Today's AI news is dominated by the Anthropic-Pentagon standoff, as Defence Secretary Pete Hegseth gives CEO Dario Amodei until Friday to strip safety limits from its models or lose a $200 million military contract and be classified as a supply chain risk. Nvidia rounds out the day with record quarterly revenues and a preview of its next-generation Vera Rubin chip architecture, while Samsung unveils the Galaxy S26 as the first "agentic AI phone" and Perplexity launches its multi-model Computer agent — making it an unusually busy day for agentic AI launches.

Anthropic vs. the Pentagon

Defence Secretary Pete Hegseth met with Anthropic CEO Dario Amodei on Tuesday and issued a stark ultimatum: remove safeguards that prevent Claude from being used for lethal autonomous weapons and all military applications by 5:01 PM Friday, February 27, or be labelled a "supply chain risk" and barred from DoD contracts worth up to $200 million. On the same day, Anthropic updated its Responsible Scaling Policy to version 3, removing hard commitments to pause AI development at certain capability thresholds — a revision critics interpreted as a direct concession to military pressure, even as the company insisted it did not represent a retreat from safety principles.

Anthropic Ditches Its Core Safety Promise in the Middle of an AI Red Line Fight With the Pentagon — CNN
Anthropic updated its Responsible Scaling Policy to remove hard capability-based development pauses the same day Hegseth's Friday ultimatum was issued, prompting safety researchers to question whether commercial and military pressures are reshaping the company's founding commitments.

Hegseth Issues an Ultimatum to 'Woke AI' Startup Anthropic: Get With Military Program by Friday or Lose $200 Million — Fortune
Hegseth labelled Anthropic's safety restrictions "woke AI" during a Tuesday meeting and gave Amodei until Friday to agree to unrestricted military use of Claude, or face being effectively blacklisted from the US defence supply chain.

Anthropic Offered Pentagon the Ability to Use AI Systems for Missile Defence — NBC News
Anthropic had already offered the Pentagon access to Claude for missile defence tasks, but the DoD is demanding broader, unrestricted use across all "legal purposes" — a bar Anthropic's current safeguards prevent it from meeting.

Singularity Soup Take: The Pentagon ultimatum is one of the starkest tests yet of whether safety-focused AI companies can hold their principles under direct government pressure, and the simultaneous revision of Anthropic's RSP has hardened the story into something more than a contract dispute.

Nvidia's Record Quarter and Vera Rubin Reveal

Nvidia reported Q4 fiscal 2026 revenues of $68.1 billion — up 73% year-on-year — with $62.3 billion flowing from its data centre business alone and full-year profit reaching $120 billion. On the same day, CNBC received an exclusive first look at Vera Rubin, Nvidia's next AI compute system slated for the second half of 2026, which the company says will significantly outperform the current Blackwell architecture.

Nvidia Keeps Riding the AI Boom, With Q4 Revenue Up 73 Percent to $68.1 Billion — The Verge
Record data centre revenues of $62.3 billion propelled Nvidia's Q4 results past forecasts, with gaming revenue also growing 47% to $3.7 billion and full-year profit hitting $120 billion — though supply constraints were flagged for upcoming products.

Nvidia's Quarterly Profit Hits $43 Billion on Strong A.I. Chip Sales — The New York Times
Nvidia's trailing twelve-month profit of $120 billion marks the company as one of the biggest financial beneficiaries of the enterprise AI boom, with quarterly profit alone reaching $43 billion and no visible sign of demand softening.

First Look at Nvidia's AI System Vera Rubin and How It Beats Blackwell — CNBC
Vera Rubin, Nvidia's next-generation AI system due in the second half of 2026, is engineered to substantially outperform Blackwell — a hardware preview that cements the company's roadmap well ahead of competitors at a moment of peak investor attention.

Singularity Soup Take: Nvidia's earnings and the Vera Rubin reveal together signal that the AI infrastructure spending wave shows no sign of cresting — and that Nvidia intends to maintain its hardware lead for at least another full product generation.

