Renting a GPU in Singapore used to be a way around American export controls. As of Monday, Washington says it isn't. Turns out the cloud has borders after all. It's Tuesday, November 17th, and this is Lantern AI Daily Briefing. Coming up: the Commerce Department's rule on remote chip access, AMD finishing its World Labs purchase with Fei-Fei Li as chief scientist, Alibaba's Qwen4-Max against an independent leaderboard, and Hugging Face moving its inference routing onto OpenRouter. Tap follow so the next episode finds you.
Ars Technica, on the Commerce Department's new rule:
In a summary of the rule change, the Commerce Department said it was closing a loophole that previously allowed foreign rivals like China, Russia, and Iran to bypass export curbs on certain advanced computer chips by instead renting access to those chips from data centers in Indonesia, Malaysia, Singapore, Thailand, and Brazil.
This is the policy branch of the Nvidia story we've followed since Huang resigned last Thursday. Katie McQue at NPR draws the line directly. Alice Huang, the woman at the center of the scandal and no relation to Jensen, is a former executive at Megaspeed International, a Singapore data center company that operates in Malaysia and is suspected of selling remote access to US AI chips to Chinese companies. Megaspeed told NPR it conducts legitimate and lawful business and has hired outside counsel for an internal investigation. And note the reversal. NPR points out that before this scandal, the Trump administration had resisted adding rules on foreign access. In May of last year it said it would not enforce Biden's diffusion framework and planned to rescind it, calling it ill-conceived and counterproductive. Now it's closing the same kind of gap. Congress is piling on. The Remote Access Security Act passed the House back in January, three hundred sixty-nine to twenty-two, and The Verge describes it as stalled in the Senate. Senate Banking chair Tim Scott has written to Huang and Commerce Secretary Howard Lutnick asking them to testify publicly, and Representative Yvette Clarke told NPR both are expected before Congress this month and next. Then there's the evidence question. MIT Technology Review reports that ex-OpenAI researcher Will DePue published findings last Thursday suggesting Chinese AI companies had far more access to Blackwell chips than previously estimated. DePue says senior researchers at OpenAI and Anthropic verified it. No independent analysis has confirmed it, and the magazine stresses it doesn't show Nvidia helped anyone get the chips. For builders: if you buy or resell compute out of Southeast Asia, read the rule text when it lands, not the summary. And Nvidia reports earnings today with all of this hanging over the call.
Erik van Klinken, writing in Techzine Global:
The World Labs team remains committed to conducting fundamental model research, according to the announcement. At the same time, the researchers will collaborate with AMD engineers to translate insights from new model capabilities into AI infrastructure for customers and developers.
AMD announced the all-stock deal on September 28th at about eight point two billion dollars, and it closed Monday. Li becomes executive vice president and chief scientist, reporting directly to Lisa Su. Co-founders Justin Johnson and Ben Mildenhall stay on to lead the team inside AMD. World Labs builds models that generate interactive 3D environments. Its product Marble makes persistent worlds from images, video or text, and the lab shipped a new world model called Atlas on November 12th, four days before closing. Van Klinken puts the price next to the cap table: eight point two billion for a startup that raised about a billion. And Arry Hashemi, reporting on the close, notes that one of the backers of World Labs' February round was Nvidia. So AMD just bought a lab its biggest rival had a stake in. The logic AMD gives is co-design. Models that reason about 3D space and train robots in simulation will stress hardware differently, and owning the lab means seeing that early. Van Klinken notes this is ground where Nvidia is already big, with Omniverse. Same day, AMD got a demand-side headline too. TensorWave, an AMD-backed neocloud, says it signed a six-year deal to give OpenAI four hundred fifty-seven megawatts of IT capacity on MI455X in Helios racks, starting in the first half of 2027. That's TensorWave's press release. Fagbemi Daniel's write-up is the useful check: no contract value, no per-site schedule. A megawatt is a power allocation, not a running GPU.
Pasquale Pillitteri, on Monday's Qwen4-Max launch:
In Alibaba's tables the model sits close to Opus 5.5 on several tests, while the independent index places it far behind.
Numbers first. Artificial Analysis gives Qwen4-Max forty-three on its Intelligence Index, up nine on Qwen3.8 Max, but sixteen behind Claude Opus 5.5 and three behind Kimi K3.1 Pro. Alibaba's own sixteen-benchmark table gives Qwen four firsts and Opus seven. The DeepSWE result, seventy-five point seven to seventy-four point two, is where the on-par talk started. And the losses are the ones I care about. Terminal-Bench 4.0, forty-eight and a half against fifty-nine point six for Opus and sixty-seven point two for Kimi. ProgramBench, thirty-one and a half against sixty-five. If your agent lives in a shell, that's the story. The real argument is price. Alibaba cut the list to a dollar fifty in and four fifty out per million tokens, and cost per task on the index fell sixty-seven percent, to three twenty-seven. But in the same table, GPT-6.1 Sol scores fifty-five for a dollar twenty-seven a task. Two caveats from Pillitteri. The index ran at xhigh reasoning, while the API defaults to high, and there's no speed data yet. Also, these weights aren't downloadable. Alibaba promised open versions of the smaller Qwen4 models, no date. Cheap and proprietary is a different pitch than cheap and open.
Célina Hanouti and colleagues, on the Hugging Face blog:
Today, we are switching Inference Providers from our own custom router to OpenRouter's. Requests made to language models now go through OpenRouter, which picks the best provider for each call based on live price, speed and reliability, and retries on another provider if the first one fails. Your code stays the same.
Automatic failover inside the same call is the line I like. Before Monday, Hugging Face picked one provider and handed you its error. OpenRouter says its rebuilt routing cut capacity errors forty-six percent and improved time to first token by almost seventeen percent. You can still steer it with a suffix on the model id, fastest, cheapest, or a named provider, and turn providers off in settings. Routing is cache-aware, so a conversation stays where its prompt cache lives. Hugging Face says it adds no markup on provider prices. OpenRouter's side adds a trace option that drops each request into a Hugging Face Storage Bucket you own. Read the breaking change, though. For language models, you can no longer bring your own provider API key. Everything bills through Hugging Face. If you negotiated a rate directly with a provider, that path just closed for chat models. The context Hugging Face cites: open-weight models went from about twenty-seven percent of OpenRouter tokens in October 2025 to nearly seventy-two percent by the end of this September, at roughly a quarter the price. The gap hasn't closed, though. On Artificial Analysis, Kimi K3.1 Pro scores forty-six to sixty-five for Claude Fable 5.2, at about a seventh of the cost per task.
Four hundred fifty-seven megawatts with no site list is precisely the sort of number our sibling show, The Data Center Daily, takes apart. It's a daily briefing on AI compute: hyperscaler capex, the power grid, semiconductor supply, and energy markets reshaped by intelligence at scale. It's in the same feed directory where you found us.
One for the road: if you route through Hugging Face with your own provider key, check your error logs this morning. Tomorrow we'll pick through Nvidia's earnings and watch for Anthropic's IPO pricing. Thanks for spending a few minutes with Lantern AI Daily Briefing, a show from Lantern Podcasts. Talk to you Wednesday.