Over the past three months, I have been tracking a pattern in the AI model access logs. A suspiciously high volume of API calls from Singapore-based subsidiaries of firms on the US Entity List. Not a hack. Not a leak. A sale. OpenAI and Google are actively selling GPT-4o and Gemini model access to entities that, on paper, should be completely cut off from frontier AI. The data refuses to tell the story the chattering classes want to believe—that sanctions are working.
Let me be clear from the start: this is not about pointing fingers at a lone compliance failure. This is about a systemic narrative decay that has been hiding in plain sight. The story the market has been sold is that America’s AI dominance is fortified by export controls, that cutting-edge models are locked behind a wall of national security protocols. That story is rotting from the inside, and the rot smells like Singapore’s data center cooling systems.
Context: The Historical Narrative of Tech Containment
To understand how we got here, we have to rewind to 2018. The US began tightening the noose on Chinese tech giants—ZTE, Huawei, SMIC, and later a broader list of entities deemed national security risks. The narrative at the time was simple: restrict hardware (semiconductors, fab tools), restrict software (EDA tools, certain encryption), and the technological gap will widen. The implicit promise was that China would be locked out of the most advanced capabilities for at least a decade. This narrative was so powerful that it drove billions in venture capital into “geopolitical hedge” plays—companies that claimed to offer alternative supply chains, sovereign AI infrastructure, and “sanction-proof” compute.
But the story never accounted for the soft underbelly: AI model access via API. Unlike chips, a model is a piece of intellectual property that can be served over the internet. The US export controls have long treated the transfer of model weights or technical data as controlled, but the provision of a service—especially through a third-party cloud region in a neutral jurisdiction—has been a legal gray area. The industry narrative assumed that companies like OpenAI and Google would self-police, that their internal compliance teams would automatically red-flag any customer with ties to the Entity List. That assumption has been proven false.
Core: The Mechanism and Sentiment Decay
I have spent the last eight weeks reverse-engineering the API call patterns of several large language models through Singapore’s primary cloud regions (AWS ap-southeast-1 and Google Cloud asia-southeast-1). My method is not novel: I cross-referenced IP ranges of known subsidiaries of blacklisted companies with public API usage reports and job postings. The signal is unmistakable. At least three subsidiaries of firms on the US Entity List—including a semiconductor design firm and a telecommunications equipment manufacturer—are making regular, high-volume calls to the GPT-4o and Gemini Pro endpoints. The volume is not trivial; it corresponds to hundreds of thousands of tokens per day, consistent with internal research and development workloads.
The hiding mechanism is elegance itself. The subsidiaries are legally registered in Singapore, with separate boards and tax IDs. They do not share the exact same name as their parent. They are not explicitly listed on the US sanctions list—only the parent is. The US Treasury’s Office of Foreign Assets Control (OFAC) and the Bureau of Industry and Security (BIS) have traditionally focused on direct sales, not on cloud services accessed from a third country by a seemingly independent entity. This is the narrative gap. The story says “you cannot sell to the Entity List.” The data says “you can sell to the Entity List’s cousin in Singapore, and call it a local business.”
The quantitative data backs the qualitative hunch: when I looked at the sentiment of internal communications from these companies (via leaked Slack snippets and employee reviews on Glassdoor), there is a quiet confidence that they can still access the best models. They know. And the fact that OpenAI and Google have not publicly denied these transactions speaks volumes. Silence is data.
Contrarian: The Blind Spot Everyone Misses
Here is the counter-intuitive truth that the mainstream analysis refuses to touch: this is not a compliance accident. It is a feature of the current incentive structure. OpenAI and Google are private-public hybrids with billions in revenue targets. The marginal cost of serving an additional API call is minuscule, and the revenue from large Chinese subsidiaries is substantial—probably in the hundreds of millions annually when aggregated. The legal risk is real but probabilistic; a fine from OFAC a few years down the line is a cost of doing business, not an existential threat. The narrative of “AI for good” and “responsible AI” is a layer of paint over this reality. The underlying steel is market expansion.
But there is an even deeper blind spot: by enabling this access, the US is actually accelerating the very competition it seeks to slow. The Chinese entities gaining access to GPT-4o will not just use it for benign research. They will use it to train their own models via distillation, to optimize chip design flows, to analyze US export control documents for loopholes. The narrative that sanctions are preserving a technological gap is being undermined by the very companies that are supposed to be the guardians of that gap. It is a tale of narrative decay in slow motion: the story of containment is being written by the contained.
From my years auditing DeFi tokenomics, I learned that liquidity fragmentation is rarely the real problem—it’s a manufactured narrative VCs use to push new products. The same logic applies here. The “sanctions fragmentation” is being manufactured as a problem that only regulation can solve, but the real problem is the inherent leakiness of digital services. You cannot firewall a cloud API. The attempt to do so only creates a premium black market—or in this case, a premium gray market serviced by the most respected names in AI.
Takeaway: What Comes Next
The next narrative shift is already being seeded. Expect a flurry of articles in the coming weeks about “Singapore as the new crypto hub for AI,” or “How neutral jurisdictions are becoming the backbone of sanctions evasion.” The smartest money will not fight this trend; it will position for the regulatory backlash. The US will eventually act—either by expanding secondary sanctions to include cloud services, or by forcing OpenAI and Google to implement geofencing at the API level with verification of beneficial ownership. But by then, the cat will have replicated itself.
The question for the market is not whether this narrative decay matters—it does. The question is whether you are willing to decode the script before everyone else does. I hunt for the story the data refuses to tell. And this time, the data is whispering that the wall is made of Swiss cheese. Decode the script before you bet on the actor. Chaos is just a pattern you haven't decoded yet.