Malaysia AI Chip Export Controls and the Phone in Your Pocket

A Malaysian data center operator who buys NVIDIA H100 modules from a distributor in Singapore now files paperwork before those chips can physically reach the building. The hardware might never touch US soil again after the original export. Doesn’t matter. As of July 14, Kuala Lumpur wants an application, an end-user certificate, a full chain of custody, and thirty days of lead time before anything moves. That thirty-day window is the part of Malaysia’s AI chip export controls that actually changes behavior, and it’s the part everyone skated past while reading the headline.

Malaysia’s Ministry of Investment, Trade and Industry issued it as Directive No. 1/2025, under Section 12 of the Strategic Trade Act 2010, covering the export, transshipment, transit, or re-export of US-origin high-performance AI chips. No grace period, not even for shipments already in the air. It doesn’t list part numbers. It leans on the US classification architecture instead, the ECCNs that BIS introduced in October 2023 for advanced accelerators, 3A090 for the silicon and 4A090 for the assemblies, plus anything swept in by the Foreign Direct Product Rule. The practical line is the one US controls already drew around datacenter interconnect: the 600 GB/s-and-up class that catches H100, H200, A100, the MI300X, and every successor anyone ships in volume. Your phone sits nowhere near it, which matters later.

Intel’s Penang campus, a stack of OSAT houses feeding NVIDIA’s supply chain, free trade zones, bonded warehouses. Malaysia runs the world’s third-largest semiconductor packaging and testing operation by revenue, behind Taiwan and South Korea, and something like 13 to 18 percent of global chip shipments pass through its logistics at some point, so the leak was always going to surface here. The route was almost embarrassingly clean. A US manufacturer ships to Singapore or Malaysia under a tier that didn’t need a China-direct license, the chips sit in a Malaysian FTZ where no domestic re-export control applied, and from there they move on to China through Hong Kong shell companies and research institutes that exist mostly on letterhead. BIS documented it for years. Earlier in 2025 a Singapore smuggling case put roughly $390 million of servers carrying NVIDIA silicon on that exact route. Closing it needed the destination country to run its own export-control machinery, and Malaysia had the STA 2010 framework sitting there the whole time, just never aimed at AI transshipments with this kind of granularity.

An STP application means identifying the end user with documentary evidence, declaring chip model and ECCN and quantity and aggregate compute, mapping the custody chain from US fab to Malaysian port of entry to final destination, and producing an end-use certificate signed by the ultimate consignee. Hyperscalers run just-in-time. NVIDIA’s lead time on H100 and H200 SXM5 modules in Q2 was already 8 to 14 weeks, so bolt a mandatory thirty-day window onto the front of that and anything routed through Malaysia runs a 14-to-18-week procurement cycle minimum. That notice period is the part that actually bites. The compliance overhead is real money too, somewhere between fifteen and a hundred thousand dollars a transaction depending on whether you’re importing directly, transshipping through an FTZ, or re-exporting onward, with re-exports the ugliest because they pull in full MITI review plus BIS coordination. Those figures are rough, extrapolated from EU Dual-Use and US BIS license processing, but the order of magnitude holds.

Biden’s AI Diffusion Rule, published in January, set up a tiered country system: close allies in Tier 1 with basically no license, most of the world including Malaysia in Tier 2 under compute caps, China and Russia and the arms-embargoed crowd in Tier 3 near-locked-out. The Trump administration rescinded it in May, before the main compliance deadline even hit, calling the tiers too complex and too abrasive with allies. The ECCN classifications and the FDPR stayed, but the Tier 2 compute caps evaporated, and for a stretch of roughly May through July the transshipment route was arguably more open than it had been under Biden, the US loosening this exact path by accident. That same window is when the H20 ban flipped back to a green light. Malaysia’s directive closes the gap with domestic controls that, in the thirty-day sense, run more operationally stringent than the rule they’re effectively replacing. Nobody in the prior US framework asked for a month of advance notice.

