The US is debating whether to strangle open-source AI in the cradle. Chamath Palihapitiya dropped a number that should stop every policymaker cold: forcing American enterprises to use closed-source models will push per-million-token costs to $26-$56, while offshore competitors, armed with open weights, pay $0.50-$1. That's a 26-to-50x structural disadvantage. Jack Dorsey and David Sacks have now joined the chorus, warning that a single-minded focus on security via restriction will backfire economically and make the nation less safe. As a blockchain due diligence analyst who spent years dissecting sharding faults and stablecoin contagion, I see the same pattern here: policy driven by fear, not data, and a failure to model second-order effects.
Context
The debate is not new. Since the rise of advanced LLMs, Washington has oscillated between promoting innovation and imposing guardrails. The latest flashpoint is whether to restrict the export or distribution of open-weight models. Proponents argue that unfettered access to powerful AI code enables malicious actors — from cybercriminals to state-backed hackers — to weaponize capabilities at scale. Opponents, including Palihapitiya, Sacks, and Dorsey, counter that the economic cost of such restrictions far outweighs the security benefits. The numbers cited are dramatic: a 50x cost gap in inference, a network attack asymmetry where American defenders pay $56 per million tokens while adversaries spend $1, and a rising tide of Chinese models — like Moonshot AI's Kimi K3, which just topped a coding benchmark — that shrink the performance lead of US frontier systems.
Core: The Structural Cost Disparity
Let's dig into the cost math because that's where the policy leverage breaks. Palihapitiya's claim of $26-$56 vs $0.50-$1 is not a rounding error — it's a tectonic shift. At scale, if AI is to underpin the next wave of economic activity, American companies will be paying a tax that no competitor in Europe or Asia faces. This is not theoretical. From my audit of MakerDAO's collateral thresholds, I learned that small structural flaws — like a single oracle feed — can cascade into systemic failure. Here, the structural flaw is the assumption that closed-source safety justifies a 50x markup. The cost does not buy better capability; it buys perceived control. But control is illusory when open-weight models of comparable quality are a download away for anyone outside US jurisdiction.
The blockchain context makes this worse. AI agents built on decentralized platforms — the kind Block is deploying with its Goos
