Reading

A court in São Paulo, Brazil, has issued an injunction prohibiting US tech giant Meta from using content belonging to member companies of three Brazilian press associations to train its artificial intelligence (AI) models. The ruling directly targets Meta's previous attempt to force media organizations to license their content for free in exchange for continued platform access through modified terms of service.

According to heise online, Judge Guilherme Madeira Dezem of the 44th Civil Court in São Paulo granted the interim relief requested by the National Newspaper Association (ANJ), the Brazilian Association of Radio and Television Broadcasters (Abert), and the National Association of Magazine Publishers (Aner). These associations alleged that Meta's new terms of service for business users made the "free and unrestricted provision of copyright-protected material" a condition for continuing to use its platforms. The judge stated in the ruling that this requirement constituted an obvious abuse of rights.

In addition to suspending the execution of the relevant clauses, the court also banned Meta from sanctioning or penalizing media companies that refuse to accept these new terms. If Meta violates the injunction, it will face a fine of 1 million reais per day (approximately 162,000 euros). This decision is seen as an initial victory for the Brazilian media industry in protecting intellectual property and maintaining the sustainability of professional journalism.

A+ Content is the rich media widget on Amazon product detail pages, where brands use combinations of images, text, and video to tell their product stories. The challenge is that checking each asset in isolation—such as verifying image clarity or grammar errors—misses cross-module quality issues: for example, a feature list might advertise a 64-ounce capacity while a lifestyle photo displays a single-serving model. To address this, the Amazon team renders the entire page as a single snapshot, allowing the vision-language model to evaluate the content holistically, much like a customer would.

A core architectural decision is to not fine-tune the vision-language model. The team reasoned that fine-tuning a base model with only a few thousand labeled samples requires GPU training infrastructure and carries a risk of overfitting. As an alternative, they keep the VLM and its scoring rubric fixed. The model outputs scores from 0 to 10, confidence values, and natural language rationales across four dimensions: Visual Storytelling, Feature Presentation, Layout, and Conversion Optimization. These structured outputs are then used as features to train a CatBoost gradient-boosted decision tree classifier to predict expert-labeled high-quality or low-quality tags. Because only the classifier is retrained, updates required as annotations accumulate or criteria change can be performed as a short CPU-only task.

While vision-language models excel at capturing overall semantics and visual composition, they are less reliable for fine-grained typographic readability; they often overlook layout signals such as text density, line spacing, and character crowding. To capture these details, the team runs an OCR model in parallel to extract text blocks and their positions from the snapshot, calculating four readability metrics: Text Density, Character Density, Spacing, and Overall Readability. The fusion model inputs eight features—the four VLM scores and the four OCR metrics—into the same CatBoost classifier. On a 20% holdout set, the classifier using only VLM features achieved 89% accuracy and an F1 score of 0.90, while the one using only OCR features reached 82% and 0.83. The fused model achieved 94% accuracy and an F1 score of 0.95, representing a 5 percentage point improvement over the VLM-only approach.

Production validation of this method included several key steps: using SHAP to explain the contribution of each feature to decisions, feeding the most negative features along with product category and best practices into a language model to generate prioritized improvement suggestions, and running global A/B experiments that adjusted the visibility of low-quality content based on model judgments, which resulted in a statistically significant increase in Ordered Product Sales. Sellers who revised their content using these suggestions saw an average quality score increase of 1.4 points (on a scale of 10). It is important to note that all figures above are self-reported by Amazon, based on internal annotations and holdout sets, without independent replication. The annotations covered 10 product categories, and whether this method can be generalized to other multi-module content evaluation scenarios remains unverified.

Bruno Tafur and Tushar Madan of Databricks published an extension guide for application builders in November 2026, discussing how to maintain fast response times and high availability as a Databricks App grows from a few users to hundreds or thousands. The core judgment of the article is that the intuitive answer 'add more compute resources' is only effective once you know where requests spend their time and whether the application itself is built to scale.

The authors divide the scaling process into four stages: design, build, test deployment, and operations. In the architecture design stage, the most critical decision is dividing application state into two categories. Soft state is temporary working data, such as in-memory context, caches, and ongoing work, which can be rebuilt from hard state if lost; hard state consists of business records, checkpoints, and user preferences, which must be persisted in Lakebase or Unity Catalog tables so that any instance can recover the session. This classification determines whether horizontal scaling is truly feasible—if session state exists only in memory, users will lose their work when instances restart or migrate.