Japan put export controls on 23 categories of chipmaking gear back in 2023, aimed at EUV-adjacent lithography and coordinated with Washington and the Dutch. The Netherlands leaned on ASML to stop shipping its more advanced DUV systems to China without a license, with the FDPR extending US reach over the rest. South Korea has Samsung and SK Hynix under quiet guidance on HBM sales to Chinese AI-silicon shops, the HBM3E flowing toward Huawei’s Ascend program drawing particular attention. Singapore tightened re-export controls in Q1, vaguer than what Malaysia just did. Malaysia is one node in a coalition that now covers every layer: design through US ECCNs, manufacturing equipment through Japan and the Netherlands, memory through Korea, logistics and packaging through Malaysia and Singapore. The chokepoint stopped being a single point a while ago. It’s a mesh now, and that mesh is what Beijing has to plan around.

Thailand runs serious assembly through the Suvarnabhumi cargo hub with nothing resembling an STP regime, so that route went softer than Malaysia’s overnight. Vietnam is growing fast on Intel assembly in Ho Chi Minh City with no equivalent framework either. India is the awkward one, a close US partner with a young export-control apparatus and a data-center buildout that makes it a more and more attractive waypoint. And none of it touches the cleanest dodge going, cloud arbitrage. A Chinese entity renting H100 time through a Malaysian provider’s API moves no physical chip and trips no permit at all. Permits regulate the box on the loading dock, not the API call, and that gap is the obvious next front. Its own article, really.

The Nothing Phone 3 shipped into this exact climate in July, and almost nobody connected the two. The Snapdragon 8s Gen 4 inside it is fabbed on TSMC’s N4P, a 4nm node that still depends on EUV layers from the same ASML scanners the Dutch keep out of China. That dependency decides whether the SoC can be built at all, the kind of thing a consumer-hardware analyst should carry in the model and almost never does.

The Phone 3 does not run the Snapdragon 8 Elite. It runs the Snapdragon 8s Gen 4, the SM8735, a tier down, and Nothing took real heat for it after Carl Pei billed the phone as the company’s first true flagship. The 8s Gen 4 is built on N4P with ARM’s licensed Kryo cores, one Cortex-X4 prime at 3.2 GHz over seven Cortex-A720s, a classic big.LITTLE layout rather than the homogeneous all-Oryon arrangement that sits in the actual 8 Elite. The Adreno 825 GPU and the updated Hexagon NPU are both solid, the NPU up around 44 percent on the 8s Gen 3, plenty for the on-device inference and the AI tricks Nothing is selling. It’s a genuinely good chip, just not the fastest Android silicon of 2025, and the gap shows up exactly where a $799 buyer would look.

A 6.67-inch LTPO AMOLED wraps around that chip, ramping 1.5 to 120 Hz at 2800×1260 and peaking at a genuinely bright 4,500 nits, paired with a 5,150mAh silicon-carbon cell on 65W wired and 15W wireless, and a triple 50MP rear array led by a 1/1.3-inch main sensor with OIS alongside a 50MP 3x periscope and a 50MP ultrawide, plus a 50MP front camera. IP68, 218 grams, Gorilla Glass 7i over the front and Victus on the back, 12GB with 256GB or 16GB with 512GB of UFS 4.0, $799 in the US and £799 in the UK.

Nothing ditched the linear LED strips from the Phone 1 and 2 and dropped a monochrome micro-LED dot-matrix into the upper-right of the rear glass, 489 individually lit LEDs running as a tiny secondary screen: caller ID, app alerts, a clock, a battery gauge, a camera countdown, even toy mini-games you cycle with a dedicated Glyph Button under the glass. The part that lifts it past gimmick is the public SDK. Nothing exposes the matrix as a programmable surface with every one of those 489 LEDs addressable, so developers and tinkerers write their own widgets and animations against it instead of waiting for Nothing to ship them.

A vapor chamber finally appears here too, a first for the Phone line, sandwiched with graphite spreaders over the SoC and memory to keep the 8s Gen 4 from throttling under sustained load. Nothing OS 3.5 runs on Android 15, a heavily reworked SystemUI and a monochrome design system enforced at the framework level rather than bolted on as a theme, with the on-device AI features running locally through Qualcomm’s QNN runtime on the Hexagon. The support commitment is the part the early skeptics got wrong: five years of OS updates and seven years of security patches. That matches Samsung and Pixel on security outright and trails them only on OS-version longevity, so the durability knock people reached for on day one mostly doesn’t land.

