Wednesday, November 18, 2026 Edition no. 38 · Evening

Artificial intelligence news from Brazil, in English

Noticiário do Joãozinho: Edition no. 38 · Nov. 18, 2026 · Evening

Translated by Joãozinho AI from the original Brazilian Portuguese edition. Facts, numbers and sources are the same as in the original; quotes, video titles and source names stay in their original language. Read the original in Portuguese

Listen to the spoken edition

Length: In Brazilian Portuguese

Download the MP3

Top storyEdition no. 38 · Nov. 18, 2026 · Evening

Update: Anthropic sets IPO price at $83 per share, above the range of $70 to $78, and is valued at $1.5 trillion on the eve of its Nasdaq debut

Anthropic set the price of its initial public offering at $83 per share, above the preliminary range of $70 to $78, according to The Information, late Wednesday (11/18). The price values the owner of Claude at $1.5 trillion on a fully diluted basis, and Class A shares will begin trading on the Nasdaq this Thursday (11/19) under the ticker ANT. The offering consists of 825 million shares: 801.35 million sold by Anthropic itself and approximately 23.65 million by current shareholders. Meanwhile, AMD, Broadcom, Google, Micron, Microsoft, NVIDIA, and TSMC are collectively purchasing $26 billion in Class N shares, which carry no voting rights, at the IPO price. Earlier, CNBC reported that the company was considering increasing the offering due to demand, and Bloomberg noted it was signaling a price above the range despite days of tech stock declines. The figures Anthropic published for the IPO show revenue of $17.0 billion in the third quarter of 2026, more than 12 times that of the same period last year, and a net loss of $23.3 billion, of which $21.9 billion is a non-cash expense. The BBC noted that the company was founded in 2021 by seven individuals who left OpenAI, including siblings Dario and Daniela Amodei, chief executive and president; at the top of the old range, each founder's stake would already be worth at least $18.5 billion. Vanda Research estimates approximately $250 million in net purchases by retail investors on the first day, double what it had forecast for SpaceX's IPO.

PDF (pt-BR)

Models

1 story

High importanceNov. 18, 2026

Update: Google launches Gemini 4 Neon on the API, its fourth model from the Gemini 4 family in four weeks

Google released Gemini 4 Neon this Wednesday (11/18), positioning it as a frontier model for software development, knowledge work, and science. It is already available on the Gemini API, Google AI Studio, and the AI Pro and Ultra subscriber apps, where it replaces Argon and 3.1 Pro; 3.5 Flash-Lite becomes the "Auto" option for paying users. The price is $3 per million input tokens and $15 per million output tokens, with a 1-million-token window and four reasoning levels (low, medium, high, and ultra-high). It accepts text, images, video, audio, and PDFs, supports Computer Use, code execution, file search, and queries to Google Search and Maps; it does not generate images or audio. In the Google DeepMind blog, the company states that Neon ranks 2nd on the Vals Index (70.9%), 3rd on the Artificial Analysis Intelligence Index, and costs at least 36% less per task than competing models at the highest tier; according to 9to5Google, it leads Artificial Analysis's programming agent index with 58.6 points.

Research

7 stories

High importanceNov. 18, 2026

Axiom Math Startup Says It Has Verified 15,000-Page Proof of Classification of Finite Simple Groups with AI

Axiom Math announced the formal verification, automatically conducted by a swarm of AI agents, of the proof for the Classification of Finite Simple Groups (CFSG), one of the largest achievements in 20th-century mathematics, according to IEEE Spectrum. The original proof spans approximately 15,000 pages, spread across more than 500 papers by around 100 authors over half a century. The formalization resulted in 43 million lines of code, comparable to the size of the Linux kernel. In parallel, the company formalized the Stacks project, a reference work in algebraic geometry spanning 7,500 pages, which translated to 16 million lines of code.

High importanceNov. 18, 2026

Calif startup uses GPT-6 Cyber, still not public, to find critical flaw in Java; Oracle releases fix without announcement

According to MobileSyrup, cybersecurity startup Calif used GPT-6 Cyber, an OpenAI model not yet available to the public, to identify a flaw in Java triggered by a sequence of "A" characters followed by an equals sign. When exploited, the flaw can grant full and unauthorized access to a device. Java is used by 90% of Fortune 500 companies, including Amazon and Apple, as well as banks such as J.P. Morgan Chase and Bank of America, and hospitals. Oracle, Java's owner, made no public statement but released a fix on Tuesday (11/17).

