LecturaInvestigaciónRadarMarco de inversión
Iniciar sesión / Registrarse
TodoTecnologíaProductosInversiónBolsaOtros

Lectura

2026-10-21154 publicaciones

Actualizar

Eka Robotics Raises $150M Series A at $1.45B Valuation

Tema · Eka机器人融资Material
2026-10-21 08:13 GMT+8

US-based robotics dexterity startup Eka Robotics announced the completion of a $150 million Series A funding round at a valuation of $1.45 billion. The round was led by Andreessen Horowitz (a16z) and F-Prime Capital, with participation from Obvious Ventures and E14 Fund. This brings the total capital invested in the company to approximately $161 million.

Eka's core offering is a "vision-force-action" artificial intelligence model for robotic manipulation. Unlike traditional robots that rely solely on camera images, Eka's custom grippers incorporate force sensors that provide real-time feedback on grip strength. This allows the robot to adjust its hold mid-task if an object begins to slip. The company claims its current robots can handle about 1,500 picks per hour of objects outside their training data, compared to a human baseline of roughly 600.

Despite impressive technical metrics, industry experts remain cautious about actual deployment capabilities. Kris Hauser, a professor at the University of Illinois Urbana-Champaign, noted that Eka's demo had a "wow factor" due to its ability to adapt to various objects without retraining, but he questioned whether the system could run reliably on an assembly line for a year without failure. Currently, Eka is conducting pilots with unnamed industrial customers for pick-and-place and wire-handling tasks.

This funding reflects continued investor enthusiasm for Embodied AI. Eka plans to use the new capital to expand its model to more tasks and accelerate customer deployments. As AI extends from pure software to physical interaction, robots capable of fine manipulation are becoming significant investment targets in the private market.

Fuentes:siliconangle.com

Investigación

Qualcomm Snapdragon Summit: Reconstructing Architecture from Chip Supplier to Agent Infrastructure

Tema · 高通智能体架构转型Material
2026-09-28 17:15 GMT+8

At the September 2026 Snapdragon Summit in Maui, USA, Qualcomm President and CEO Cristiano Amon presented a core judgment: the starting point of the smartphone experience is shifting from "apps" to "intent." This is not merely a change in interaction but implies that the underlying logic of chip design must be reconstructed—moving away from optimizing for single-device performance toward serving personal agents that flow across devices.

To support this shift, Qualcomm launched the sixth-generation Snapdragon 8 Elite and Elite Extreme dual flagship platforms. The key change lies not in simple frequency boosts but in the redivision of computational units: the CPU is positioned as the agent's "planner," breaking the 5GHz barrier for the first time; meanwhile, the Hexagon NPU expands its shared memory by 50% to reduce frequent accesses to external DDR, enabling efficient on-device execution of MoE models with up to 30 billion parameters. This architectural adjustment directly addresses the needs of agents for long contexts, concurrent tasks, and low-latency response.

Beyond mobile SoCs, Qualcomm extends this logic to PCs and wearables. PCs equipped with the Snapdragon X2 series are viewed as "productivity nodes for agents," capable of running complex reasoning locally for extended periods. The second-generation Snapdragon Audio Platform Elite attempts to transform earbuds from playback devices into "persistent perception entry points" for agents, using integrated Wi-Fi 6E and low-power AI compute to capture environmental context in real time. Amon emphasized that future competition is no longer about single-item compute power but about who can organize perception, reasoning, planning, and execution into a continuously running system with lower power consumption.

This strategic pivot marks Qualcomm's attempt to evolve from a semiconductor company providing processors to a "full-stack solution provider for agents" covering endpoints, edge, and data centers. By acquiring Modular to bolster software capabilities and previewing HBC (High-Bandwidth Compute) coprocessors for mobile platforms, Qualcomm is building a compute continuum across different device form factors. However, realizing this vision faces challenges: despite significant improvements in hardware efficiency, the reliability of agents in real-world scenarios, battery life under continuous operation, and whether users truly need such complex cross-device coordination remain to be validated by the market.

