• The Neuron
  • Posts
  • 😺 Nvidia reportedly supplied China via Asia

😺 Nvidia reportedly supplied China via Asia

PLUS: Samsung's AI data-center push and a physics proof

Welcome, humans.

So apparently the first-ever AI Music Awards happened in Los Angeles on Friday, and the show leaned hard on actual humans. Variety reports the night used AI announcers, lip-syncing, and human stand-ins, while repeated technical problems kept interrupting the ceremony.

Quick setup, because "AI artist" is a weird phrase. Someone writes the lyrics, an AI tool makes the song and the singing voice, and the "artist" is a named character with a face. Organizers had promised AI acts like Xania Monet would perform live. But a character with no body can't walk onto a stage, so people filled in.

Somewhere, Milli Vanilli is finally feeling understood.

Here’s what happened in AI today:

  • 😸 Nvidia reportedly kept Chinese AI firms supplied via Asia.

  • 📰 Samsung reportedly courted Anthropic and OpenAI as data-center tenants.

  • 📰 OpenAI's GPT-6 Astra helped prove a Standard Model construction.

  • 🍪 Apple opened a 200,000-person AI health study app.

  • 🎓 One compromised MCP server can hijack your AI agent.

😸 Nvidia reportedly kept Chinese AI firms supplied through Southeast Asia

The Wall Street Journal reported that Nvidia cultivated Asian data-center partners that rented its advanced chips to Chinese customers. One of them, Megaspeed, which has opened its own outside-lawyer probe, reportedly arranged nearly $12B in chip orders across 2024 and 2025.

Why does renting matter? Export controls police shipments, and customs can inspect a crate. But a Singapore data center can legally buy chips, then sell time on them to anyone who pays. If the renter logs in from China, the chips never cross a border. The computing does. Nobody inspects a login. That gap is why Congress is reviving bipartisan proposals to restrict China's remote access to overseas compute.

Here's what else landed over the weekend:

All of this is press reporting, not findings. A second WSJ story adds a loop worth knowing. Nvidia has backed customers who then buy its chips, and it has reportedly begun pulling back.

The first test is Tuesday's earnings call, Colette Kress's first as interim CEO. Listen for China sales, customer backing, and any hint on the CEO search.

Why this matters: Supporters of tight chip controls argue that cutting-edge AI rewards raw computing power and will shape economies and militaries, so staying ahead is a security question. Skeptics reply that Chinese labs still build competitive models, and that every restriction costs American companies too.

For a casual AI user, it shows up in your apps and your wallet. Hugging Face's spring report put Chinese open models at 41% of global downloads, so more of the tools you touch may run on them. That can mean lower prices and more choice, plus security and data questions buyers are still sorting out.

The race runs hot because both governments treat AI as a pillar of economic and military strength, and chips are the chokepoint. Washington wants the world running on American hardware and software, Beijing is pushing its own alternative, and many countries in the Global South see Chinese open models as a path to AI independence. Whichever stack becomes the default shapes prices, standards, and who sets the rules.

FROM OUR PARTNERS

Why start from scratch when you can start with what’s proven?

Every enterprise AI project starts with the same choice: a blank page, or something already proven. Most teams take the blank page, and most stall there. The Lenovo AI Library is the other option: a curated catalog of enterprise-ready AI agents and validated use cases, built on proven Hybrid AI solutions, governed by design, and ready to customize. Why start from scratch when you can start with what’s proven?

🎓 AI Skill of the Day: Audit the trust chain behind your AI agent

VentureBeat's write-up of security testing attributed to Palo Alto Networks' Unit 42 has a finding worth taking seriously if your AI agent connects to more than one tool. In a setup with five MCP servers feeding one agent, a single compromised server tricked the agent into misusing the other, trusted tools 78.3% of the time. (MCP, the Model Context Protocol, is the standard that lets an agent plug into things like your email, a database, or a deployment tool. Each plug-in is called a "server.")

One caveat first. VentureBeat links that 78.3% to a Cycode blog post rather than a Unit 42 report, and it doesn't say how many trials were run or how "success" was defined. Read it as a warning flare about the direction of the risk.

The mechanism is what matters. The article calls it cascading trust. An agent treats whatever each tool sends back as trustworthy working notes. A compromised server can hide instructions in those notes, and the agent carries them into its later calls, including calls to servers nobody touched. Picture a shared notebook where every coworker's page gets read before you send anything. Someone slips a fake instruction onto one page, and you follow it when you write to your boss. A research tool that only reads can still influence what your email or deployment tool does next. "This server can't deploy" is cold comfort if it can talk the agent into asking another server to deploy for it.

