➀ Xscape Photonics has secured $44 million in Series A funding; ➁ The round was led by IAG Capital Partners; ➂ Cisco Systems Inc. and Nvidia Corp. participated in the round.

➀ Kioxia introduces the XD8 series of PCIe 5.0 SSDs for data centers; ➁ The series offers sequential read speeds up to 12.5GB/s and write speeds up to 5.8GB/s; ➂ Kioxia's third generation of E1.S SSD models with NVMe 2.0 compliance and support for OCP NVMe SSD v2.5 specification.

➀ AMD unveiled a series of AI products including a new flagship AI chip, server CPU, AI network card, DPU, and AI PC mobile processor; ➁ The flagship AI chip AMD Instinct MI325X GPU uses HBM3E high bandwidth memory, with AI peak computing power reaching 21PFLOPS; ➂ AMD disclosed the latest AI chip roadmap with the MI350 series using CDNA 4 architecture, expected to be launched next year; ➃ AMD launched the world's first AI network card supporting UEC and the Pensando Salina 400 DPU, doubling performance; ➄ The fifth-generation EPYC server CPU supports up to 192 cores and 384 threads, with a top price of 14813 USD per core; ➅ AMD's third-generation commercial AI mobile processor Ryzen AI PRO 300 series is expected to be on the market with more than 1 million units by 2025.

➀ AI is transforming data center infrastructure with innovations like advanced connectivity and chiplets; ➁ The shift to optical connectivity is enhancing performance and efficiency; ➂ Chiplet ecosystems are enabling scalable and cost-effective solutions for AI workloads.

➀ Supermicro introduces its datacenter-scale liquid cooling solution for Nvidia Blackwell servers; ➁ The solution supports up to 96 B200 GPUs per rack and includes advanced cooling components; ➂ The technology aims at ultra-dense AI servers and features rapid deployment capability.

➀ Supermicro announces a complete liquid cooling solution for data centers, aiming to reduce TCO and energy costs. ➁ The solution includes CDUs, cold plates, CDMs, cooling towers, and management software. ➂ Supermicro has delivered over 2,000 liquid-cooled racks since June 2024. ➃ The solution is designed for high-wattage servers with latest GPUs and CPUs, improving efficiency and reducing costs. ➄ Supermicro's liquid cooling technology can save up to 40% energy compared to air cooling.

➀ OCP is an open-source hardware project aiming to revolutionize data center hardware; ➁ OCP NICs are standardized, easy to use, and open for fast technological updates; ➂ OCP NICs offer a variety of interface speeds and are compatible with different server configurations; ➃ The development of OCP NICs is aligned with the evolution of server CPUs and network bandwidth requirements; ➄ OCP NICs are designed for space-intensive server environments.

➀ JFS Laboratory, a Chinese government-backed company, has developed a silicon photonics chip, marking a first for the country; ➁ The chip is expected to have advantages over traditional chips in high-performance applications like AI and HPC; ➂ The breakthrough is significant in the context of sanctions against China and the inability to produce high-performance processors.

➀ Marvell's visibility of accelerated growth became clearer after the Q1 2025 earnings release; ➁ The Data Center (DC) division saw a sequential revenue growth of 7.9%, accelerating from 6.7% in Q1 2025; ➂ No signs of slowing down in数据中心 investments, with recent highlights including Microsoft's agreement and BlackRock's AI fund; ➃ Marvell's order volume of 800G PAM and 400ZR data center interconnect products remained stable in Q2 2025, and production of 800G DSP for AEC will increase; ➄ Analysts expect Marvell's EBITDA to grow 27% in FY25 and 60% in FY26, with a current growth rate of about $26 billion.

➀ Rambus details its new HBM4 memory controller with speeds up to 10Gb/s and bandwidth of 2.56TB/sec; ➁ HBM4 offers 33% faster bandwidth than HBM3E and twice the speed of HBM3; ➂ SK hynix and Samsung are expected to mass produce HBM4 memory by the end of 2025.

➀ NVIDIA's new GB200 NVL72 AI server faces significant challenges due to its 132kW TDP; ➁ This makes it the highest-power-consuming server in history; ➂ The development of the server is experiencing delays due to technical issues.

➀ IBM opened its first quantum data center outside of the US in Ehningen, Germany, with Chancellor Olaf Scholz in attendance; ➁ The center houses two IBM Quantum Eagle-based systems and an upcoming IBM Heron system, enhancing IBM's global quantum network; ➂ The German government is supporting the development of quantum technologies and the center is expected to foster innovation and business growth.