Fuentes:geekpark.net

Investigación

Musk: SpaceX-Tesla JV Terafab Plans Chip Mass Production in Q3 2028, Admits Timeline Is 'Extremely Challenging'

Tema · Terafab芯片项目Material
2026-10-21 08:08 GMT+8

Elon Musk stated during a SpaceX conference call on Tuesday local time that Terafab, a chip manufacturing joint venture between SpaceX and Tesla, plans to achieve mass production in the third quarter of 2028. The project aims to produce 1 terawatt of computing power annually for artificial intelligence and robotics businesses. Musk admitted this timeline is "extremely challenging," saying, "I've left a large margin of error because we are not sure if we can meet it."

The planned integrated facility in Texas will handle advanced logic and memory chip manufacturing, packaging, and testing within the same campus. As an intermediate milestone, Musk revealed an internal goal to produce the first test chip using static random access memory (SRAM) circuits at the Austin R&D base around May or June next year and get it working properly. Chipmakers typically manufacture SRAM circuits first to refine their production processes.

Reports note that building a frontier logic chip fab in the United States usually takes three years or more, making Musk's schedule "almost absurdly fast." Additionally, Terafab faces difficulties procuring sufficient precision semiconductor equipment from suppliers like ASML, Applied Materials, and KLA. When asked if SpaceX had placed orders for EUV lithography tools from Dutch firm ASML, Musk did not answer directly, only thanking suppliers for accelerating deliveries. Intel announced its participation in the project in April, though specific collaboration details remain undisclosed.

Fuentes:ithome.com

Investigación

Micron Announces Investment in New York Critical Minerals Processing With Indium Corporation

Tema · 美光美国供应链投资Material
2026-10-20 14:32 GMT+8

Micron Technology, Inc. (Nasdaq: MU) announced on October 20, 2026, that it will provide strategic funding to materials science company Indium Corporation to support the expansion of its refining and manufacturing operations in Rome, New York. The move aims to restart domestic processing of critical minerals essential for advanced semiconductor manufacturing.

Under the agreement, the two companies will enter into a long-term supply arrangement giving Micron access to high-purity minerals. This action is part of Micron's plan, announced in July, to invest up to $3 billion to strengthen the U.S. semiconductor supply-chain ecosystem and secure a reliable domestic supply of critical manufacturing materials.

Ben Tessone, Senior Vice President and Chief Procurement Officer at Micron, stated that advanced memory manufacturing depends on a reliable supply of critical minerals. He noted that partnering with Indium Corporation expands U.S. capacity for critical minerals production and builds a more resilient supply chain for American semiconductor manufacturing. The investment is expected to reduce reliance on concentrated global refining capacity, addressing a growing national security priority.

The announcement coincided with Micron's annual Supplier Awards event. New York Governor Kathy Hochul and U.S. Senator Charles E. Schumer welcomed the partnership, highlighting it as a step toward solidifying New York State's role as a key hub in the U.S. semiconductor supply chain.

Fuentes:investors.micron.com

Investigación

Former Qihoo 360 Security Chief Guhe Joins Z.ai; Company Claims GLM Model Found 23 Microsoft Vulnerabilities

Tema · 智谱GLM漏洞发现争议Resumen rápido
2026-10-20 19:30 GMT+8

Chinese artificial intelligence model developer Z.ai has announced the hiring of Yuki Chen, widely known in the cybersecurity community by his alias "Guhe." Chen previously headed the vulnerability research institute at Qihoo 360, a major Chinese cybersecurity firm, before working as an independent researcher. He is a prominent figure in the global white-hat hacking community, having been ranked among Microsoft’s "Most Valuable Security Researcher" five times and receiving two "Pwnie Awards."

In a statement, Z.ai said that Chen recently helped the company uncover 23 cyber vulnerabilities in Microsoft’s Windows operating system using its flagship GLM models. The company noted that Microsoft acknowledged these discoveries. However, the announcement did not specify the extent of human involvement versus the model's autonomous contribution in identifying these flaws.

