From AT&T's forced open-sourcing to the dispute over gene sequencing: History has proved that the truly huge profits of AI all lie in the "last mile".
Shenyi Translation is a translation team under 36Kr, focusing on technology, business, workplace, life and other fields, with an emphasis on introducing new foreign technologies, new perspectives and new trends.
Editor's note: Open source will not destroy AI investment, but it will crush the exorbitant profits of closed source. When models become as ubiquitous as water and electricity, the only entities that can truly make money are the "complementary assets" that are deeply rooted in the physical world. This article is translated and compiled.
In May 1851, millions of people walked through the corridors of a glass building to catch a glimpse of what the future would look like. The Great Exhibition of the Works of Industry of All Nations brought together the most representative innovations from around the world at that time, including steam hammers, telegraphs, reapers, and flush toilets. Anyone could get close to these technologies and try to reverse-engineer them, and countries also sent their top engineers to do their utmost to absorb core technologies from their competitors. Prince Albert, the initiator of the exhibition, firmly believed that the widespread dissemination and diffusion of technology was the key to accelerating economic progress.
A century and a half later, economist Petra Moser organized the exhibition catalog from that year, as well as nearly 15,000 inventions from that exhibition and the 1876 U.S. Centennial Exposition, into a database. Crucially, these inventions came from countries with and without patent protection systems. Her findings stunned staunch advocates of strong intellectual property protection: the patent system had no impact on the total amount of innovation, it only changed the direction of inventors' attention. At that time, Switzerland had no patent protection system, so innovators flocked to fields such as scientific instruments and food (Nestlé was founded in Switzerland in 1866), where confidentiality, time-to-market advantage, and complementary assets were sufficient to give them a competitive edge. Under a strong intellectual property protection system, innovation was more widely distributed. The total amount of innovation was the same in both cases, the only difference was the distribution across different fields.
Today, the same game and tension have once again become the core of the debate between closed-source and open-source weight artificial intelligence. Anthropic and OpenAI argue that "distillation attacks" from Chinese companies not only threaten the industry's ability to raise funds for next-generation models, but also endanger U.S. national security. Supporters of open-source weights counter that widespread diffusion is essential to building a competitive and dynamically innovative artificial intelligence market.
However, Moser's research evidence shows that the current debate is starting from the wrong point. Open-source weights are unlikely to change the overall level of investment in artificial intelligence, it only changes the direction of investment: what to develop, who develops it, and who captures the returns. Closed-source labs may continue to push the cutting edge of general capabilities, while open-source weights allow experimentation to spread widely to enterprises and fields that can combine machine intelligence with scarce data, distribution channels, and tacit knowledge. The same logic of resource allocation also affects security issues. The real core question is not whether open source is dangerous at an abstract level, but at what points diffusion can strengthen the defenders' power, and at which links harmful capabilities can truly be contained.
Two Opposing Visions for the Machine Intelligence Market
The economic argument against open-source weights is simple: model distillation is intellectual property theft that undermines innovation incentives. If the U.S. government does not step in and use all available means to stop it, not only will domestic labs in the U.S. be unable to raise the funds needed for the next round of large-scale training, but China will also free-ride and catch up. For people who hold this worldview, open-source weights are essentially an extreme "decelerationist" stance.
The opposing view focuses on the laws of diffusion of general-purpose technologies in the economic system throughout history, and concludes not only that open-source weights fundamentally promote competition and technological acceleration, but also that without open source, the United States will quickly fall behind. Openness and full experimentation are exactly why we maintained our leading position in the Internet era, and this time is no exception.
Security risks have made this discussion even more complicated. Dario Amodei has been strongly opposing open-source weights for many years on the grounds that they pose huge security risks. He believes that once intelligence crosses a certain critical threshold, human society will be unable to defend against malicious actors, and out-of-control models could cause devastating existential crises. Only by controlling access and setting strict safety guardrails can the country home to "data center geniuses" be allowed to exist. Worse still, because open-source weights cannot be "recalled", we could suddenly be plunged into a doomsday scenario without any warning or means of response.