➀ This letter discusses the updates and activities of STH in Q3 2024, including the challenges and achievements in filming at the Paris 2024 Olympics, the upcoming server reviews, and the expansion of the STH team and channels. ➁ The author reflects on the growth of the STH YouTube channel and the introduction of the Axautik Group for more specialized content. ➂ The letter concludes with a call for team expansion and a reminder to subscribe to STH's newsletter and YouTube channel for more content.

➀ Yuanxin's core products include tunable laser diodes with 13 dBm output power and 50 nm tuning range, and LNOI modulators with 3 cm device length and 40 GHz RF bandwidth.➁ Wuhan Anpai Optoelectronics offers LNOI MZM modulators with 100 GHz EO bandwidth and Y-waveguide modulators for chip integration and miniaturization.➂ Neo Optoelectronics presents single-wavelength 400Gbps modulators with bandwidths exceeding 110GHz and 1.6T/800G DPIQM modulators with 70GHz bandwidth and 25dB extinction ratio.➃ Suzhou Yilaiwei develops silicon photonics chips with D2W bonding technology and achieves 70GHz electrical-optical modulation bandwidth.➄ Guoguang Optoelectronics successfully打通8-inch low-loss SIN platform and develops TFLN/SIN heterojunction technology for 800G/1.6T data center applications.

1. Data center revenue grew 154% year-over-year to $26 billion, driven by strong demand for Hopper, GPU computing, and networking solutions. 2. Blackwell is expected to begin production ramp-up in the fourth quarter and continue into FY2026. 3. The trend of continued purchases of the best-in-class compute for AI is likely to continue as next-generation models require 10 to 20 times more compute power to train with significantly more data. 4. Cloud service providers account for 45% of Nvidia's data center revenue, with consumer, Internet, and enterprise companies making up more than 50%. 5. Management highlights accelerated computing and generative AI as major computing transitions, justifying ongoing capital expenditures and long-term sustainability of GPU demand.

➀ Broadcom has announced the general availability of Sian2, its 200Gbps per lane PAM-4 DSP PHY that supports 100Gbps electrical and 200Gbps optical interfaces. Sian2 and Sian DSPs are designed to enable pluggable modules with 200G/lane interfaces for next-generation AI clusters. ➁ The migration from 400G/800G links with 100G/lane optics to 800G/1.6T links with 200G/lane optics is necessary for higher bandwidth, lower power, lower latency, and lower cost. ➂ Sian2 and Sian DSPs are optimized for 800G and 1.6T optical module platforms, doubling the bandwidth with lower power, lower latency, and lower cost per bit to facilitate AI data center scale.

Vespertec has launched its AI lab environment in Stockport, featuring the latest NVIDIA Grace Hopper architecture and the SN5600 switch. The Test Drive programme allows customers to explore and validate cutting-edge computing technologies. The lab is equipped with the NVIDIA GH200 Grace Hopper Superchip and a PCIe-based GPU server, linked by the NVIDIA Spectrum-4 SN5600 switch and Edgeware interconnect. The remote access feature enables users from various backgrounds to leverage the technology without geographical constraints.

➀ Google is reportedly planning to establish a hyperscale data center near Ho Chi Minh City, Vietnam, marking its first such investment in the country. ➁ The project, if approved, aims to be operational by 2027 and could significantly impact Vietnam's digital infrastructure. ➂ The decision is driven by Vietnam's growing digital economy and the rapid expansion of YouTube's user base in the country, despite challenges like power supply shortages and unstable network infrastructure.

1. NVIDIA reported a record $30 million in revenue, up 15% sequentially and 122% year-over-year, driven by its Data Center division. 2. Despite exceeding expectations, NVIDIA's share price dipped pre-market due to a perceived slowdown in growth rate. 3. NVIDIA's Data Center business, especially in AI, continues to be the major revenue driver, with upcoming launches like Blackwell expected to boost future earnings.

➀ NVIDIA's dominance in AI GPU market extends to 2027 with Rubin Ultra AI GPU featuring 576GB of HBM4 memory; ➁ Competitors include AMD, Intel, Google, and others with their respective AI chip developments; ➂ TSMC's advanced CoWoS-L packaging technology and N3 process node support NVIDIA's next-gen AI GPUs.