LesenForschungRadarAnlageframework
Anmelden / Registrieren
AlleTechnikProdukteInvestierenBörseSonstiges

Lesen

2026-10-21154 Beiträge

Aktualisieren

Anthropic and Nscale in Talks for Second $45 Billion Compute Deal, Sources Say

Wesentlich
2026-10-21 10:25 GMT+8

According to TechNews, citing people familiar with the matter, AI cloud computing company Nscale is in talks with Anthropic for a new large-scale computing resource partnership, potentially worth up to $45 billion. If finalized, the deal would correspond to the second facility at Nscale's Monarch data center campus in West Virginia, USA.

The planned new building is expected to provide 460 megawatts (MW) of computing power and will utilize NVIDIA's latest Vera Rubin AI chips. Previously, Anthropic reached another approximately $45 billion cloud computing rental agreement with Nscale in August of this year. These negotiations target additional expansion within the same campus, indicating deepening cooperation between the two parties.

Sources noted that Anthropic also holds an option to purchase the third planned building at the Monarch campus. Discussions are ongoing, and deal amounts and terms may change; neither party has commented publicly. Nscale mentioned in its planned US IPO filings that Anthropic-related infrastructure contracts could reach up to approximately $44.6 billion, subject to delivery and service conditions.

Quellen:technews.tw

Forschung

Google and Meta Open-Source TorchTPU to Enable Native PyTorch on TPUs, Challenging Nvidia's Dominance

Thema · TorchTPU开源项目Wesentlich
2026-10-21 10:38 GMT+8

The PyTorch teams at Google and Meta announced TorchTPU, an open-source AI development stack, on Tuesday. The core goal of this tool is to make it easier for PyTorch to execute on Google TPUs, thereby attracting the application development community and challenging Nvidia's long-standing dominance in the AI compute market.

TorchTPU was the most significant product among several announced by Google at the PyTorch 2026 conference. It enables PyTorch to run natively on Google Cloud TPUs, allowing developers to use their familiar tools—such as vLLM, SGLang, llm-d, TorchTitan, Ray, TorchRec, HuggingFace, and LiteRT—to train, deploy, and serve models.

According to the TorchTPU webpage documentation, the tool was co-developed by Google and PyTorch. It allows existing PyTorch workloads to run on Google's own TPUs with minimal code changes, using familiar APIs. TorchTPU provides a TPU PyTorch Aten core implementation, compilation caching for XLA execution, support for distributed training (including data parallelism and tensor parallelism), examples for various models (such as Llama, Qwen, ResNet, DLRM, minGPT, and UNet), as well as debugging and benchmarking tools.

Google stated that developers can first develop and debug in eager mode, then optimize TPU utilization and reduce memory usage via torch.compile to improve performance in production environments. Additionally, Google has partnered with Hugging Face to allow open-source models on its platform to run natively on TorchTPU. In the Transformers v5 library, deploying models to Google Cloud TPUs now only requires setting device_map="tpu".

By enabling native PyTorch support for Cloud TPUs through TorchTPU, Google aims to break Nvidia's massive CUDA software ecosystem. Notably, Amazon Web Services (AWS) has also adopted a similar strategy, recently announcing TorchNeuron to lower the barrier for switching AI models to Trainium chips. This indicates that cloud providers are collectively leveraging software compatibility to weaken Nvidia's moat.

Quellen:ithome.com.tw

Forschung

Meta Details Prometheus Cluster Architecture: Over 11% GPUs Held in Reserve, Training Interruptions Cut by 99%

Wesentlich
2026-10-21 10:40 GMT+8

US tech giant Meta disclosed operational details of its "Prometheus" supercomputing cluster located in New Albany, Ohio, on October 20, 2026. The cluster, designed to support large-scale artificial intelligence model training, is expanding toward a scale of 1 million GPUs. Meta stated that to maintain training continuity, the cluster incurs significant redundancy costs, with more than 11% of GPUs currently idle, either held as hot spares or awaiting replacement due to failure.

In terms of hardware configuration, Prometheus utilizes Nvidia GB200/GB300 systems, where each NVLink domain contains 72 GPUs, but only 64 are actively used for training, while the remaining 8 serve as immediate hot spares. Meta explained that a single GPU failure could force an entire rack out of the cluster; the hot spare mechanism prevents such cascading effects, thereby reducing downtime. Additionally, the cluster employs distributed checkpointing, which asynchronously saves model states to multi-tier storage when larger failures occur, preventing the need to restart training from scratch.

