AI Weekly: 03/23/26
The Pentagon calls Anthropic an 'unacceptable risk,' Elon Musk unveils Terafab chip manufacturing plans, and Amazon's Trainium lands a $50B OpenAI deal
Good morning and welcome to this week’s edition of AI Weekly! The Anthropic vs. DOD saga escalated this week as the Pentagon called Anthropic’s safety “red lines” an “unacceptable risk to national security” and revealed it’s actively developing alternative LLMs to replace Claude.
In infrastructure news, Elon Musk unveiled “Terafab,” a chip manufacturing collaboration between Tesla and SpaceX, while Amazon gave an exclusive tour of its Trainium chip lab following a $50 billion deal with OpenAI. Meanwhile, Nvidia’s networking division quietly posted $11 billion in quarterly revenue, surpassing Cisco’s entire annual networking business.
In coding/devtools, Cursor admitted its new Composer 2 model was built on top of Chinese startup Moonshot AI’s Kimi, and Apple cracked down on vibe coding apps in the App Store.
More on this past week’s top AI headlines below!
- ZG
Here are the most important stories of the week:
AGENTS
An AI agent at Meta went rogue and exposed sensitive company and user data to unauthorized employees for two hours, prompting a “Sev 1” classification — the second-highest severity level for security incidents. Link.
The incident occurred when an engineer asked an AI agent to help analyze a technical question on an internal forum; the agent posted a response without permission and gave faulty advice that led to the data exposure.
Meta’s safety and alignment director Summer Yue described a separate incident where her own OpenClaw agent deleted her entire inbox despite being instructed to confirm actions first.
Despite the incidents, Meta remains bullish on agentic AI, having recently acquired Moltbook, a Reddit-like social platform for AI agents to communicate with each other.
Tools for Humanity launched AgentKit, a tool that lets commercial websites verify real humans are authorizing AI agents to make purchases, using World ID iris scans and Coinbase’s x402 blockchain protocol. Link.
AgentKit integrates Sam Altman’s World ID with the x402 protocol developed by Coinbase and Cloudflare, allowing AI agents to transact on behalf of verified humans without requiring human intervention at each step.
Chief Product Officer Tiago Sada compared the system to delegating “power of attorney” to an AI agent for purchasing decisions.
The tool addresses growing fraud concerns as major platforms including Amazon, Mastercard, and Google have already introduced automated buying capabilities.
IMAGE/VIDEO
AI inference platform Fal is raising $300–$350 million at an $8 billion valuation, nearly doubling from $4.5 billion just three months ago, as annualized revenue doubled to $400 million. Link.
Fal serves 3 million developers including Adobe, Canva, and Shopify, providing cloud infrastructure for running AI models that generate images, video, and audio.
The company was founded in 2021 by former Amazon and Coinbase engineers and previously raised $314 million across three rounds in 2024 alone.
The dual-tranche fundraising structure lets earlier investors get quick markups, though the approach has raised concerns about startup valuation bubbles in the AI inference space.
CODING/DEVTOOLS
Cursor admitted its new Composer 2 coding model was built on top of Moonshot AI’s open-source Kimi 2.5, a Chinese model backed by Alibaba, after an X user discovered Kimi’s model ID buried in Composer 2’s code. Link.
Cursor VP Lee Robinson acknowledged the open-source base but said only about one-quarter of the compute in the final model came from Kimi, with the rest from Cursor’s own training.
Co-founder Aman Sanger called it “a miss to not mention the Kimi base in our blog from the start,” noting they would be more transparent with future models.
The revelation is notable given the framing of AI as a U.S.–China “arms race,” and comes as Cursor reportedly exceeds $2 billion in annualized revenue following a $2.3 billion raise at a $29.3 billion valuation last fall.
Apple blocked vibe coding apps Replit and Vibecode from releasing updates on the App Store, citing rules that prohibit apps from running code that changes how they or other apps function. Link.
Apple’s crackdown targets apps that help users create web apps outside the App Store — a potential threat to Apple’s 15–30% commission structure on in-app purchases.
Since being blocked in January, Replit’s mobile app dropped from #1 to #3 on Apple’s free developer tools chart; the company is privately valued at $9 billion.
Other apps with similar features, including Vercel’s v0, Snap, and Canva, have continued releasing updates without reported issues, raising questions about inconsistent enforcement.
INFRASTRUCTURE
Elon Musk unveiled “Terafab,” a joint Tesla–SpaceX chip manufacturing facility to be built near Tesla’s Austin headquarters, saying semiconductor manufacturers aren’t making chips fast enough for his companies’ AI and robotics needs. Link.
Musk said the goal is to manufacture chips supporting 100 to 200 gigawatts of computing power per year on Earth, plus a terawatt in space.
“We either build the Terafab or we don’t have the chips, and we need the chips, so we build the Terafab,” Musk said at an event in downtown Austin.
Musk has no background in semiconductor manufacturing and provided no timeline, but has a track record of pursuing ambitious industrial projects despite skepticism.
Amazon gave an exclusive tour of its Trainium chip lab in Austin following a $50 billion deal with OpenAI for 2 gigawatts of Trainium computing capacity, with 1.4 million chips now deployed across three generations. Link.
Anthropic’s Claude runs on over 1 million Trainium2 chips, and Project Rainier — one of the world’s largest AI compute clusters — went live in late 2025 with 500,000 chips.
Trainium3, a 3-nanometer chip produced by TSMC, costs up to 50% less to run than traditional cloud servers for comparable performance.