The article provides a specific retailer scenario to demonstrate load estimation: during the busiest hour of Black Friday, 200 out of 2,000 store managers are using an inventory query app simultaneously, with each performing two queries per minute, each executing three SQL statements. The result is 1,200 statements per minute, averaging 20 per second. The authors emphasize that this estimates query rate rather than concurrency; testing needs to validate burst traffic and growth, and database connections should be budgeted separately.

Regarding connection budgeting, the article provides an actionable formula: when each worker maintains a connection pool, the upper limit of connections is the number of instances multiplied by the number of workers per instance multiplied by the maximum pool size. For example, with five instances, eight workers each, and a pool limit of ten, up to 400 connections can be opened. After adding other clients and connections opened on behalf of users, this total must be compared against database documentation limits, leaving headroom for deployments. For scenarios requiring queries to Lakebase on behalf of users, each user requires an independent Postgres role, and Lakebase limits branches to 500 roles; adding instances does not change this upper limit.

The core recommendation for the build phase is to use asynchronous code. Most applications spend time waiting for external services—SQL queries, model calls, and external APIs. Blocking calls occupy an execution slot for the duration of the wait, meaning a few slow calls can exhaust capacity. When handlers await model and database calls, one worker can continue serving other requests while waiting. The article also recommends using asyncio.gather or Promise.all to launch independent queries in parallel, making request latency close to that of the slowest query rather than the sum of all queries, while limiting parallelism to avoid overwhelming the backend.

A GPU performing slightly below target specifications may not trigger errors, Xid events, or abnormal utilization metrics on monitoring dashboards. However, it can slow down an entire distributed training job to the speed of its slowest rank, with operations staff potentially discovering the issue only hours later through user complaints. In an article published on November 5, 2026, NVIDIA Technical Blog authors Sowmyan Soman and Kai Wombacher use this problem as a starting point to explain why GPU cluster readiness validation cannot be performed just once at delivery.

Their argument is that while a cluster passes readiness checks at handover, the original certification no longer reflects the cluster's actual state as nodes are replaced, drivers are patched, and capacity is expanded. This visibility gap wastes compute power on failing nodes and leaves healthy nodes idle due to conservative policies. The authors propose measuring cluster efficiency using goodput (the volume of successfully completed work) rather than utilization, and argue that maintaining readiness requires continuous validation throughout the operational lifecycle—after configuration changes, during capacity growth, and before repaired nodes return to service.

The article maps the three open-source components of NVIDIA DSX OS to these three verification moments. AICR (AI Cluster Runtime) provides version-locked configuration recipes, recording compatible versions and dependencies for components such as networking, storage, and drivers as auditable files, rendered as deployment artifacts usable by Argo CD, Flux, or Helm. NVCRE (Cluster Readiness Engine) runs NCCL communication tests and training workloads prior to cluster delivery, determining whether nodes meet standards via configurable performance thresholds, and bisecting suspect node ranges upon failure. NVSentinel continuously monitors GPU, node, and network health signals during cluster operation, publishes node statuses, and isolates, drains, and remediates faulty nodes when enabled.

The true incremental value lies in the automatic integration mechanism between these three components. When a node joins the cluster or returns after remediation, NVSentinel’s lifecycle controller automatically creates a ValidationRequest, gates scheduling using cordon or taint, and then invokes NVCRE to run tests. If the tests pass, the controller removes the gate; if they fail after retries, the node remains isolated. This means repaired nodes do not silently rejoin the pool. The authors specifically note that for situations such as driver upgrades, operational procedures, or fabric changes that do not trigger lifecycle events, AICR’s drift detection (aicr diff --fail-on-drift) can fail within a CI pipeline, thereby triggering a certification run within the same change process.

The article also clearly delineates the responsibility boundaries of each component. NVCRE is responsible only for testing and attribution—recording which nodes failed and why, without altering node status; NVSentinel is responsible for isolation and repair. The authors explain that this separation is intended to prevent two systems from cordoning the same node simultaneously, which would lead to double isolation and desynchronized cleanup. NVSentinel is installed in monitoring mode by default, reporting faults merely as node conditions without executing cordon, eviction, or repair actions; operations teams enable automated handling only after trusting the quality of the reported signals.

US chipmaker Qualcomm publicly exhibited the world's first 6G terminal prototype compliant with the latest 3GPP definitions at the ninth China International Import Expo (CIIE), which opened on November 5. This marks the first time the prototype has been displayed physically to the public outside of a laboratory setting.

