Tracing the immutable logic of open-source AI, the latest push to restrict model weights reveals a paradox: the cure may be deadlier than the disease. Jack Dorsey, Chamath Palihapitiya, and David Sacks have publicly warned that limiting open-source artificial intelligence will economically gut American enterprises while failing to stop the spread of dangerous capabilities. Their argument, grounded in cold hard data on pricing disparities and benchmark performances, cuts through the noise of Washington security debates.
Context: The Open-Source AI Crossroads
The United States faces a critical policy juncture. On one side, national security hawks argue that advanced open-source AI models—those with publicly available weights—enable malicious actors to weaponize artificial intelligence at low cost. They point to scenarios where terrorists or rogue states fine-tune models for cyberattacks or bioweapon design. On the other side, a coalition of tech leaders contends that restricting open-source will only hand the advantage to foreign competitors, particularly China, while raising the cost of AI for American businesses to unsustainable levels.
Palihapitiya, the billionaire venture capitalist, ignited the debate by calculating a stark cost gap. He stated that closing open-source AI would force U.S. enterprises to pay between $26 and $56 per million tokens for proprietary API access, while their overseas rivals—using open-source models deployed on cheap compute—could pay just $0.50 to $1.00 per million tokens. That is a 26-fold to 56-fold disadvantage. 'If AI truly underpins future economic activity,' Palihapitiya argued, 'this situation is untenable.' His numbers, though not sourced to a specific model tier, align with industry observations of the pricing chasm between closed APIs and self-hosted open-source alternatives.
Core: The Economic and Competitive Anatomy
Forensic autopsy of this cost asymmetry reveals more than just a pricing issue. It is a structural deformation of the market. U.S. companies that rely on high-margin, closed-source AI services will either absorb the cost, slashing profits, or pass it on to consumers, losing price wars. Meanwhile, foreign firms and startups can leverage open-source models to build equivalent products at a fraction of the cost. The result is not a level playing field but a tilted one where American innovation is taxed at the point of inference.
Palihapitiya didn't stop at economics. He extended the logic to national security. 'You have a situation where the U.S. spends tens of dollars per million tokens to defend systems, while adversaries attack with costs measured in pennies,' he said. This asymmetry makes defense economically unsustainable. If open-source is restricted, only the U.S. defense sector pays the premium; attackers—who ignore export controls—will still use cheap open-source models. David Sacks, a prominent venture capitalist and former COO of PayPal, offered a counter-narrative: instead of limiting open-source, the U.S. should embrace 'AI-powered cyber defense' as the only viable shield against AI-powered attacks. His solution relies on speed and scalability, not restriction.
Silence in the code speaks louder than audits. The technical underpinning of this debate is the narrowing gap between open-source and closed-source model capabilities. Palihapitiya cited that open-weight models have dramatically reduced the performance distance from proprietary systems. Concrete evidence comes from China: Moonshot AI's Kimi K3 model recently ranked first on a programming benchmark, outperforming all U.S. models. 'Other releases,' Palihapitiya said, 'show that the capability gap between U.S. and Chinese systems is further shrinking.' This implies that even if the U.S. restricts its own open-source models, foreign open-source alternatives of comparable power will emerge, rendering the restriction economically painful but strategically futile.
Where logic meets the fragility of human trust, the discussion enters the realm of so-called 'Mythos' capabilities. Sebastian Mallaby, a Council on Foreign Relations senior fellow, warned that Anthropic's Claude Mythos model has already raised concerns about 'Mythos-level' cyber capabilities, a term he used to describe offensive AI abilities beyond current public understanding. 'The world will soon go from almost nobody having this capability to almost everybody having it,' Mallaby predicted. This diffusion is inevitable because open-source weights can be copied and shared globally. Policy walls cannot contain them.
Contrarian: The Security Blind Spot of Restriction
Decoding the silent language of the policy debate reveals a critical blind spot. The advocates for restriction assume that limiting U.S. open-source models reduces global risk. But the opposite may be true. By crippling U.S. firms' ability to deploy and experiment with open-source AI, the policy weakens America's own defensive capabilities. Attackers will still obtain advanced models from abroad—or from leaked weights—while U.S. defenders are forced to use expensive, less flexible closed systems. The net effect is a degraded defense and an unchanged offense.
Furthermore, the argument overlooks the dual-use nature of open-source AI itself. Open-source models are not just weapons; they are the primary platform for safety research, red-teaming, and alignment work. Restricting them stifles the very communities that find vulnerabilities and propose fixes. In my experience auditing DeFi protocols, the most resilient smart contracts are those subjected to public scrutiny. The same principle applies to AI. Opening the code—while implementing responsible governance like usage policies and audit trails—can actually enhance security by distributing the detection of flaws.
Takeaway: The Fork in the Code
The architecture of freedom, compiled in billions of parameters, now faces a choice. Either the U.S. treats open-source AI as a national asset to be fortified, or it treats it as a liability to be restricted. The former requires investing in AI-driven defense, fostering open-source ecosystems, and competing on capability rather than on artificial scarcity. The latter risks consigning American businesses to a 50x cost penalty while failing to slow down adversaries. The immutable truth is that code, once released, cannot be recalled. The question is not whether dangerous AI capabilities will spread—they will—but whether the U.S. will be the strongest player in that new world. Based on the numbers, the path of restriction leads only to economic hemorrhage and strategic irrelevance. The silent witness of the market will soon deliver its verdict.