Medium importanceNov. 18, 2026

OpenAI says its models improved AMD's GPU design by up to 14%, equivalent to a year of work, in less than five months

OpenAI states that, working with AMD, its models achieved an improvement of up to 14% in PPA (power, performance, and chip area) for AMD's flagship GPU design in less than five months, according to The Decoder. The company says this equates to roughly a year of progress over a design that AMD engineers had already been optimizing for years, and that some of the changes have been incorporated into a tapeout sent to manufacturing. The same method was applied to Jalapeño, OpenAI's first proprietary inference chip, and in tests with open RISC-V cores, the GPT-6.2 Astra achieved the best results across model generations. OpenAI is also working with Synopsys on a model called GPT-Synopsys, which uses the company's EDA tools. The figures are from OpenAI itself.

Medium importanceNov. 18, 2026

Harvard study in Nature Sustainability shows that 54% of energy from the largest US data centers comes from gas and coal

Gianluca Guidi and Francesca Dominici, from Harvard, identified 403 large-scale data centers in the US using satellite imagery and cross-referenced estimated energy demand between May 2024 and April 2025 with emissions from each regional grid, according to The Decoder. Consumption ranged between 68 and 99 TWh per year, with a central estimate of around 82 TWh, close to 1.8% of the country's total electricity, with more than half concentrated in Virginia, Oregon, Ohio, and Iowa. About 54% of this energy came from fossil fuels: 38% natural gas and 15% coal. Emissions reach around 25.7 million tons of CO2 per year, comparable to Denmark's, two to three times a previous estimate for 2018.

Medium importanceNov. 18, 2026

AWS Opens Tools Connecting AI Agents to Isabelle Proof Assistant via MCP to Write Hallucination-Free Proofs

AWS published three open tools in the AutoCorrode suite (I/Q, I/R, and Isabelle Assistant) that connect an AI agent to the Isabelle proof assistant via the Model Context Protocol (MCP), according to the Amazon Science blog. The idea is to interleave proof generation with computer checking, keeping the agent on track and preventing hallucinations. I/Q allows the agent to interact with an open Isabelle session on the user's machine, and I/R enables it to develop proofs as the person continues working in the assistant. With Isabelle Assistant, Claude Opus 4.6 independently produced readable and error-free proofs for classic examples, such as the irrationality of √2 and the integral of sin(x).

Low importanceNov. 18, 2026

Apple Releases FastFrames, a Video Model That Generates 480p at Interactive Latency From Text

Apple has published FastFrames-L on Hugging Face, a text-to-video model with 14 billion parameters, along with a smaller version, FastFrames-S, featuring 1.3 billion. According to the model's page, it is a two-step autoregressive diffusion model distilled from Wan2.1-14B that generates 832-by-480 pixel video at 81 frames at 16 per second. Apple reports approximately 35 milliseconds per frame block for FastFrames-S on a GPU. The model runs on Linux with an NVIDIA GPU, and it is licensed under Apple's research license rather than a commercial open-source license.

Low importanceNov. 18, 2026

NVIDIA Launches Daviator, Accelerating Quantum Chemistry Calculations in VASP by Up to 22 Times

NVIDIA introduced Daviator, a GPU kernel library that replaces the sequential processing of VASP's k-points, a program widely used in density functional theory (DFT) materials simulation, with batch operations, according to the company's technical blog. In small systems with dense k-point grids, hundreds of small GPU calls per step caused the card to perform worse than a CPU. With Daviator, VASP runs up to 22 times faster and at up to 75% lower cost than on a CPU; a mixed-precision mode yields about 2 times more performance on GPUs like the RTX PRO 6000. The feature is available through the NVIDIA ALCHEMI VASP microservice.

Products

2 stories

Medium importanceNov. 18, 2026

Mistral launches Vibe Desktop for macOS and Windows, with multiple agent tasks running in parallel

Mistral has released a desktop application for Vibe, its AI agent formerly known as Le Chat and previously available only on the web, via command line, and on mobile, according to The Decoder. In the app, users can submit multiple tasks at once and watch the agent work on them in parallel, isolated by Git worktrees, both in a local repository and in a cloud sandbox. The working mode now accesses local files in addition to code. The app is available for macOS and Windows, with a Linux version promised; Mistral did not mention computer control.