Critics of Amodei point out that without open-source weights, the machine intelligence market will quickly become highly concentrated; and it is far too much of a coincidence that Anthropic's commercial interests perfectly align with both AI security goals and U.S. national security priorities. Security researchers also often jailbreak to bypass the protections of closed-source models, and Anthropic and OpenAI's models have been widely exploited by hackers, even for intrusions into sensitive government infrastructure. They argue that closed-source models only give us the illusion of safety, and at best they only buy us a false sense of "having plenty of time".
Both sides worry that today's wrong moves will lead us to an irreversible situation. And both sides sincerely believe that their solution is the only way to ensure security (or the safest path under realistic conditions). Fortunately, economic laws do not favor either side.
The De Facto Defunct Proprietary System
Anthropic and OpenAI have long been the main targets for other labs to catch up with the industry frontier. Anthropic has even publicly requested Congress to severely punish Alibaba, accusing it of launching unscrupulous illegal attacks to steal its technology. If Chinese labs can replicate its model performance at low cost in just a few months, how can U.S. labs continue to raise funds for expensive training?
The crux of Anthropic's appeal is that apart from fake accounts and systematic API abuse, the patent-like proprietary system it expects has never actually existed. Model distillation is a widely recognized legitimate practice in the industry, which is fundamentally different from the espionage and trade secret theft faced by U.S. defense, aerospace and chip contractors in the past. Chinese labs did not directly steal the core weights, they simply used prompt engineering to make U.S. models act as "mentors" for their own models.
Content generated by models is not protected by copyright law, and AI labs usually grant ownership of output content to their customers. Although terms of service can explicitly prohibit customers from using these outputs to train competing models, this raises an obvious question: if the technology is so advanced, why can't labs use technical means to prevent distillation? The answer is that neither the request content nor the request source can directly reveal the attacker's intent. Each independent request looks no different from that of a normal customer, and organized teams can spread calls across a large number of batch-registered accounts, aggregation platforms, and different jurisdictions. Without adding cumbersome usage barriers across the entire platform, ban actions will only degenerate into an endless cycle of "ban account - re-register". This trade-off has already been clearly demonstrated on Fable. Due to the restrictions it imposed on assisting cutting-edge AI R&D, which sparked strong dissatisfaction among customers, Anthropic had to urgently adjust its rules within a few days. Risk control measures substantial enough to produce real effects are bound to harm a large number of legitimate normal work cases. In the long run, preventing paying customers from using output data for training is commercially unworkable, if restrictions are too tight, core high-frequency users will directly switch to open-source weight models.
Finally, ruling that the core technologies that have driven tremendous progress in artificial intelligence over the past decade are illegal in isolation will put top labs in an extremely awkward position, because they themselves are currently vigorously invoking the "fair use" principle to conduct large-scale data extraction and absorption of massive web pages, media content and books on the Internet.
Taking a step back, is the fact that top AI models can act as mentors to improve the performance of lagging models really harmful to the U.S. artificial intelligence industry? Even if the government can use soft power, policies and diplomatic means to force such barriers, do we really want to see that outcome?
Monopoly, or No Monopoly?
Richard Nelson in 1959 and Nobel laureate Kenneth Arrow in 1962 both deeply explored this classic dilemma. Knowledge is non-excludable and non-rivalrous: my use of a certain idea does not prevent you from using the same idea; and once an idea is made public, you can no longer exclude others from using it. Therefore, the total social value created by an idea often far exceeds the returns that the inventor can capture, which raises a classic concern: certain ideas with extremely high social value cannot be fully explored and funded from the very beginning.
Model weights - at least when existing separately from peripheral supporting software - behave very much like "ideas".
For society as a whole, this evolves into a core intertemporal contradiction: once an idea is born, society wants it to spread as widely as possible. After all, its marginal cost is almost zero, just like downloading model weights. But before the idea is born, society must convince explorers that they can capture generous returns to recoup their R&D costs.
