Here is what the valuation spreadsheets won’t tell you: the entire AI industry is built on a single, fragile assumption—that closed-source models can maintain a six-month lead forever.
I learned this lesson the hard way in 2017, when I spent 72 hours auditing the Gnosis Safe multi-signature contract. I found twelve critical flaws in a project that had raised $12 million on the promise of “trustless security.” The code wasn’t malicious—it was rushed. The same pattern repeats today: billions poured into AI labs, while open-source alternatives run on commodity hardware for 99% lower cost.
Context: The Fragile Architecture of Today’s AI Giants
Let’s step back. The current AI narrative rests on three pillars: proprietary data, massive compute, and talent concentration. Venture capitalists value OpenAI, Anthropic, and Inflection at $150B+ combined, assuming these moats will persist. But the crypto world has seen this movie before. In 2017, Ethereum’s “world computer” was supposed to be unassailable—until EOS, Tron, and later Solana ate its lunch. The pattern is identical: a centralized narrative meets decentralized commoditization.
Brian Armstrong of Coinbase recently warned that “AI bubble valuations mirror the dot-com and crypto manias.” He’s right, but his analysis misses the deeper structural force: open-source code’s ability to collapse the cost of a digital good to near zero. In crypto, we saw this with smart contract platforms (Ethereum’s fees drove users to L2s and competitors). In AI, the same force is pulling inference costs down by 99%, with models like Llama 3.1 and Mistral Large running on consumer PCs.
Core: My Technical Autopsy of the AI-Crypto Parallel
Based on my experience auditing DeFi protocols during the 2020 Summer, I can tell you that the most dangerous moment in a hype cycle is when everyone believes the cost structure is permanent. Here is what the charts won’t show you:
First, the scaling law inflection point is already here. Training a model on 1 trillion tokens yields diminishing returns per billion dollars. Open-source communities (with universities, startups, and hobbyists) are reproducing 90% of the capability of GPT-4 for $10 million instead of $1 billion. I’ve seen this in crypto: in 2018, building a decentralized exchange cost $5 million in engineering; by 2022, a Uniswap fork cost $50,000. The same commoditization curve is now hitting AI.
Second, inference cost is the new gas fee. In Ethereum, you had high gas fees during congestion—but L2s like Arbitrum cut costs by 90% while preserving security. In AI, deploying a model on a GPU cluster is like paying Layer 1 gas. Open-source models, quantized and optimized with vLLM or TGI, bring inference to “Layer 2” level: cheap, fast, and local. The user doesn’t care about the underlying architecture; they care about output quality and price. When price drops 99%, the old model—charging $0.01 per query—becomes extinct.
Third, fragmentation destroys unit economics. Nikhil Kamath of Zerodha said “countries will run domestic copies of AI models.” This is exactly what happened with crypto: after China banned exchanges, local versions of Uniswap and Binance clones proliferated. Each regional fork dilutes the value of the original product. If every country runs its own Llama 4, who pays OpenAI for its API? Only those requiring “specialized tasks” like discovering new physics—a market too small to support a $150B valuation.
Contrarian: What If the AI Bubble Bursts Differently?
Here is where my skepticism turns toward a more nuanced view. Not all bubbles burst the same. The dot-com crash wiped out companies with no revenue, but Amazon survived. In crypto, the 2018 bear market killed 95% of ICOs, but Bitcoin and Ethereum emerged stronger. The same could happen in AI: the pure-play model labs may collapse, but infrastructure providers (GPUs, data centers, energy) will thrive. This is my read of the market—not a blanket “AI is overvalued,” but rather a warning that value will shift from closed-source model providers to the open-source ecosystem and the hardware enabling local inference.
I’ve seen this shift before. In 2021, during the NFT bubble, I refused to mint speculative PFP projects. Instead, I launched “On-Chain Diaries”—a curated collection of 50 digital artifacts that told stories of Beijing’s interaction with blockchain. The project failed financially, but it taught me that authentic, small-scale utility survives hype. AI’s equivalent will be domain-specific, locally-deployed models that serve niche communities (e.g., an Indian farmer using a Kannada-language LLM for crop advice, not ChatGPT).
But there’s a blind spot in the doomsday narrative: security alignment. Open-source models are harder to regulate and more prone to misuse for deepfakes or automated cyberattacks. Governments may mandate API-level controls, preserving a premium for trusted, closed-source AI providers. Anthropic’s Constitutional AI may become a necessary insurance policy for enterprises, just as centralized exchanges still hold institutional trust despite DeFi’s superiority. This could keep some valuation—but not for all.
Takeaway: Follow the Fear, Not the Chart
I’ve been called a pessimist. But I prefer “structural realist.” The fear I see isn’t that AI will lose its value—it’s that the current winners are priced for perfection in a world where open-source gravity is accelerating. Five years from now, I believe we’ll look back on 2024–2025 with the same astonishment we view the 2017 ICO mania: how did we pay those multiples for something that could be replicated for pennies?
If you can’t build a closed-source model that stays six months ahead of open-source for less than $100 million, you don’t have a moat—you have a donation. The bubble will pop. But when it does, the infrastructure beneath will rise. Follow the fear, not the chart.
If you can read a smart contract, you can see the future. The same logic that made DeFi eat CeFi is now eating AI. Watch the costs, not the hype.
