The AI Dispatch — March 1, 2026

Federal Response

Anthropic Refuses Pentagon Demand to Drop AI Safeguards; Administration Blacklists Company

Defense Secretary labels Anthropic a “supply chain risk” after company declines to enable autonomous weapons and mass surveillance. Claude surges to No. 1 in App Store.

Sources: CNN Business, CBS News, TechCrunch, The Hill

The Trump administration on Thursday ordered all federal agencies to immediately cease business with Anthropic, the San Francisco-based artificial intelligence company, after the firm refused a Pentagon directive to remove safety restrictions that prohibited the use of its Claude model for autonomous weapons targeting and warrantless mass surveillance of American citizens. The executive action, which took effect at midnight Eastern time, marks the first time a sitting administration has attempted to blacklist a major domestic technology company over its refusal to comply with military demands.

The confrontation had been building for weeks. According to documents reviewed by CNN, the Department of Defense in early February sent Anthropic a formal letter demanding that the company modify Claude’s acceptable use policy to permit “unrestricted deployment in national security contexts,” including autonomous lethal targeting systems and bulk communications monitoring. Anthropic’s leadership, led by Chief Executive Dario Amodei, declined in a written response that cited both constitutional concerns and the company’s long-standing position that current AI systems lack the reliability required for life-or-death autonomous decisions.

Defense Secretary Pete Hegseth escalated the dispute by publicly designating Anthropic a “supply chain risk” to national security — a classification normally reserved for foreign adversaries and entities like Huawei or Kaspersky Lab. The designation empowers the administration to bar federal contractors from using Anthropic products and could affect the company’s eligibility for government-adjacent work across the intelligence community. Anthropic said late Friday that it would challenge the designation in federal court, calling it “an unprecedented and unconstitutional attempt to coerce a private company into abandoning safety principles.”

The public backlash to the administration’s move was swift and commercially significant. Within 48 hours of the blacklist announcement, Anthropic’s Claude chatbot rocketed to the No. 1 position on Apple’s App Store, surpassing both ChatGPT and TikTok, with the company reporting record-breaking signups. Technology executives, constitutional scholars, and civil liberties organizations issued a flurry of statements in support of Anthropic, with several warning that the government’s actions set a dangerous precedent for the regulation of private AI development.

Venture Capital

OpenAI Closes $110 Billion Round, Valuation Hits $840 Billion

Amazon leads $50B tranche; SoftBank and Nvidia each contribute $30B in one of the largest private capital raises in history.

Sources: Mercury News, Advisor Perspectives

OpenAI announced Saturday the close of a $110 billion funding round that values the company at approximately $840 billion, cementing its position as the most highly valued private technology firm in history. Amazon Web Services led the raise with a $50 billion commitment, while SoftBank’s Vision Fund and Nvidia each contributed $30 billion. The round dwarfs the company’s previous $6.6 billion raise from October 2024 and underscores the extraordinary pace at which capital continues to pour into foundation model development.

As part of the agreement, Amazon Web Services will become OpenAI’s exclusive cloud infrastructure distributor for enterprise customers, a deal that expands the existing partnership by an estimated $100 billion in compute commitments over the next eight years. The arrangement gives Amazon a direct pipeline to OpenAI’s models while providing Sam Altman’s company with the vast GPU clusters it needs to train next-generation systems. Industry analysts noted that the round’s sheer scale reflects not just investor confidence in OpenAI, but a broader conviction that the race for artificial general intelligence will require capital expenditures previously associated only with nation-state infrastructure projects.

The funding comes at a pivotal moment for OpenAI, which completed its conversion from a nonprofit to a for-profit benefit corporation in January. Critics, including several former board members, have questioned whether the new corporate structure provides adequate safeguards against the pursuit of profit at the expense of safety. OpenAI has maintained that the transition was necessary to attract the capital required to compete with well-funded rivals across the United States and China.

Alignment Research

Anthropic Paper: AI Failures Grow Incoherent, Not Systematically Misaligned

A major new paper from Anthropic’s alignment science team presents what the authors call the “hot mess” theory of AI failure. Drawing on extensive empirical analysis of frontier model behavior across thousands of tasks of escalating difficulty, the researchers found that as problems become harder, model errors are overwhelmingly dominated by incoherence — essentially random, unpredictable variance — rather than by coherent, systematically misaligned behavior. In statistical terms, the variance term swamps the bias term as task complexity increases.

The findings carry significant implications for the AI safety community, which has devoted considerable attention to the threat of “scheming” — the hypothesis that sufficiently capable models might deliberately pursue hidden goals. While the paper does not dismiss that possibility entirely, it argues that the empirical evidence points to a more prosaic and in some ways more challenging problem: that advanced AI systems are more likely to fail in messy, unpredictable ways than to execute coherent adversarial strategies. The authors suggest the safety field should rebalance its focus from defending against deceptive superintelligences toward building robust systems that fail gracefully under uncertainty.