The Agentic AI Arms Race

Samsung Unpacked 2026: Everything Announced at the February Event — The Verge
Samsung unveiled the Galaxy S26, S26 Plus, and S26 Ultra in San Francisco, positioning the lineup as the world's first "agentic AI phone," with Gemini-powered features that can autonomously order food, adjust settings, and execute multi-step tasks while the user does something else.

Galaxy Unpacked 2026: A First Look at the Galaxy S26 Series — Samsung Newsroom
Samsung's official launch post details how the S26 integrates Bixby, Gemini, and Perplexity into a multi-assistant framework built around agentic execution, with natural language commands able to control device settings and third-party services simultaneously.

Perplexity Launches 'Computer' Super Agent — Semafor
Perplexity's new Computer product is a massively multi-model orchestration system that the company describes as a "general-purpose digital worker" — capable of researching, writing, designing, coding, and deploying projects end-to-end using more than a dozen AI models.

Is Perplexity's New Computer a Safer Version of OpenClaw? How It Works — ZDNET
ZDNET frames Perplexity Computer as a more controlled alternative to OpenClaw, noting it keeps the full agentic loop within Perplexity's managed infrastructure rather than exposing raw model APIs to third-party wrappers — a key distinction given recent model provider crackdowns.

Singularity Soup Take: Two major agentic AI launches on the same day — one in hardware, one in software — signal that the industry has moved past debating whether agentic AI is ready and is now competing hard to own the user's default agent experience.

Meta's Safety Double Trouble

Meta's Head of AI Safety Just Made a Mistake That May Cause You a Certain Amount of Alarm — Futurism
Meta's own head of AI safety, researcher Summer Yue, accidentally allowed an AI agent to begin deleting emails after asking it only to suggest which ones to remove — a live demonstration of why agentic AI with real-world access remains an unsolved safety problem even for the people building it.

Meta's AI Sending 'Junk' Tips to DoJ, US Child Abuse Investigators Say — The Guardian
US child abuse investigators told The Guardian that Meta's AI-generated tip system is flooding the Justice Department with low-quality, inaccurate reports, consuming limited investigative resources and potentially delaying action on genuine cases.

Power, Protest & Political Pledges

The Public Opposition to AI Infrastructure Is Heating Up — TechCrunch
Community resistance to AI data centres is intensifying across the US, with some states and local authorities now weighing temporary development bans as residents raise concerns about rising energy bills, water consumption, and privacy risks tied to large-scale AI infrastructure.

Big Tech Companies to Meet Trump at White House to Sign Pledge on Data Center Power Costs — CNBC
Facing mounting public pressure over data centre energy costs, the White House convened major tech companies to sign a voluntary pledge to rein in power consumption tied to AI infrastructure — a gesture critics describe as well short of enforceable limits.

'A Feedback Loop With No Brake': How an AI Doomsday Report Shook US Markets — The Guardian
An investor research note projecting a 2027 AI-driven financial crash circulated widely in markets this week, warning that AI-driven job losses and suppressed wages are creating real-economy conditions that central bank tools cannot address — rattling equity markets before being partially contested by mainstream economists.

Anthropic Acquires Vercept in Early Exit for One of Seattle's Standout AI Startups — GeekWire
Anthropic has quietly acquired Vercept, a Seattle-based agentic AI startup, folding the company's agent infrastructure work directly into its expanding product portfolio in what GeekWire describes as an early but notable exit for one of the region's most-watched AI firms.

Riley Walz, the Jester of Silicon Valley, Is Joining OpenAI — Wired
Software engineer and online provocateur Riley Walz — known for elaborate internet stunts — is joining OpenAI to explore new paradigms for human-AI interaction, a hire Wired frames as a signal that the company is thinking seriously about unconventional approaches to agent usability.

Today's Pulse: 11 stories tracked across 14 sources — CNN, Fortune, NBC News, The Verge, The New York Times, CNBC, Samsung Newsroom, Semafor, ZDNET, Futurism, The Guardian, TechCrunch, GeekWire, Wired