ALETH / AI-BRIEF / 2026-02-21 / PENTAGON THREATENS ANTHROPIC WITH BLACKLISTThe Aleth Briefs link to the sources behind the stories and show how the week unfolded.
The week in five lines:
The Pentagon threatened to blacklist Anthropic over limits on military use.
Google’s Gemini 3.1 Pro led Artificial Analysis’s index and doubled on ARC-AGI-2.
OpenAI neared a round that could top $100bn at a valuation above $850bn.
Nvidia reportedly planned a $30bn OpenAI stake with no deployment milestones.
Two AWS outages, one lasting 13 hours, were tied to Amazon’s AI coding tools.
Browse by day:
Weekend: Doubao 2.0 · Paramount · OpenClaw · UK chatbot law
Monday: Qwen3.5 · Pentagon threat · Irish DPC v X
Tuesday: Sonnet 4.6 · Meta-Nvidia · Adani · Moonshot
Wednesday: World Labs · Ineffable · Sarvam · OpenAI-TCS
Thursday: Gemini 3.1 Pro · Nvidia-OpenAI · US DHS-Palantir · Agent hit piece
Friday: AWS outages · Claude Code Security · ggml
Weekend (14-15 February)
ByteDance launched its Doubao-Seed-2.0 models for complex real-world tasks.
ByteDance says the models still trail top global models on some hard coding tests, but they match Western models on benchmarks at a fraction of the price. Seed2.0 Pro costs about $0.47/m input and $2.37/m output tokens through the API, against $5/m and $25/m for Claude Opus 4.5 (Anthropic). It powers ByteDance’s Doubao chat app.
Seed 2.0 official launch · ByteDance Seed
Bytedance’s Seed2.0 adds price pressure on Western AI models · The Decoder
Paramount Skydance sent ByteDance a cease-and-desist letter accusing Seedance and Seedream of “blatant infringement” of its franchises.
ByteDance said it is “taking steps to strengthen current safeguards” against misuse by users. On Tuesday Warner Bros. followed with its own letter, demanding that ByteDance stop training on its characters. It said Seedance comes pre-loaded with them as “a deliberate design choice”.
Paramount sends ByteDance cease-and-desist letter over Seedance · Variety
Warner Bros. blasts ByteDance for AI videos of Superman and Batman · Variety
OpenClaw creator Peter Steinberger is joining OpenAI, and the agent project will move to a foundation.
Steinberger said OpenClaw will “stay open and independent”, with OpenAI already sponsoring it. Sam Altman said Steinberger will drive OpenAI’s next generation of personal agents, which he expects to “quickly become core to our product offerings”.
OpenClaw, OpenAI and the future · Peter Steinberger
Peter Steinberger is joining OpenAI · Sam Altman on X
The UK government will force all AI chatbot providers to meet the illegal-content duties in the Online Safety Act.
The government is tabling an amendment to the Crime and Policing Bill to bring chatbots that fall outside the Act into scope. It is also taking powers to act within months on a consultation, launching next month, that will consider a minimum age for social media. “The action we took on Grok sent a clear message that no platform gets a free pass,” Keir Starmer said.
PM: “No platform gets a free pass”: action to keep children safe online · GOV.UK
Monday 16 February
Alibaba released open weights for Qwen3.5-397B-A17B, a natively multimodal model that activates 17B of its 397B parameters.
It pairs linear attention with a sparse mixture of experts (MoE) and supports 201 languages and dialects, up from 119. Alibaba puts its decoding throughput at 8.6x that of Qwen3-Max (Alibaba) at 32k context and 19.0x at 256k. Artificial Analysis scored it 45 on its Intelligence Index, third among open-weights models behind GLM-5 (Zhipu AI) and Kimi K2.5 (Moonshot AI).
Qwen3.5: towards native multimodal agents · Alibaba Cloud
Qwen3.5-397B-A17B: everything you need to know · Artificial Analysis
The Pentagon threatened to label Anthropic a supply chain risk.
A senior Pentagon official told Axios that Defense Secretary Pete Hegseth is “close” to cutting ties with Anthropic. Anyone doing business with the US military would then have to drop Anthropic. Axios notes that the penalty is usually reserved for foreign adversaries. “We are going to make sure they pay a price for forcing our hand like this,” the official said. Anthropic is willing to loosen its terms but wants to keep its bans on mass spying on Americans and on weapons that fire with no human involvement.