This appointment comes amid growing concerns over AI-assisted cyberattacks and autonomous agents acting beyond their instructions. Z.ai is positioning its GLM series, particularly the latest GLM-5.3, as highly capable in vulnerability discovery. The company claims GLM-5.3 scored 84.5% on the CyberGym benchmark, surpassing competitors like Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol, though these results are self-reported by the vendor.

Fuentes:scmp.com

Investigación

OpenAI Claims Breakthrough in Legacy Language Migration with 97.9% COBOL Accuracy

Material
2026-10-21 07:45 GMT+8

OpenAI released a research report on October 20 announcing enhanced capabilities for its GPT-6 Astra and other models in building and migrating "low-resource programming languages" such as COBOL, Fortran, and Visual Basic. Although these languages have fewer public code samples and documentation, they widely support critical systems including banking, ATMs, weather forecasting, and nuclear safety analysis.

To measure this capability, OpenAI introduced two benchmarks: MultiLangBench, which requires models to build a SQLite-compatible database engine from scratch in 20 languages; and another benchmark requiring models to migrate COBOL programs to modern languages like C++, Java, Python, or Rust while preserving original business logic.

The results show that GPT-6 Astra achieved an average test pass rate of 87.4% across the 20 languages in MultiLangBench, a significant improvement over the previous generation GPT-5.6 Sol's 26.8%. Specifically, the COBOL pass rate rose from 10.9% to 72.1%, Fortran from 21.4% to 83.5%, and Visual Basic from 27.5% to 89.4%. The lower-cost GPT-6 Luna achieved an average pass rate of 40.3%, with estimated API costs reduced by approximately 78% compared to GPT-5.6 Sol.

In the more challenging COBOL migration benchmark, GPT-6.1 Sol reached an accuracy of 86.6% at low reasoning intensity, far exceeding GPT-5.6 Sol's 5.4%, with costs reduced by 39%. At higher reasoning intensity, accuracy reached 97.9%, approaching the evaluation ceiling.

OpenAI stated that this progress helps enterprises maintain legacy systems, fix defects, and advance software modernization. However, the company also emphasized that using AI migration results in production environments still requires expert review and testing against specific business rules and dependencies.

Fuentes:ithome.com

Investigación

Goldman Sachs: Global data center power demand to rise 170% by 2030, US grid delays cited as biggest AI bottleneck

Tema · AI数据中心电网瓶颈Material
2026-10-21 07:15 GMT+8

In a report published on October 19, Goldman Sachs warned that the pace of power infrastructure expansion is failing to keep up with the rapid growth of the artificial intelligence industry, identifying it as the primary bottleneck for AI development. The report projects that global data center power demand will surge by 170% between 2025 and 2030, with more than 60% of this increase occurring in the United States.

Currently, approximately 2,600 gigawatts (GW) of generation and energy storage projects in the US are waiting for grid interconnection, a capacity more than double the existing installed base. However, the median time from application to commercial operation is about five years. Additionally, the construction of new high-voltage transmission lines has slowed significantly, environmental permitting takes three to four years, and supply chains for large transformers face delays of three to five years.

Local community opposition also poses a significant hurdle. Since January, 86 related projects have been cancelled or severely delayed, with 76 of these directly linked to resident concerns over noise, traffic congestion, water usage, and potential electricity price increases. Goldman Sachs estimates that total investment in AI computing, data centers, and power infrastructure will reach approximately $7.6 trillion between 2026 and 2031. Persistent power constraints could extend the payback period for these massive investments.

Fuentes:etnews.com

Investigación

Microsoft CEO Nadella: Using Excel Agents to Question Formulas and Simulate Scenarios for ROIC Assessment

Tema · 英伟达微软本地AI硬件Resumen rápido
2026-09-28 13:29 GMT+8

In the latest episode of the Sources Podcast, Microsoft CEO Satya Nadella detailed how he uses Cowork and Excel agents in his daily work. Unlike his previous emphasis on macro visions, this time he focused on specific micro-operations: he does not just accept the results given by the model, but "operates it, questions it, and keeps reasoning around the results."

Nadella gave an example: when the data center team sends a complex spreadsheet, he opens it, locates a cell, asks the Excel agent to examine the formula used in that cell, and requests it to generate five worksheets covering different scenarios. He calls this ability to complete analysis through reasoning "causal inference," identifying it as one of the important advances in Microsoft's existing AI systems.