Meta acknowledged that signal integrity and power management posed severe challenges as the cluster scaled. With over 600,000 NVLink connections within the system, even minor signal degradation could trigger recovery processes. To address this, Meta introduced continuous telemetry and L11 diagnostic tests, and used software called Dynamo Dimmer to gradually adjust GPU power consumption to handle power spikes. After a year of engineering optimization, training interruptions in Prometheus decreased by 99% compared to mid-2025, dropping from nearly 1,000 per day per 8,000 GPUs to about 11. Meta aims to further reduce the interruption rate to approximately two per day per 8,000 GPUs.

Quellen:technews.tw

Forschung

Askwho Casts AI reads Scott Alexander's 'Against Against Handover': Distinguishing Successionism from Handoverism

Wesentlich
2026-10-21 04:59 GMT+8

The Askwho Casts AI podcast released an audio reading of Scott Alexander's article 'Against Against Handover' from the Astral Codex Ten blog on October 20, 2026. This episode is not an original interview but converts a deep philosophical essay into an accessible audio format, suitable for readers interested in systematic explorations of AI governance controversies.

In the original text, Alexander first clarifies a common conceptual confusion: he distinguishes 'Successionism' (the belief that AIs should succeed humans as rulers of Earth, potentially overriding human interests) from 'Handoverism' (the idea that humans might be best served by putting aligned AIs in positions of control). He argues that many critics fail to make this distinction, incorrectly lumping AI safety researchers who find the latter plausible with anti-humanists.

To concretize this argument, Alexander constructs five hypothetical governance visions, ranging from weakest to strongest. The first is 'Kemalist AI', where military AIs are aligned to democracy itself rather than any individual, preventing elite coups. The second is 'President Claude', imagining a perfectly aligned, incorruptible AI running for US President and outperforming human candidates on measurable metrics. The third is 'Parliament of Angels', where each citizen has an AI representative acting in their best interest within a direct democracy. The fourth is 'The Night Watchman', describing an archipelago of self-governing communities protected by a superintelligence that enforces basic order and exit rights. The fifth is 'God', a benevolent digital superintelligence arranging the universe directly.

Alexander references the Book of Samuel, where Israelites demanded a human king, to analogize humanity's attachment to being ruled by its own kind. He suggests that rejecting AI public servants while accepting robot garbage collectors stems from irrational nationalist or identity-based preferences rather than practical utility. He acknowledges the speculative, science-fiction nature of the discussion but emphasizes it as a necessary thought experiment for long-term AI governance.

Quellen:askwhocastsai.substack.com

Forschung

Zhang Xiaojun Interviews Ecovacs CEO Qian Cheng: Succession and Robotics Management Philosophy

Wesentlich
2026-10-21 08:00 GMT+8

On October 21, 2026, 'Zhang Xiaojun Jun | Business Interview' released Episode 155, featuring David (Qian Cheng), CEO of Ecovacs. The episode's core theme revolves around "Father and Son, Samples and Textures of Second-Generation Succession, Robots, and My Inner Pursuit," aiming to present the realities and thoughts faced by individuals when first-generation Chinese enterprises transition from entrepreneurship to succession.

Born in 1990, Qian Cheng was once the youngest CEO of an A-share listed company in China, holding dual identities as both a "second-generation entrepreneur" and a "successor." Host Zhang Xiaojun notes in the introduction that Qian Cheng displays a "calmness" inconsistent with his age but is deeply attentive to operational details. In the program, Qian shares his experience of capturing 30% market share in Germany within six months and reflects on the idea that "the peak of success is the peak of ignorance" after becoming CEO.

Regarding technology implementation issues of interest to AI industry readers, the episode outline mentions the topic "Today, AI has not yet reached the state of stealing meat," suggesting the guest has specific judgments about the current commercial competition landscape of AI in robotics. Additionally, the content covers comparisons between "desire and restraint," "farming and nomadic" management styles, and the dynamics of father-son succession.