Semafor reported that an Anthropic official discussed the Maduro raid with a Palantir executive at a routine check-in, and the executive reported the conversation to the Pentagon. Anthropic disputes that account.
Pentagon threatens to label Anthropic’s AI a “supply chain risk” · Axios
Palantir partnership is at heart of Anthropic, Pentagon rift · Semafor
Ireland’s Data Protection Commission (DPC) opened a large-scale GDPR inquiry into X over sexualised images made with Grok, including of children.
As X’s lead EU regulator, the DPC will examine whether X met its duties on lawful processing, data protection by design and impact assessments. It notified X on Monday. The Center for Countering Digital Hate (CCDH) estimated that over 11 days from 29 December, Grok made about 23k sexualised images of children.
Data Protection Commission opens investigation into X · DPC
Grok floods X with sexualized images of women and children · CCDH
Tuesday 17 February
Moonshot AI is seeking a $10bn valuation, about a month after raising $500m at $4.3bn, Bloomberg reported.
The Kimi developer is expanding a round already backed by Alibaba and Tencent. Existing backers have committed >$700m.
China AI startup Moonshot targets $10 billion valuation · Bloomberg Law
Adani plans about $100bn of green AI data centres in India by 2035.
It aims to grow its data centre capacity from 2 GW to 5 GW and is putting $55bn into renewables. It expects the plan to draw $150bn more into related industries.
Adani commits $100bn to sovereign AI infrastructure · Adani Group
Anthropic released Claude Sonnet 4.6 with a 1M token context window in beta, at Sonnet 4.5’s price of $3/m input and $15/m output tokens.
It becomes the default for Free and Pro users. Claude Code testers preferred it to Sonnet 4.5 about 70% of the time and to Opus 4.5 59% of the time. Artificial Analysis put it top of GDPval-AA with an Elo of 1633, ahead of Claude Opus 4.6. In a business simulation, Anthropic’s system card found it “comparably aggressive” to Opus 4.6, “including lying to suppliers and initiating price-fixing in some cases”.
Introducing Sonnet 4.6 · Anthropic
Claude Sonnet 4.6: new leader in GDPval-AA · Artificial Analysis
Claude Sonnet 4.6 system card · Anthropic
Nvidia and Meta signed a multiyear partnership covering millions of Blackwell and Rubin GPUs and Nvidia CPUs.
Nvidia called it the first large-scale deployment of its Grace CPUs on their own, and Vera CPUs could follow at scale in 2027. Meta is also adopting Spectrum-X Ethernet, and Nvidia’s confidential computing for WhatsApp.
Meta builds AI infrastructure with NVIDIA · NVIDIA
Wednesday 18 February
Former DeepMind researcher David Silver is raising about $1bn in a Sequoia-led seed round for London start-up Ineffable Intelligence, the FT reported.
The round would value the company at about $4bn pre-money, and Nvidia, Google and Microsoft are also in talks to invest. PitchBook says it would be the largest seed round ever for a European start-up. Silver worked on AlphaGo and AlphaStar at DeepMind and helped develop Gemini. Ineffable will build on his reinforcement learning work rather than training mainly on internet text. In a paper on the “Era of Experience” last year, Silver and Richard Sutton predicted that experience will “ultimately dwarf the scale of human data used in today’s systems”.
British scientist raising $1bn for new AI lab in Europe’s biggest seed round · FT
Fei-Fei Li’s World Labs raised $1bn from AMD, Nvidia, Autodesk and others.
Autodesk put in $200m and will advise World Labs, working with it at the research and model level. World Labs’ first product, Marble, builds persistent 3D worlds from images, video or text.
World Labs announces new funding · World Labs
Autodesk invests $200 million in World Labs, secures advisor role · Autodesk
Indian start-up Sarvam AI launched two large language models (LLMs) at the India AI Impact Summit, and said both will be released as open source.
Sarvam-30B was pre-trained on 16T tokens and has a 32,000-token context window. Sarvam-105B handles 128,000 tokens. On Friday Sarvam released its Indus chat app in beta on iOS, Android and the web.
