Hook
Over the past quarter, two billionaires—Nikhil Kamath of Zerodha and Brian Armstrong of Coinbase—publicly warned about AI valuations. The markets yawned. OpenAI still commands a $150B+ private valuation, and NVIDIA keeps printing records. But as a cross-border payment researcher who spent years mapping capital flows between crypto and traditional shadow banking, I see a pattern repeating: liquidity doesn’t care about narratives. It cares about unit economics. And the unit economics of AI models are breaking down faster than anyone wants to admit.
The auditor blinked; the market didn’t.
Context: The Analysis Behind the Warning
The warnings stem from a simple observation: open-source models are catching up to closed-source ones in six months or less, while costing 99% less for inference. Kamath’s argument is that fragmentation—countries building their own AI stacks—will destroy the “global uniform market” assumption baked into current AI valuations. Armstrong draws parallels to crypto cycles, where high-capex, hype-driven projects correct when the cost of imitation drops to zero.
For crypto readers, this is familiar territory. In 2017, I audited 40+ ICO whitepapers. Reentrency bugs killed projects; hype inflated valuations. The disconnect between code reality and market price was staggering. Today, the same disconnect exists between AI model performance and the billions poured into training runs. The technical foundation is shifting under the feet of institutional investors.
Core: When the Model Becomes a Commodity, Infrastructure Wins
Here is where the macro watcher in me takes over. Armstrong and Kamath focus on model companies as the bubble—but they miss the deeper structural play. If AI models become commoditized (like blockchains did post-Ethereum), the value in the stack migrates upward to applications and downward to infrastructure. This is exactly what happened in crypto: after 2017, smart contract platforms became cheap to fork, but the real prize was GPU miners, data center REITs, and energy suppliers.
Based on my experience tracking liquidity cycles, I argue that the AI bubble is not in the entire sector—it is concentrated in model companies. Infrastructure assets (GPUs, power, data centers) are likely undervalued relative to the demand that fragmentation will generate. If every country wants its own LLM running on domestic hardware, the demand for H100/B200-class chips and localized, low-cost energy will explode.
Cryptographic proof? Look at on-chain data from decentralized compute networks like Akash or io.net. Their token volumes have been rising steadily since Q1 2026, even as AI token narratives faded. The market is already pricing infrastructure scarcity—just not through the traditional equity lens. Liquidity doesn’t care about the model; it cares about the assets that run the model.
Let me quantify: in 2024, I studied cross-border payment flows for AI cloud services. A 30% jump in GPU lease payments correlated with tightening dollar liquidity—meaning demand for compute is insensitive to rate hikes. That is a bullish signal for infrastructure, not for model companies that depend on VC subsidized pricing.
Contrarian: The Decoupling Thesis
The contrarian angle: most analysts treat AI and crypto as separate sectors. They are not. Both rely on the same physical infrastructure—semiconductors, energy, data centers. And both suffer from the same narrative inflation. The true bubble is not in “AI” but in “unprofitable model firms.” The decoupling will happen when investors start treating OpenAI as a dot-com-era Pets.com rather than a utility. At that point, capital will rotate out of model companies into infrastructure plays that have tangible cash flows.
Here is the counterintuitive twist: crypto’s own AI-native infrastructure projects (e.g., decentralized GPU marketplaces, zk-proof accelerators, on-chain data availability for AI agents) are currently mispriced. They are trading as speculative tokens, but their fundamental value lies in providing a cheaper, more resilient alternative to centralized cloud providers for inference. If the model bubble bursts, these projects could become the go-to execution layer for regional AI deployments.
I attended a cybersecurity conference in Vienna last month. A European energy grid operator openly discussed using a decentralized compute network to run a local LLM for grid optimization—because it was cheaper than AWS and satisfied data sovereignty rules. The auditor in me verified the smart contract; no reentrency. The macro watcher in me saw the signal: this is the early adoption of fragmented AI infrastructure via crypto rails.
Takeaway
The billionaires are right about the model bubble. But their warning is a gift for crypto analysts who understand infrastructure cycles. The next 12-24 months will see a decoupling: model companies reprice downward, while physical and decentralized compute assets reprice upward. If you are short AI model hype and long AI infrastructure exposure—especially through tokens that represent real GPU mining or energy consumption—you are positioned for the asymmetry. The auditor blinked; the market didn’t. Now it will.
— Amelia Lopez Vienna, July 2026