According to reports from Zhidongxi and official CIIE news, the prototype serves as a key test and verification platform for moving towards 6G commercialization. It has already been used in joint 6G technology tests and Interoperability Device Testing (IoDT) with Chinese industry partners including China Mobile and ZTE. In September, Qualcomm and China Mobile completed IoDT based on the latest 3GPP-defined radio frequency specifications, achieving uplink and downlink link interconnection between devices from different vendors.

Regarding performance parameters, tests conducted by Qualcomm and ZTE in their Xi'an laboratory late last year showed that the prototype utilized 400 MHz of downlink bandwidth and 400 MHz of uplink bandwidth. Its downlink bandwidth is four times that of current 5G spectrum, with a peak downlink rate of approximately 14 Gbps. Beyond wired radio frequency calibration tests, Qualcomm stated it will gradually move into Over-The-Air (OTA) testing phases.

Currently, the 3GPP 6G standard is still evolving, with the first version, Release 21, expected to be finalized by 2029. At the Mobile World Congress (MWC) in March, Qualcomm reached a consensus with nearly 60 global partners, including about 20 Chinese ones, to push for 6G commercialization starting from 2029. Qualcomm plans to continuously configure and upgrade this prototype based on standard developments to accumulate practical experience.

SemiAnalysis released the first cross-model results for TPUv7 (Ironwood) on the InferenceX/AgentX agent inference benchmark on November 5, 2026. AgentX simulates real-world agent workloads characterized by multi-turn interactions, long contexts, and high prefix reuse, which differ significantly from single-turn chat scenarios. This release covers four models: Meta's Muse Spark Open (816B parameters), Zhipu AI's GLM-5.2/5.3 (744B parameter MoE), Alibaba's Qwen3.5 397B, and Moonshot AI's Kimi K3 (2.8T parameters).

At FP8 precision, the results for each model must be viewed separately. On GLM-5.2, within an interaction range of 60 to 300 tokens per second per user, TPUv7 delivers 1.2 to 2.7 times more tokens per dollar than the B200, B300, and AMD MI355X. On Qwen3.5 397B, SemiAnalysis reports that TPUv7 FP8 performance per dollar is 1.67 times that of the GB300 NVL72 FP4—this marks the first AgentX model result submitted by Google itself. On Kimi K3, results submitted by Inferact, founded by the vLLM founding team, show that TPUv7 delivers 1.15 to 2.5 times more tokens per dollar than the GB200, B200, B300, and MI355X. Muse Spark Open was submitted by RadixArk, maintained by SGLang, achieving competitive decode cost-effectiveness even on a GPU-optimized architecture theoretically unfriendly to TPUs.

However, SemiAnalysis notes that NVIDIA FP4 still maintains a lead in pure performance per dollar. On GLM-5.2, at 100 tok/s/user, the GB300 and GB200 FP4 deliver 3.6x and 3.5x more tokens per dollar than TPUv7, respectively. This is because FP4 halves the weight bytes, and small-batch decode is memory-bandwidth bound. TPUv7 currently lacks native FP4 support, with this feature expected to arrive with TPUv8 in Q1 2027. NVIDIA submitted FP4 counter-results on Qwen3.5; while these exceeded TPUv7 FP8 in decode performance, they came at the expense of prefill, resulting in lower end-to-end performance than TPUv7 FP8.

SemiAnalysis's core argument is that TPUv7's competitiveness stems not just from hardware specifications but from ecosystem collaboration with upstream maintainers such as vLLM (Inferact, RedHat), SGLang (RadixArk), llm-d, PyTorch, and Ray. Meta has already deployed TPUv7 in production using SGLang-TorchTPU. Kimi K3.1 Pro received TPUv7 support on its release day. SemiAnalysis argues that NVIDIA's moat is being eroded from its software ecosystem, citing bureaucratic processes—such as a three-month legal review required for driver feature approvals—as a structural risk.

It is important to emphasize that the submitters of the above results vary: Qwen3.5 was submitted directly by Google, Muse Spark Open by RadixArk, and Kimi K3 by Inferact. All data are first-party results analyzed and published by SemiAnalysis and have not yet been independently reproduced. Regarding NVIDIA, preview submissions for the Rubin NVL72 suggest its performance per dollar could reach 6 times that of TPUv7, but SemiAnalysis points out this is an unfair comparison; Rubin should be benchmarked against TPUv8i rather than TPUv7. For readers making compute procurement decisions, these figures provide directional reference, but independent verification should be requested before making commitments.

