Zero One Investment Research Daily Intelligence Brief, August 26, 2026

OpenAI's 700W chip claims to beat NVIDIA's 1,400W GB300 on throughput per watt

Top Market Signals

OpenAI's first chip claims more work per watt

OpenAI presented the first published benchmarks for Jalapeño, the inference chip it co-developed with Broadcom (AVGO US), at the Hot Chips conference on 25 August. On SemiAnalysis's InferenceX suite, systems built on the 700W part delivered 1.5 to 1.9 times more throughput per kilowatt and 1.7 to 3.6 times lower end-to-end latency than NVIDIA's (NVDA US) GB200 and GB300 rack systems.

Hardware product lead Richard Ho said the tests covered three open models: OpenAI's own gpt-oss-120b, DeepSeek R1 and Moonshot AI's Kimi K2.5. OpenAI deploys Jalapeño 128 chips to a rack, and a full pod of 2,048 chips carries 1.7 exaFLOPS of 4-bit compute and 27.5TB of HBM4 (high-bandwidth memory), with 15.4TB per second of memory bandwidth per package. SemiAnalysis, which ran the suite alongside OpenAI engineers in OpenAI's lab, described the part as beating every NVIDIA, AMD and Google chip it has been able to test.

The comparison carries limits OpenAI states itself. Results were normalized to each accelerator's published package power, and Jalapeño's measured sustained draw stayed at or below 550W. An appendix using all-in utility power, 1.18kW for Jalapeño against 2.55kW for the GB300, narrows the gaps, as does running the GB300 with multi-token prediction, where the peak efficiency lead falls to roughly 1.5 times. Jalapeño was not tested against Vera Rubin, and it does not train models.

Ho said OpenAI is not slowing its NVIDIA deployments, and answered a question about selling the chip to other companies by pointing to how hard OpenAI finds it to build enough volume for itself. A second generation is in deep development and a third is under way. Each package pairs its compute die with six HBM4 stacks totaling 216 GiB, drawn from a supply where Samsung, SK Hynix and Micron (MU US) have sold capacity through 2027. Limited deployment in OpenAI's own data centers starts later this year and widens in 2027, which is when the efficiency claim gets tested against Vera Rubin rather than against Blackwell.

Linked stocks: AVGO US, NVDA US, MU US, 005930 KS

NVIDIA's photonic switch scales to 512,000 GPUs

NVIDIA (NVDA US) used Hot Chips 2026 to set out the architecture around Spectrum-X Ethernet Photonics, the co-packaged optics switch our 18 August Blade flagged as entering volume production. The Multiplane design splits each server's connections into independent planes, each running a two-tier network, so a single AI factory can reach 512,000 GPUs without adding a third network tier.

The resilience numbers are the other new disclosure. In an eight-plane setup, losing one plane still leaves about 90% of total bandwidth, and hardware recovery runs 11 times faster than software load balancing. Against networks using conventional pluggable optical modules, NVIDIA claims five times better power efficiency, five times longer continuous operation and 1.3 times faster data center deployment.

Commercial Times also puts numbers on the Taiwanese division of labor. TSMC (2330 TT) manufactures the silicon photonics components, integrating the photonic and electronic integrated circuits with its advanced packaging platform. ASE Technology's (3711 TT) SPIL unit handles chip-level packaging, assembly and test. Foxconn (2317 TT) integrates the switch into a complete rack-ready network system. As Vera Rubin volumes build, Quanta's (2382 TT) QCT unit, Wistron (3231 TT), Wiwynn (6669 TT) and Inventec (2356 TT) pick up switch, SuperNIC and rack integration work.

Volume production moves co-packaged optics from demonstration into commercial deployment, and it moves where the difficulty sits. Fiber array units, external lasers, microlenses, fiber coupling and active alignment become the hard parts, with Largan Precision (3008 TT) among the Taiwanese firms named as holding the relevant technology. Test moves earlier as well, from final inspection to the wafer and bare-die stage, covering optical power, wavelength, insertion loss, coupling efficiency, signal integrity and bit error rate. Yield at that test stage is what decides whether the Taiwanese content here holds.

Linked stocks: 2330 TT, 3711 TT, 2317 TT, NVDA US

NVIDIA's dedicated inference racks reach volume production

NVIDIA (NVDA US) said on 24 August that racks built on Groq 3 LPX, its low-latency inference accelerator, have entered full volume production. Each rack carries 256 Groq 3 LPU chips and pairs with the Vera CPU and Rubin GPU, with first deployment at cloud provider Nebius later this year.

NVIDIA licensed Groq's inference technology last year and hired its core team, then launched the Groq 3 LPU chip in March 2026. The company is explicit that LPX does not replace GPUs. Vera Rubin NVL72 handles broader training, inference and context processing, while LPX targets interactive speed and token generation for agent workloads such as coding, where the constraint is producing tokens at very low latency while holding a growing context.

On Artificial Analysis testing with the Gemma 4 31B model at a 100,000-token context, Groq 3 LPX processed 3,400 tokens per second, the highest recorded for that model. NVIDIA says response speed reaches up to four times that of the nearest competing platform. Each chip carries 500MB of on-chip SRAM to reduce the memory bottleneck, produced by Samsung Electronics (005930 KS), and the GPU is fabricated by TSMC (2330 TT). Analysts see Foxconn (2317 TT) as the largest assembler of these racks.

The ramp pulls high-speed board materials with it. Commercial Times reports that LPX volume production is lifting demand for high-end copper-clad laminate, low-roughness copper foil and printed circuit boards, with combinations pairing M9 resin with quartz glass fabric and M8 with lower-loss glass cloth. Elite Material (2383 TT) was among the first to qualify and ship M9, and analysts expect M9 to move from a single-digit share of its revenue in 2H26 to a double-digit share in 2027. The company's monthly capacity plan runs from 6.15m sheets at end-2026 to 9.3m by end-2027 and 11.1m by end-2028.

