The open-weights fight echoes Microsoft's campaign against Linux: a controlled platform is being defended with safety language while openness threatens its position. Anthropic's warning carries more weight because released model capabilities can be hard to withdraw. Open weights may still make humanity safer by distributing AI power beyond a handful of companies.
On July 24, 2026, 25 technology companies launched Open Weights and American AI Leadership. Nvidia, Microsoft and Meta signed that day. By July 27, OpenAI had joined. Anthropic had not.
By August 3, Microsoft's list had passed 270 signatories. Anthropic and Apple were still absent. Apple was a notable omission. Among frontier-model labs, Anthropic was the conspicuous holdout.
Anthropic sells controlled access to Claude. Open weights let organizations run models on their own infrastructure, which can reduce dependence on a provider for some work. That creates a plausible commercial incentive beside a safety concern Anthropic has held for years. I think both can be true. The concern can be sincere while the controlled platform remains commercially useful.
Anthropic's missing signature is a signal because its safety position and business model point in the same direction. The tension resembles the open-source fight, although the risks around model weights are harder.
Microsoft already learned this lesson
In August 1998, an internal Microsoft memorandum called Halloween Document I described open-source software as a "direct, short-term revenue and platform threat." The memo also admitted that open projects could achieve or exceed commercial quality. Its author wrote that "FUD tactics can not be used to combat it."
The public rhetoric became harsher anyway. In 2001, Steve Ballmer called Linux "a cancer" while arguing about the GNU General Public License.
History gave Microsoft a fairly dry answer. Every TOP500 supercomputer has run Linux since November 2017. Microsoft acquired GitHub in 2018, maintains Azure Linux, and now hosts the open-weights letter.
Commercial incentives can coexist with sincere security concerns. I think they still shape which concerns get amplified and which remedies feel acceptable. Microsoft understood Linux well enough to fear it, then spent years defending the platform Linux threatened.
Microsoft privately recognized open source as credible while publicly attacking its risks. Its later embrace of Linux shows how easily incumbent protection can hide inside a security argument.
AI makes the analogy harder
Anthropic's position deserves more care than Ballmer's rhetoric. The company published its Responsible Scaling Policy in September 2023, long before this dispute. In his July 27 response, Dario Amodei rejected a blanket ban, called open-weight models without dangerous capabilities "a public good," and supported safety tests for every sufficiently capable model, open and closed.
Open weights also differ from open source. Weights are learned model parameters. A weights-only release falls short of the Open Source AI Definition when it omits required data information or complete training-and-runtime source code.
A hosted provider can change future service calls, monitor use and withdraw access. Once weights are copied, the developer cannot force every operator to update. Distributed software can remain unpatched too, but a model copy may preserve a capability after its maker decides that capability is unsafe.
Heartbleed keeps the openness case honest. The vulnerable code entered the public OpenSSL repository on December 31, 2011 and was publicly disclosed in April 2014. Full source visibility still left a gap of more than two years.
I think the safety argument is stronger this time. I don't know whether broad access will favor attackers or defenders in the most dangerous domains. The evidence does not settle it yet.
Open weights create a withdrawal problem that open-source slogans cannot solve. Anthropic has a legitimate case: scrutiny helps only when defenders can find and contain a dangerous capability before others keep it running.
Decentralization may be the safety feature
The open-source story did more than expose bugs. It distributed the authority to inspect and alter foundational infrastructure. Linux could outlive any vendor's strategy because no single vendor owned its future.
Open weights offer a narrower version of that freedom. An organization can run a model on its own infrastructure and modify it without asking the original provider. Bad actors gain the same freedom. Distribution creates resilience and misuse risk at the same time.
Concentrated AI has its own failure mode. A small group of companies can decide who gets access and what outsiders may inspect. Those companies can improve safeguards quickly, yet humanity's dependence on them grows with every closed capability.
I think the strongest safety case for open weights is decentralization. Faster bug discovery would help, but shared power is the larger prize. An AI world with many capable operators is harder to control from one boardroom or one state.
The open-source world eventually made shared inspection and shared repair part of its security model. Open weights may fail to repeat that result because released capabilities can survive an attempted withdrawal. That is where the analogy breaks.
Still, I think distributing AI power may become a safety feature for humanity. It would give society more places to inspect and challenge the systems it depends on. I don't know if that benefit will outweigh the new risks.
Who knows? The safest AI future may turn out to be the one no company can close.
Open weights may never reproduce open source's patch-and-repair advantage. Their larger contribution could be decentralization: enough independent control that humanity's AI future cannot be owned by a single gatekeeper.