This episode is positioned as a sample for observing corporate succession in China. For professional readers interested in governance structures of AI hardware companies, founder succession plans, and long-term business philosophy, this interview provides a personal perspective and management insights from a frontline CEO. Note that the program explicitly states "This content does not constitute investment advice."

Quellen:xiaoyuzhoufm.com

Forschung

NTT and Four Partners Launch Optical Chiplet Mass Production Tech with NEDO Support

Thema · NEDO光芯片小片项目Wesentlich
2026-10-21 18:30 GMT+8

NTT, NTT Innovative Devices, Japan Semiconductor, Furukawa Electric, and Shinko Electric Industries announced on October 20, 2026, that their research theme for the production technology of optical chiplets has been adopted by the New Energy and Industrial Technology Development Organization (NEDO) of Japan. This adoption marks the formal start of intensive development towards mass-producing optoelectronic fusion devices.

As AI computing expands, power consumption in data communication between semiconductor chips has become a critical challenge. Optoelectronic fusion technology is expected to provide high-speed, low-power solutions. The five companies have been advancing research under NEDO's support since 2023. The newly adopted project aims to establish a consistent mass production technology covering everything from optical device manufacturing and inspection to optical connection and package implementation.

Under the project's division of labor, Japan Semiconductor will develop silicon photonics mass production techniques suitable for heterogeneous material integration. Furukawa Electric is responsible for compound optical semiconductor mass production. Shinko Electric Industries will handle packaging and process-in-line inspection technologies, while NTT Innovative Devices focuses on manufacturing, inspection, and implementation technologies required for mass production. NTT acts as an advisor, promoting corporate collaboration and supply chain formation.

The project targets a more than tenfold improvement in productivity compared to current R&D stages. The initiative seeks to enable the high-speed, large-capacity, and low-power data transmission required for next-generation AI computing infrastructure, while also strengthening Japan's domestic manufacturing base and accelerating the early practical application of optoelectronic fusion devices.

Quellen:eetimes.itmedia.co.jp

Forschung

Raja Koduri, Who Led HBM Accelerator Development at AMD and Intel: AI Chip Competition Is Shifting from Compute to Data Movement

Thema · Koduri内存瓶颈论Wesentlich
2026-10-21 07:59 GMT+8

Raja Koduri made a judgment in the fourth installment of SK Hynix's AI Ecosystem series: the center of gravity in AI semiconductor competition is shifting from raw compute to data movement. He was involved in the development of an HBM-based GPU at AMD in 2015, and later led the development of Ponte Vecchio at Intel, which integrated 47 chiplets and four HBM2e stacks into a single package. These projects repeatedly confirmed a fact for him—signal integrity, power delivery, thermal management, and mechanical stability are the critical determinants of actual performance, not the floating-point throughput of a single chip.

His argument begins with physical constraints: a single multiplication or addition requires only a few femtojoules of energy, but reading data from off-chip DRAM can cost hundreds or even thousands of times more. The farther data must travel from the compute device, the greater the time and energy costs. For decades, computer architectures have relied on caches to amortize this cost—bringing data closer to the compute device and reusing it multiple times, spreading the initial retrieval cost across many operations. This worked well for conventional computing workloads.

But transformers have challenged this assumption. As model parameters have grown significantly, it has become difficult to keep all weights on-chip; during inference, each weight may be used only a limited number of times before new data is required, sharply reducing data reuse and making it more difficult for caches to sufficiently offset the cost of initial retrieval. As context length increases, the KV cache expands proportionally, while attention operations increase both computation and memory accesses. Systems must repeatedly read data from DRAM or HBM. Koduri notes that even optimization techniques such as FlashAttention, PagedAttention, and speculative decoding—which reduce the amount of data movement—cannot entirely eliminate the physical cost of moving data.

His conclusion: memory is becoming the starting point of AI architecture design, rather than an afterthought once computation is complete. HBM compresses the distance between compute and memory from centimeters to millimeters through vertical stacking and through-silicon vias; advanced packaging and hybrid bonding further shorten die-to-die connections; and processing-in-memory places some computation directly where the data resides, reducing energy consumption by 10 to 100 times for specific workloads. The competitive standard shifts accordingly—no longer whose chip has higher compute throughput, but who can simultaneously satisfy bandwidth, energy efficiency, reliability, and thermal management within a limited space.