Medium importanceNov. 18, 2026

AWS Lambda Now Runs GPU-Powered Functions for Serverless Inference and Media Processing Management

AWS Lambda now accepts GPU-enabled EC2 G-series instances for event functions running on Lambda Managed Instances, according to the AWS News Blog. Users can run machine learning inference, media processing, rendering, and scientific computing without managing drivers, CUDA versions, AMIs, or clusters. Functions are published as container images up to 32 GB, with 3 to 48 GB of accelerator memory, execution times up to 90 minutes, and support for up to 48 concurrent requests per environment.

Business and Investment

2 stories

High importanceNov. 18, 2026

Palo Alto Networks acquires Applied Compute, specializing in open model customization, for $4.5 billion in stock

Palo Alto Networks has finalized an agreement to acquire Applied Compute, a company that helps businesses train, serve, and refine proprietary models based on open-weight models, in an all-stock transaction valued at $4.5 billion, according to a filing submitted to the SEC this Wednesday (11/18). The deal’s completion is contingent on regulatory approval and is expected in Palo Alto's third fiscal quarter of 2027. The company stated it will offer the platform to customers so they can build, fine-tune, and run open models with built-in security across public clouds or sovereign environments, while leveraging its research team to train proprietary models of all sizes. Applied Compute’s CEO, Yash Patil, will become Palo Alto’s co-head of AI. According to The Information, the startup was valued at $3.25 billion earlier this month.

Medium importanceNov. 17, 2026

Cerebras combines AMD GPUs and its giant chips in the same inference and says it delivers five times more tokens per megawatt

Cerebras has launched disaggregated inference, already in production: GPUs handle the prompt (the prefill phase) and Cerebras' CS-3 systems generate the responses (the decode phase). In production deployment, AMD Instinct GPUs handle the prompt and a software called KV Portal moves the model state from one side to the other via RDMA. According to the company, in a service with GPT-OSS-120B measured on 11/14, the arrangement delivered five times more throughput per megawatt than a CS-3-only configuration, without losing generation speed. AWS, Fireworks AI, Gimlet Labs and General Compute have early access, and Cerebras aims to scale the model to hundreds of megawatts by the end of 2027.

Regulation

3 stories

High importanceNov. 18, 2026

GAO report, U.S. Congress oversight agency, shows Claude, ChatGPT, Gemini and Grok helped assemble “viable” attack plans in test

The Government Accountability Office (GAO), a nonpartisan agency that audits the government for Congress, released on Wednesday (11/18) a report requested by Senator Maggie Hassan. According to ABC News, investigators used Claude, Grok, ChatGPT and Gemini, employing techniques already known to bypass security safeguards, and within up to 12 hours obtained “feasible” information to plan attacks. Each platform was given one of two scenarios: two developed attack plans using improvised explosives in major cities, including dates, times and targets linked to public events, and suggested ways to evade police; the other two provided instructions for attacking power grid substations, citing drone and weapon models. The tests were conducted in April, and the report does not attribute the results to each company. Anthropic, Google, OpenAI and SpaceXAI told the GAO that they have improved their safeguards since then; OpenAI stated to ABC that the test used an older version of ChatGPT. Hassan wants hearings on AI safety next year and a law requiring safeguards that cannot be bypassed in this manner before launch.

Medium importanceNov. 18, 2026

On the eve of the IPO, Future of Life Institute letter accuses Anthropic of racing toward self-improving AI

A group of signatories convened by the Future of Life Institute (FLI) published a statement against Anthropic this Wednesday (11/18), accusing the company of recklessly pursuing recursive self-improvement (RSI), a process in which AI systems improve themselves or build their successors. Signatories include, among others, writers Steven Pinker and Rutger Bregman, as well as co-founders Igor Babuschkin of xAI and Thomas Wolf of Hugging Face. The letter calls for curbing Anthropic’s political influence and rejecting its bet on technological dominance. The FLI cites an independent METR report stating that the company’s pursuit of RSI accelerated AI development by nearly two times, and notes that its IPO filing (S-1) reveals Claude already writes more than 90% of its own code. Max Tegmark, FLI co-founder, labeled Anthropic, the company that positions itself as responsible, as the most culpable among AI firms.