Sarvam AI launches 30B and 105B open-source models · Business Today
India’s Sarvam launches Indus AI chat app · TechCrunch
OpenAI will be the first customer of Tata Consultancy Services’ (TCS) HyperVault data centre business, starting with 100 MW and possibly scaling to 1 GW.
OpenAI announced the agreement as part of OpenAI for India, a nationwide initiative. India has >100m weekly ChatGPT users. Tata Group plans to roll out ChatGPT Enterprise to its staff, starting with hundreds of thousands of TCS employees.
Introducing OpenAI for India · OpenAI
Thursday 19 February
Google released Gemini 3.1 Pro, taking the lead on Artificial Analysis’s index at less than half the cost of Claude Opus 4.6.
Gemini 3.1 Pro scored a verified 77.1% on ARC-AGI-2, more than double Gemini 3 Pro’s score. It is rolling out in the Gemini app with higher limits for Google AI Pro and Ultra subscribers, and to developers through the Gemini API and Vertex AI. Artificial Analysis put Gemini 3.1 Pro Preview 4 points ahead of Claude Opus 4.6 (Anthropic) on its Intelligence Index, at less than half the cost to run. It leads 6 of the index’s 10 evaluations, and its AA-Omniscience hallucination rate fell from 88% to 50%. On GDPval-AA it still trails Sonnet 4.6, Opus 4.6, GPT-5.2 (xhigh) and GLM-5 (Zhipu AI).
Gemini 3.1 Pro: a smarter model for your most complex tasks · Google
Gemini 3.1 Pro Preview: the new leader in AI · Artificial Analysis
Wired reported the US Department of Homeland Security (DHS) had signed a five-year purchase agreement worth up to $1bn with Palantir the week before.
The agreement lets DHS agencies, including Immigration and Customs Enforcement (ICE) and Customs and Border Protection (CBP), buy Palantir software without separate competitive bidding.
DHS opens a billion-dollar tab with Palantir · Wired
Nvidia plans a $30bn stake in OpenAI’s new round, the FT reported.
The FT said the investment would replace Nvidia’s September 2025 pledge to invest up to $100bn as its systems came online in OpenAI’s data centres. CNBC reported the $30bn as separate from that agreement. Nvidia’s CFO had said in December that the earlier deal had not passed the letter of intent stage. Bloomberg reported the round could top $100bn, at a valuation above $850bn including the new money.
Nvidia and OpenAI abandon $100bn deal in favour of $30bn investment · FT
OpenAI funding on track to top $100 billion in latest round · Bloomberg Law
Nvidia is in talks to invest up to $30 billion in OpenAI, source says · CNBC
The operator of the AI agent that published a hit piece on matplotlib maintainer Scott Shambaugh came forward anonymously, calling it a social experiment.
The operator ran the agent on OpenClaw in a sandboxed virtual machine and said they had not reviewed its blog post before it went up. Shambaugh wrote that the agent’s GitHub account is no longer active.
An AI agent published a hit piece on me: the operator came forward · The Shamblog
Friday 20 February
The FT linked at least two AWS outages to Amazon AI coding tools.
In mid-December, engineers let Amazon’s Kiro agent make changes, and it decided to “delete and recreate the environment”, causing a 13-hour interruption to a customer-facing system. A second incident involved Amazon Q Developer. Amazon blamed “user error - specifically misconfigured access controls - not AI”. It said the December incident hit only AWS Cost Explorer in one of its 39 regions.
Amazon service was taken down by AI coding bot · FT
Correcting the Financial Times report about AWS, Kiro, and AI · Amazon
Anthropic launched Claude Code Security, which scans codebases for vulnerabilities and suggests patches for human review.
It opens as a limited research preview for Enterprise and Team customers, and open-source maintainers can apply for free, expedited access. Using Claude Opus 4.6, Anthropic’s team found >500 vulnerabilities in production open-source code.
Making frontier cybersecurity capabilities available to defenders · Anthropic
The founding team of llama.cpp, ggml.ai, is joining Hugging Face.
The ggml team will keep leading and maintaining ggml and llama.cpp full-time, and the projects stay open and community driven. The team will also work on integration with Hugging Face’s transformers library. llama.cpp is the most widely used software for running AI models locally on a laptop or phone.
ggml.ai joins Hugging Face to ensure the long-term progress of local AI · GitHub
ALETH · OPEN ACCESS RESEARCH