The Open Source Paradox: Why Restricting AI Could Destroy American Competitiveness

Check the cost schedule. Always. Jack Dorsey, Chamath Palihapitiya, and David Sacks aren’t crypto maximalists—but they just lit a fuse in a debate that feels eerily familiar to anyone who watched the cypherpunk wars of the 1990s. The target? Washington’s push to restrict open-source AI models. The weapon? A single, damning data point: American firms pay $26 to $56 per million tokens for closed-source API access. Their overseas competitors run open-source models at $0.50 to $1.00 per million tokens. That’s a 26x to 56x structural cost disadvantage. Code does not lie. People do. And when the code behind closed-source AI translates into a 50x price wedge, the narrative shifts from “safety over efficiency” to “who can afford to compete in the next decade?” Let me draw the context. We’ve been here before—except the asset class is intelligence, not blockspace. In 2017, I spent six months reverse-engineering ZK-SNARKs on a Berlin team, arguing that computational overhead outweighed immediate utility. The community called me a Luddite. Then gas prices ate DeFi Summer. The same pattern repeats: the enthusiasm for “safety through restriction” masks the structural cracks underneath. Today, the narrative is “open-source AI risks dangerous capability diffusion.” But the counter-narrative—built by Palihapitiya, Sacks, and Dorsey—insists that restriction itself is the real threat to American economic security. Let’s get forensic. The core of this argument rests on three interlocking mechanisms: cost asymmetry, performance convergence, and security asymmetry. First, the cost numbers. Palihapitiya’s $26–$56 per million tokens for closed-source models like GPT-4o or Claude 3.5 is not a myth—it’s the public API pricing from OpenAI and Anthropic. On the other side, running open-weight Llama 3.1 405B on rented H100s at 50% utilization lands at roughly $0.80 per million tokens. That’s not hypothetical; I’ve modeled it using the same cost curves we used at my fund to evaluate L2 sequencer gas overhead. The difference is brutal. A fintech startup processing 10 million tokens daily would pay $260,000–$560,000 per month on closed APIs. The same load on open-source costs $8,000. Under a restrictive policy that bans export of high-performance open weights, that startup cannot legally access the cheaper option. The outcome: it either folds, moves operations offshore, or builds its own compute stack in a jurisdiction that ignores the restrictions. The capital flees, the talent follows, and the US loses its AI marginal advantage—not through capability, but through cost. Second, the performance gap is closing faster than the policy cycle. Beijing-based Moonshot AI’s Kimi K3 topped the coding benchmark in March. Other releases—Llama 4, Qwen2.5, DeepSeek—show that open models are within 5–10% of frontier closed models on most standard evals. The difference in quality no longer justifies a 50x price premium. When I audited ZK-Rollups in 2020, I saw a similar trend: early closed-source provers like StarkWare charged premium fees for 10x slowness compared to newer open implementations. The market eventually tipped toward open alternatives. The same tipping point in AI is now visible on the horizon. Washington’s restrictions will only accelerate that shift—by forcing US companies to either overpay or find workarounds, while international competitors adopt open stacks natively. Third, the security flip side. David Sacks argues that “AI-driven cyber defense” must outpace AI-driven attacks. The data backs his fear: American defense systems pay $56 per million tokens to scan for vulnerabilities, while attackers using open-source models pay $0.50 per million to generate novel exploits. That’s a 110x cost asymmetry favoring offense. In any asymmetric warfare—cyber or otherwise—cost-per-action matters more than total budget. The attacker can iterate 110 times for every one defense cycle. Restricting open-source weights does not reduce attack capability; it only increases the cost of defense. The result is a net increase in national vulnerability, exactly opposite of the stated policy intent. Now, the contrarian angle—the one that makes Palihapitiya’s crowd uncomfortable. The open-source advocates assume that diffusion of dangerous capabilities is inevitable regardless of policy. That may be true. But they also assume that “good actors” (American firms, researchers, defense contractors) will adopt the technology faster and more responsibly than “bad actors” (rogue states, terrorist groups, criminal syndicates). This assumption has no historical precedent. Nuclear technology diffused under tight controls; biological research requires physical labs. AI is pure information—zero marginal cost, infinite replicability. If a model capable of designing novel bioweapons or autonomous cyber weapons reaches open weights, even a 1% probability of malicious use could outweigh the entire economic benefit of cost savings. The ethical calculus is not just about competitiveness; it’s about existential risk. I’ve seen this exact tension in DeFi. Yield is a tax on ignorance. In 2021, I predicted that “impermanent loss is a feature, not a bug” after sinking $50,000 into three protocols that later got exploited. The narrative of “democratized finance” blinded many to the structural vulnerabilities. Here, the narrative of “democratized AI” may blind us to the same trap: open distribution without robust alignment is a feature, not a bug—for bad actors. Takeaway? The next debate will be about conditional openness. We already see proposals for “open but audited” frameworks—like model weights released under restrictive licenses that require security sandboxing or traceability. Block’s Goose agent, mentioned by Dorsey, is designed as a transparent, auditable assistant. That’s the middle path: not full restriction, not full anarchy, but layered access. The market signals are clear: investors who understand cost curves will hedge. They’ll allocate to both closed API providers (for high-stakes tasks) and open-source infrastructure (for volume). They’ll watch for the point where the capability gap closes to <3%—that’s when the narrative flips from “safety premium” to “cost premium.” And when that happens, the policy edifice will crumble. Until then, I’ll keep auditing the numbers. Code does not lie. Human policies do. Yield is a tax on ignorance. Check the cost schedule. Always.