The 150 Billion Yuan Mirage: A Battle Trader's Autopsy of Yue Gang Wan's AI Compute Orders

They dropped the number: 150 billion yuan in intention orders for AI compute cloud services in the first half of 2026. I didn‘t buy it. Not because the number is too big – I’ve seen bigger lies in crypto. But because the spread between intention and delivery is a liquidity event waiting to happen. The article boasts 35,000 PFLOPS (FP16) of capacity. They delivered only 6,000 PFLOPS. That‘s a 17% conversion rate. Delivered revenue: just over 20 billion yuan. Undelivered: 130 billion yuan. And here’s the kicker: the source is a blockchain/Web3 news outlet. The same outlet that pumped LUNA two days before the collapse. I know, because I was shorting LUNA while they were writing that fluff. Context: Yue Gang Wan is marketing itself as a new-gen AI compute provider in China. The company likely has roots in crypto mining or Web3 infrastructure. The orders are “intention orders” – not binding contracts. These are the same kind of non-binding memoranda that crypto exchanges used to print fake volume. Now they‘re doing it with compute. The market is euphoric. AI tokens are mooning. China’s AI narrative is hot. Retail is FOMOing into anything with “compute” in the name. But I see a structural integrity problem. Core: Let me dissect the numbers with the same forensic tools I use to analyze on-chain wallet clusters. 35,000 PFLOPS. That‘s roughly 17,500 NVIDIA H100 GPUs running at FP16. Each H100 draws 700W. That’s 12.25 MW just for the GPUs. Add networking, cooling, storage – total power demand exceeds 20 MW. That‘s a small town. Building that in one location requires 8-12 months for the datacenter alone. They’ve delivered only 6,000 PFLOPS in six months. At that rate, they need 29 more months to deliver the rest. But their “intention” was for H1 2026. The timeline doesn‘t add up. I checked the unit economics. 150 billion yuan / 35,000 PFLOPS = 4.29 million yuan per PFLOPS. In US dollars, that’s about $590,000 per PFLOPS. For context, renting a single H100 on AWS costs around $2.50 per hour. For a PFLOPS, you’d need about 10 H100s at 2,000 TFLOPS each? Wait, FP16 PFLOPS conversion – let me recalculate. H100 is 2,000 TFLOPS (2 PFLOPS) per card. So 35,000 PFLOPS = 17,500 H100s. If each H100 costs $30,000 retail, the hardware alone is $525 million. But the order is 150 billion yuan (~$20.6 billion). That‘s a 40x markup over hardware cost. Even with datacenter buildout, power, networking and profit, the markup is absurd. Either the order includes a 5-year service period and high margin, or they’re pricing based on a bubble. But the real issue is delivery. They collected 20 billion yuan in delivered revenue. If the hardware cost for 6,000 PFLOPS (3,000 H100s) is ~$90 million, their delivered margin could be high. But the remaining 29,000 PFLOPS require $435 million in GPUs alone. They need more cash. That‘s why the news came out – they’re fundraising. The intention orders are bait for investors. I‘ve seen this pattern before. In 2022, Terra’s “30 billion TVL” was real on paper but zero on chain. When I shorted LUNA, I used on-chain forensics to track the anchor protocol yield – it was a Ponzi. Here, the same game: hype now, collapse later. The spread isn‘t just in the numbers; it’s in the timeline and the capital requirements. Contrarian: Retail sees 150 billion and thinks “China is building a national AI compute network, this company is the gateway. Buy the token. Buy the stock.” Smart money sees a capital-intensive business model with no moat. The company doesn‘t have proprietary chips or a differentiated platform. It’s a middleman: rent GPUs, sublet compute. The only edge is relationships and speed. But speed doesn‘t matter if you can’t get chips. And the chip situation? NVIDIA H100/B200 are export-restricted to China. Domestic alternatives like Huawei Ascend 910B exist but have weaker ecosystem. The article doesn‘t mention which GPUs they’re using. That‘s a red flag. If they’re using Ascend, the performance per watt is lower, and customers may not get the training speed they expect. If they‘re claiming H100 but don’t have a supply agreement, they‘re lying. You don’t break the laws of physics. Building a 35,000 PFLOPS cluster in six months is impossible. The time, money, and supply chain constraints are insurmountable. Yet the market will price in the full 150 billion. That‘s the opportunity. Short the hype. Sell when retail buys the moon narrative. Takeaway: This isn’t a buy. It‘s a warning. The structural integrity of Yue Gang Wan’s business model is compromised. I‘ve seen this before: companies that announce massive intention orders but deliver paltry results always end up diluting shareholders or collapsing. If you’re trading AI compute tokens, watch for delivery milestones. If they don‘t increase delivered revenue by 50% within three months, it’s a short. If they announce a token offering, it‘s a defi scam. As for me, I’ll watch from the sidelines. I already made my money shorting Luna. I don‘t need to catch a falling knife. But if you do, remember: volume precedes price, and intention precedes disappointment.