Billionaire AI Warnings Echo Crypto's Past: Is the Bubble Already Bursting?

On a recent stage in Bengaluru, two titans of industry—Nithin Kamath, founder of India's largest brokerage Zerodha, and Brian Armstrong, CEO of Coinbase—delivered a stark warning to the AI investment community. Kamath said he sees 'no reason to pay current valuations for private AI companies," comparing the frenzy to the dot-com crash and crypto bubble. Armstrong doubled down: "The costs of running open-source models are now 99% lower than their proprietary counterparts, and the gap in capability is only six months. This is unsustainable." The audience, a mix of venture capitalists and tech founders, fell silent. I know that silence. I heard it in 2017 when I flagged the token distribution vulnerabilities in EOS’s whitepaper, only to watch investors pile in anyway. This is a narrative that is about to collapse—and anyone who has been through the ICO wild west can smell it. The core of the argument is not new, but the data is. Open-source models like Llama 3 and Mistral are closing in on GPT-4 and Claude 3.5 in benchmark tests, while costing a fraction to run. Armstrong pointed out that top AI labs spend billions on training runs, yet their inference pricing is being undercut by open-source alternatives that can run on standard consumer hardware. This mirrors what I saw in DeFi Summer 2020: Uniswap’s AMM offered a permissionless liquidity mechanism that undercut centralized exchanges, not by being better, but by being cheap and good enough. The market shifted within weeks. The same dynamic is unfolding in AI. The difference is the stakes—and the valuations. Private AI companies command multi-billion-dollar valuations based on the assumption of a global, unified market with high margins. But Kamath argued that the future is fragmented: nations will build their own models, tokens, and energy infrastructure. That kills the global monopoly thesis. My experience auditing blockchain protocols taught me to look for structural vulnerabilities hidden under hype. Here, the vulnerability is in the unit economics. A closed AI company spends $10 billion on a training run for GPT-5. It charges $20 per million tokens for API access. An open-source fork of Llama 4, fine-tuned by a community, runs on a $5,000 GPU cluster and charges $0.20 per million tokens for hosted inference. The quality difference? Perhaps 5% on MMLU. For 90% of enterprise use cases, that 5% is irrelevant. The only safe havens are hyper-specialized domains—discovering new physics or writing cancer drug candidates—but those are tiny markets. The rest will flee to cheaper alternatives as soon as trust is established. And trust, as I keep saying, is the only currency that matters. But the contrarian angle is where it gets interesting. The same fragmentation that threatens centralized AI giants creates an opening for blockchain-based AI infrastructure. Decentralized compute networks like Akash Network or io.net are already offering GPU access at 30-50% below AWS. If open-source models become the standard, the next battle is not for who trains the best model, but who provides the cheapest, most resilient inference layer. This is a crypto-native problem. I’ve written before about how liquidity fragmentation in DeFi was a manufactured narrative; here, fragmentation is real, and it demands a trustless settlement layer. Imagine a DAO that spins up a country-specific AI model, paying for compute with a native token, and auditing the model weights on-chain. This is not fantasy—it is the logical endgame of the Armstrong-Kamath thesis. Of course, there are blind spots. The warning assumes that the cost curve for open-source will continue to fall while capability rises. That has been true for the past two years, but what if the next generation of proprietary models—say, GPT-6—introduces a new architecture that resets the gap? Or what if scaling laws return with a vengeance? I have seen this in blockchain: when Ethereum transitioned to Proof of Stake, many thought it would kill L2s. Instead, it spurred them. Technical revolutions do not follow linear projections. The AI bubble could inflate further before it pops. Kamath gave a five-year horizon, but in crypto, 2021 bull market euphoria turned to 2022 despair in less than 18 months. The same compressed timeframes could apply here. The takeaway is not to sell everything, but to realize that the next narrative shift is already visible. The winners will not be the model providers, but the infrastructure that enables fragmented, local, open-source AI. For blockchain investors, this means paying attention to decentralized compute networks, zero-knowledge proofs for model verification, and any protocol that bridges AI inference with on-chain settlements. The noise is loud, but I have been filtering it for 25 years. Truth over hype. Always. Noise filtered. Signal preserved. The signal says: the AI bubble is real, but its collapse will birth a new ecosystem where crypto plays a leading role. Prepare accordingly.