explainx.ai0k
TrendingAI News TodayPathwaysSkills
Pricing
explainx.ai

Upskill in AI — 16 free pathways, live workshops & bootcamps, and 50+ courses from practitioners. Plus the skills, tools, and MCP servers to practice on.

follow us

follow on google

Add explainx.ai as a preferred source

corporate training

support@explainx.ai

get started

Find your pathTake Free Evaluation

community

Join the community

learn

mind: share how you thinkpathways — start freeworkshopsbootcampscoursescompare Explainxcertificationsmock testsexplainx universitycorporate traininglearn skills & mcp

discover

skillsmcp serversexplainx mcptoolsmdx readeragentsllmsdesignsdictionarypeopleagi trackerfelony benchranks

company

aboutvisionmissionteaminstructorsteach on explainxpartnershipscommunityhackathonscareers

content

daily AI newsstate of AI — live resultsblogreleasespromptsgeneratorsresource libraryfor LLMsexplainx.ai kids

solutions

all solutionsdeveloper upskillingmarketing upskillingproduct manager upskillingleadership upskilling

newsletter · weekly

Get AI news, tools, and insights in your inbox.

supportcontactprivacytermsdata rightshow we create contentsubmission guidelines

© 2026 AISOLO Technologies Pvt Ltd

explainx.ai

On this page

  • TL;DR: what is claimed and what is verified
  • What Tom's Hardware reports
  • Nvidia's statement, word for word
  • How DCGM Exporter exposes hardware
  • The enforcement backdrop
  • Why the B300 is the chip people want
  • What people are asking
  • What would turn the claim into established fact
  • What this means for builders and buyers
  • Bottom line
  • Related reading
← Back to blog

explainx / blog

32 Nvidia B300 GPUs on China Telecom's Network? What Lava Found and What Is Unproven

Nvidia, Export Controls, AI Chips, China, Claimed vs Verified

Part of AI Chips and Infrastructure

Security firm Lava says it spotted 32 export-barred Nvidia B300 GPUs on China Telecom's network via open DCGM monitoring. What is proven, and what is not.

Oct 23, 2026·8 min read·Yash Thakker
add explainx.ai
go deep
32 Nvidia B300 GPUs on China Telecom's Network? What Lava Found and What Is Unproven

An Israeli security firm says it can see 32 export-barred Nvidia Blackwell Ultra B300 GPUs on the network of China Telecom, a state-owned operator. The finding, relayed on October 23, 2026 by Tom's Hardware from a report by The Information, rests on a surprisingly mundane clue: GPU monitoring endpoints left open to the internet.

Nvidia did not deny the finding. It said diversion is negligible. Neither the firm nor Nvidia has shown who bought the chips, how they got there, or where in the network they sit. This post separates what the evidence supports from what the headlines imply.

Weekly digest3.6k readers

Catch up on AI

Curated AI updates on agents, skills, and MCP — delivered to your inbox. Unsubscribe anytime.

TL;DR: what is claimed and what is verified

table · 2 cols
QuestionAnswer
Who found it?Lava, an Israeli security firm, per Tom's Hardware citing The Information
What was found?32 Nvidia Blackwell Ultra B300 GPUs reporting from China Telecom's network
How?Nvidia's DCGM Exporter, reachable online, sometimes without authentication
Where exactly?Lava cannot confirm the location inside the network
How did they get there?Unknown
Nvidia's responseDid not deny; says diversion is under 0.5 percent and a drop in the bucket
Proof of smuggling by China Telecom?No

What Tom's Hardware reports

According to Tom's Hardware, Lava says it identified 32 B300 chips within China Telecom's network. The chips are barred from export under US controls, which allow sales only to whitelisted buyers who agree not to pass them on to sanctioned parties, including those in China.

The article is careful about what is unknown. It says "it's unknown how B300 chips found their way into China Telecom's server racks, nor can Lava confirm exactly where within the company's network they are located." The chips were identified via Nvidia's own DCGM Exporter, software data center operators install to monitor server health and performance. It is designed for internal use but "can oftentimes be accessed beyond the data center's network," and "doesn't always require password authentication."

Nvidia's statement, word for word

The response in Tom's Hardware is worth reading closely. An Nvidia spokesperson said: "Governments, researchers, and media have all found insignificant diversion of controlled US products, less than one half of one percent in recent court cases, a drop in the bucket compared to the ocean of domestic compute China already has." The statement adds that with "tens of millions of GPUs in service in approved countries, this report again demonstrates that smuggling is a nonfactor."

Two observations. First, the statement does not say the 32 chips are not B300s or that they are not in China. It argues scale. Second, the "less than one half of one percent" figure is attributed to recent court cases, which critics on social media argue measures only prosecuted diversion rather than all diversion. That critique is an argument, not a measurement, and no one has published a full estimate we could verify.

How DCGM Exporter exposes hardware

DCGM, Nvidia's Data Center GPU Manager, collects health and performance metrics such as temperature, utilization, memory and the GPU model. The Exporter publishes those metrics in a format monitoring systems scrape. If the endpoint is reachable from the internet and not protected, anyone who finds it can read the GPU model and count, and Internet-wide scanning can find such endpoints.

That is why this finding is plausible without anyone breaking in. It is passive observation of what the operator's own telemetry says. It is also why the limits are real:

  • Telemetry says what a GPU reports about itself; it does not prove ownership, purchase route or legal status.
  • A network address maps to an organization, but not necessarily to a specific building or tenant, especially for telecom operators that lease space and run cloud services.
  • A scan is a snapshot. It cannot tell whether the chips arrived last week or a year ago.

For any operator, the takeaway is practical: do not expose monitoring endpoints. Anything that tells the internet your GPU count and model is also an inventory leak. This holds for AI labs and enterprises, not just for smuggling cases.

