The 15 Billion Yuan Question: When AI Compute Intent Orders Smell Like DeFi’s Ghost

But here’s the number the press release won’t tell you: 87 percent exists only in the realm of intent. Yuegangwan Intelligent Computing just announced 15 billion yuan in AI compute cloud service intent orders for the first half of 2026. The headline screams $2 billion-plus of future revenue. But the footnote screams something else: only 20 billion yuan delivered — a 13 percent conversion rate. I hunt for the story the data refuses to tell. And this data tells me we are watching the same playbook that burned ICO investors in 2017 and DeFi degens in 2020. Context: The AI Compute Narrative Reaches Crypto’s Ear The AI gold rush has flooded into crypto. Decentralized compute networks like Render, Akash, and io.net have seen their valuations soar on the promise of “filling the GPU gap.” Venture capital is pouring into centralized compute providers that promise to bridge Web2 and Web3. Yuegangwan Intelligent Computing — a company with a name that hints at a Greater Bay Area connection and a background shrouded in blockchain whispers — fits this mold perfectly. The company announced that it had secured intent orders totaling 35,000 PFLOPS (FP16) of AI compute power, valued at over 15 billion yuan. Deliveries so far: 6,000 PFLOPS and approximately 20 billion yuan in recognized revenue. On the surface, it looks like a rocket ship. But the word “intent” is doing heavy lifting. In crypto, we know what intent means: a promise with no binding obligation, a non-binding memorandum of understanding, a press release dressed as a contract. Core: Deconstructing the Narrative Mechanism Let’s reverse-engineer this “success.” First, the numbers don’t hold up to basic sanity checks. 35,000 PFLOPS of FP16 compute is an enormous amount. To put it in perspective: training a GPT-4-class model requires around 2.1e25 FLOPs. At that scale, you could train it in less than a day. But building that compute requires roughly 17,500 NVIDIA H100 GPUs (if we assume 2 PFLOPS per card in FP16). That’s $500 million worth of hardware at today’s prices — before networking, cooling, power, and data center build-out. The company says it has delivered only 6,000 PFLOPS. That implies about 3,000 H100s deployed in six months. Not a small feat, but it’s 17 percent of the target. The gap is 29,000 PFLOPS — 14,500 GPUs. Where do they come from? In 2017, I spent six weeks reverse-engineering ICO vesting schedules. I discovered a project that promised 500,000 TPS but had only deployed a testnet with 50 nodes. The ratio was exactly the same: announce big, deliver small, raise more capital. I called it the Tokenomics Paradox — mathematical elegance overriding human greed. Here, the paradox is compute: you can promise the moon if you haven’t bought the rocket fuel yet. Second, the pricing math reveals a margin squeeze. At 15 billion yuan for 35,000 PFLOPS, the unit cost is about 428,000 yuan per PFLOPS. If we assume a three-year contract (common in cloud deals), that’s about 143,000 yuan per PFLOPS per year. Compare that to a typical H100 instance on AWS: roughly $30 per hour for an 8-GPU instance (16 PFLOPS), which works out to about $1.6 million per PFLOPS per year if you run 24/7. The Yuegangwan price is an order of magnitude lower — which means either they are selling at razor-thin margins, or the “intent” is not for three years but for much shorter periods, or the service is not guaranteed to be H100-class. Third, the capital requirement to deliver the remaining 130 billion yuan worth of compute is staggering. Even at 50% gross margin, they would need to invest over 65 billion yuan in infrastructure. Where is that money coming from? The company’s funding history is opaque. My suspicion: this press release is targeted at venture capital or even a future SPAC listing. I saw the same pattern during DeFi Summer 2020, when protocols announced “$10 billion TVL” only to crater when liquidity incentives ended. During my DeFi Liquidity Illusion Exposé, I discovered that Aave’s projected APY was driven by volatile token emissions, not real revenue. Here, the emissions are “intent orders” — tokens of promise that can vanish the moment the market turns. But the most telling detail: the article source is a blockchain/Web3 news outlet, not a traditional financial wire. That means Yuegangwan Intelligent Computing is pitching itself to the crypto audience. The company likely has a token or plans to issue one. The narrative is clear: we have massive compute demand; buy our token to participate in the future of AI. This is the exact same playbook as Filecoin, Storj, and every “decentralized infrastructure” project that has ever existed — except this one is centralized and opaque. Chaos is just a pattern you haven’t decoded yet. And the pattern here is narrative decay: the gap between the story (15 billion intent) and reality (13% delivered) will widen as more data emerges. Contrarian: Is This Actually a Good Sign? Let me play the devil’s advocate. Perhaps the intent orders represent genuine demand from Chinese enterprises desperate for GPU access. The US export restrictions have created a shortage of H100 and B200 chips in China, driving up prices and demand for any compute provider that can deliver. Yuegangwan might have secured a supply chain advantage — maybe through Huawei’s Ascend chips, which are less powerful but available at scale. The 6,000 PFLOPS delivered could be a proof of concept, and the 29,000 PFLOPS could be locked in with signed contracts that are simply not disclosed. If that’s the case, the company could become a top-tier AI compute provider in China within 18 months. The hype would be justified. But the odds are stacked against it. My experience tracking Terra’s narrative decay — that 100,000-reader autopsy I wrote in 2022 — taught me that when the gap between promise and delivery is this large, the narrative decays faster than the code. Terra promised algorithmic stability; it delivered a death spiral. Yuegangwan promises 35,000 PFLOPS; it delivered 6,000. Moreover, intent orders are non-binding. In the deal world, fewer than 40% of MOUs convert into binding contracts. Combine that with the need for massive capital expenditures, and you get a business that depends on continuous fundraising. If the crypto market turns bearish, if AI demand softens, or if export controls tighten, those intent orders evaporate. Decode the script before you bet on the actor. The script here is: announce huge numbers → raise capital → build infrastructure → deliver only a fraction → announce more intent orders to keep the story alive. It’s the same script as a failing ICO, just with GPUs instead of tokens. Takeaway: The Next Narrative Shift So what comes next? The market will eventually wake up to the metric that matters: delivered compute utilization vs. announced intent orders. Projects that report actual revenue from active compute will outperform those relying on inflated MOUs. The narrative will shift from “compute capacity booked” to “compute capacity used.” Watch for Yuegangwan’s next quarterly update: if delivered compute hasn’t doubled, the story is already decayed. I don’t just read the headlines; I read the footnotes. And the footnote here reads: “Intent orders are subject to final contract execution.” That’s a polite way of saying: they might never happen. I hunt for the story the data refuses to tell. The data says 15 billion yuan. The story it refuses to tell is that 87% of that is vapor. Until the company proves otherwise, treat this as narrative, not reality.