Koduri distills the future competition into three questions: How close can data be placed to the compute device? When movement is unavoidable, how quickly and efficiently can data be transferred? And how much computational value can be extracted once data has been retrieved? He acknowledges that PIM faces severe thermal-management challenges—dissipating heat when logic circuits sit close to temperature-sensitive DRAM is extremely difficult, and bandwidth, power efficiency, reliability, and thermal performance must all be achieved simultaneously within a limited space. The concept is straightforward, but turning it into a practical product requires system-level co-design spanning memory, compute, packaging, thermal design, and software.

Quellen:news.skhynix.com

Forschung

Cisco AI Security Team: Generic Test Banks Cannot Find the Risk of Agents Being Exploited Through Normal Function

Thema · 思科AI智能体测试Kurzfassung
2026-10-21 04:59 GMT+8

Agents are no longer just features inside applications. They hold credentials, call tools, and write to systems of record. This means their security testing cannot follow the logic of traditional software scanning: run a test suite, read the results, fix what it found.

The Cisco AI security product team, in a blog post dated October 20, 2026, drew a specific distinction: a support agent talked into refunding the wrong person is not a jailbreak. It is the agent doing its job, except the beneficiary is now an attacker. A test bank written for every agent finds only what is wrong with every agent; it will not find the error unique to this particular agent.

This distinction explains why testing must start from context. A healthcare scheduling agent and a financial advisory agent fail in different ways, because they serve different purposes, reach different tools, and carry different permissions. The adaptive red-teaming process Cisco describes works as follows: first study the agent's purpose, the tools it can reach, and the guardrails already in place; then derive the outcomes that agent must never produce, and pursue them across many turns, pivoting when one approach is blocked. The question stops being whether the agent fails a known attack and becomes what an attacker who reads your documentation would try.

Cisco also noted that standardized testing and adaptive testing each have a role: the former provides breadth against public frameworks (OWASP Top 10, MITRE ATLAS, NIST), yielding wide coverage, comparable results, and something an auditor can repeat; the latter provides depth on the agents that matter most. No one can hand-write a test plan for every agent in an enterprise, but if the goals are derived from each agent's own context, no one has to.

It should be noted that the above is Cisco's own description of its product methodology, with no independent verification data, comparative experiments, or third-party evaluation results attached. The Cisco–Workday partnership—Workday Agent Passport is currently in Early Access, with Cisco AI Defense serving as its launch assessor—is the first public practice case of this method, but its actual security effectiveness has not yet been independently verified.

Quellen:blogs.cisco.com

Forschung

a16z Crypto Partner Eddy Lazzarin Argues Against Pausing AI Development, Advocating Cybersecurity and Accountability Tools

Thema · AI暂停开发之争Wesentlich
2026-09-24 15:00 GMT+8

Eddy Lazzarin, General Partner at US venture capital firm a16z's crypto division, joined host Theo Jaffee on the MTS podcast for a debate over whether AI development should be paused. Eddy argued that the current AI safety conversation places too much emphasis on speculative superintelligence risks while underestimating the real costs of delaying useful technology. Rather than treating every AI failure as evidence of an alignment problem, he made the case for relying on familiar tools such as cybersecurity, accountability, liability, market incentives, and stronger technical controls. The episode also covered whether AI models can develop reputations for trustworthiness, the risks of concentrating oversight among a small group of evaluators, and why Eddy thinks the collision between Silicon Valley's AI debates and broader politics could fundamentally reshape the conversation over the next year. a16z and its affiliates may maintain investments in the companies discussed, as disclosed by the show.

Quellen:a16z.simplecast.com

Forschung

Linux Foundation Study: ~25% of New Open-Source Pull Requests Are Low-Quality AI Slop

Wesentlich
2026-10-20 21:18 GMT+8

The Linux Foundation published a study on October 20, 2026, examining the impact of generative AI on open-source software development. The report notes that while GenAI has the potential to cut coding time in half and increase productivity (with 71% of maintainers agreeing it lowers barriers to entry), it also introduces significant quality challenges.

The research found that approximately 25% of incoming pull requests are estimated to be low-quality 'AI slop'. Additionally, 59% of project maintainers reported growing GenAI-based contributions. This data highlights the trade-off facing the open-source community: the potential increase in contribution volume versus the rising burden of code review and declining quality.

