ALETH / AI-BRIEF / 2026-03-21 / NVIDIA SEES A TRILLION DOLLARS IN CHIP SALESThe Aleth Briefs link to the sources behind the stories and show how the week unfolded.
The week in five lines:
Nvidia forecast at least $1tn of Blackwell and Rubin revenue from 2025 to 2027.
A Supermicro co-founder was charged over AI servers smuggled to China.
A Meta employee exposed user data to staff after following an AI agent’s advice.
OpenAI agreed to buy Astral, maker of the uv and Ruff Python tools, for Codex.
The UK dropped plans to let AI train on copyrighted work unless creators opt out.
Browse by day:
Weekend: ByteDance pauses Seedance 2.0 launch
Monday: Nebius · Britannica · xAI · Nvidia GTC · Nscale · Comet
Tuesday: UK £2.5bn · OpenAI-AWS · Copilot · GPT-5.4 mini · H200 · Anthropic
Wednesday: UK copyright · Micron · Meta agent · Xiaomi MiMo
Thursday: Uber-Rivian · Astral · Cursor Composer 2 · Supermicro
Friday: White House AI framework
Weekend (14-15 March)
ByteDance suspended the global launch of its Seedance 2.0 video model after copyright disputes with Hollywood, The Information reported on Saturday.
ByteDance planned a mid-March release through its BytePlus cloud and a standalone app outside China. There is no new date. Disney, Netflix, Warner Bros. Discovery, Paramount Skydance and Sony had sent cease-and-desist letters.
ByteDance suspends video AI model after copyright disputes · The Information
Monday 16 March
Meta signed an AI infrastructure deal with Nebius worth up to about $27bn over five years.
Amsterdam-based Nebius will provide $12bn of dedicated capacity on one of the first large Nvidia Vera Rubin deployments from early 2027. Meta will also buy up to $15bn of further capacity that Nebius does not sell to other customers. On Wednesday Nebius priced $4.0bn of convertible notes, upsized from $3.75bn.
Nebius signs new AI infrastructure agreement with Meta · Nebius
Nebius announces pricing of upsized $4.0 billion convertible notes · Nebius
Encyclopaedia Britannica and its Merriam-Webster subsidiary sued OpenAI, alleging it scraped nearly 100,000 Britannica articles to train its models.
The complaint also targets ChatGPT outputs containing “full or partial verbatim reproductions” and ChatGPT’s use of Britannica articles in retrieval. A Lanham Act claim says ChatGPT attributes invented information to the publisher.
Complaint, Encyclopaedia Britannica v. OpenAI · CourtListener
The dictionary sues OpenAI · TechCrunch
Three Tennessee teenagers filed a class action against xAI, alleging its technology was used to make sexually explicit images of them.
The complaint says the perpetrator used a third-party app that licenses xAI’s Grok model, with the images generated on xAI’s servers. The plaintiffs seek class-action status, damages and an injunction.
Complaint, Doe 1 v. X.AI Corp. · CourtListener
Teens sue Musk’s xAI over AI-generated nonconsensual nudes · NPR
Nvidia said it sees at least $1tn of Blackwell and Rubin revenue from 2025 to 2027.
Jensen Huang told the GTC keynote in San Jose that he now sees “at least $1 trillion” of revenue from 2025 to 2027. A year earlier he had cited about $500bn of high-confidence purchase orders for Blackwell and Rubin through 2026. CFO Colette Kress confirmed the figure covers only Blackwell and Rubin plus networking. It excludes Groq LPUs and standalone CPUs.
At GTC Nvidia also launched the Groq 3 LPX inference rack and the Vera CPU for agentic AI. It formed the Nemotron Coalition with labs including Mistral AI, Cursor and Perplexity to build open frontier models.
NVIDIA GTC 2026: Live Updates on What’s Next in AI · Nvidia
Jensen Huang says company has one trillion dollars in orders through 2027 · DCD
NVIDIA Vera Rubin Opens Agentic AI Frontier · Nvidia
NVIDIA Launches Nemotron Coalition of Leading Global AI Labs · Nvidia
UK AI infrastructure firm Nscale agreed to buy American Intelligence & Power and its West Virginia campus, which it plans to grow to 8 GW by 2031.