Medium importanceNov. 18, 2026

Muslim writer sues xAI after users used Grok to remove her hijab in images

Writer and activist Amani Al-Khatahtbeh filed a lawsuit in the U.S. District Court for the Northern District of California against xAI, accusing the company of intentionally distributing nonconsensual sexual images, according to Wired. X users asked Grok, the AI embedded in the platform, to alter photos of her, which generated images of her without a hijab and wearing revealing clothing. She is seeking up to $250,000 per violation of the California Civil Code. Her lawyer uses defective product law to argue that xAI designed Grok in a way that sexually exploits women and children. A January report from the Center for Countering Digital Harm estimated approximately 3 million photorealistic sexualized images generated by Grok on X in 11 days.

Open Source

3 stories

High importanceNov. 18, 2026

Sashiko, Google’s Open-Source Code Reviewer Powered by Gemini, Performs 1,200 Reviews Daily on Linux Kernel and Cuts Late Fixes by 35%

Google detailed how it built Sashiko, an open-source agent-driven code review system based on Gemini that monitors over 100 Linux kernel mailing lists and performs approximately 1,200 reviews daily, according to the Google Cloud Blog. The rate of fixes issued within 60 days after a patch was merged dropped to 2.8% in Linux 7.1 and 2.2% in Linux 7.2, representing a 35% reduction according to Google. Sashiko is referenced in over 2,000 mainline kernel commits and more than 500 official Linux CVEs this year, with over 1,500 kernel developers already responding to its reports. In a company test, the version powered by Gemini 4 Carbon caught 75.8% of bugs that human reviewers missed.

Medium importanceNov. 17, 2026

SenseNova publishes Loop-9B, an open 9-billion-parameter model that reuses layers and fits in about 10 GB

SenseNova released the official version of SenseNova-Loop-9B on 11/17 under an Apache 2.0 license, and added AWQ and GGUF versions on 11/18, according to the model's page on Hugging Face. The model accepts text and images and was designed for reasoning, coding, tool use, and multimodal tasks. It has about 9 billion parameters, but repeats the middle layers in a looping block: 32 layers executed 56 times per pass, equivalent to about 14.6 billion. The company says it ties with Qwen3.8-27B on AutomationBench and scores 59.55 on Terminal-Bench 2.1; the numbers are from them. With 4-bit weights, the page estimates about 10 GB of memory with a context of 64,000 tokens.

Medium importanceNov. 18, 2026

Cloudflare unlocks the weights for Clairic, a 27B data extraction model, and launches 9B Clairic-flash

Cloudflare has released two open-weight data extraction and prediction models: an updated version of Clairic with 27 billion parameters, and the lighter, faster Clairic-flash with 9 billion, according to the company's blog. Both run on Workers AI and full weights are available on Hugging Face. Until December 1, 2026, both cost $0.10 per million input tokens on Workers AI; after that, Clairic reverts to $0.24.

Agents and Code

3 stories

High importanceNov. 17, 2026

Official Microsoft Agent Framework package on PyPI gains two versions with credential stealer, published 33 minutes apart

Versions 1.26.1 and 1.26.2 of the agent-framework package, the Python metapackage for Microsoft Agent Framework to build AI agents, were published on PyPI on Tuesday (11/17), 33 minutes apart, by someone who used valid publishing credentials, according to Aikido Security. The malicious code ran every time Python started, even without the package being imported, and collected environment variables that looked like secrets, credential files for Azure, npm, Git, Docker, and Hugging Face, as well as tokens for Azure DevOps pipelines and, in the second version, GitHub Actions. The data was sent to a webhook.site address. Both versions were removed around 9:20 PM (UTC) the same day, and version 1.26.0 became the latest again. The recommendation is to pin version 1.26.0, look for the agent_framework_meta.pth or af_runtime_diag.pth files, and rotate all secrets that the affected machines could access.

Medium importanceNov. 17, 2026

Test with 27,000 executions shows that the tool surrounding the model changes a code agent's cost by up to five times, without improving success rates

Runta, in partnership with Baseten, published FrontierHarness Eval v1.1, which compares “harnesses,” the tools that wrap around a model in a code agent. There were 22 configurations across 18 harnesses, five open models (GLM-5.3, GLM Flash, DeepSeek V4.1 Flash, Muse, and Kimi-K3), and 62 terminal tasks, totaling 27,280 executions and approximately 83 billion tokens. With the model fixed, the most expensive harness cost between 1.8 and 5 times more per task than the cheapest, and cost and success rates barely correlated. The best cost-effectiveness varies by model: Codex CLI ranked first with GLM-5.3 and 16th with Kimi-K3. According to the study, 64.5% of spending went toward attempts that failed. The test only used open models served by a single provider.

