Angel Investor and All-In Co-Host Jason Calacanis: The Safety Narrative Aims to Restrict Open Source
Angel investor and All-In co-host Jason Calacanis stated that investors in cutting-edge model companies will quickly shift to a regulatory capture model. He predicts that vertical AI companies, governments, and enterprises, the largest token buyers, are embracing open source, which puts pressure on and caps the returns from investing in closed cutting-edge models. The goal of the entire safety mobilization is to block open source.
He wrote that if safety is truly the goal, there should be disclosure; open source is the ultimate disclosure process. He claims to have discussed this position on his podcast for two years. The concurrent context includes discussions between Anthropic and OpenAI about slowing down, on-site evaluations, and nationwide mandatory safety regulations. Calacanis attributes the migration of buyers to their unwillingness to hand over intelligence and data to private labs, stating that open source tokens are "dark tokens": their usage does not appear in the financial statements of model companies but only in cloud bills from AWS, Azure, GCP, etc.
He previously warned startups against exchanging free or discounted cutting-edge tokens for equity and data usage, stating that platforms will observe which applications perform best and then incorporate those features into their own products. He cited examples of some startups migrating inference to open source models on their own hardware, reducing costs by about 50% to 90%, with quality close to that of most tasks. Podcast partner David Sacks outlined the path: first, establish an AI regulatory body similar to the FDA; if political resistance is strong, shift to a self-regulatory organization like FINRA; then require open source and closed models to adhere to the same standards; once open source is released, it cannot be retracted or centrally monitored, and it will be pushed out of the compliant market under "equal treatment."
Chinese open source weight models like Kimi and Qwen are approaching the performance of closed cutting-edge models in some evaluations, with cheaper inference, intensifying this debate. Anthropic has previously accused industrial-scale distillation of constituting intellectual property theft to Congress. Calacanis and Sacks oppose punishing U.S. developers with bans, arguing that the source should address foreign entities, or else companies will be locked into the token tax of closed models. Brad Gerstner believes that open source and closed models are not a zero-sum game, and the U.S. needs revenue from both sides to fund annual computing infrastructure costs of about $1 to $2 trillion.
In market mechanisms, this is a battle for channels rather than the release of new models. Buyers are governments, industry software, and enterprise IT that want self-hosting and data to stay in-house; sellers are closed laboratory shareholders pricing APIs and future IPOs. Beneficiaries are open source inference, weight mass production, and privatized deployment tools in the cloud; those under pressure are laboratories whose valuations are tied to writing "safety standards equally applicable to open source" into legislation. Funding is shifting from closed interfaces billed by token to dark inference within enterprise clusters that do not keep accounts, making growth on reports appear as if demand has disappeared, while in reality, demand is just changing tracks.
Source: Public Information
ABAB AI Insight
Calacanis's Launch portfolio heavily bets on application-layer startups, whose largest variable costs are inference. Cutting-edge labs exchange free quotas for data usage, equating to using capital expenditures to acquire application-layer innovation. Open source drives the same layer costs down to self-owned GPUs or cloud bidding instances, shifting investors' exit multiples from "renting intelligence" to "owning weights." His repeated emphasis on this over two years indicates that he views safety hearings merely as a policy shell for price wars.
The capital path is the gross margin of regulatory text services. To maintain high-priced tokens, closed models must prevent open source from legally deploying equivalent capabilities; the easiest way is not to win evaluations but to ensure open source fails the "pre-release testing, retractable, and monitorable" three hurdles. The FINRA-style standard organization described by Sacks would be funded by labs that can afford evaluation bills. The motivation is to protect the trillion-dollar public narrative while framing the price impact of Chinese open source weights as a matter of national security. Resource allocation points to congressional hearings, standard committees, and safety assessment contracts, rather than the training of open source weights themselves.
A similar structure is seen in how Microsoft used security certifications to squeeze unverified Linux distributions during the server era, and how payment networks use compliance licenses to block crypto settlements. The price collapse after DeepSeek has already demonstrated this: as evaluations approach, the premium on closed APIs drops first. The industry phase is about control: computing power is still expanding, but who has the right to publish weights is becoming more decisive than who has larger parameters for profit pools.
The structural judgment is a transfer of pricing power. The mechanism is that enterprise buyers use open source as sovereign insurance, governments use open source for controllable deployment, and vertical software uses open source to avoid handing core processes to potential competitors. If closed labs write safety standards as "must be able to shut down models," the technical characteristics of open source will be defined as original sin. Disclosure and shutdown are antonyms, yet they are packaged within the same safety discourse.
ABAB News · Cognitive Laws
Once safety standards require shutdown capability, open source will be deemed non-compliant.
Dark tokens will eat into financial statements but fatten cloud bills.
Free quotas are not charity; they are hooks for harvesting application layers.