Linked stocks: NVDA US, 2317 TT, 2383 TT, 2330 TT

Global Unichip adds Microsoft and Meta after Google

Global Unichip (3443 TT) has won custom-chip orders from Microsoft (MSFT US) and Meta (META US), adding to the Google project it already runs. That gives the Taiwanese design-service house work from three of the largest cloud service providers at once.

On the Microsoft side, Liberty Times reports that Global Unichip is working on both the Maia 200 AI accelerator and the Cobalt 200 data center processor. Maia 200 is already operating in Microsoft data centers and the Cobalt line is scaling quickly. Microsoft's demand on Global Unichip strengthened over the past quarter and is expected to more than double next year, contributing at least US$250m of revenue in 2027.

The Meta work is earlier. Supply chain tracking points to a next-generation Meta ASIC (application-specific integrated circuit) project led by the Rivos team, possibly the MTIA 600, with tape-out as early as 1H27 and revenue emerging from late 2027.

The same pattern is visible one project over. Morgan Stanley wrote this week that as Google's TPU v10 grows more complex, MediaTek (2454 TT) stays the lead coordinator but has room to hand individual blocks to several design-service suppliers, with Global Unichip confident of participating and Google's customer-owned-tooling model opening a path for Alchip (3661 TT). Tape-out timing is what turns any of this into revenue, and the Meta project has not taped out yet.

Linked stocks: 3443 TT, 2454 TT, 3661 TT, MSFT US

Samsung moves compute into low-power memory

Samsung Electronics (005930 KS) detailed the industry's first LPDDR5X-PIM at Hot Chips 2026, putting processing logic directly inside low-power DRAM. In inference tasks the company says the part runs 2.28 times faster than standard LPDDR5X.

Processing-in-memory places a small block of logic next to the DRAM cells so basic calculations happen where the data already sits, taking the processor out of the path. Samsung demonstrated the idea in 2021 and piloted it through HBM stacks in AMD (AMD US) accelerators. The advance here is that every memory bank gets its own logic block, where HBM-PIM forced Samsung to cut banks to make room.

The reason to attach this to LPDDR5X rather than to HBM is cost. Samsung opened its presentation on the point that memory makes up the bulk of an AI chip's cost and that its share keeps rising. Micron (MU US) told the same conference a day earlier that HBM wafer demand is getting worse, and Samsung, SK Hynix (000660 KS) and Micron have sold HBM capacity through 2027.

Samsung says LPDDR5X-PIM has moved from concept to a validated product, and it is already working on LPDDR6X-PIM, hoping for an initial JEDEC specification this year. Two things would move this beyond a conference demonstration: a named customer, and controller support broad enough that switching between standard reads and processing reads does not land on each buyer to solve.

Linked stocks: 005930 KS, 000660 KS, MU US, AMD US

Intel's enhanced 18A wins a first outside customer

Socionext, the Japanese design-service company, disclosed that it will use Intel's (INTC US) 18A-P process to develop custom system-on-chip products. Intel has kept 18A largely for its own products, so a named external commitment to the enhanced version is the first evidence that the node can sell outside the company.

Socionext said the work answers rising demand from data center, edge computing and high-performance computing customers, with a high-performance computing chiplet as the first project. It framed the arrangement as its own design expertise combined with Intel Foundry's process and advanced packaging roadmap, aimed at workload-specific system-on-chip designs.

Intel put 18A-P into risk production in June, describing it as up to 9% faster than 18A at the same power. Socionext did not disclose when the resulting chip goes into volume production, and on normal development timelines it is unlikely to ship ahead of Intel's own 18A-P products. What turns a single disclosure into a foundry business is a second and third named customer on the same node, and neither company has put a production date on this one.

Linked stocks: INTC US, 2330 TT, AMD US

先聲 First Word: Exclusives from Chinese-Language Sources

Silicon wafer prices rise across all three diameters

A fund manager quoted by Economic Daily News says Taiwan's wafer makers, including Formosa Sumco (3532 TT), GlobalWafers (6488 TT) and Wafer Works (6182 TT), have raised 6-inch, 8-inch and 12-inch wafer prices by an average of 10% or more, the first broad increase in over three years. The claimed shortfall is specific: above 1m pieces a month in the 12-inch high-end epitaxial wafers built for AI and high-performance computing. (26 Aug 2026) Source: 台股夯實腳步 矽晶圓產業營運成長動能看旺

A TSMC chemical-tubing supplier is building in Hsinchu

Japan's Nichias and a Taiwanese development partner plan a specialty plastics plant in Hsinchu producing PFA tubing, the piping that carries ultrapure water, slurries and strong acids through chip manufacturing, with production targeted for early 2027. PFA tubing was a genuine bottleneck on fab expansion in 2021 and 2022, when lead times stretched to a year and TSMC (2330 TT) stepped in to route suppliers straight to upstream material makers. (26 Aug 2026) Source: 台積關鍵供應商 Nichias 將來台設廠 計劃明年投產 PFA 管材

Intel's x86 share slips below 70% for the first time

Intel's (INTC US) combined share of x86 processors across PCs and servers fell to 69.3% in 2Q26, below 70% for the first time, while AMD (AMD US) reached a record high, according to market research cited by DigiTimes. The AI build-out is giving Intel data center growth at the same time as it erodes the processor position that data center revenue was built on. (26 Aug 2026) Source: 資料中心復甦難掩CPU霸權失守　英特爾市佔首跌破7成、超微攀上歷史新高