Tech Digest – February 28, 2026
AI & State Power
Pentagon Designates Anthropic a National Security Risk — OpenAI Steps In Hours Later
The U.S. Secretary of War designated Anthropic a “supply-chain risk to national security” after the company refused to grant the military unrestricted access to its Claude models. The dispute centered on two red lines: mass domestic surveillance and fully autonomous weapons. Anthropic stated that “no amount of intimidation or punishment” would change its position. Hours later, OpenAI announced it had reached its own agreement to deploy on the Pentagon’s classified network, claiming the same red lines — though an Under Secretary noted the contract still flows from “all lawful use.” Over 680 employees at Google and OpenAI signed an open letter titled “We Will Not Be Divided,” demanding refusal of mass surveillance and autonomous killing applications. Anthropic’s $200 million Pentagon contract has been cancelled, with a six-month transition period. The supply-chain risk designation — typically reserved for foreign adversaries like Huawei — would bar any Pentagon contractor from using Claude.
Note: This is the first time the U.S. government has used a supply-chain risk designation against an American technology company. Any European institution using or evaluating Claude needs to watch this closely — not for the politics, but for the procurement precedent. If a major AI provider can be cut off from an entire government ecosystem overnight, vendor diversification just moved from “best practice” to “operational necessity.”
Sources: Axios, TechCrunch, Fortune, Anthropic, We Will Not Be Divided (open letter)
20,000 AI-Generated Comments Killed a California Climate Rule — And Nobody Caught It in Time
Southern California’s top air quality authority rejected a two-year effort to phase out gas-powered appliances after receiving over 20,000 opposition comments generated through CiviClick, an AI-powered advocacy platform. A public affairs consultant took credit for the campaign. When agency staff contacted a sample of purported commenters, at least three said they had never written to the agency. The board rejected the proposal 7-5. Researchers at the University of Pittsburgh described it as the next step in “digital astroturfing.”
Note: Any institution that accepts public comments — on zoning, budgets, procurement, environmental policy — now faces this. The cost of generating thousands of plausible, individualized comments is approaching zero. Detection methods haven’t caught up. If your public consultation process doesn’t have verification built in, it may already be compromised.
Sources: Phys.org / LA Times, Governing
California and Colorado Now Require OS-Level Age Verification — Including Linux
California passed a law requiring all operating systems — including Linux distributions — to collect birth dates at account setup. Colorado is advancing similar legislation. The requirements apply at the OS level, not the application level, meaning every device sold or deployed in these states will need built-in identity collection at first use.
Note: This creates a compliance burden for any institution deploying or procuring devices in these jurisdictions. It also signals a regulatory direction: age verification is moving from platforms to infrastructure. For EU institutions watching U.S. digital regulation for early signals, this is one.
Infrastructure & Capital
OpenAI Raises $110 Billion at $730 Billion Valuation — Amazon Leads with $50 Billion
OpenAI closed the largest private funding round in history: $110 billion at a $730 billion pre-money valuation. Amazon invested $50 billion, Nvidia $30 billion, and SoftBank $30 billion. As part of the Amazon deal, OpenAI committed to consuming at least 2 GW of AWS Trainium compute and will spend an additional $100 billion on AWS over eight years. OpenAI also reported 900 million weekly ChatGPT users and said its Codex programming tool tripled its user base to 1.6 million. The company is now targeting roughly $600 billion in total compute spend by 2030.
Note: To calibrate: total U.S. venture capital invested in all startups across all of 2023 was $170 billion. OpenAI just raised two-thirds of that in a single transaction. The circular nature of these deals — Nvidia and Amazon invest in OpenAI, OpenAI spends the money buying Nvidia chips and AWS services — means the real question isn’t whether these numbers are sustainable. It’s how deeply institutional AI tooling will be locked into the infrastructure these deals create.