S. Krishnan, Secretary of India's Ministry of Electronics and Information Technology (MeitY), stated in a November 4 interview that under the "India Semiconductor Mission 2.0," the country intends to develop an indigenous graphics processing unit (GPU) within the next 24 months. "We should have a GPU chip of our own. This will not happen overnight. It is a project that will require approximately 24 months, and we intend to work towards that objective," Krishnan said. The initiative aims to build a self-reliant hardware ecosystem and reduce dependence on foreign chip suppliers.

Beyond GPUs, India plans to bring its SHAKTI and VEGA processors, based on the open-source RISC-V architecture, to commercial markets. These CPU designs allow local developers to build processors without relying on proprietary instruction sets. Previous efforts include the C-DAC developed VEGA ET301 processor used in the Aries 3.0 development board for IoT applications. Additionally, the government's Design-linked Incentive (DLI) program has supported 24 companies in areas such as microcontrollers, drone flight controllers, and smart meter chips.

Krishnan compared the current AI boom to the Year 2000 (Y2K) bug transition, suggesting Indian IT firms could benefit by helping global businesses integrate AI operations, similar to how they capitalized on software services demand at the turn of the century. While NITI Aayog estimates that AI adoption could make about 2 million existing jobs redundant, it also projects the creation of 4 million new roles. India's semiconductor market is currently valued at approximately $50 billion and is expected to approach $100 billion by 2030.

However, the GPU project remains in its early stages. Krishnan did not identify a specific manufacturer, chip supplier, or builder for the project. In contrast, India has made more concrete progress in wafer fabrication, with ASML partnering with Tata Electronics to equip India's first commercial semiconductor fab in Dholera, Gujarat, and Intel announcing a strategic partnership with the Tata Group. Developing a competitive AI accelerator presents significantly higher technical barriers than general-purpose CPUs, making the 24-month timeline highly ambitious.

Researchers often face an apparent contradiction: they must ensure the security of sensitive data while allowing results to serve society as openly as possible. If this contradiction is treated as a switch—either fully open or fully closed—it is indeed difficult to manage. The alternative approach proposed by Manish Parashar, Chief AI Officer at the University of Utah, is to treat it as a knob: having the capability to turn this knob becomes critically important.

In an interview with theCUBE’s Sovereign AI series, Parashar stated, "In research, we are always trying to balance research security with doing open research for societal good." Depending on what you are working on, you may need to lean toward one end or the other. He further emphasized, "We don’t want everything tightly controlled, nor do we want everything completely open; we want to ensure we have the ability to right-size it depending on what we are working on."

How does this knob turn in practice? The University of Utah’s Redtail AI supercomputer provides a concrete scenario. Hewlett Packard Enterprise (HPE) helped architect and deploy this turnkey system, but daily operations are managed entirely by the University of Utah Center for High Performance Computing—admin access, resource allocation, user onboarding, monitoring optimization, and capacity adjustments are all handled internally by the university. Parashar describes this system using five dimensions of sovereignty: territorial, legal, technological, operational, and financial. The university retains operational and legal responsibilities, while funding relies on partnerships with the state government and donors.

At the infrastructure level, Parashar explicitly stated that he does not pursue a single physical location: "Our overall ecosystem is hybrid, with on-premises deployments and cloud-hosted solutions; we always view it as a hybrid system, not requiring everything to be in one data center or within one state." This hybrid architecture is precisely the prerequisite for the knob to turn—if all computing were locked in one place, adjustment would be impossible.

The practical effect of turning the knob is reflected in capacity reallocation. The University of Utah’s existing restricted environments support sensitive and regulated research, including healthcare projects. With Redtail online, a significant amount of open research has moved to this new facility, thereby freeing up capacity in the original data centers used for regulated work. Parashar noted that the next step is entering facilities that possess not only capacity but also compliance with emerging regulatory controls, which is currently the focus. In other words, moving open research outward is itself an operational maneuver to create space for secure research.

Financial sustainability is key to whether this model can operate long-term. Parashar mentioned that expansion planning began before the Redtail contract was signed. The university is engaging federal agencies such as the U.S. Department of Energy and the National Science Foundation to discuss investments in the existing ecosystem. Simultaneously, it is demonstrating value to the Utah state government to persuade them to invest, arguing that it benefits both the state and its citizens, while also reaching out to additional philanthropic donors. This diversified funding source is itself part of operational sovereignty—avoiding reliance on a single vendor or a single grant channel.

