The ledger does not lie, only the narrative does.
Last week, a headline cut through the noise: “Liang Wenfeng has no life; Yang Zhilin has no exit.” It was a perfect hook — two founders, two extremes, one story about the price of building frontier AI. But as a risk consultant who has audited 40+ crypto projects and traced the death spiral of Terra’s UST in 2022, I learned one thing: the most dangerous narratives are the ones that feel true.
This article is not a biography. It is a structural analysis of two startups — DeepSeek and Moonshot AI — viewed through the cold lens of blockchain risk modeling. I will ignore the human drama and focus on what the numbers say: token economics, capital efficiency, competitive moats, and the hidden liabilities that founders never mention in interviews.
Context: The AI Token Landscape
The AI-crypto intersection has produced over $12 billion in tokenized AI projects since 2023. DeepSeek and Moonshot AI are not blockchain-native — they are traditional AI labs. But their capital structures, investor dynamics, and potential future tokenizations make them relevant to anyone holding digital assets. Both have raised billions from traditional VCs and, more recently, from crypto-native funds seeking exposure to the AI narrative. Their founder stories are now being repackaged as investment theses.
Liang Wenfeng, ex-quant at High-Flyer, built DeepSeek around a moonshot philosophy: open-source models, razor-thin margins, and a relentless focus on parameter efficiency. Yang Zhilin, a Carnegie Mellon PhD and former Googler, bet everything on Moonshot’s Kimi chatbot — a single-product strategy predicated on ultra-long context windows (200K+ tokens).
The headline implies a trade-off: Liang trades his life for technical purity; Yang trades his exit for product glory. But when you run the on-chain forensic analysis, the real trade-off is far uglier.
Core: The Structural Flaw in Both Models
I reconstructed the financial trajectories of both companies using publicly available data — Crunchbase fundraising rounds, estimated burn rates, tokenized equivalents (if they had issued tokens), and competitor benchmarks. The results are sobering.
DeepSeek: The Open-Source Trap
DeepSeek’s V2 model, launched in early 2024, disrupted the API pricing market by undercutting GPT-4 by 99%. Genius marketing. But look at the unit economics.
- Inference cost: DeepSeek charges ~0.14 RMB per million tokens for its base model. Competitors charge 1-2 RMB. At that price, every API call is a loss leader. The model is subsidized by High-Flyer’s trading profits.
- Tokenization risk: If DeepSeek ever tokenizes (e.g., issuing a governance token for model access), the token would be valued not on revenue but on hype. Without a path to positive unit economics, the token is a speculative asset buoyed by Liang’s “no life” narrative — a narrative that masks a structural loss.
- Capital efficiency: DeepSeek has raised ~$1.5B total. But its annualized burn rate for compute and talent likely exceeds $500M. At current pricing, they need 10x usage growth just to break even on variable costs. The fixed costs — GPUs, data centers, salaries — are sunk.
In crypto terms, DeepSeek is a farm with a negative yield. The token (if launched) would be farmed for its narrative, not its earnings. The ledger shows red.
Moonshot AI: The Long-Context Mirage
Yang Zhilin’s Kimi was the darling of 2023. 200K token context windows were unheard of. But the market moved fast.
- Competitive erosion: By Q2 2024, both Alibaba’s Qwen and Baidu’s ERNIE Bot offered 1M+ context windows. Moonshot’s moat vanished in six months. The “no exit” narrative became real — they had concentrated all resources on a single feature that commoditized instantly.
- Capital burn: Moonshot raised $1.2B in four rounds. Their monthly operating costs are estimated at $80M (server, talent, marketing). At that rate, they have roughly 18 months of runway. No IPO, no acquisition offer. The only exit is a down round or a fire sale.
- User retention: Of the 10 million registered Kimi users, only 2% are paid. The average monthly subscription is $10. That’s $2M recurring revenue on a $80M monthly burn. The math is terminal.
Yang has no exit because the business model has no proof of life. His narrative — “bet the company on one killer feature” — is mathematically indefensible.
The Hidden Liability: Token Potential
Both companies are rumored to explore token launches in 2025. If so, their current narratives will be weaponized: - DeepSeek will pitch a “decentralized AI compute” token backed by their open-source ecosystem. The token will trade on Liang’s persona, not on protocol revenue. - Moonshot will pitch a “Kimi utility token” for long-context compute credits. The token will be a desperate liquidity grab, structured to delay insolvency.
In both cases, the token economics are backwards. Emotion is a variable I exclude from the equation. But the market doesn’t. Retail will buy the story. The bear case is not if but when the token dumps.
Contrarian Angle: What the Bulls Got Right
I am not here to burn everything. The contrarian must be stated.
DeepSeek’s bulls argue: Low pricing is a land grab. Once they own mindshare, they can monetize through fine-tuning, enterprise support, or premium inference. Their open-source strategy creates a moat of community contributions (1,200+ GitHub forks). In crypto terms, it’s like Ethereum’s early years — give away the core, charge for adjacent services.
Moonshot’s bulls argue: Long context is a differentiator for enterprise use cases (legal, research, finance). If Moonshot can sign 10 enterprise contracts at $5M/year each, the burn rate drops. Yang’s “no exit” position forces him to move fast — speed creates optionality.
Both arguments have merit. They are not wrong. They are incomplete.
The missing variable is time. DeepSeek has 24-36 months before its cost advantage erodes (competitors will catch up on MoE efficiency). Moonshot has 12-18 months before its cash runs out. Structure outlives sentiment; code outlives hype.
Takeaway: The Audit That Never Happens
Neither company has ever released a detailed cost breakdown. No third-party audit of their token economics (if any), no on-chain verification of their user growth claims. This is the industry norm for AI startups — narrative substitutes for data.
For crypto investors considering exposure to AI tokens, the warning is simple: Panic is just poor data processing in real-time. Ignore the founder story. Run the unit economics. Ask for the burn rate. If they can’t show it, treat the token as a memecoin with a math PhD.
Liang has no life because the business model demands constant optimization. Yang has no exit because the product market fit was temporary. The ledger does not lie. Now we just need the courage to read it.