The enforcement backdrop

This story arrives on top of several others we have covered. In August a Taiwan indictment named nine people over B300 exports; see Taiwan B300 smuggling. A contractor later pleaded guilty in a related case, covered in Super Micro contractor guilty plea. Taiwan has moved to criminalize AI chip and server exports to China, as in Taiwan's Foreign Trade Act proposal. On the US side, see the Treasury COINS Act proposed rule and the permitted-chip channel in Chinese tech giants' RTX Pro 5500 orders.

Tom's Hardware also notes that Nvidia previously confirmed it was working on a tool that would let its chips be located, but it has not been released. Location verification is the policy idea that this finding bears on most: if chips can report where they run, enforcement does not have to depend on tips, audits or court cases.

Why the B300 is the chip people want

Tom's Hardware describes the Blackwell Ultra B300 as one of the most powerful data center GPUs, with 50 percent more memory and FP4 compute than the previous-generation B200. Nvidia announced it in March 2025, pitching it as built for reasoning models. Demand for such parts is a long-running story: Tom's Hardware recalls a 2025 report that Chinese companies smuggled about $1 billion worth of Nvidia AI chips in three months, when the B200 was the most sought-after part and distributors were already preparing to sell B300 systems. Nothing in the finding says the 32 chips are connected to those earlier cases.

The size of the find also deserves perspective. Thirty-two GPUs is about four eight-GPU servers. That is meaningful as proof that the part is in the country but small next to the tens of thousands of accelerators in a large training cluster. It supports the claim that diversion happens; it does not measure how much.

What people are asking

Is this a hack? No. Based on the reporting, Lava read information that the operator's own monitoring endpoints published to the open internet. The weakness is misconfiguration, not an intrusion.

Could the telemetry be spoofed? In principle a server could report a different GPU name, but we found no suggestion of that in the coverage, and no evidence either way. This is one reason independent corroboration matters.

Does Nvidia know where its chips are? Not in real time, according to the reporting that a location tool was promised but not released. Nvidia can trace serial numbers through its sales channel, which is how investigations usually proceed.

Could the chips be legal? Export controls cover sales and transfers, so a B300 inside China would normally need an authorization that is not available for this part. We did not find any claim of a license, and neither company has said so.

What happens next? Possibilities range from nothing to an inquiry. The clearest signals would be a statement from US export authorities or a follow-up from Lava.

What would turn the claim into established fact

  • Lava's own write-up. Methodology, timestamps, IP ranges and the GPU identifiers it saw, so others can assess the scan.
  • Independent confirmation. A second researcher or a government agency matching the telemetry.
  • Chain of custody. Serial numbers traced through Nvidia and distributor records to a buyer. Only Nvidia and its partners can do this.
  • China Telecom's or Beijing's statement. None was found in the coverage we could open.
  • US action. A Commerce Department inquiry, an entity-list change or a charge would be the clearest sign that officials consider it credible.

What this means for builders and buyers

Most readers do not run export-controlled hardware, but a few practical points follow.

  • Audit your telemetry. If you run DCGM Exporter or similar tools, bind them to internal interfaces, put authentication or a reverse proxy in front, and verify from outside your network that the port is closed.
  • Expect compliance pressure on supply chains. Server makers and distributors face more checks after the Taiwan and Super Micro cases. That can mean longer lead times and extra paperwork for legitimate buyers, which feeds the tight market described in our post on B300 rental prices.
  • Treat single-scan stories with care. A telemetry finding is a lead. Wait for corroboration before treating any one firm's scan as a verdict about a company.

Bottom line

Verified: a named security firm says it identified 32 B300 GPUs reporting from China Telecom's network, and Nvidia's response, on the record, does not dispute the finding. Not verified: where the chips are, who bought them, how they were shipped, and whether China Telecom itself owns them. The more durable lesson may be the method: open GPU telemetry can show where restricted hardware ends up, which strengthens the case for location verification.

Details are accurate as of October 23, 2026 and rest largely on press reports of Lava's findings; check for Lava's own publication.

Related reading

  • Taiwan B300 smuggling indictment
  • Super Micro contractor guilty plea over AI server smuggling
  • Taiwan moves to criminalize AI chip exports to China
  • Treasury COINS Act proposed rule
  • Nvidia versus Huawei China AI chip share
  • B300 rental prices at record highs
  • Trump export controls and remote access
  • Sources: Tom's Hardware report, Digitimes
Spotted something out of date? Let us know.
Yash Thakker

Written by

Yash Thakker

Yash is an AI expert with over 300K learners. Join his workshops →

View Yash Thakker in People in AI →

Related posts

Oct 13, 2026

Nvidia 34%, Huawei 28%: China AI Chip Share, Claimed vs Verified

A report circulating on October 13 says Nvidia holds 34% of China data-center AI chips against Huawei at 28%, a six-point gap. Here is what the IDC numbers say, what we could and could not confirm, and what it means for builders.

Oct 22, 2026

ByteDance, Alibaba and Tencent Order 3 Million Nvidia RTX PRO 5500 Cards: Reuters, Claimed vs Verified

Reuters reported on October 20, 2026, citing two anonymous sources, that ByteDance, Alibaba and Tencent ordered more than 3 million Nvidia RTX PRO 5500 workstation cards at about $11,400 each, over $35 billion at sticker price, while Nvidia told partners it may fill only 5 to 10 percent. We lay out the numbers, the supply squeeze, and why a purchase order is not revenue.

Oct 22, 2026

Were the Microducks at GTC Berlin Human-Piloted? Claimed vs Verified

After Jensen Huang's October 21 keynote, X posts claimed four people were manually operating the Microduck robots. Nvidia's own recap says they learned their skills in simulation. Here is what each side actually says, and how to tell teleoperation from policy.