German semiconductor startup Quantum Diamonds has announced a €152 million investment in a new production facility in Munich, Bavaria, with the first phase opening in July 2026. This move marks a critical transition from laboratory research to industrial manufacturing for the company's core product: "quantum diamond microscopes," which use atomic defects in synthetic diamonds to detect internal flaws in multi-layered AI chips without damaging them.

As AI chip architectures become increasingly complex, traditional inspection methods struggle to identify microscopic internal errors, leading to lower yields and higher costs. Quantum Diamonds' technology locates these defects by measuring weak magnetic fields within the chip. According to the company, the method has been tested with nine of the world's ten largest semiconductor manufacturers, including TSMC and Samsung. Since April 2026, the first systems have been installed and operational at Eurofins EAG Laboratories in California and Integrated Service Technology (iST) in Taiwan, indicating that the technology has moved beyond the research phase into actual industrial application.

The funding structure reflects European industrial policy's emphasis on local supply chain security. Of the €152 million investment, €91 million comes from a financing mix: €15 million from venture investors like World Fund, and €76 million provided by the German federal government and the State of Bavaria under the European Chips Act. Previously, Quantum Diamonds had considered relocating to Arizona, USA, but a subsidy rate of approximately 50% ultimately kept the startup in Germany. Daria Saharova, Managing Partner at World Fund, stated that the company has the potential to become "Europe's next ASML," although it currently lags significantly behind the Dutch lithography giant in scale and technological maturity.

For AI industry investors, the Quantum Diamonds case highlights a new trend in Europe's semiconductor revival: focusing on "picks and shovels" strategies. While large-scale wafer fab projects like Intel's Magdeburg plant have faced setbacks, German startups specializing in specific segments—such as materials, interconnects, or inspection—are receiving dual support from public funds and venture capital. Companies like Black Semiconductor (optical interconnects), FMC (ferroelectric memory), and NextGO Epi (gallium oxide wafers) have secured similar levels of funding. These firms do not produce chips directly but provide key equipment and technologies that improve chip manufacturing yields and efficiency. However, challenges remain: the technology currently focuses on individual chip inspection, and proving it can rapidly and reliably process entire wafers containing hundreds or thousands of chips is the crucial hurdle for mass production.

China's leading artificial intelligence lab Moonshot AI is seeking access to advanced Nvidia chips through data centers located in Europe, aiming to navigate around U.S. export restrictions. According to a CNBC report published Thursday, four people familiar with the matter stated that discussions took place this year regarding Moonshot renting capacity from neocloud providers hosting Nvidia graphics processing units (GPUs) at European facilities. The sources requested anonymity due to the private nature of the information.

Nvidia's effective monopoly over the most powerful chips has made U.S. export controls a key instrument in Washington's effort to maintain its advantage over China in the AI race. American AI labs such as OpenAI and Anthropic are engaged in training and pricing battles with Chinese developers, including Moonshot AI, which produces the Kimi family of models. While U.S. law restricts the shipment of Nvidia's most advanced chips to China, remote access via overseas data centers offers a loophole, as current regulations do not grant authorities broad power to control corporate access to these chips through cloud services hosted abroad.

Moonshot was interested in leasing capacity powered by Nvidia's Blackwell systems, which are currently subject to export controls banning their import into China. Two sources indicated that Moonshot focused on clusters for inference—serving models to end users—in Europe, while another source noted interest in training clusters as well. One source described talks about arrangements where Moonshot would rent access to clusters of 8,000 or more Blackwell GPUs hosted in the region, characterizing the Chinese AI lab as "very aggressive" in its pursuit of computing power.

To lease Nvidia computing power in Europe, a Chinese AI company typically rents capacity from non-Chinese neoclouds or data center operators capable of purchasing advanced chips. One source mentioned a proposal for Moonshot to pay 30% to 50% more upfront than a non-Chinese offtaker would. Arrangements with Chinese customers can be more lucrative upfront, as neoclouds and financing firms look to hedge against a potential future ban on remote access agreements.

This development follows a July statement by the White House alleging that Moonshot had accessed Nvidia's Blackwell chips via a facility in Thailand, a claim Moonshot did not comment on at the time. Additionally, the Financial Times reported on Tuesday that Nvidia told cloud providers they must sell at least 51% of their services to non-Chinese customers. An Nvidia spokesperson responded: "America cannot win the AI race unless America's AI infrastructure wins worldwide—if the world's clouds rely on foreign-made chips, the model and application layers will be foreign-built as well."