Electronics (2324 TT), which came to AI server assembly later than its Taiwanese peers, is reported to have won orders for two Google tensor processing unit server programs at the L10 and L11 build levels, the node assembly and full rack integration stages. Economic Daily News reports further North American cloud ASIC server orders behind them. Compal declined to comment and said customer and order information follows its exchange filings and earnings calls.

The company had already signaled the shape of it without naming the customer. Compal's chairman said earlier that hyperscaler and enterprise customers were expected to begin volume production on new L10 and L11 lines by the end of this year, and that AI server revenue should grow sharply next year to reach 30% to 40% of total revenue.

Google's build-out is what creates the orders. The company has been placing custom silicon work with MediaTek (2454 TT), Broadcom and Marvell while continuing to buy NVIDIA and AMD GPU servers, so the self-designed rack business grows alongside the merchant one rather than replacing it. Institutional investors cited by the paper say the TPU codenamed Zebrafish enters wafer production at TSMC in 2H26 and reaches assemblers before year end, which would put Compal's programs into volume production next year. A second TPU codenamed Sunfish may enter production at the same time, which would give Compal both.

Capacity is being built to match. Compal has AI server sites in Taiwan for research and production, Vietnam for surface-mount work, and Texas for L6, L10 and L11 assembly, and has budgeted NT$18bn of capital spending this year against NT$8bn to NT$10bn last year, mostly on new AI server lines. Another North American cloud ASIC program is still in planning.

Linked stocks: 2324 TT, GOOGL US, 2454 TT, 2330 TT

先聲 First Word: Exclusives from Chinese-Language Sources

LandMark wins indium phosphide back-end orders

LandMark Optoelectronics (3234 TT) is reported to have won back-end process orders for indium phosphide, the compound the lasers in high-speed optical modules are built on, and has indirectly entered the supply chains of Lumentum and Applied Optoelectronics. Its self-compiled July result turned to profit after a first-half loss, and it plans to double back-end capacity by year end against 3Q26 and double again next year. Indium phosphide back-end capacity is one of the few visible relief valves in the 1.6T optical build-out. (24 Aug 2026) Source: 光環傳拿下 InP 後段製程訂單 間接打入美光通訊鏈

MediaTek and Ericsson hold outdoor positioning inside 30cm

MediaTek (2454 TT) and Ericsson (ERICB SS) completed what they describe as the first end-to-end test of 3GPP-standard high-precision outdoor positioning, holding error within 30cm over a commercial 5G network using MediaTek's newest modem and the device's own built-in antenna. The satellite correction data travels inside the existing 5G network rather than through a separate positioning system or extra hardware, which lowers the cost of precise location for self-driving cars, drone fleets and factory robots. (24 Aug 2026) Source: 聯發科揪伴 瞄準 5G 應用

China slows germanium and quartz shipments to Taiwan

China has been restricting exports of germanium and high-purity quartz to Taiwan through customs measures such as stepped-up inspection and delayed clearance rather than a formal ban, according to reports carried by Economic Daily News. TSMC (2330 TT) said it does not expect a material effect on its operations, and Topco Scientific (5434 TT) said its quartz unit sources raw material mainly from Germany and secondarily from Japan. Restriction by administrative friction is harder to measure and harder to appeal than a published control list. (24 Aug 2026) Source: 大陸限制對台鍺、石英等出口 半導體業3大觀點估影響有限