The Monarch Compute Campus covers up to 2,250 acres in Mason County, with 2 GW of initial power due by H1 2028. Nscale also signed a letter of intent to provide Microsoft with 1.35 GW of Vera Rubin compute in tranches from late 2027, and will use Caterpillar gas generators for power.
Nscale Acquires American Intelligence & Power Corporation · Nscale
Nscale and Microsoft announce 1.35GW of Vera Rubin GPUs · PR Newswire
A US appeals court suspended the injunction barring Perplexity’s Comet agent from shopping in password-protected Amazon accounts.
The Ninth Circuit granted a temporary stay on Monday. On Wednesday it stayed the injunction pending appeal, expedited the case and set it for hearing in June.
Orders of 16 and 18 March, Amazon v. Perplexity, Ninth Circuit · CourtListener
Tuesday 17 March
Chancellor Rachel Reeves announced a record £2.5bn for UK AI and quantum computing, including up to £1bn to buy quantum computers.
The Treasury said the UK will “achieve the fastest AI adoption in the G7”. Quantum takes £2bn of the package.
UK will win AI race as Chancellor sets out economic ‘big choices’ · GOV.UK
OpenAI signed a deal with AWS to sell its AI products to US government agencies for classified and unclassified work.
OpenAI said it covers Amazon Bedrock in AWS GovCloud and in AWS Classified Regions for Secret and Top Secret workloads. On Wednesday the Financial Times reported that Microsoft is weighing legal action against Amazon and OpenAI over whether AWS can offer OpenAI’s Frontier without breaching its Microsoft deal. Microsoft said it is “confident that OpenAI understands and respects the importance of living up to [its] legal obligation”.
OpenAI expands government footprint with AWS deal, report says · TechCrunch
Microsoft weighs legal action over Amazon-OpenAI cloud deal · Financial Times
Microsoft put former Snap executive Jacob Andreou in charge of a unified Copilot, freeing Mustafa Suleyman to focus on frontier models.
Satya Nadella named Andreou EVP, Copilot, in a memo. Suleyman said he will focus “all my energy on our Superintelligence efforts” and deliver models within five years.
Announcing Copilot leadership update · Microsoft
OpenAI released GPT-5.4 mini and GPT-5.4 nano, smaller versions of GPT-5.4 for coding, subagents and computer use.
OpenAI says GPT-5.4 mini runs >2x faster than GPT-5 mini and approaches GPT-5.4 on SWE-Bench Pro and OSWorld-Verified. It costs $0.75/m input and $4.5/m output tokens, against $0.25/m and $2/m for GPT-5 mini. API-only GPT-5.4 nano costs $0.20/m and $1.3/m, against $0.05/m and $0.40/m for GPT-5 nano.
Jensen Huang said Nvidia is restarting manufacturing of its H200 chips for China after receiving purchase orders from “many customers”.
“We’ve been licensed for many customers in China for H200,” Huang told reporters at GTC. Late on Tuesday Reuters reported that Beijing had approved H200 purchases by several Chinese companies.t
Nvidia restarting production for H200 chips for sales in China · Axios
Chinese authorities approve Nvidia’s H200 AI chip sales, source says · Reuters
The Justice Department told the court in Anthropic’s lawsuit that the company could disable Claude or alter it during warfighting operations.
The Tuesday filing argues that the supply chain risk designation did not breach Anthropic’s First Amendment rights. A declaration by Pentagon official Emil Michael says Anthropic employs foreign nationals “including many from the [People’s] Republic of China”. Judge Rita Lin set a hearing for the following Tuesday.
In a Friday filing Anthropic said it “does not maintain any back door or remote ‘kill switch’”. Thiyagu Ramasamy, its head of public sector, wrote that it has never been able to stop Claude working or alter its functionality during military operations.
Opposition to motion · CourtListener · Declaration of Emil Michael · CourtListener
Declaration of Thiyagu Ramasamy · CourtListener
DOJ says Anthropic can’t be trusted with warfighting systems · WIRED
Wednesday 18 March
The UK government dropped its plan for an AI copyright exception with opt-out.
The opt-out was the government’s preferred option in its consultation, and most respondents rejected it. A report published on Wednesday says a broad exception with opt-out “is no longer the government’s preferred way forward”. The government will not change copyright law until it is confident reforms meet its aims. It will gather evidence and will not intervene in the licensing market for now.