Beyond internal table analysis, Nadella mentioned a second advance: combining the corporate environment with the global environment. He deployed a "data runner" in Microsoft Fabric that automatically retrieves cloud service provider data published by the U.S. Securities and Exchange Commission (SEC), imports it into semantic models, and has programming agents organize the results into dashboards. This allows him to view return on invested capital (ROIC) at various levels in real time, continuously aggregating external public filings with his own internal analyses.

This first-person account offers a unique high-level perspective on enterprise AI adoption. It suggests that for large tech companies, the value of AI lies not only in automating content generation but in serving as an interactive analytical partner, helping decision-makers perform hypothesis testing and causal inference in complex data environments. However, this remains Nadella's personal usage experience; actual effectiveness still depends on product stability and accuracy.

Fuentes:ithome.com

Investigación

TrendForce: AI Demand Tightens Supply, Forecasts ~20% QoQ Rise in Q4 2026 DRAM Contract Prices

Tema · AI内存涨价冲击Material
2026-10-21 07:17 GMT+8

Market research firm TrendForce, citing multiple reports, forecasts that DRAM contract prices will rise by approximately 20% quarter-over-quarter in the fourth quarter of 2026. This rate of increase represents a notable slowdown compared to the second quarter (~80%) and third quarter (~40%), but the DRAM market remains at high price levels with persistent supply tightness and strong demand.

The core driver behind the price increases is identified as the continued expansion of artificial intelligence inference demand. Larger models, longer context windows, and growing KV cache requirements are driving up capacity and bandwidth needs for server DRAM and enterprise SSDs. Simultaneously, major memory manufacturers are prioritizing production capacity for high-end products like High Bandwidth Memory (HBM), further tightening the supply of traditional DRAM.

Realistic reasons for the slowing growth include a higher base from previous price hikes, weakened purchasing power in end markets, and the gradual implementation of price caps in some long-term procurement agreements. Regarding HBM, due to tight supply, media reports suggest that NVIDIA is considering reducing the HBM stack layers for its next-generation GPU, Rubin Ultra, from 12 layers to 8. If other conditions remain unchanged, this would reduce HBM capacity per GPU by about 33%, but theoretically allow the same DRAM chip supply to support 50% more GPU shipments. However, because the cost proportion of base dies is higher in 8-layer HBM, the cost per Gb is expected to be 10% to 20% higher than 12-layer HBM in 2027.

Fuentes:ithome.com

Investigación

Intel and AMD Announce Collaboration on oneDNN Optimizations for PyTorch

Tema · TorchTPU开源项目Material
2026-10-21 00:50 GMT+8

US semiconductor companies Intel and Advanced Micro Devices (AMD) jointly announced on October 20, 2026, that they will collaborate on the deep learning neural network library oneDNN to simplify AI development on x86 architecture CPUs. According to the announcement, AMD will merge performance optimizations from its proprietary downstream fork, ZenDNN, back into the open-source oneDNN project, establishing it as the default acceleration library for mainstream AI frameworks like PyTorch on the x86 platform.

Previously, Intel led the development of oneDNN to support its hardware, while AMD maintained ZenDNN, an API-compatible but distinct branch optimized for its Zen processor designs. This fragmentation required developers to perform additional tuning for different vendors. The new collaboration signifies a move toward convergence in the software stack between the two competitors. Additionally, AMD will join the UXL Foundation, backed by Intel, Samsung, Qualcomm, and others, as a steering member to promote open standards. This initiative is expected to reduce engineering complexity for deploying AI applications and improve cross-platform compatibility.

Fuentes:phoronix.com

Investigación

Anyscale and Google Cloud Announce Native Ray Support for PyTorch on TPUs

Tema · TorchTPU开源项目Material
Verificado 2026-10-21 07:26 GMT+8

On October 20, 2026, compute platform company Anyscale and Google Cloud jointly announced that the open-source distributed computing framework Ray now natively supports running PyTorch workloads on Google Tensor Processing Units (TPUs). This integration builds on Google's TorchTPU project introduced last year and KubeRay on Google Kubernetes Engine (GKE), aiming to provide a unified control plane for platform engineering and machine learning teams to schedule and manage distributed training, vLLM inference, and reinforcement learning tasks across GPUs and TPUs.

