ABMedia, citing Commercial Times, reported that Intel’s next-generation PC processor roadmap could deepen its reliance on Taiwan Semiconductor Manufacturing Co. across the full 2 nm family. The report said Nova Lake desktop chips are expected to use a hybrid process approach combining Intel 18A with TSMC N2P, while Razor Lake may move further to TSMC N2X, a higher-performance extension of the N2 platform aimed at high-frequency CPU, AI and HPC workloads. Supply-chain sources also said Nova Lake desktop parts may be the first to introduce bLLC, or Big Last-Level Cache, with notebook adoption possibly deferred to the later Razor Lake-HX generation. The article also highlighted a strategic reading from Citrini analyst Jukan, who argued that Intel’s TSMC allocation should not be viewed only as a response to concerns around Intel 18A yields. In his view, maintaining or expanding wafer capacity at TSMC could also limit the room available to AMD for server CPU production, putting pressure on AMD’s shipment timing, scale and expansion pace. On the supplier side, the report named Taiwanese companies such as Phoenix Silicon International and Chung-Hsin Electric & Machinery-linked China Sand, saying tighter process requirements at 2 nm and 1.6 nm are lifting demand for CMP-related materials and tools.

Ajinomoto, best known as a Japanese food company, said its electronic materials business delivered outsized profit in the latest quarter, driven by ABF, or Ajinomoto Build-up Film, a key insulating layer used in IC substrates for advanced processors. The company said the segment generated JPY 33.1 billion in revenue and JPY 19.1 billion in business profit, implying a 57.7% margin. While the unit accounted for about 8% of group revenue, it contributed 32% of profit. According to the company, ABF holds more than 95% of the global market for insulating films used in high-performance semiconductors. The material is used in substrates for high-end CPUs and GPUs, and demand has benefited from larger AI accelerator packages, more substrate layers, tighter line widths, and stricter thermal expansion requirements. Ajinomoto said growth came from an improved product mix rather than price increases. The report also outlined competitive and technological pressure, including efforts by Sekisui Chemical, U.S. startup Thintronics, and longer-term interest in glass substrates. In response, Ajinomoto is expanding capacity in Japan, with investment already underway in Gunma and Kawasaki and more spending planned through 2030.

China has begun allowing limited shipments of Nvidia H200 AI chips into the country, with approvals handled through a case-by-case process led by the National Development and Reform Commission, according to a Financial Times report cited by ABMedia. Two people familiar with the matter said ByteDance and Tencent each received about 10,000 H200 processors in recent weeks, while several other Chinese technology groups are expected to secure import approvals of similar size. The easing is narrow rather than broad. Lenovo and other Nvidia partners have told Chinese customers they can again place orders for AI servers containing H200 chips, but companies still need individual approval. The arrangement also runs into a practical constraint: Chinese regulators have told companies they can deploy the chips in Hong Kong, yet the city’s data center capacity is tight and power supply limits make near-term expansion difficult. The report says the move reflects Beijing’s attempt to keep leading domestic AI developers training frontier models while still backing local chipmakers such as Huawei. Although Chinese firms are shifting faster toward domestic chips for inference, they continue to rely heavily on Nvidia hardware for the more demanding training workloads behind advanced AI models.

Tapeout.net is pitching an unusual idea: building an on-chain version of an integrated circuit by mapping core chip components to token standards. In the project’s design, ERC-1155 tokens represent transistors, while ERC-721 tokens represent the circuits that connect them. The article argues that this mirrors how real-world chips are built from basic units, then combined into logic gates, arithmetic units, and eventually complex processors such as GPUs and CPUs. It also points to a blockchain-native equivalent of a chip clock. In conventional hardware, transistors and circuits alone do not run unless a clock frequency drives the system. The author says block production on a blockchain can serve that role, with each new block acting like a recurring pulse for an on-chain chip. Tapeout was first launched on BNB Chain and is already live, with enthusiasts experimenting by minting “transistors” and designing their own “chips.” While the piece describes the project as primitive in terms of clock speed and current functionality, it presents the idea as a new line of thinking for crypto: whether semiconductor logic, and eventually parts of AI hardware, software, and compute, could be explored through decentralized architectures.

