bottleneck

According to the Electronic Times, TSMC plans to further expand the outsourcing scale of the CoW (Chip-on-Wafer) segment in its CoWoS advanced packaging, entrusting more packaging orders to ASE and other OSAT manufacturers. CoWoS is TSMC's 2.5D packaging technology that connects AI processors such as GPUs with HBM through an intermediary layer, where CoW is the step of combining the chip with the intermediary layer, and WoS is the step of packaging the substrate. Previously, TSMC mainly outsourced the WoS segment, with only a small portion of CoW being entrusted.This move aims to alleviate the packaging capacity bottleneck caused by the continuous surge in demand for AI chips. TSMC holds about 90% of the global AI chip manufacturing market, but as the demand for self-developed chips from manufacturers like Nvidia and AI data centers skyrockets, capacity has become difficult to meet. Industry insiders expect OSATs to significantly increase equipment investment, especially in post-process areas such as wafer cutting and bonding. Several laser processing and bonding equipment suppliers in South Korea have been confirmed to be negotiating with OSATs regarding the supply of CoW equipment.

According to official news, Marvell Technology announced the launch of a new generation of memory solution portfolio for AI infrastructure, covering server-level AI storage, rack-level CXL memory expansion and pooling, as well as multi-rack optical interconnect shared memory, aimed at addressing the growing memory capacity and bandwidth bottlenecks in Agentic AI inference processes.Marvell stated that as AI model sizes increase, context windows extend, and KV Cache demand grows, traditional tightly coupled architectures of computing and memory are limiting AI inference efficiency. Through memory disaggregation, memory resources can be made more independent of computing resource expansion, improving GPU utilization and reducing data movement latency. The products released include:Bravera SC6 PCIe 6.0 SSD controller: Designed for AI inference storage scenarios, it helps cloud service providers migrate more KV Cache to high-performance SSDs, enhancing infrastructure efficiency. This product features an architecture compatible with multi-vendor NAND and is expected to begin sampling in the fourth quarter of 2026.Marvell Structera X memory expansion solution: Based on CXL technology, it supports rack-level memory expansion and resource pooling, helping data centers share and allocate memory resources more flexibly, reducing AI infrastructure costs.Marvell Photonic Fabric optical interconnect memory solution: Constructs a shared memory architecture across multiple racks using optical interconnect technology, supporting up to 32TB warm KV Cache offloading and helping AI inference clusters enhance throughput capacity.Marvell stated that the Photonic Fabric solution can achieve a 2 to 3 times increase in Token throughput under existing data center space and power consumption constraints, supporting larger scale models and longer context AI applications.Marvell executive Will Chu stated that AI infrastructure is transitioning from a single server architecture to a system where computing, memory, and connectivity operate in synergy, and in the future, memory needs to expand more independently to enhance resource utilization and Token efficiency.As the demand for AI Agents and large model inference continues to grow, memory capacity, bandwidth, and data transfer efficiency are becoming new focal points of competition in AI infrastructure, following computing power.

According to CryptoSlate, Solana increased the block computation unit limit from 60 million to 100 million on July 29, a 66.7% increase. This adjustment provides more parallel processing space for unrelated transactions, reducing the block occupancy rate of a single popular account from the original 20% to 12%.However, the computation limit for writing to the same account in each block remains at 12 million CU, and the incremental data limit for accounts is still 100 MB, which means the congestion bottleneck for popular accounts has not been fundamentally resolved. If a large number of transactions need to be written to the same popular account (such as a high-volume DEX or liquidation contract), they are still subject to this limit. The officials did not provide actual comparative data on transaction confirmation rates, fees, block propagation, or replay before and after the upgrade.

Tom Lee stated that as AI "bottleneck" related stocks adjust, AI downstream assets are strengthening, including SMH, DRAM, MAGS, IGV, as well as AI downstream related ETHA and Ethereum ETH.He pointed out that over the past month, Ethereum has outperformed DRAM by about 5,500 basis points, or 55%.

On-chain detective ZachXBT stated that during the migration from the traditional private messaging system to XChat, the X platform encountered multiple technical bugs. At the same time, the reliability of the platform's search function declined, affecting information retrieval efficiency. More importantly, the X platform's algorithm has a "traffic bottleneck" for cryptocurrency content, making it easier for content to be restricted within the cryptocurrency vertical, thereby reducing the ability to reach a broader audience.