Read the full paper

Benchmarks

Gemini 3.1 Pro Posts Record 77.1% on ARC-AGI-2 Benchmark

Source: TechCrunch

Google DeepMind’s Gemini 3.1 Pro has shattered the previous ceiling on the ARC-AGI-2 benchmark, a test of novel abstract reasoning widely regarded as one of the most rigorous measures of general intelligence in AI systems. The model scored 77.1 percent, a staggering 46-percentage-point leap from the 31.1 percent achieved by its predecessor just months earlier — the largest single-generation gain in reasoning performance ever recorded on the benchmark.

The jump is all the more striking given the benchmark’s design: ARC-AGI-2 tasks require solvers to identify patterns in grids they have never seen before, a challenge that has historically resisted the kind of pattern-matching at which large language models excel. Gemini 3.1 Pro also posted a 94.3 percent score on GPQA Diamond, a graduate-level science reasoning test, placing it at or above the estimated average of domain experts. While benchmark performance does not directly translate to real-world capability, researchers at DeepMind said the results indicate that current architectures may have considerably more headroom than critics had predicted.

Mergers

SpaceX Acquires xAI in $1.25 Trillion Deal for Orbital AI Data Centers

Source: CNBC

SpaceX has agreed to acquire Elon Musk’s artificial intelligence venture xAI in a transaction valued at $1.25 trillion, making it the largest corporate merger in history and signaling a dramatic convergence of the space and AI industries. Under the terms of the all-stock deal, xAI’s Grok model family, its Memphis supercomputer cluster, and its 3,400 employees will be folded into a new SpaceX division dedicated to building orbital data centers powered by solar energy.

Musk, who controls both companies, said the merger reflects his conviction that Earth-based infrastructure cannot keep pace with the energy demands of next-generation AI training. The orbital data center concept envisions modular computing platforms deployed in low Earth orbit, drawing power from continuous solar exposure and cooled by the vacuum of space. Regulatory analysts noted that the transaction raises complex antitrust questions, as it effectively consolidates the world’s dominant private launch provider with one of the largest AI model developers under a single corporate umbrella. The Federal Trade Commission is expected to open a formal review.

Open Source

Alibaba’s Qwen3.5 Brings 397B Open-Weight Multimodal Model to 201 Languages

Source: SiliconAngle

Alibaba Cloud has released Qwen3.5, an open-weight multimodal model boasting 397 billion parameters and support for 201 languages, making it the most linguistically diverse frontier model ever published. The model, which handles text, images, and video inputs, is built on a mixture-of-experts architecture that activates only 17 billion parameters per inference pass, dramatically reducing the compute cost of running a model of its scale.

The release intensifies the open-weight AI race that has seen Chinese technology companies emerge as formidable competitors to their American counterparts. Qwen3.5’s language coverage extends well beyond the major world languages typically prioritized by Western labs, encompassing dozens of languages spoken in Southeast Asia, Central Asia, and Sub-Saharan Africa. Alibaba is distributing the model under a permissive Apache 2.0 license, inviting commercial and academic use without restriction. Researchers at several leading universities said early evaluations show Qwen3.5 performing competitively with proprietary Western models on standard multimodal benchmarks, particularly in multilingual reasoning tasks where its language breadth gives it a structural advantage.

Workforce

Block Cuts 4,000 Jobs as Dorsey Predicts Industry-Wide AI Displacement

Source: CNN Business

Block, the financial technology company formerly known as Square, is eliminating roughly 4,000 positions, reducing its workforce from approximately 10,000 to fewer than 6,000 employees. Chief Executive Jack Dorsey said in a company-wide memo obtained by CNN that artificial intelligence systems had rendered many roles redundant and predicted that most technology companies would follow a similar trajectory within the next twelve months.

“This is not a restructuring. It is an acknowledgment that the nature of work in technology has fundamentally changed,” Dorsey wrote, adding that Block’s AI coding and customer service agents now handle tasks that previously required entire teams. The cuts span engineering, customer support, and middle management. While Dorsey’s prediction of industry-wide displacement within a year drew considerable attention, several labor economists pushed back sharply. “We’ve heard this forecast before with every major wave of automation,” said MIT economist Daron Acemoglu. “The displacement is real, but the timeline is almost certainly longer and more uneven than executives want to believe.”

In Brief

Around the Industry

Disney Invests $1B in OpenAI, Licenses 200+ Characters to Sora

Disney has become the first major content licensing partner for OpenAI’s Sora generative video platform, committing $1 billion and granting access to more than 200 characters spanning the Disney, Marvel, Pixar, and Star Wars franchises. The three-year agreement also brings ChatGPT to Disney employees for internal productivity use.

Source: OpenAI

vLLM 0.16.0 Ships With 30% Throughput Gain

The popular open-source inference engine has released version 0.16.0 with a 30 percent throughput improvement, async scheduling, a WebSocket realtime API, and a comprehensive XPU platform overhaul. The release includes 440 commits from 203 contributors.