Medium importanceNov. 18, 2026

KPMG tests over 100 AI agents in the audit of a company listed in London, and they sometimes ask better questions than auditors

KPMG UK used over 100 AI agents to examine the accounts of a large company listed in the country, alongside a conventional audit team, according to European Business Magazine, which reports on the Financial Times. The agents performed tasks at all stages of the audit, not just administrative work, and at times investigated financial issues more deeply and asked better questions than humans. The technology took about four months to develop. During the test, an auditor had to intervene when one agent failed to correctly match regulatory requirements with the company's documentation.

Local Models

1 story

High importanceNov. 18, 2026

Compact RTX Spark desktops, with up to 128 GB of unified memory for local AI, hit stores starting at $4,400

The first compact RTX Spark desktops from ASUS, Dell, and MSI began selling this Wednesday (11/18), according to NVIDIA's blog. They feature the RTX Spark N1X chip, with a 20-core ARM Grace CPU, a Blackwell GPU with 6,144 CUDA cores, and up to 128 GB of unified memory, delivering up to 1 petaflop of AI performance over a 10-gigabit network. NVIDIA says you can run models with 120 billion parameters locally, without paying per token to a cloud API, keeping the code and files on your PC. According to Clubic, the starting price in the U.S. is $4,400; in France, the ASUS ProArt GR1X with 64 GB goes for €5,815 and the MSI EdgeMesa N AI+ with 128 GB goes for €8,190. Acer, GIGABYTE, HP, and the Microsoft Surface RTX Spark Dev Box will follow.

Finance

2 stories

Medium importanceNov. 17, 2026

NVIDIA's stakes in companies reach $65.5 billion, up from $22 billion in January

NVIDIA reported in its third fiscal quarter 2027 earnings, released on Tuesday (11/17), that its company stakes are worth $65.5 billion, compared to $22 billion reported in January, according to TechCrunch. In the last quarter alone, the company invested $16 billion in equity, and over the first nine months of the fiscal year, $58 billion, compared to $5 billion in the same period last year. Among its investments are Anthropic, OpenAI, Thinking Machines, Reflection, and the neoclouds Lambda, Nscale, and Crusoe; according to PitchBook, it backed nearly 350 startups in about three years. NVIDIA also recorded $118 billion in guarantees, including $105 billion for OpenAI's $500 billion data center in Ohio.

Low importanceNov. 18, 2026

Coinbase Blocks $4.7 Million Sent to Scammers Using Elon Musk Deepfake in Fake Tesla Broadcast

Coinbase blocked $4.7 million in cryptocurrency transfers to scammers who orchestrated a fake live Tesla broadcast on YouTube, featuring an Elon Musk deepfake promising to double whatever viewers sent, according to BleepingComputer. The broadcast drew more than 500,000 viewers. The company stated that its fraud prevention systems blocked approximately 91 percent of the transfer attempts through a fully automated detection and intervention process, with no human involvement.

Games

3 stories

High importanceNov. 18, 2026

Epic to Remove Blueprints from Unreal Engine 6 and Migrate Everything to the Verse Language, Focusing on AI Agents

According to IEEE Spectrum, Epic Games will remove Blueprints, Unreal’s visual programming system, in Unreal Engine 6. All code will move to Verse, a text-based language that replaces both Blueprints and the C++ game framework of Unreal Engine 5, because language models handle text code more effectively. UE6 also replaces the object-oriented actor system with Scene Graph, becoming more modular and exposing its systems via an API so AI tools can access them. The developer preview on November 9 introduced Unreal Canvas, a collaboration feature inspired by visual workspaces.

Medium importanceNov. 18, 2026

Google DeepMind Opens Genie 4, Creating Interactive 3D Worlds with Sound and Allowing Weather, Characters, and Story Changes via Text

Google DeepMind has released Genie 4, which it calls its most advanced world model for generating interactive 3D worlds, previously available only as a research preview, according to The Decoder. Users can modify the world in real time using text commands for weather, characters, and plot, while the model generates accompanying sound. The Project Genie update also includes sessions of up to three minutes, 360-degree panoramas, linked worlds, an agent that constructs worlds from spoken descriptions, and a gallery to share and remix creations. Genie 4 is available to Google AI Ultra subscribers, while AI Pro users can join a waitlist; all content carries the SynthID watermark.