According to Forbes, although the trading volume of stablecoins has exceeded $1 trillion in the past year, most activities are still concentrated in the crypto-native space (trading, arbitrage, and inter-protocol settlement), with limited applications in everyday commercial payments. WasabiCard CEO Ray Yang pointed out that the transfer of funds is no longer the core issue; licensing, compliance, risk management, and banking capabilities are the key foundations for achieving widespread adoption.Forbes noted that while stablecoin settlement can significantly enhance cross-border payment efficiency, each market has different compliance standards, licensing requirements, and banking relationships, making the construction of localized compliance in each market both slow and expensive, which contradicts the instant global settlement that stablecoins advocate. Currently, the stablecoin market has surpassed $320 billion, and industry discussions are shifting from whether stablecoins can replace existing networks to how they can be integrated into existing networks.Forbes believes that the challenge of the last decade was to get funds flowing, while the challenge of this decade is to ensure that global payments operate in compliance and at scale within a fragmented regulatory environment.

The "New Stock God" Serenity published an analysis stating that as Co-Packaged Optics (CPO) technology is expected to enter large-scale deployment in the second half of 2027, Sivers Semiconductors (SIVE) may play both a bottleneck and a key node role in the industry.It pointed out that there are signs of tight supply for Continuous Wave (CW) lasers. Affected by previous orders from NVIDIA, the capacities of companies such as Sumitomo Electric, Furukawa Electric, and Win Semi are highly saturated. Meanwhile, Sivers, which adopts a fab-lite model, has effectively secured a significant amount of CW laser supply by locking in capacity with foundries like Win Semi in advance.The analysis believes that multiple CPO routes, including ASIC projects from Ayar Labs, Jabil, Marvell Celestial, and other large-scale cloud providers, are highly dependent on Sivers' laser solutions, lacking mature alternative sources in the short term, which positions it as a structural "bottleneck" in the entire ecosystem.Additionally, Sivers is the default reference laser design solution for GlobalFoundries, with relevant ecosystem participants including AMD and several CPO chip suppliers. Aside from vertically integrated companies like NVIDIA and Broadcom, most ASIC and commercial CPO projects are likely to revolve around Sivers.Serenity expects that as the CPO market size grows from nearly zero to between $81 billion and $91 billion in about a year and a half, Sivers is likely to replicate Lumentum's growth path and may grow into a company with a market value of around $75 billion in the coming years. However, the above views only represent the personal judgment of market analysts.

At the "Crypto 2026: From Cryptocurrency to Smart Economy" themed forum held in Hong Kong, guests including KiteAI's Asia Pacific head Laughing, Sentient's Asia Pacific head Anita, Brevis content director Nic Tang, Mentis product head Jerry, and Predict.fun's Asia Pacific BD head Ah Huang Ricardo engaged in a roundtable discussion on the theme "AI × Crypto --- The Foundation of the Next Generation Digital Economy."Regarding whether the outbreak of open-source AI Agents (such as "raising lobsters") is a short-term sentiment or a long-term trend, Laughing believes this marks a paradigm shift of AI from "conversational" to "executive," signaling the beginning of a long-term trend. However, Agents lack a physical identity, making it difficult for merchants to trust their trading behavior, and issues of data leakage and uncontrollability remain obstacles that must be overcome for scaling. Jerry pointed out, based on his own usage experience, that current Agents face security risks such as memory fuzziness, token consumption in dead loops, and accidental file deletion. Although newly emerged Agents have improvements, they still have shortcomings. Ah Huang Ricardo stated that behind the short-term sentiment is real user demand, but truly reproducible profitable trading Agents are still very rare, and there is a long way to go.Addressing core bottlenecks such as trust, security, and verifiability, Nic Tang emphasized that the current execution process of Agents is like a black box, where users cannot verify whether their decisions are executed as promised. Zero-knowledge proof technology can mathematically prove the source of an Agent's output and compliance with behavior, while not exposing privacy. Anita added that Sentient is committed to building an open and collaborative AI agent network, promoting the deep integration of AI and blockchain in identity, data, and incentive layers, providing foundational support for the next generation digital economy.Regarding the application scenarios that are likely to scale first, the guests generally believe that the programming field is relatively mature, and Agents in prediction markets perform better than humans in specific tracks due to their rationality and lack of emotions. In the short term, they are optimistic about API Marketplace and cross-platform consumer shopping Agents, while in the medium term, they are optimistic about content creation AI (such as AI influencers). Agent-to-Agent interaction and the Agent/Skill Marketplace are seen as important future trends.