Source: GitHub

NVIDIA Debuts Nemotron 3 Open Agentic Model Family

Nvidia has announced Nemotron 3, a family of open models purpose-built for multi-agent systems. The Nano variant is available now under an open license, with Super and Ultra tiers expected in the first half of 2026. The models employ a novel hybrid latent mixture-of-experts architecture optimized for agentic workflows.

Source: Nvidia Newsroom

MCP Ecosystem Expands: Cloudflare Edge, GitHub Repo Autonomy

The Model Context Protocol ecosystem continues to gain momentum. Cloudflare has launched an edge orchestration server, while GitHub’s MCP integration now enables agents to execute, test, and commit code autonomously within repositories. The MCP registry is approaching general availability.

Source: MCP

40% of Healthcare Workers Encounter Unauthorized ‘Shadow AI’ Tools

A new survey has found that 40 percent of healthcare workers report encountering unauthorized artificial intelligence tools in clinical settings, a phenomenon researchers are calling “shadow AI.” Industry leaders are labeling 2026 “the year of governance” as hospitals and health systems scramble to establish policies that can keep pace with clinician-adopted AI tools already in widespread use.

Source: Healthcare Dive

Culture & Society

Cultural Shift

‘Proof of Human’ Emerges as Value Signal in AI-Saturated World

Source: Communicate Online

Omnicom’s cultural intelligence unit Backslash has released its 2026 Edges report, arguing that audiences are actively and increasingly seeking evidence of human authorship in a world saturated by AI-generated content. The report identifies “proof of human” as one of the defining cultural tensions of the year, noting that consumers are developing what researchers describe as an intuitive skepticism toward content whose provenance is uncertain.

Several major brands have already begun to respond. Patagonia has been testing “100% human-made” badges on its marketing materials and reports 20 percent higher engagement on certified human-authored content compared to unlabeled equivalents. iHeartMedia has adopted a “guaranteed human” tagline for select programming. Apple TV’s critically acclaimed series “Pluribus” includes end credits that specifically state “This show was made by humans” — a declaration that has drawn both praise and parody online.

The trend reflects a broader cultural renegotiation of what it means for creative work to carry value. As generative AI becomes capable of producing competent prose, imagery, and video at near-zero marginal cost, the Backslash researchers argue, the scarcity and authenticity of demonstrably human labor is being reconceived as a premium attribute rather than a baseline assumption.

AI Cultural Convergence Study Finds Autonomous Systems Default to Generic Outputs

Researchers have found that when text-to-image and image-to-text AI systems are chained together autonomously — with outputs from one model feeding directly into another without human intervention — the resulting content converges rapidly on a narrow set of generic themes regardless of the diversity of the original prompts.

The study has significant implications for industries that rely on AI for creative production at scale, including advertising, entertainment, and journalism. Left unchecked, the researchers warn, autonomous AI pipelines may produce an increasingly homogeneous cultural output that erodes rather than enriches the diversity of human expression.

Source: The Conversation

Policy & Safety

Global Safety

International AI Safety Report Warns of Escalating Biological, Cyber Risks

Source: International AI Safety Report

A landmark international assessment led by Turing Award laureate Yoshua Bengio and contributed to by more than 100 experts from over 30 countries has concluded that frontier AI systems now exceed PhD-level scientific performance across multiple domains while simultaneously lowering the barriers to cyberattacks and biological weapons development. The report, which represents the most comprehensive multilateral evaluation of AI risk to date, calls for urgent societal resilience measures to address what the authors describe as a rapidly widening gap between capability and governance.

The assessment is notable both for the breadth of its authorship and the starkness of its conclusions. Unlike previous reports that have hedged on the timeline for catastrophic AI risks, the Bengio-led group argues that several categories of risk — particularly those involving the synthesis of biological agents and the automation of offensive cyber operations — have already crossed thresholds that demand immediate policy responses from governments worldwide.

U.S. Regulation

Trump AI Preemption Order Approaches March 11 Deadline

Source: Gibson Dunn

The Department of Commerce faces a March 11 deadline to publish its review of state-level artificial intelligence regulations that the Trump administration considers burdensome to innovation. The executive order, signed in January, also requires the Federal Trade Commission to issue a formal AI policy statement and has established a Department of Justice AI Litigation Task Force charged with challenging state laws deemed to impede federal AI priorities.

Legal analysts have noted, however, that the executive order’s practical reach may be more limited than its rhetoric suggests. An executive order cannot independently override state legislation without congressional action, and several states — including California, Colorado, and Illinois — have signaled they will defend their AI statutes in court if challenged. The coming weeks are expected to set the stage for a prolonged federal-state conflict over who holds ultimate regulatory authority over the development and deployment of artificial intelligence systems in the United States.