The same $799 buys a Galaxy S25 carrying the full Snapdragon 8 Elite, a service network Nothing can’t touch, and one more year of OS updates on top. Carl Pei calling the Phone 3 our first true flagship set the bar at the top of Android, and the chip lands a tier below it. Plenty of design-led companies took that exact swing and lost: HTC, Sony Mobile, Essential, the list runs long. What Nothing has that none of them quite held onto is a brand identity that actually coheres, plus a hardware hook in the Glyph Matrix that no rival patches in with a software update. Whether that defends $799 against a phone with faster silicon and a marginally longer update window, perso, I’m not convinced. It’s still the most credible swing anyone’s taken at the top of Android in years.

The OnePlus Nord 5 that landed July 8 runs the Snapdragon 8s Gen 3 on TSMC N4P, a Cortex-X4 prime, an Adreno 735, a 6.83-inch 144Hz panel, and a 6,800mAh cell on 100W charging for around the equivalent of $450. Qualcomm’s ladder flexes the other way with it. The Phone 3 sits one rung up the 8s tier at nearly double the price, which is the whole tension in a single comparison. A mid-ranger in mid-2025 hands you performance that cost north of $800 in 2022, and that compression, du coup, is the most underrated thing happening in Android right now. Oryon stays reserved for the 8 Elite and above, so neither of these phones gets it, and for most buyers that matters far less than the spec snobs want it to.

Qualcomm pulled its Snapdragon Summit forward to September 23, about a month ahead of the usual October slot. Apple’s A19 arrives in the iPhone 17 that same month and MediaTek’s Dimensity 9500 follows in Q4, so Qualcomm clearly wants the benchmark narrative set before either of them shows up. The leaked next chip, the SM8850, reads as TSMC N3P, Oryon V3 in the same 2+6 layout, a prime core near 4.6 GHz, an Adreno 840, an NPU pushing toward 65 TOPS, and a Geekbench 6 single-core north of 4,000. Treat that last number with real skepticism, the sourcing is thin. But if it holds it would be the first time a Qualcomm chip matched Apple on single-core in the same generation instead of chasing it a year late.

Apple’s A19 rides N3P as well, the same node as that leaked SM8850, which makes the 5 percent performance bump at iso-power over N3E the least interesting thing about it. Apple’s historical edge was being first to each new node, and this generation that edge just disappears, both chips on the same TSMC process. The earlier Summit reads as brutally logical once you see that. Put SM8850 scores above 4,000 into the press on September 23, weeks before iPhone 17 ships, and the whole media cycle ends up framing the A19 against Qualcomm instead of the reverse. Whether the silicon backs the leaks is a separate question, but the calendar alone says Qualcomm thinks it can win the framing for once. I’m a declared Snapdragon Insider, so salt my optimism to taste; the structural read holds either way.

TSMC’s N3 and N4 capacity depends on ASML EUV tools the Dutch won’t send to China and US EDA software under US control. That capacity yields the 8s Gen 4 in the Phone 3, the 8 Elite in the S25, the SM8850 still to come. Those chips move through Malaysian logistics that now answers to the STP directive on anything re-exported, then land in your hand running billion-parameter models locally on the Hexagon. The same coalition writing these export controls is the one pinning TSMC’s advanced nodes to Taiwan, Korea, and eventually the US and Japan.

H100 NVLink runs 900 GB/s. The 8s Gen 4 in the Phone 3 moves about 77, a full order of magnitude under the 600 GB/s line the controls are drawn around, which is the whole reason the consumer chips walk free. That line got drawn to catch the things that train frontier models at datacenter scale and leave the thing in your pocket alone, and today it’s well aimed. But the Hexagon already does on the order of 45 TOPS on the 8 Elite, the SM8850 is reportedly chasing 65, and on that slope consumer NPUs cross 100 inside a couple of generations. Run a small Llama at 4-bit on current hardware and you get a usable read-speed token rate with first-token latency low enough that on-device inference stops being a stunt. The boundary between a consumer AI chip and a regulated accelerator keeps drifting toward the regulators. Give it three years and someone in Kuala Lumpur or Washington gets to decide whether the phone in your hand needs a permit too, and franchement, I don’t love where that ends.