Nvidia said in a technical blog post on Aug. 19 that cuML and cuVS now include multi-GPU UMAP support, allowing large-scale vector-data dimensionality reduction to run across multiple GPUs while preserving embedding quality and cutting runtime. The company said that on a DGX system with eight H100 GPUs, testing on the MIRACL dataset, which contains 106 million vectors and about 870GB of data, the end-to-end process finished in 8 minutes. Nvidia said the result represents up to a 74x speedup versus an estimated CPU implementation. It also said a prior CPU approach could not complete the full workload even with 2TB of memory. The method splits data into multiple clusters, builds local k-nearest-neighbor graphs in parallel across GPUs, and then merges them into a global graph to bypass single-GPU memory limits.

PANews published a detailed beginner’s guide to Tapeout by BruceBlue, framing the product less as a trading screen and more as a hands-on on-chain logic environment. The piece argues that new users should not start by chasing prices or asking which “chip” is worth minting. Instead, they should first understand what a processor does, how NAND and LATCH components are used, and how those parts connect to circuit creation and on-chain calls. The guide walks through a minimum-risk path: use a separate experimental wallet on BNB Smart Chain, review processor-level data before connecting a wallet, mint only very small amounts when supply remains, and treat the marketplace as an order-book venue rather than a one-click buy interface. It also stresses that the protocol is currently upgradeable and only plans to seal later, so users should verify every signature request and not rely on claims of immutability. From there, the article moves to the Tapeout canvas, where users can test simple examples in the browser before tapeout consumes components on-chain. It closes by warning readers not to misread the site’s “on-chain BTC miner” page as a mining yield product and offers a final checklist covering website verification, network selection, contract review, spending limits, and whether the user can absorb a total loss.

Liquid cooling is moving from an optional feature to core infrastructure for high-end AI servers as chip power consumption climbs, according to TrendForce. The research firm said liquid-cooling penetration for AI chips is projected to rise from about 33% in 2025 to 53% in 2026, then approach 60% in 2027. It linked that shift to rapidly evolving processors from NVIDIA, AMD and Google, with single-chip thermal design power already exceeding 1 kW and full rack-scale systems reaching several hundred kilowatts. TrendForce said NVIDIA remains the clearest force behind broader adoption. Its Vera Rubin platform has adopted a fanless, full-liquid-cooling design that extends beyond GPUs and CPUs to include CX9 NICs, busbars, power boards and optical transceiver modules. Even with a possible delay to the Kyber NVL144 platform, the firm expects overall GPU rack shipments to grow more than 30% year over year, while NVIDIA GB/VR rack shipments could double in 2026. AMD is also expanding from standalone GPU products to full AI platform systems, with its Helios rack-scale solution set to become a focus from the second half of 2026 and larger-volume shipments expected in 2027. TrendForce added that more than 80% of Google’s AI servers already use liquid cooling. On the supply side, Taiwanese companies including AVC, Auras and Jentech have secured positions in cold-plate and heat-spreader segments.

Data from GMGN showed that BEHEMOTH, a meme coin on BSC, briefly pushed its market capitalization above $2.7 million before falling back to $1.04 million, with a 24-hour gain of 8,251%. The token’s surge came after a developer said they had completed what was described as the first real on-chain 4-bit processor on BNB Chain. The processor, also named Behemoth, consists of 2,251 logic gates including NAND gates and flip-flops, and was said to have been cycle-by-cycle verified against performance specs comparable to Intel’s 4004 from 1971. The project has already run a Fibonacci sequence calculation, producing 1, 1, 2, 3, 5, 8, and 13. Because its execution is synchronized with BNB Chain block production, the processor runs one cycle every 0.45 seconds, or about 2.22 Hz. The developer said the full protocol will be made public so others can build their own on-chain processors and let them call one another to form more complex computing systems. Binance co-founder CZ later responded to questions about the project by asking, 「What is the use case?」, a remark that triggered community discussion.