AWS CEO Andy Jassy said Trainium is already a multibillion-dollar business; Apple publicly praised the chips in 2024, and AWS recently partnered with Cerebras Systems to integrate their inference chip on Trainium servers.
Nvidia’s networking division reported $11 billion in quarterly revenue — a 267% year-over-year increase — generating more than $31 billion for the full year and surpassing Cisco’s entire annual networking business. Link.
The division was built on Nvidia’s $7 billion acquisition of Israeli company Mellanox in 2020 and includes NVLink, InfiniBand Switches, Spectrum-X, and co-packaged optics switches.
At GTC 2026, Jensen Huang announced the Rubin platform with six new chips for “AI supercomputers,” a new Inference Context Memory Storage platform, and more efficient Spectrum-X Ethernet Photonics switches.
The networking unit has become Nvidia’s second-largest revenue driver behind compute, reflecting the growing importance of data center connectivity as AI clusters scale.
ROBOTICS
Amazon acquired Rivr, a Zurich-based autonomous robotics startup formerly known as Swiss-Mile, whose four-legged wheeled robots navigate streets and stairs to deliver packages from their torsos. Link.
Rivr was valued at $110 million in an August 2024 funding round; the robots work alongside human drivers, handling some packages while couriers shuttle others to nearby locations.
Amazon has planned $200 billion in capital expenditures for 2026 and already deploys over 1 million robots in warehouses, having acquired robot maker Kiva Systems for $775 million in 2012.
The acquisition complements Amazon’s broader robotics investments including Agility Robotics (humanoid robots), Physical Intelligence (robotics AI models), and its self-driving unit Zoox.
POLICY/GOV’T/ETHICS
The Department of Defense filed a 40-page response calling Anthropic an “unacceptable risk to national security,” arguing the company might disable its AI or alter behavior during warfighting operations if its safety “red lines” are crossed. Link.
Anthropic had negotiated contract terms prohibiting mass surveillance of Americans and refusing use in fully autonomous lethal weapons; the Pentagon argued private companies shouldn’t dictate military technology use.
Constitutional rights lawyer Chris Mattei criticized the DOD’s argument as speculative, with no evidence of actual wrongdoing; many tech companies and legal groups filed amicus briefs supporting Anthropic.
A preliminary injunction hearing was scheduled for the following Tuesday in the ongoing legal battle over Anthropic’s supply-chain risk designation.
The Pentagon revealed it is “actively pursuing multiple LLMs” to replace Anthropic’s Claude, with engineering work underway and operational availability expected “very soon.” Link.
Chief digital and AI officer Cameron Stanley said the Department is building government-owned AI environments as alternatives following the breakdown of Anthropic’s $200 million Pentagon contract.
OpenAI has secured its own Pentagon agreement, and the DOD signed a deal with Elon Musk’s xAI to use Grok in classified systems.
Defense Secretary Pete Hegseth’s supply-chain risk designation bars Pentagon contractors from working with Anthropic, effectively cutting the company off from the defense ecosystem.
The Trump administration released a legislative framework establishing a singular federal AI policy that would preempt state AI laws and shift child safety responsibility to parents rather than platforms. Link.
The framework calls for a “minimally burdensome national standard” and draws a hard line against states regulating AI development, calling it an “inherently interstate” issue that needs uniform rules.
White House AI czar David Sacks, a venture capitalist described as an “accelerationist,” shaped the innovation-first approach; the framework is notably missing liability frameworks, independent oversight, or enforcement mechanisms.
The framework also seeks to prevent states from “penalizing AI developers for a third party’s unlawful conduct involving their models,” effectively providing a liability shield for AI companies.
OTHER
OpenAI’s first ChatGPT advertisers report they can’t prove the ads work, having spent only 15–20% of their committed budgets as ads weren’t shown frequently enough despite a $60 CPM price tag on par with NFL games. Link.
OpenAI required a $200,000 minimum commitment per advertiser and offered a low-tech buying process using phone calls, emails, and spreadsheets rather than a self-serve platform.
ChatGPT has approximately 920 million weekly active users, but only about 5% are paying subscribers; OpenAI projects $17 billion in consumer revenue this year including ads and subscriptions.
OpenAI is addressing the issues by expanding ads to all free U.S. users, introducing a self-serve ad manager, and partnering with Criteo for better ad buying and targeting.
Jeff Bezos is seeking $100 billion for a new fund to acquire companies in aerospace, chipmaking, and defense and modernize them with AI through his startup Project Prometheus. Link.
Bezos serves as co-founder and co-CEO of Prometheus alongside former Google executive Vik Bajaj; the company launched with $6.2 billion in initial funding.
Prometheus builds AI models to improve manufacturing and engineering, and the new fund would acquire companies to deploy those models at scale.
Bezos recently traveled to Singapore and the Middle East to raise capital for the fund, which would be one of the largest acquisition vehicles in history.
DoorDash launched a standalone “Tasks” app that pays delivery couriers to submit videos and photos to train AI and robotic systems, with tasks like filming themselves washing dishes while wearing a body camera. Link.
The footage will train DoorDash’s in-house AI models and those of partners in retail, insurance, hospitality, and technology; pay is shown upfront and varies by effort and complexity.
DoorDash has more than 8 million Dashers who can reach almost anywhere in the U.S.; the app is available in select locations, excluding California, New York City, Seattle, and Colorado.
The move mirrors Uber’s similar program announced late last year that lets drivers earn extra income by completing small jobs like uploading photos to train AI models.