Historically, using TPUs with Ray required adapting to a Single-Program, Multiple Data (SPMD) execution model designed primarily for JAX, where a single Ray worker process per host VM managed all attached TPU chips. This update introduces a native PyTorch experience with one worker per TPU chip: each Ray worker requests a single logical TPU device, similar to setting num_gpus=1 on GPUs. Ray automatically isolates and assigns individual TPU devices to workers via TPU_VISIBLE_CHIPS. Developers can migrate existing PyTorch code by simply changing the device string from "cuda" to "tpu" and switching the communication backend from "nccl" to "tpu_dist".

For large-scale training scenarios, Ray Train's TorchTrainer now supports automatic configuration of TPU slices. By specifying use_tpu=True along with the target topology (e.g., "4x4") and accelerator type (e.g., "TPU-V6E"), the system automatically reserves atomic TPU slices, ensuring all training workers are gang-scheduled onto contiguous physical TPU slices over high-speed interconnects, preventing resource fragmentation. Additionally, combined with the vLLM-TorchTPU plugin, developers can deploy OpenAI-compatible online inference services on TPU slices using the same declarative API.

Fuentes:anyscale.com

Investigación

ICBC President Liu Jun: AI Accountability Should Follow Benefit and Decision

Tema · 工行AI受益证明机制Material
2026-10-21 07:00 GMT+8

Liu Jun, president of the Industrial and Commercial Bank of China—the world's largest lender by assets—laid out a specific judgment about AI deployment responsibility at a panel discussion during the Shanghai Bund Summit: commercial banks and central banks sit in asymmetric positions when facing AI risk. Central bankers focus on oversight and policy, while commercial lenders that deploy AI at scale bear direct financial and legal liability. Liu's own words: 'If ICBC deploys AI tools towards different business lines, and all the benefits contribute to the IT balance sheet and income statements, it means I take the responsibility of the AI deployment. If something bad happens, I'm going to have to pay the victim and give the final answer to the market.'

At the same panel, as banks begin adopting agentic AI—systems that can plan and carry out multi-step tasks—Arthur Yuen Kwok-hang, deputy chief executive of the Hong Kong Monetary Authority, echoed Liu's concern about accountability and offered a concrete observation: institutions typically split one application across six to eight agents, but if each agent and the human supervisor are responsible for only part of the task, 'nobody is responsible for the whole thing.' Liu also noted that ICBC has deployed hundreds of AI models and given them more autonomy in data-heavy work, while keeping human teams in client-facing positions.

To address this fragmentation, Liu described a framework ICBC introduced called 'proof of benefit'—an approach that directly ties accountability to those who reap the financial returns and make independent decisions within the system. The logic is that whoever benefits bears responsibility, rather than diffusing liability vaguely across developers, deployers, and users. Yuen revealed that Hong Kong's de facto central bank had unveiled an updated AI governance framework the previous week, focused on accountability, resilience, and trust, and would publish a discussion paper on supervising agentic AI, exploring accountability tied to intent, a registry of material AI agents, and ways to intervene. Yuen said the Hong Kong government aims to foster an environment in which financial institutions can safely 'fail fast.' 'We, as supervisors, become more facilitative rather than regulatory,' he said.

It is worth noting that the specific operational details of the 'proof of benefit' framework—how 'benefit' is defined, how 'independent decision' is determined, and how decision chains are traced when agents execute autonomously—were not elaborated in the reporting. The HKMA's governance framework has been published, but the discussion paper on agentic AI supervision remains forthcoming and has not yet formed binding rules. Both Liu and Yuen identified the difficulty of assigning responsibility, but the actual effectiveness of their proposed solutions remains to be verified.

Fuentes:scmp.com

Investigación
Siguiente →
Iniciar sesión / Registrarse
LecturaInvestigaciónRadarMarco de inversión