Joseph Schumpeter spent his entire life grappling with this contradiction, affirming on the one hand the value of monopoly rents in stimulating early investment, and on the other hand emphasizing the necessity of startups and "creative destruction" to break monopolies and drive subsequent explosive growth. In the end, he reached a slightly contradictory compromise conclusion: the best strategy society can adopt is to first grant it a temporary monopoly position, and then fully open up competition.
Since then, all research in the field of innovation economics has essentially been endlessly debating how long this monopoly should last and what mechanisms should be relied on to constrain it. However, there is a key variable here that can significantly shift the balance between open source and closed source.
The Compound Interest Effect of Ideas
The vast majority (if not all) of innovations are the result of some form of idea recombination. In fields that are highly dependent on the cumulativeness of knowledge and require reprocessing of existing achievements, the duration and intensity of rights granted to first movers must be carefully weighed against the additional costs caused by later explorers delaying their work. No matter how abundant the resources of leading AI labs are, the number of promising paths they can explore is ultimately limited, and this also applies to the field of AI safety.
Artificial intelligence has been the product of rapid distillation and refinement from start to finish. Today's stunning models benefit not only from decades of continuous accumulation in the field of neural networks in academia during the "AI Winter", but also from Google's public release of the Transformer architecture. From this perspective, open-source weight models, as the "students" of closed-source models, are a direct continuation of this technological lineage. The output of models is exactly the key R&D raw material that drives the entire ecosystem forward.
The purest natural experiment on how restricting access to basic R&D elements inhibits innovation comes from another race: the contest between for-profit company Celera and the publicly funded Human Genome Project in DNA sequencing. After Celera took the lead in completing the sequencing of certain genes, it restricted external access through charging, limiting redistribution, and commercial licensing. Economist Heidi Williams found that these restricted genes attracted 20% to 30% less subsequent research and product development than comparable genes that were fully public from the start. Although these restrictions ended within two years when public projects independently completed sequencing of the same genes, the resulting gap persisted for a long time. As late as 2009, genes sequenced by Celera still lagged behind other genes in subsequent R&D output.
When the value of subsequent R&D is extremely high, even small early frictions can produce compound negative cumulative effects. Accordingly, removing frictions can have a profound impact on the rate and direction of innovation. In the field of biomaterials, Jeff Furman and Scott Stern found that when access to R&D elements became more convenient, the cumulativeness of research increased by 57% to 135%. In the study of genetically engineered mice, Fiona Murray, Philippe Aghion and their collaborators studied the scenario after the National Institutes of Health (NIH) negotiated to remove DuPont's restrictions on hundreds of Cre-lox and Onco strains, and abolished penetrating royalties and cumbersome reporting requirements that hindered academic access. The results showed that subsequent research increased significantly. More critically, research directions showed a trend of diversified development, as new researchers joined and explored far more brand-new technological paths. Upstream, the cultivation of new engineered mice was not negatively affected at all.
Overall, as long as the benefits from technological accumulation and extensive exploration are sufficiently significant, openness is the most advantageous strategy. This also means that when technological development is still in a stage of high uncertainty (like AI today), society gains the most dividends from openness. When technological evolution is reduced to deterministic execution along known paths, the harm of closed source will be greatly reduced. Even in this specific scenario (for example, a drug molecule with clear clinical benefits that is urgently awaiting commercialization), society grants it a temporary monopoly in exchange for public disclosure of relevant technical information. The patent system itself was originally designed to promote technology dissemination and avoid the risk of technology being completely hidden away due to trade secret protection.
