Here is the elegant, almost insulting simplicity of what happened.
The United States banned Nvidia's most advanced AI chips from being sold to China. The logic was straightforward: no chips, no compute; no compute, no frontier models. Choke the hardware, slow the rival. Three years of policy rested on it.
So Chinese AI firms simply stopped buying the chips. Instead, they rented them, renting time on Nvidia hardware sitting in data centers in Thailand, Malaysia, and Singapore. The chips never crossed into China. The Chinese engineers never owned them. They just logged in from Beijing and ran their training jobs on machines parked a border away.
And by the letter of the law, that was fine.
America's export controls stop chips from crossing a border. They say nothing about a Chinese engineer logging into a chip that never moves. The ban guarded the door. China used the window.
~3 years
America's lead on the physical hardware and it didn't stop China
3 hubs
Thailand, Malaysia & Singapore host the rented Nvidia compute
GB300
The banned Nvidia chip Moonshot was accused of renting in Thailand
0
Chips that actually crossed into China, which is why it was legal
Why renting was legal and buying wasn't
The distinction sounds like a technicality because it is one and that's exactly why it worked.
US export controls govern physical objects crossing physical borders. A chip is a thing; shipping that thing to China is banned. But renting remote access to a chip that stays put in Thailand isn't shipping anything. As one compute-policy researcher put it plainly to CNBC: the export regime controls physical AI chips, not remote access to them.
So a whole grey market grew up around the gap. Southeast Asian data centers stocked with Nvidia hardware that America was happy to sell to Thailand or Malaysia quietly became the compute engine for Chinese AI. Industry watchers now believe this offshore access is a major reason Chinese models kept leaping forward all year, even as the official story insisted they'd been cut off.
This month it got specific. A senior White House official publicly accused Moonshot AI, one of China's hottest labs of training its new Kimi K3 model on banned Nvidia GB300 chips through a facility in Thailand, just days after the model launched.
Here's the whole board in one glance, three different channels, three different rules:
Banned outright | Nvidia's most advanced chips (GB300 / Blackwell class) cannot be sold to China at all |
Sold, but capped | Less-advanced H200 chips allowed to ~10 approved Chinese firms, up to 75,000 units each |
The workaround | Rent the banned chips remotely, via data centers in Thailand, Malaysia & Singapore, legal, because nothing is bought or shipped |
This week's response | US enforcement arm opens a systematic review of offshore rentals; new legislation being drafted |
Keep this straight, it matters
That Moonshot claim is an accusation from a White House official, not a proven finding. Moonshot hasn't been shown in court to have done it, and renting offshore compute is, at least for now, legal so long as the Chinese firm doesn't own the hardware. So the honest framing isn't "China cheated", it's "China found a legal route the ban didn't cover, and Washington is furious about it."
Also worth being precise: America didn't lose the chip war. It's still three years ahead on the hardware itself. What it lost was the assumption that controlling the hardware means controlling who can use it. Those turn out to be very different things.
This week, Washington moved to slam it shut
The fresh news is the response. The US agency that investigates chip-smuggling the enforcement arm of the export-control regime has now started systematically reviewing exactly these offshore-rental arrangements, according to reporting from Bloomberg and CNBC. Legislation to close the loophole is being drafted.
But closing it is genuinely hard, and the difficulty is the whole point. To stop a Chinese engineer from renting a chip in Thailand, you have to police not the chip but every possible use of it, in every country, forever.
You'd have to track where compute is consumed, not just where hardware is sold. That's a fundamentally different and much bigger surveillance problem and every month it takes to solve, Chinese models keep training.
You can control where a chip is sold. Controlling where its power gets used, in every country, forever, is a different problem entirely and nobody has solved it yet.
Three takeaways, even if you never touch a chip
1. Borders don't contain compute anymore, and that cuts both ways.
The lesson under this whole story is that AI capability is fungible and location-independent. Power can be rented from anywhere, by anyone, instantly. That's the same force that lets a two-person startup in Lagos or Lahore access the exact same frontier compute as a company in San Francisco. The playing field is flatter than the headlines suggest, use it.
2. Any moat built on "they can't get access to X" is already leaking.
China just demonstrated, at the scale of a superpower rivalry, that access restrictions crumble the moment the thing can be rented instead of owned. If your competitive advantage is exclusive access to a tool, a model, or a dataset that someone else could rent their way around, it's weaker than you think. Build your moat out of things that can't be accessed remotely: your customers, your data, your distribution.
3. Enforcement always lags the workaround.
The gap between "someone finds the loophole" and "the authorities close it" is where a lot of business gets done. That's not an invitation to break rules, it's a reminder that in fast-moving markets, the advantage goes to whoever spots the gap first and moves before the rules catch up. China played that timing at the geopolitical level. The same dynamic plays out in every young market, including yours.
🔮 The Bottom Line
America spent three years and enormous political capital building a wall around its best AI chips. China didn't climb the wall or tunnel under it. It just rented a room on the other side and logged in.
The ban wasn't wrong, exactly it did keep the physical chips out of China. It was just aimed at the wrong thing. It controlled the hardware and assumed that controlled the power. In an era when compute is something you rent by the minute from anywhere on earth, that assumption had a hole in it big enough to drive a frontier model through.
Washington is patching it now. But the deeper lesson is already loose in the world: in the age of AI, what you own matters far less than what you can reach. Plan accordingly.
📧 Forward this to 3 entrepreneur friends who need to see this opportunity
