Tracing the liquidity ghosts through the ICO fog.
DeepSeek plans to list on Shanghai’s STAR Market by Q2 2027. The news broke. Markets blinked. The AI winter narrative cracked, just slightly. But read between the lines: this is not a victory lap for open-source AI. It’s a liquidity event dressed in technical ambition. A macro signal wrapped in a valuation spreadsheet.
I’ve seen this pattern before. In 2017, I spent months modeling on-chain fund flows during the Ethereum ICO boom. Sixty percent of initial liquidity recycled within four hours. Fake organic demand. The crash came when the liquidity ghosts evaporated. Now, DeepSeek’s IPO feels eerily similar—a massive capital raise predicated on future revenue that does not exist today. The only difference is the asset class: AI models instead of ERC-20 tokens.
Context: The DeepSeek Thesis
DeepSeek is not just another Chinese AI lab. It’s a technical outlier. The company built DeepSeek-V3 and R1, models that rival GPT-4o and Claude 3.5 on math, code, and reasoning benchmarks. They did it with a fraction of the compute—thanks to Mixture-of-Experts (MoE) architectures and aggressive training optimizations. Their API prices are 1/50th of OpenAI’s. They open-source nearly everything. Developers love them. VCs crave them.
But here’s the rub: DeepSeek has no stable B2B revenue. No enterprise SaaS product. No clear path to profitability. Their model weights are free. Their API is subsidized. Their only moat is talent and algorithmic efficiency. That’s a fragile foundation for a potential $5 billion valuation.
The IPO prospectus will likely highlight “model development” and “compute infrastructure” as use-of-proceeds. Standard playbook. But behind the scenes, the real driver is liquidity: early backers (including the quant hedge fund 幻方量化) need an exit. The STAR Market offers a premium multiple for “national AI champions.” The timing is deliberate—slip before the next US chip ban hits harder.
Core: The Macro-Liquidity Mirage
DeepSeek’s story is a macro-liquidity play, not a technology one. Let’s break the components.
First, the global liquidity map. Since mid-2023, the Fed’s reverse repo facility drained, M2 money supply expanded slowly, and risk assets rallied. AI stocks led the charge. In China, the government is flooding the market with cheap credit to revive the economy. The STAR Market is a beneficiary of this liquidity wave. DeepSeek will ride that wave—until the crest breaks.
Second, the arbitrage of attention. DeepSeek’s open-source strategy is brilliant for ecosystem growth, but terrible for unit economics. Every token served at $0.27 per million costs real electricity and GPU hours. The company is effectively buying market share with investor money. This is not sustainable unless they can pivot to a “freemium” model—like offering advanced fine-tuning, private deployment, or compliance audit services for enterprises. But such pivots take time. The IPO window is 2027. That’s only two years away.
Third, the chip constraints. US export controls throttle DeepSeek’s access to NVIDIA H100/B200 clusters. They rely on Huawei Ascend chips (910B, 910C) and leased overseas cloud (Singapore, Middle East). The IPO cash will fund massive domestic compute procurement. But if the next generation of US sanctions blocks even the leased cloud path, DeepSeek’s training pipeline hits a wall. The yield of open-source is not measured in dollars, but in attention—and attention does not pay for H100s.
Fourth, the valuation disconnect. Compare DeepSeek to listed AI peers in China: SenseTime (valuation cratered from HK$130B to $30B), Cambricon (still unprofitable at $20B market cap). DeepSeek will need to show annual revenue >$500M by 2026 to justify a $5B+ valuation. Current estimates put their API revenue at <$50M. The gap is massive. Investors will be buying a promise, not a business.
Tracing the liquidity ghosts through the ICO fog. The ghosts are the recycled VC money moving from one AI startup to another, creating an illusion of organic growth. DeepSeek’s IPO is the liquidity event that allows the cycle to continue. But when the next macro headwind arrives—a recession, a credit crunch, a geopolitical flashpoint—the fog clears. Then only real revenues survive.
Contrarian: The Decoupling Thesis That Everyone Misses
The bullish consensus says DeepSeek will become China’s OpenAI. The contrarian view: DeepSeek is a liquidity shelter, not a long-term winner. The decoupling may happen not between China and the US, but between hype and fundamentals.
Consider the Terra collapse. In 2022, I published a critical analysis of Terra’s seigniorage mechanism three days before the crash. The logic was simple: algorithmic stability is fragile; when confidence breaks, the death spiral is inevitable. DeepSeek’s model is not algorithmic, but it shares a structural flaw: it relies on continuous capital inflows to subsidize its low-cost operations. Label it “open-source AI as a service,” but the economics are identical to a growth-at-all-costs startup. Once the IPO proceeds are spent, and the next funding round is not guaranteed, the business must stand on its own. If it can’t, the stock will collapse.
Another blind spot: regulation. China’s Cyberspace Administration requires AI models to pass safety assessments, avoid “improper content,” and ensure political alignment. DeepSeek has done this for its Chinese products. But its open-source weights can be downloaded and fine-tuned by anyone—including for disinformation or deepfakes. The company may face regulatory blowback if its models are misused. That risk is not priced into the IPO hype.
Censored compute is the new counterparty risk. If DeepSeek cannot scale its training due to chip sanctions, its technical lead erodes. Competitors like Zhipu AI, Baidu, or even ByteDance will catch up. The IPO cash only delays the inevitable commoditization of base models.
Takeaway: Positioning for the Cycle
DeepSeek’s IPO will be a landmark event. It will attract retail frenzy, state-backed funds, and strategic investments from cloud providers. But for the macro-aware investor, the real signal is not the IPO itself—it’s the liquidity cycle it represents. We are late in the bull market for AI hype. The next bear phase will punish companies without real revenue. DeepSeek is a high-conviction short candidate for the 2028-2029 timeframe, or a risky long for the momentum traders.
For crypto natives, the lesson is stark: the same patterns that drove ICO mania and DeFi yield farming are now driving the AI IPO. The names change. The math does not. When the liquidity ghosts vanish, who will be left holding the bag?
I’ll be watching the STAR Market filings for signs of insider selling, revenue disclosure, and chip supply agreements. Until then, the fog remains thick.