So review the chain, not just each link. Before you connect another server:

  1. List every tool the agent can call and flag the dangerous pairings, like a tool that reads the open web sitting next to one that can send email.

  2. Give each server short-lived, narrowly scoped credentials, such as read-only mailbox access instead of full access. The article also suggests pulling credentials when a call happens instead of storing them at startup, so you can revoke them.

  3. Require approval for irreversible actions, and log the complete sequence of tool calls, not only the final one.

Paste this into a chat with your architecture notes or tool list:

Map the trust chain for this AI agent. For every connected tool, list its credentials, allowed actions, data it can place in model context, and which other tools could act on that data. Then identify the smallest permission changes and approval gates that would contain one compromised server.

Start with the answer's riskiest pairing, any tool that reads untrusted content (the web, inbound email) living in the same agent as a tool that can't be undone. That's where your first approval gate belongs.

Have a specific skill you want to learn? Request it here.

FROM OUR PARTNERS

Adapt your workspace to your business

With Builderlab, teams can build and share apps for recurring work and use plugins to change how the workspace looks and functions. Buzz is open source, so companies can inspect, extend, and host the workspace themselves.

📰 Around the Horn

In case you haven’t noticed…

  • Samsung's IT arm reportedly entered early talks to sign Anthropic and OpenAI as tenants for the AI data centers Samsung wants to build, run, and power.

  • OpenAI and physicist Nathaniel Craig published a candidate lattice construction for the Standard Model (physics' theory of particles) that GPT-6 Astra helped develop and prove.

  • Pax Silica members, including Japan, Australia, Israel, India, and the U.K., agreed to build a joint list of key AI technologies and supply chains that need more protection.

  • Bloomberg reported Apple's first smart glasses will use Siri AI to operate existing iPhone apps, with a WWDC 2027 reveal and late-2027 shipment targeted.

  • SK Hynix committed about $1.05B over five years to strengthen Korean suppliers of chip materials, parts, and equipment.

  • The Atlantic tested ChatGPT on 32,245 personal messages, raising consent questions when one person's AI analysis pulls in other people's private conversations.

Want absolutely EVERYTHING that happened in AI this week? Click here!

🍪 Treats to Try

*Asterisk = from our partners (only the first one!). Advertise to 700K+ readers here!

  1. Qwen4-Max is Alibaba's new flagship model for coding and agent work, available through its Token Plan subscription and apps like Qwen App.

  2. Apple's Health AI Study app lets U.S. adults 21 and older with an iPhone 15 Pro or later ask health questions and check the linked peer-reviewed citations behind the answers.

  3. Spiik records meetings, lectures, and calls on your iPhone, iPad, or Mac, and on Mac it captures both sides of a Zoom call without a bot joining.

  4. Talyn runs Claude Code and Codex in parallel cloud sandboxes, then keeps your GitHub pull requests moving by fixing failed checks, merge conflicts, and review comments—free plan, then $40/mo.

  5. Polylane for Security reviews pull requests headed to production for exploitable flaws, shows how attacker input reaches the bad line of code, and suggests the fix.

  6. ByteAsk Cloud Agents takes a bug report from Slack, Jira, or your browser, tests a fix in a sandbox, and opens a draft pull request with the results attached.

  7. Gavel walks you through a coding agent's changes in a local diff viewer, with the agent explaining its choices next to the code before you approve.

  8. db-branch creates MySQL and PostgreSQL database copies on your own servers in seconds, so each AI agent or pull request can test against its own data (open-source core, beta).

UPCOMING EVENT: The Neuron IRL in San Francisco

The Neuron is going IRL on November 18 at 4:30 PM PT! Join us in San Francisco for drinks, bites, special guests, and a live recording of The Neuron podcast. Huge thanks to Slack and Alumni Ventures for making it happen.

It’s free, but space is limited. Save your spot!

😹 Monday Meme

Blue pill or red pill?

New from The Neuron: AI Explained

New episodes air every week on Wednesdays: Spotify | Apple Podcasts | YouTube

A Cat’s Commentary

That’s all for now. If you want to get featured above, fill out the poll below and tell us how we did today!

What'd you think of today's email?

Login or Subscribe to participate in polls.

Btw: We just launched a robotics newsletter! Sign up for it here.

P.S. Love the newsletter, but only want to get it once per week? Don’t unsubscribe—update your preferences here.