Two billionaires just rang the alarm on AI's valuation bubble — and they're drawing direct lines to crypto's own boom-bust cycles.
Brian Armstrong, Coinbase CEO, and Nikhil Kamath, Zerodha founder, didn’t sugarcoat it. In a roundtable hosted by BeInCrypto, they laid out a chilling thesis: the AI industry is sitting on a valuation time bomb. And the detonator? Open-source models that are catching up faster than anyone expects.
I’ve been in this game since the 2017 ICO mania. I remember the three sleepless nights auditing whitepapers in Tokyo, chasing the next 100x. That same FOMO energy is now flooding AI. But Armstrong and Kamath say the party’s about to end. And they’ve got the data to back it up.
Context: Why Now?
The AI investment narrative has been running on rocket fuel. OpenAI, Anthropic, Inflection — private companies raising billions at eye-watering valuations. The pitch: proprietary models will dominate, creating a trillion-dollar market. But the two billionaires see a structural flaw.
Armstrong, who rode the crypto rollercoaster from Mt. Gox to Coinbase’s IPO, knows a bubble when he smells one. He pointed out that open-source models are now only six months behind the best closed-source offerings. And they cost up to 99% less to run. That’s not a gap — that’s a chasm about to collapse.
Kamath, who built Zerodha into India’s largest brokerage by democratizing trading, sees a familiar pattern. “AI is following the crypto trajectory,” he said. “First, centralization and hype. Then fragmentation and local copies. No reason to pay current valuations.”
Core: The Numbers Don’t Lie
Let’s break it down. Armstrong’s core claim: open-source inference costs are 1% of closed-source. He cited real infrastructure data — running a Llama 3.1 405B on cutting-edge hardware is trivial compared to GPT-4o’s API costs. For enterprise customers, the math is brutal. Why pay $10 per million tokens when you can self-host for $0.10?
And the capability gap is shrinking. “Six months ago, open-source models were a year behind. Now it’s six months. Next year, maybe three,” Armstrong said. The implication? Closed-source labs are spending billions on training runs that open-source communities can replicate for a fraction of the cost using architectural innovations like Mixture-of-Experts and state-space models.
Kamath added a geopolitical layer. He sees every major country building its own national AI model — domestic copies, localized tokens, local energy. “Fragmentation is inevitable. Unified market assumptions are dead.” That means the total addressable market for a single global AI company is far smaller than the current narrative suggests.
But here’s the kicker: both billionaires aren’t just talking theory. They’re acting on it. Armstrong admitted he’s positioned to short AI company valuations through private market derivatives. Kamath said he’d rather invest in infrastructure — GPUs, data centers, energy — than in model providers.
Contrarian: The Blind Spots Everyone Misses
Now, let’s pump the brakes. The mainstream narrative is that scaling laws will continue forever. That bigger models will keep outperforming smaller ones. But Armstrong’s argument hints at a diminishing return curve. If the next GPT-6 delivers only a 5% improvement over GPT-5 while costing 10x more to train, the economics flip.
Yet there’s a contrarian angle the billionaires didn’t fully explore: what if a new architectural breakthrough — like reasoning-time compute or an alternative to Transformers — re-ignites the scaling law? Then closed-source labs could leap ahead again, rendering open-source obsolete. That’s the “paradigm shift” risk.
Another blind spot: enterprise stickiness. Big companies don’t switch to open-source overnight due to compliance, security, and vendor lock-in. Custom fine-tuned models on closed-source APIs create switching costs. The migration might take years, not months. That could give closed-source players a longer runway to adjust pricing.
And then there’s the infrastructure play. Kamath’s “local energy” thesis misses that most countries don’t have the chip supply or talent to build their own frontier models. They’ll still rely on cloud providers. The fragmentation might be less than imagined.
Still, the core warning stands. I’ve seen this movie before. During DeFi Summer 2020, I watched Uniswap’s open-source AMM code get forked by every new chain. The original lost pricing power. The same is happening to AI models. Open source ate software. Now it’s eating AI.
Takeaway: What Comes Next?
The real alpha isn’t in betting on closed-source models. It’s in the picks and shovels — GPU manufacturing (NVIDIA, AMD), energy (nuclear, solar), edge inference chips, and local deployment infrastructure. The bubble might take years to pop, but when it does, the correction will be violent.
Will the next crypto winter mirror an AI winter? Or will infrastructure plays keep the market warm? One thing is certain: speed is the only currency that matters here. Those who read the signals early will survive the purge.