Quellen:phoronix.com

Forschung

ZDNet Cites WSJ: Anthropic’s Potential Compute Contracts Could Reach $750 Billion

Thema · Claude企业采用率Wesentlich
2026-10-21 08:30 GMT+8

ZDNet Korea, citing The Wall Street Journal’s reporting on preliminary materials prepared for Anthropic’s initial public offering (IPO), reports that contracts to secure computing resources could total up to $750 billion over the next several years. This article relies on the ZDNet report, without directly checking the WSJ article or those preliminary materials, and does not establish that a filing has been formally submitted to a regulator. The report concerns potential long-term compute costs, rather than a total already spent.

The relay report states that existing contractual obligations for infrastructure spending exceed $600 billion. If additional contracts currently under negotiation are finalized, this amount could increase by another $150 billion. These agreements are primarily with cloud providers such as Google Cloud, Amazon Web Services (AWS), and Microsoft, spanning terms of 7 to 10 years. Notably, many contracts are described as difficult to cancel mid-term, implying that Anthropic would remain liable for substantial fixed costs even if future AI demand growth slows.

The report breaks down the providers as follows: cloud computing contracts alone account for approximately $450 billion in spending obligations, with Google Cloud representing $111 billion, AWS $110 billion, and Microsoft $31 billion. Additionally, Anthropic has entered into Nvidia chip rental agreements worth up to $86 billion with Neocloud providers Nscale and Lambda, and holds obligations exceeding $160 billion for renting Google chips via Broadcom.

The report also describes increased actual infrastructure spending. In the first half of 2026, Anthropic invested approximately $9 billion in computing capacity and related infrastructure, nearly four times the $2.3 billion spent in the same period last year, and surpassing the full-year total of $7.3 billion from 2025. Despite revenue surging from $1.1 billion in the same period last year to $16.3 billion in the first half of this year, net losses during the same period reached $80.7 billion, with about $72 billion attributed to book losses from value changes in financial instruments issued during fundraising.

ZDNet reports that the preliminary materials state that despite these massive investments, the company still faces compute shortages and is shifting towards establishing its own data centers through long-term leases or partnering with external operators to reduce reliance on traditional cloud providers. Simultaneously, increasingly organized opposition to data center construction in the United States has been cited as a new risk factor, prompting the company to consider overseas expansion. These reported points describe Anthropic’s contractual obligations and infrastructure risks; they do not represent direct verification of the preliminary materials by this site.

Quellen:zdnet.co.kr

Forschung

Metastone Releases TokenFleet Scheduling Framework: Claims 32.4% Reduction in P95 First-Token Latency for GLM Tests

Kurzfassung
2026-10-20 16:01 GMT+8

On October 20, 2026, Chinese compute operator Metastone released a distributed scheduling framework named TokenFleet at the CCF National High Performance Computing Academic Conference. The framework aims to address the simultaneous occurrence of request queuing and resource idling in GPU clusters by optimizing request routing, cross-datacenter capacity coordination, and mixed-load governance to improve effective output from existing computing power.

TokenFleet comprises three core technologies: TokenPilot handles intelligent inference routing by estimating both cache reuse benefits and queuing costs during instance selection; FleetMind focuses on cross-datacenter capacity coordination by aggregating model availability and network status across sites to select execution locations; and SLOMax manages production according to service requirements through classified queues and capacity feedback for mixed loads.

According to tuning data provided by Metastone for the GLM5.3 model, under tested GPU cluster and load conditions, adopting predictive latency routing reduced P95 first-token latency by 32.4% and P50 first-token latency by 71.3%. Additionally, with flow-control coordinated scheduling, total output throughput increased by approximately 1.5%, while Goodput (qualified output throughput) rose by about 29.4%, with the qualification rate increasing from roughly 67% to 86%.

Metastone emphasized that these figures are vendor self-tested benefit estimation examples, and actual gains require verification against specific workloads and service thresholds. As an independent third-party full-stack compute operator, Metastone currently manages over 20,000P+ of computing resources, and its team has won three Gordon Bell Prizes.

Quellen:qbitai.com

Forschung
Weiter →
Anmelden / Registrieren
LesenForschungRadarAnlageframework