Sources: Bloomberg, CNBC, OpenAI
Nvidia to Unveil Dedicated Inference Chip with Groq Architecture at GTC
Nvidia is preparing to launch a dedicated inference processor at its GTC conference on March 16, integrating technology from Groq’s Language Processing Units. The chip addresses a growing gap: Nvidia dominates AI training, but competitors like Amazon’s Inferentia and Google’s TPUs have been gaining ground on inference — the process of actually running AI models in production. Nvidia paid $20 billion to license Groq’s architecture and hired its founder. OpenAI has committed 3 GW of dedicated inference capacity on the new platform, reportedly driven by dissatisfaction with GPU inference speeds for its Codex programming tool.
Note: The training vs. inference split matters for procurement. Training is a one-time cost; inference is the ongoing operational expense every time someone uses an AI tool. Dedicated inference hardware that cuts energy consumption and latency changes the total cost of ownership for any AI deployment — including the institutional tools arriving in the next 12-18 months.
Sources: WSJ, SiliconANGLE
Bezos’s Project Prometheus Seeks Tens of Billions to Remake Manufacturing with AI
Jeff Bezos’s AI venture Project Prometheus, which launched with $6.2 billion at a $30 billion valuation, is now raising “tens of billions” more to acquire and transform manufacturing businesses with AI. The company has hired over 120 researchers from OpenAI, DeepMind, and Meta, and is building what insiders describe as a “manufacturing transformation vehicle.” Talks with ADIA and JPMorgan are underway. Bezos serves as co-CEO — his first operational role since leaving Amazon in 2021.
Sources: Financial Times, PYMNTS
Workforce & Adoption
Agent Users Now Outnumber Autocomplete Users 2-to-1 at Cursor — While AI Agents Start Taking Exams
Cursor co-founder Michael Truell reported that agent users on the platform now outnumber autocomplete users 2-to-1, inverting last year’s ratio. The shift signals a move from AI-assisted work to AI-driven work in software development. Meanwhile, 404 Media reported on an agent called “Einstein” that attends university lectures, writes papers, and takes exams on a student’s behalf. On the other end, small U.S. law firms are branding themselves “Claude-Native,” claiming the general-purpose model outperforms every specialized legal AI tool. Anonymous reports from FAANG companies describe unscheduled all-hands meetings announcing 25% workforce reductions tied to AI investments — though these remain unverified single-source claims.
Note: The Cursor ratio inversion is the clearest data point yet on how AI coding tools are being used in practice. Last year, AI suggested the next line; this year, it writes the feature. The speed at which this shifts determines how fast institutional software procurement changes — and how fast the skills required of in-house IT teams evolve.
Sources: Michael Truell (Cursor), 404 Media, Zack Shapiro (Claude-Native law firms)
Regulatory Acceleration
FDA Approves Lung Cancer Drug 44 Days After Filing
The FDA approved Hernexeos (zongertinib) for first-line treatment of HER2-mutant lung cancer just 44 days after the application was filed, under its new National Priority Voucher pilot program. The drug showed a 76% tumor response rate in clinical trials — compared to 30-45% for standard care. The program aims to compress approval timelines from 10-12 months down to one or two. This is only the second approval issued under the pilot.
Note: Forty-four days. The standard timeline is 10-12 months. This isn’t a one-off — it’s a pilot program designed to be repeated. When regulatory processes that took a year compress to six weeks, every institution’s assumption about “how long things take” needs updating. The question for public sector planners: which of your own approval processes could be restructured this aggressively?
Sources: FDA, Fierce Pharma
Data & Privacy
Bright Data Offers SDK That Turns Smart TVs Into Web-Scraping Proxy Nodes
Bright Data, a web data company, is offering an SDK that turns internet-connected smart TVs into web-crawling proxy nodes for AI training data collection. The product monetizes idle screen time by routing data collection requests through consumer devices — often without meaningful user awareness. The Verge described it as evidence that in the race for AI training data, every connected device is becoming a potential compute resource.
Note: This is what the appetite for training data looks like when it reaches consumer hardware. Any institution managing networks, procurement specifications, or data governance policies should understand that the devices in your buildings may already be participating in data infrastructure you didn’t consent to.
Sources: The Verge
