Jack Dorsey Backs Open-Source AI Warning as US-China Cost Divide Widens

Dorsey Adds Weight to Warning That AI Restrictions Could Cost US Economy

Jack Dorsey, the co-founder of Twitter and a longtime advocate for open-source technology, has publicly aligned himself with a stark warning from venture capitalist Chamath Palihapitiya: US export controls or other constraints on open-source artificial intelligence would saddle American companies with crippling cost disadvantages. Dorsey responded "yes" to a post in which Palihapitiya detailed a looming economic divide between US firms, which would pay between $26 and $56 per million tokens to access proprietary AI models, and foreign competitors tapping similar capabilities through open-weight systems for as little as $0.50 to $1 per million tokens.

The argument comes as Chinese labs have rapidly narrowed the performance gap with Western models. Most recently, Moonshot AI, a Beijing-based startup, released Kimi K3, which topped coding benchmarks this month and sent shares of US chip makers tumbling. Palihapitiya described that trajectory as unsustainable if AI is truly meant to underpin the future economy, framing the situation bluntly: intelligence cannot be both the engine of the economy and a luxury import.

The warning extends to national security. Palihapitiya pointed out that paying a premium for intelligence to protect American systems while adversaries can launch attacks at a fraction of the cost creates the same imbalance. David Sacks, the former PayPal COO, amplified the message, agreeing with researcher Sebastian Mallaby that dangerous cyber capabilities—highlighted around Anthropic's Claude model—are on the verge of being universally accessible regardless of any single nation's controls. Sacks had previously forecast that Chinese models would acquire advanced cyber capabilities within months, and he argued that Washington's phased release of advanced models like GPT-5.6 demonstrated that classification alone cannot curb the progress of foreign actors.

The exchange unfolds as US policymakers debate whether to mandate pre-release testing for frontier AI models. While some officials see that as a way to manage security risks without ceding ground to China, Sacks and Palihapitiya now represent a growing faction that believes restrictions, not transparency, pose the greatest threat to American competitiveness. Dorsey's one-word endorsement is notable in part because of his work on Goose, an open-source AI agent developed by his payments company Block, and his public commitment to open AI architectures.

Inside the Open-Source vs. Proprietary AI Standoff

The Token Cost Gulf Is Not an Academic Exercise

Palihapitiya's figures—$26–$56 per million tokens for US-based proprietary models versus $0.50–$1 for open-source alternatives—encapsulate a structural cost gap that, if maintained, could erode the ability of American enterprises to build on AI at scale. Even as open-weight models have closed much of the performance gap with systems like GPT-4.5, the pricing chasm has not narrowed in step. For a US business handling billions of tokens per day, that differential translates into millions of dollars in added operating expenses, forcing a choice between eroded margins and loss of access to the cheapest intelligence.

China's Narrowing Capability Gap Magnifies the Risk

The benchmark-topping performance of Moonshot AI's Kimi K3 illustrates how quickly a combination of open research, publicly released weights and domestic innovation can produce models that rival proprietary Western systems. When such models also enjoy a 50-to-100-fold price advantage, the commercial incentive shifts dramatically. The chip stock turbulence following Kimi K3's release suggests investors are already pricing in a world where AI performance is not sufficient moat—and where globally available, low-cost models could commoditize large portions of the inference market.

The Security Argument Cuts Both Ways

National security officials often cite proliferation risk as the reason to control open-source models. But Palihapitiya and Sacks invert that logic: if adversaries already have or will soon obtain equivalent capabilities—as the Anthropic Claude episode implies—then hobbling US access only guarantees that American defenders will face faster, cheaper attacks than they can match. Mallaby's framing that we are passing from a world where almost nobody had "Mythos-level" cyber power to one where almost everyone does underscores the fragility of a control-based approach. Sacks' alternative—AI-powered cyber defense—reflects a broader view that the strategic imperative is not to wall off the technology but to build the systems that can survive in an open-intelligence environment.

Washington's Policy Crossroads

The push for pre-release testing of frontier models, already scaled with the phased rollout of GPT-5.6, is running headlong into the economic and technical reality that code can travel. Once a competent model is downloadable anywhere, any singular regime of evaluation becomes a speed bump rather than a barrier. The rift between those who see control as prudent and those who see it as a self-inflicted wound will define the next wave of AI regulation—and, if the Sacks-Palihapitiya thesis proves correct, will shape which nation's corporations can afford to lead the coming economic transformation.

What the Debate Means for US Companies and Investors

• Stress-test cost models using the $26–$56 vs. $0.50–$1 token range. US companies building AI-dependent products should map their inference volumes against this cost spread and identify breakpoints where domestic proprietary services become commercially unviable compared with open-weight alternatives.
• Reassess exposure to Chinese AI benchmarks. The Kimi K3 event shows that a single release can move chip stocks and signal capability shifts. Technology investors and hardware-dependent firms should monitor upcoming releases from DeepSeek, Moonshot AI, and other Chinese labs as leading indicators of competitive pressure.
• Evaluate open-source model integration now, while policy paths are uncertain. Enterprises that begin piloting open-weight models for non-sensitive workloads can build internal expertise and optionality before potential US restrictions narrow the menu. The Goose agent from Block (Dorsey’s company) offers one concrete example of a production open-source AI system.
• Watch for second-order effects on chip demand. A world in which most inference runs on inexpensive, globally available models could shift demand away from premium, datacenter-scale hardware toward more cost-efficient silicon, with implications for Nvidia, AMD, and the broader semiconductor supply chain.
• Lobby for policy that weighs economic competitiveness alongside security. For industry groups and large AI players, the Palihapitiya-Sacks frame provides a concrete narrative—backed by price data and benchmark results—to argue that unilateral restrictions may accelerate the very shifts they seek to prevent.