At the "Crypto 2026: From Cryptocurrency to Smart Economy" themed forum held in Hong Kong, Tencent Cloud's industry architect Alan Nie delivered a keynote speech titled "Cloud + AI Dual Engine: Tencent Cloud Empowers New Growth in Web3."Alan Nie pointed out that Web3 enterprises face three major bottlenecks in global infrastructure, R&D efficiency, and business intelligence. Tencent Cloud deeply integrates the "Cloud + AI" dual engine to provide low-latency infrastructure covering the globe. Among them, the Singapore data center is the only cloud provider in the world that offers four availability zones, and the self-developed TDSQL-C database can achieve elastic scaling in seconds, with Redis single-node performance reaching over 300,000.In terms of AI empowerment, Tencent Cloud launched CodeBuddy (AI Pair Programmer) and WorkBuddy (Personal AI Assistant), which can automatically generate code, batch process office documents, organize meeting minutes, and support multi-agent parallel execution of complex tasks. In financial scenarios, the automation research report reproduction cycle has been shortened from three days to half a day; in on-chain data analysis scenarios, storage costs have been reduced to one-tenth of the original, with query responses reaching millisecond levels.Alan Nie stated that Tencent Cloud is committed to using the "Cloud + AI" dual engine to help Web3 enterprises build the next generation of smart economic infrastructure.

According to a post by Delphi Digital, stablecoins are becoming an alternative to traditional foreign exchange trading channels. The costs of traditional foreign exchange channels are high, with comprehensive fees for remittances to Argentina or Nigeria potentially reaching up to 8%, and 81% of these costs come from maintaining the underlying banking infrastructure.However, current exit mechanisms such as access to bank accounts or interbank channels remain a significant bottleneck. Delphi Digital states that most of the "friction" exists outside of the blockchain. While the minting and burning of stablecoins can be completed in just a few seconds, significant delays occur due to the batch processing mechanisms of bank wire transfers into these systems.

The a16z Crypto article "How AI 'Judges' Expand the Scale of Prediction Markets" points out that the biggest challenge facing prediction markets is not "pricing the future," but rather determining what actually happened. Similar issues frequently arise in some small-scale events, where erroneous or opaque settlement mechanisms can undermine market trust, liquidity, and the accuracy of price signals.Industry experts suggest introducing large language models (LLMs) as arbiters in prediction markets, including on-chain rule commitments, resistance to manipulation, enhanced transparency, and increased neutrality. For example, specific LLM models, timestamps, and judgment prompts could be encrypted and recorded on the blockchain at the time of contract creation, allowing traders to understand the complete decision-making mechanism in advance. The model weights would be fixed and not easily tampered with to reduce the risk of cheating, and the settlement mechanism would be publicly auditable with no arbitrary human rulings. The AI judgment mechanism could significantly improve the efficiency and scalability of prediction market settlements while ensuring transparency and fairness.

According to CoinDesk, due to domestic regulatory restrictions, South Koreans transferred over 160 trillion won (approximately 110 billion USD) to foreign cryptocurrency exchanges last year. Analysis indicates that the delay in the implementation of South Korea's "Digital Asset Basic Law" has created a regulatory vacuum, forcing investors to turn to overseas platforms.Research has found that cryptocurrency has become a major investment asset in South Korea, with the number of investors rising to 10 million, and revenue from trading platforms like Upbit and Bithumb reaching trillions of won. However, the report shows that despite South Korean investors actively trading cryptocurrencies and increasingly turning to overseas platforms like Binance and Bybit, the growth of domestic trading platforms has stagnated.However, the regulatory vacuum has raised concerns among market participants, who worry that domestic centralized cryptocurrency trading platforms are increasingly struggling to compete with overseas platforms that offer more complex trading products.

Facing global challenges such as short-term chip shortages, long-term energy constraints, and ongoing storage space bottlenecks, BitTorrent, as the cornerstone of a decentralized internet, is paving new paths for the future of AI.It is reported that BitTorrent, through its distributed architecture, transforms a vast number of edge nodes into a pool of computing and storage resources, empowering developers to store, share, and access data without intermediaries. BTFS is not only a pillar of the decentralized data economy but also turns every device into a participant in the evolution of AI, collaborating with global resources efficiently and sustainably towards a decentralized intelligent future.

ChainCatcher news, OpenAI founder Sam Altman stated that the release of OpenAI's new version may be delayed, currently facing a computing power bottleneck, and services may sometimes slow down.Previously, Sam Altman mentioned on February 12 that he plans to launch GPT-4.5 and GPT-5 in the coming weeks or months. He also stated yesterday that a "new type of open-source weight language model" with reasoning capabilities will be released in the coming months. An open-source weight language model refers to a model that can be publicly used, downloaded, modified, or deployed. Although it is not as open as a fully open-source model, it represents a significant shift compared to the completely closed GPT-3 and GPT-4. OpenAI partially open-sourced the GPT-2 model in February 2019 and completed full open-sourcing in November of the same year.