You may wonder whether these cases from heavy R&D fields are equally applicable to artificial intelligence. After all, large-scale deployment of models requires huge investment in underlying infrastructure, which is where labs face the greatest risk. However, the history of the telecommunications industry reveals exactly the same enlightenment. An antitrust settlement in 1956 forced AT&T to open one of the most valuable patent portfolios in history for free. Calculations by Martin Watzinger and his collaborators show that once AT&T's inventions, including the transistor, were freely available for others to draw on, the whole society's inventive and innovative vitality surged by 17% within five years. Interestingly, this growth was not driven by large companies free-riding on AT&T's intellectual property, but by startups targeting entirely new incremental markets. At the same time, Bell Labs further focused on its core business, and its innovative momentum did not diminish at all: it successively developed the laser in 1957, launched the communications satellite in 1962, and developed the Unix system in 1969.
While allowing others to build compound innovations on their ideas may be frustrating for Anthropic and OpenAI, the good news is that with very few exceptions, economic progress has always evolved according to this rule. William Nordhaus's research found that innovators on average capture only about 2.2% of the total social surplus they create for the world. Artificial intelligence is no exception. This also allows us to take a more objective view of Anthropic's appeal to the U.S. government: society does not owe these labs a stricter proprietary protection system. Facts have proved that a completely different set of value realization mechanisms has long been at work.
The "Last Mile" of Innovation
When a new general-purpose technology emerges, it initially has to be awkwardly embedded into existing systems. These early "point solutions" can only unlock part of the value of the new paradigm; only by reconstructing the overall architecture from first principles can this technology demonstrate its disruptive transformative power.
This kind of system reconstruction requires control and adaptation of key complementary assets, including infrastructure, user trust, distribution networks, and regulatory compliance relationships. This provides established industry giants with a second opportunity to break through. Although new entrants may seize the initiative in the early stage through surprise technological innovation, as complementary assets gradually become the core bottleneck restricting development, traditional giants can re-stabilize their industry position through imitation, mergers and acquisitions, and rapid follow-up.
In the 1980s, Richard Levin and his colleagues at Yale asked corporate executives in the U.S. innovation field a basic question: What exactly protects your R&D results? Patent protection ranked last. What ranked at the top? It was continuous learning capabilities, technological confidentiality, first-mover time advantage, and complementary assets. This study was repeated ten years later, and the conclusions were identical. The only exceptions are the pharmaceutical and chemical industries, where a specific molecular structure can define the entire product category, and precise patent exclusivity is highly deterrent. However, in most areas of economic life, enterprises' ability to exclude others from underlying ideas is very limited.
As long as the proprietary nature of technology is weak, there is still ample room for profits in the industry; these profits will only flow to the parties who control "complementary assets" - assets that can make the original technology better, cheaper, and more agile. A classic example is EMI, the record distributor of the Beatles, which was also involved in the electronics business and held the patent for the world's first CT scanner. Its chief engineer, Godfrey Hounsfield, won the Nobel Prize for this. However, EMI barely captured any of the commercial value of this technology. Within a few years, EMI was acquired, and the CT market was firmly divided between General Electric (GE) and Siemens. What complementary assets did GE rely on? It was its mature hospital distribution channels and maintenance service networks, which are exactly the "last mile" for scanners to truly reach the market.
Artificial intelligence has been on the weak side of proprietary nature since its birth. Although the competition for general artificial superintelligence (ASI) is extremely fierce, cutting-edge research results will inevitably flow out of labs, top talents frequently move between companies according to Silicon Valley traditions, and model APIs are bound to leak valuable technical details. In all these dimensions, even the most stringent blocking measures can at best win a very slight time lead for cutting-edge labs. Moreover, given the bargaining power that top talents and large enterprise customers have at this stage, overly strict control will backfire and produce counterproductive consequences.
Of course, not all of the lab's core secrets flow into the public domain. According to estimates by Epoch AI, the final training process only accounts for 10% to 23% of the total computing power cost. Labs have accumulated a large amount of extremely valuable tacit knowledge in data cleaning and labeling, infrastructure architecture and optimization, and learning lessons from failed experiments. But the rule remains consistent - if cutting-edge labs cannot quickly master the key complementary assets needed to scale AI in the market, value will eventually shift to other fields.
This also explains why major AI labs are striving to strengthen customer stickiness by deeply binding models with interactive shells