Nvidia's Jensen Huang was close to an alleged Chinese spy

The biggest AI story of the week is not about a model. It is about who was standing next to the man who makes them — and what Washington did with what it knew.

US intelligence officials concluded that a Chinese businesswoman who had a close relationship with Nvidia chief executive Jensen Huang was gathering information for Beijing, the New York Times reported on Thursday. The woman is Alice Huang, an executive at Singapore-based Megaspeed International, a company US authorities have spent two years investigating for moving restricted Nvidia accelerators toward China. Surveillance of her, according to people briefed on the classified intelligence, turned up both her access to the Nvidia chief and her ties to the Chinese state. Nvidia and the White House have not disputed the substance on the record.

The part that should stop you is not the espionage. It is the paper trail underneath it. Officials briefed President Trump on the finding roughly a year ago, and he declined to act — reportedly wary of disrupting Nvidia's chip sales and the broader trading relationship with Beijing, the Times reported. Jensen Huang was never formally warned by US intelligence that a foreign service was cultivating a contact close to him, and surveillance of Alice Huang is ongoing. Read that sequence again: the United States concluded that the chief executive of its most valuable company was being worked, told the president, and then let it run.

Megaspeed is not a random acquaintance from a Taipei cocktail party, which is where the two appeared together at a gathering of Asian customers. It was the biggest importer of Nvidia hardware in Southeast Asia between 2023 and 2025, by C4ADS's count. Established in 2023 by Chinese gaming firm 7Road as an offshore entity in Singapore, it spent roughly $2 billion on H100 and H800 GPUs inside its first year — chips it did not buy from Nvidia directly, but through Aivres, the California-based arm of Inspur, a Chinese server maker already sanctioned for supplying military supercomputing hardware. The systems went to facilities in Indonesia and Malaysia run by Megaspeed's Speedmatrix unit, per Tom's Hardware's read of the C4ADS trade records. From there the compute was rented out — to Alibaba and other Chinese customers.

A late-2024 inspection by the Bureau of Industry and Security at Megaspeed's Malaysian site found Nvidia servers and GPUs that were supposed to be running still sealed in their shipping crates. Investigators also found a near-empty Singapore office, a small Malaysian storefront with no engineering staff, and an ownership chain that ran through shell entities in Shanghai. That is the physical signature of a pass-through, not a cloud provider. By mid-2025 Megaspeed had stopped buying Nvidia hardware altogether and abandoned plans for another $3.2 billion in systems. Alice Huang left the company and dropped out of view.

The cleanup came from the landlord, not the government. In April, Bain Capital's Bridge Data Centres terminated Megaspeed's contract at its Malaysian hub and reassigned all 68.4 megawatts of contracted power to Zenlayer, telling lenders backing a $2.8 billion credit line about the swap without giving a reason, Bloomberg reported. Nvidia, for its part, ran spot checks across Megaspeed's Southeast Asian footprint several times last year and said it would do so again.

Here is the structural problem the case exposes, and it is not the one the export-control debate usually names. Washington polices the chip. The leak is in the compute. Megaspeed's product was time on export-controlled accelerators, sold remotely to Chinese customers — and renting a GPU abroad is not an export under the rules as written. That is why a company US officials suspect of smuggling could operate for two years in the open: in a Bain-owned data center, buying through a California subsidiary of a sanctioned Chinese firm, with one of its executives seated beside Nvidia's CEO at a customer dinner. Every link in that chain was visible in public records before the intelligence community reached a conclusion about the person running it. The enforcement model depends on the customer of record telling the truth, and the customer of record was the thing being manufactured.

Two things can be true at once, and the fairness of this story depends on holding both. Nvidia told the Times it had investigated and debunked the allegation that Megaspeed was supplying chips to China, and that its executives' dealings with the company were no different from those with any other regional customer. No charges against Megaspeed or against Alice Huang have been announced. What exists is an intelligence assessment drawn from unnamed sources, not an indictment — and being the target of a foreign intelligence service is not an accusation of wrongdoing by the person targeted. Huang was the access point, not the subject. Readers skimming the headline should hold that distinction as carefully as the Times did.

What makes it consequential is the position Nvidia now occupies. It is the world's most valuable listed company, its chips are the substrate of the entire AI buildout, and its chief executive has been unusually influential in shaping this administration's semiconductor export policy. A finding that foreign intelligence had a standing channel to that executive — and that the finding was shelved — lands directly on the credibility of the policy he helped write. It also reframes the smuggling cases that have been trickling out all year: the Taiwan indictment of nine people over diverted AI servers, the arrest of a California CEO over $300 million in servers routed through Malaysia and Singapore. Those were treated as customs problems. This one says the customs problem had a relationship manager.

The company is also not standing still. Megaspeed is pressing ahead with construction of a large data center in Johor, the Malaysian state across the border from Singapore, according to the Times — the same corridor where the crates were found sealed.

Hugging Face shipped on-device speech recognition for Apple apps, and its own benchmark table puts Apple's transcriber in second place. Two new Swift packages, announced on the Hugging Face blog, keep Apple's new speech API and swap out the engine behind it: NVIDIA's Parakeet v3, which handles 25 European languages and identifies which one it is hearing without being told, Moondream's post-trained Parakeet Ultra, or Alibaba's Qwen3-ASR, which covers 30 languages and 22 Chinese dialects including Cantonese. All are converted to Core ML and published on the Hub, roughly 600 MB per Parakeet, and they download only when an app asks for them.

The numbers are the reason to care. On three-to-eight-minute LibriSpeech files, run on an M4 Max, Parakeet Unified posts a 1.43% word error rate on test-clean and 2.19% on test-other. Apple's own transcriber posts 2.12% and 3.97% on the same files, per Hugging Face's measurements. Parakeet turns an hour of audio in 20 to 40 seconds from a file, using about 100 MB of memory once the model is compiled. The caveats are honest ones, published alongside the table: one machine, a handful of datasets, and on AMI meeting recordings every engine tested — including these — went past 23% word error rate.

The package also adds something Apple's framework has no module for: speaker diarization, via NVIDIA's Streaming Sortformer, labelling up to four speakers on the same timeline even when they talk over each other. The companion package, AnyDecisionModel, answers yes-or-no, multiple-choice and rating-scale questions by reading probabilities out of a model instead of generating text, with MLX, Core AI, llama.cpp and Ollama backends. Its next-token backends cap out at 26 options and 10 levels, which is a real constraint, and its Core AI build cannot currently be combined with the speech package because of a dependency version conflict.

The takeaway for anyone building an iOS or macOS app: Apple's on-device stack is now the baseline to beat rather than the default, and the baseline is beatable by a wide margin on the accuracy that users actually notice.

What to watch: whether Megaspeed's Johor build keeps its financing, whether Commerce or the Justice Department moves on the company now that the intelligence finding is public, and whether these speech numbers hold up on an iPhone — Hugging Face has only tested macOS 27.

If a foreign intelligence service had a standing channel to the CEO of the company writing your country's chip policy, what would you expect the government to do about it? Write to us via our contact page.