Report on Copyright and Artificial Intelligence · GOV.UK
Micron reported fiscal Q2 revenue of $23.9bn, against $8.1bn a year earlier.
Micron guided fiscal Q3 revenue to $33.5bn, plus or minus $750m.
Micron fiscal Q2 2026 results, Exhibit 99.1 · SEC
A Meta employee exposed sensitive data to unauthorised staff after acting on an AI agent’s advice.
An internal agent posted an answer to a colleague’s question on a company forum without approval. An employee acted on its inaccurate advice and triggered a “SEV1” incident, Meta’s second-highest severity rating. For almost two hours, employees could see company and user data they were not authorised to view. Meta said “no user data was mishandled”.
Inside Meta, a Rogue AI Agent Triggers Security Alert · The Information
A rogue AI led to a serious security incident at Meta · The Verge
Xiaomi released MiMo-V2-Pro, a model with >1T parameters, and confirmed it was the anonymous model being tested on OpenRouter.
MiMo-V2-Pro has 42B active parameters and a 1M-token context, and is API-only, with no weights released. Artificial Analysis scored it 49 on its Intelligence Index, behind GLM-5 (Zhipu AI) at 50 and ahead of Kimi K2.5 (Moonshot AI) at 47.
MiMo-V2-Pro · Xiaomi
MiMo-V2-Pro: Everything you need to know · Artificial Analysis
Thursday 19 March
Uber will invest up to $1.25bn in Rivian for robotaxis through 2031, with $300m on signing, subject to regulatory approval.
The rest is tied to autonomy milestones. The first phase puts 10,000 autonomous Rivian R2 vehicles on Uber, starting in San Francisco and Miami in 2028. On Monday Uber and Nvidia said they plan robotaxis running entirely on Nvidia software, starting in Los Angeles and San Francisco in H1 2027 and reaching 28 cities by 2028.
Uber and Rivian partner to deploy up to 50,000 robotaxis · Business Wire
NVIDIA to Launch L4 Robotaxis on Uber Across 28 Cities by 2028 · Uber
OpenAI agreed to acquire Astral, maker of the Python tools uv, Ruff and ty.
The Astral team will join Codex. OpenAI says Codex has >2m weekly active users after 3x user growth since the start of the year. Founder Charlie Marsh said OpenAI will keep supporting Astral’s open source tools after the deal closes. The tools are downloaded hundreds of millions of times a month.
Cursor released Composer 2, a coding model priced at $0.5/m input and $2.5/m output tokens.
Cursor says the model comes from its first continued pretraining run followed by reinforcement learning (RL). A faster variant costs $1.5/m and $7.5/m. A developer then posted a Cursor model identifier pointing to Kimi K2.5 (Moonshot AI) with RL. On Friday Moonshot AI said Kimi K2.5 is Composer 2’s foundation, accessed through Fireworks AI under an authorised commercial partnership.
Introducing Composer 2 · Cursor
A developer’s post · fynn on X · Congrats Cursor on launch Composer 2 · Kimi on X
US prosecutors charged Supermicro co-founder Yih-Shyan Liaw with conspiring to divert AI servers to China.
Prosecutors allege that a Southeast Asian company bought about $2.5bn of servers in 2024 and 2025 on the defendants’ instructions. At least $510m worth allegedly went to China between late April and mid-May 2025. The defendants allegedly staged thousands of dummy servers to fool inspectors. On Friday Supermicro said Liaw had resigned from its board and named DeAnna Luna acting chief compliance officer. Its shares fell 33% that day.
Three charged over diverting AI technology to China · US Department of Justice
Supermicro Form 8-K · SEC · Super Micro co-founder indicted · CNBC
Friday 20 March
The White House published an AI policy framework asking Congress to pre-empt state AI laws.
It wants Congress to override state laws that regulate how models are developed or penalise companies for others’ use of their AI, and not to create a new federal AI regulator. It also covers protections for children and teens online. On Wednesday Senator Marsha Blackburn released a discussion draft of the TRUMP AMERICA AI Act, a federal framework that would replace state AI laws. It folds in the Kids Online Safety Act and the NO FAKES Act, and would sunset Section 230.
National Policy Framework for AI: Legislative Recommendations · White House
Blackburn releases draft of national policy framework for AI · CFO Dive
ALETH · OPEN ACCESS RESEARCH