Medium importanceNov. 18, 2026

Will Wright, creator of SimCity and The Sims, unveils MyMems, an AI-powered social simulator that turns players' memories into a toy world

Will Wright is working on MyMems, an AI-powered social simulator where players add memories, for example by telling stories out loud, which become visual elements on a personal map, according to an interview with The Verge. "We want to take your autobiography and turn it into a toy world," he said. The game uses the Behavior Engine from Artificial Agency and is scheduled for release in 2027. Wright acknowledges criticism of AI but calls it an incredible tool. Ars Technica also spoke with him about the project.

Brazil

3 stories

High importanceNov. 18, 2026

Update: Government establishes task force to assess impact of US license requirement for AI chip sales to Brazil

The government has formed a task force, led by Ministers Márcio Elias Rosa (Development, Industry, Trade and Services) and Esther Dweck (Management and Innovation), to assess the economic impact of the US rule requiring a license for selling advanced AI chips to Brazil, according to TV Sim Brasil. In a statement, the government clarified that it is not a ban but a licensing requirement, and that it is too early to evaluate its effect on imports. The rule from the BIS, under the US Department of Commerce, is expected to be published in the Federal Register on 11/19. Data centers already operating in the country have 90 days to obtain approval or stop serving clients linked to countries under embargo, and AWS, Azure and Google Cloud are exempt from individual licenses if they submit monthly reports. The government assesses that the issue could cause friction in G20 negotiations in December.

Medium importanceNov. 18, 2026

Serpro to launch AI agent factory in first half of 2027, following Serpro CODE

Serpro will have an AI agent factory as early as the first half of 2027, said Carlos Rodrigo Lima, superintendent of Artificial Intelligence at the state-owned enterprise, at the IA.Gov Summit in Brasília on Tuesday (11/17), according to Convergência Digital. The agents will be part of the IA Brasil platform and will feature cost and security controls. First comes Serpro CODE, a code generation tool that has entered phase 2, is being tested by 250 developers at the state-owned company, and will be available in the first quarter of 2027; the next step is to roll it out to all of Serpro's staff and sell it on the IA Brasil platform.

Medium importanceNov. 18, 2026

Anatel allocates R$ 375 million from the 5G auction for Telebras' AI data center, but requires disclosure of who will cover operational costs

Anatel formalized this Wednesday (11/18) the allocation of R$ 375 million to build an artificial intelligence data center that will be operated by Telebras, according to Convergência Digital. Ruling No. 325/2026 allows using balances from the 5G auction to implement the project, but not to fund its operation, and requires that the source of these funds be identified before any contracting. The 3.5 GHz Band Administrator Entity (EAF), which will carry out the implementation, must transfer the infrastructure to the federal government and cannot fund its operation; energy, maintenance, upgrades, security, and support will be handled by Telebras. The state-owned company accumulated a loss of R$ 90.7 million in the first nine months of 2026 and received R$ 169.9 million in government subsidies.

About this edition

Edition
Edition no. 38 · Evening edition
Date
Wednesday, November 18, 2026
Coverage
Covers what has been published since edition no. 37 (closed around 8 a.m. Brasília time on 11/18) until the early evening of 11/18, a day of waiting for Anthropic's IPO pricing, which debuts on the Nasdaq this Thursday (11/19), and security alerts regarding the use of AI to plan attacks and steal credentials.
Sources scanned
Automated scan of 701 sources with active feeds (687 responded, 2,807 items) conducted on the Mac mini's local AI, which grouped, filtered, and compiled fact sheets with details verified against each page's text; followed by editorial review cross-referenced with the sources and previous editions. Among the sources for this edition: CNBC, Bloomberg, The Information, BBC, IEEE Spectrum, Wired, The Verge, 9to5Google, BleepingComputer, Convergência Digital and official blogs (Google DeepMind, Google Cloud, NVIDIA, AWS, Amazon Science, Cloudflare) and SEC filings.
Criteria
Only news sourced from real, preferably primary, outlets with all links verified are included. Pages returning a 403 blocking error were retained; broken links (404) were discarded.