The Menlo Ventures partner posted a number: $2.6 billion. That is the purported 2024 combined annual revenue of five Chinese AI companies—Zhipu AI, DeepSeek, Kling (Kuaishou), Moonshot (Kimi), and MiniMax. The crypto twitter machine latched onto it as proof of Chinese AI dominance. I saw a structural vulnerability, not a strength.
From my experience auditing the Curate smart contract in 2017, I learned to look for re-entrancy—a function that calls itself before its state updates. These revenue numbers exhibit exactly that pattern. They call themselves revenue but are funded by the same capital they claim to generate. The economic re-entrancy is that the cash used to pay for inference is the same cash given by VCs to subsidize API calls. The loop never closes.
Context: The Five Pseudo-Revenue Engines
The article was sparse on details but identified five players. Zhipu ($1B), DeepSeek ($0.5B), Kling ($0.5B), MiniMax ($0.4B), Moonshot ($0.2B). Combined: $2.6B. Ranking them among the top global AI revenue generators.
Here is what the headline omitted: - Zhipu sells heavily to government entities. Those contracts are lumpy, political, and carry multi-year payment cycles. Real cash flow? Unknown. - DeepSeek prices its API at $0.5 per million tokens for the flagship model. At $0.5B revenue, that implies over 1 quadrillion tokens processed in 2024. Even with aggressive caching and quantization, the inference compute cost alone would exceed 80% of that revenue. Negative gross margin. - Kling’s revenue likely includes internal transfer pricing from Kuaishou’s advertising business. Not real external demand. - Moonshot’s $200M is an outlier compared to its user base. Its long-context model is a niche differentiator, but unit economics remain unproven. - MiniMax pulls in $400M from social and entertainment applications, but churn rates in chat apps are notorious.
Logic is immutable; incentives are the variable. The incentive for each company is to maximize reported revenue for the next fundraising round. The incentive for the VC is to propagate the narrative to mark up their portfolio. The incentive for Chinese state media is to showcase AI achievement. None of these incentives align with building a sustainable business.
Core: The Defect-Detection of Unit Economics
I built a simple model in Python, as I did for MakerDAO’s collateral crisis in 2020. Input: revenue, API price per token, average inference cost per token (based on GPU rental rates). Output: gross margin.
For DeepSeek: - Revenue = $500M - API price = $0.5/1M tokens for flagship model (as per public pricing) - Implied tokens served = 1e18 tokens? Actually: $500M / $0.0000005 per token = 1e15 tokens = 1 quadrillion tokens. - Compute cost: An H100 runs at roughly $2/hr. Each H100 processes ~2000 tokens/second for inference. 2000 tokens/sec 3600 = 7.2M tokens per hour. Cost per million tokens = $2/7.2M = $0.28. - That leaves $0.22 per million as gross profit before overhead. $0.22 1e15/1e6 = $220M gross profit. Gross margin = 44%. - That seems plausible. But this assumes 100% utilization of rented H100s. Realistically, utilization rates for random API calls are 30-50%. Adjusting for 40% utilization raises cost to $0.70 per million tokens—negative margin. - DeepSeek uses their own proprietary infrastructure. They likely have lower costs. But they also offer even cheaper models.
Structural integrity precedes market sentiment. The company's reported revenue is a function of how much compute they can burn. The more they serve, the more they lose. That is not a business. That is a subsidy.
Now apply the same to Moonshot. Their API is $1/1M tokens for the 128K context model. Similar math yields thin positive margins only if utilization is high. But Kimi's user base is mostly free-tier with heavy usage. The free tier converts to revenue only via enterprise API. Enterprise adoption? Still early.
Zhipu’s government contracts: project-based revenue often carries 20-30% margin after customization. But these projects rarely scale. Each new contract requires a new integration. The revenue is real, but the growth is linear, not exponential.
The aggregate $2.6B number gives a false sense of scale. The real metric is gross profit per unit of compute. And that metric is dangerously low.
History repeats not in price, but in pattern. We saw this in the 2021 crypto bull run. Protocols inflated their TVL through token incentives. The revenue looked real. Then the incentives stopped, and the TVL cratered. These AI companies are doing the same: burning investor capital to buy revenue. The only difference is the asset being subsidized is inference compute instead of liquidity.
Contrarian: The Decoupling Thesis—Crypto as the Escape Valve
The market consensus is that Chinese AI companies are on a path to profitability once they achieve scale. That is standard VC logic. I disagree. The structural flaw is that the underlying compute is a commodity priced by the global market. No amount of scale will change the fact that an H100 hour costs the same for everyone. The only way to win is to own the compute at zero marginal cost—that is, to have a tokenized compute network where providers compete globally.
Enter crypto. Projects like Akash, io.net, and Render provide decentralized compute marketplaces. They allow arbitrary providers to sell GPU time at market clearing prices. No centralized entity needs to subsidize infrastructure. The network effect comes from the token, not from central balance sheets.
Imagine DeepSeek migrating its inference load to a decentralized compute network. The cost per token would drop by 30-50% due to competitive bidding. More importantly, the company could align incentives: chain users pay for compute with tokens, and token holders share in the network's growth. The revenue becomes transparent, on-chain, and auditable. No more hidden subsidies.
Logic is immutable; incentives are the variable. In the current model, the incentive is to fake revenue. In a tokenized model, the incentive is to attract real usage that burns tokens. The difference is structural.
The contrarian take: The $2.6B figure is a death sentence. It signals that these companies have chosen the path of maximum centralization and vulnerability to capital markets. The more they grow using this model, the harder the eventual correction. The smart money will rotate to crypto-native compute companies that have no such re-entrancy.
Takeaway: The Audit Passed, But the Economics Failed
I have seen this movie before. In 2022, Terra’s LUNA had a $40B market cap. The revenue—from the anchor protocol—appeared sustainable. Then the audits were released. Then the incentives stopped. Then zero.
The Chinese AI companies' revenue is essentially the same "stablecoin yield" of the AI world. It pays high returns (revenue) in the short run but depends entirely on continuous capital injection. The moment VC funding slows, the API prices will need to rise by 300% to cover costs. At that point, users will switch to open-source models run on decentralized networks.
Based on my experience modeling the MakerDAO collateral crisis and the Terra collapse, I can see the exact failure mode: the cost of capital eventually exceeds the margin on compute. When that happens, the illusion breaks.
The only sustainable path is to decouple revenue from centralized capital. Tokenize the compute. Let markets set the price. Let crypto provide the transparency.
For now, the $2.6B headline is a mirage. The real signal is the structural unsuitability of centralized AI business models in a world where compute is ever-cheapening. The future belongs not to the companies with the biggest subsidies, but to those with the most aligned incentives.
History repeats not in price, but in pattern. We saw it in crypto. We are seeing it in AI. The only question is how many more quarters of fake revenue it will take before the correction. My defect-detection says: soon.
Position accordingly.
