How open infrastructure kept the network contestable
By 2002, Internet Explorer accounted for more than 90 percent of browser use by Mozilla’s historical account. Microsoft had distribution power, control of the dominant desktop operating system and contractual restrictions that a US court found had accelerated its browser’s rise.
Netscape had released its browser code in 1998. The project became Mozilla. Firefox 1.0 arrived in 2004 and restored a serious independent browser engine during Microsoft’s dominance.
Firefox did not save the Web by itself. Antitrust enforcement, W3C standards, Safari, Opera, Chrome and millions of developers also shaped the outcome. Mozilla supplied something the market badly needed: a credible alternative implementation.
That is what “saved” means here. Open infrastructure repeatedly preserved interoperability and credible exit when control of a foundational layer was narrowing. It worked alongside public funding, standards bodies and commercial investment. The history is stronger without the fairy tale.
The internet was open in more than one way
TCP/IP is an open protocol suite, not one open-source program. Its specifications could be implemented by different systems. DARPA funding, the 1983 ARPANET transition, adoption in Berkeley Unix, government procurement and later commercial implementations all helped it spread.
The Web added another mechanism. On 30 April 1993, CERN placed its basic Web client, server and common-code library in the public domain. W3C later built a standards process around interoperability and royalty-free implementation.
These distinctions matter. A public specification lets different teams build compatible products. Public-domain or open-source code gives them an implementation they can inspect, modify and distribute. A royalty-free patent policy lowers legal barriers. No single mechanism carries the whole system.
Together, they preserved credible alternative implementations and made exclusive control over connection harder.
Shared software created credible alternatives
Linux began as a small kernel project, then became the work of a distributed contributor network. Its license allowed anyone to inspect, modify and redistribute the kernel. Adoption still required hardware support, distributions, services and corporate investment. IBM’s public commitment to Linux helped make it a credible option for organisations that required long-term support.
That investment did not corrupt an otherwise pure movement. It made the alternative more credible for organisations that needed support, continuity and accountability.
Apache followed a similar path. When development of the NCSA web server stalled in 1994, a group of maintainers consolidated patches and released Apache in 1995. The project says it became the leading server in a 1996 Netcraft survey. The exact share changed over time. The structural result lasted: a deployable web server existed outside a proprietary product roadmap.
Open databases widened the same route. PostgreSQL grew from a publicly funded Berkeley research project that began in 1986. MySQL’s first version appeared in 1995, and its company adopted the GPL in 2000 while also selling commercial licenses and services.
They did not eliminate database lock-in. Schemas, extensions, operations and managed services still create switching costs. They gave developers serious alternatives without requiring permission from a single database vendor.
Open infrastructure also created large businesses. IBM acquired Red Hat, and MySQL built a company around dual licensing before Sun acquired it. The lesson is not that free software defeats commerce. Shared foundations can move competition into service, reliability and execution.
Dependency becomes visible when it breaks
Mozilla and Git began under different pressures, but both exposed the cost of a strategic dependency.
Mozilla followed a browser market that had become dangerously narrow. Releasing the Netscape code did not guarantee success. It made an independent implementation possible after the original commercial position had weakened.
Git emerged in 2005 after the Linux kernel community’s relationship with the proprietary BitKeeper system broke down and its no-cost status was withdrawn. A usable foundation appeared quickly, then a much larger community built the system people use today.
Open-source collaboration existed long before Git. Git made distributed, non-linear development work better at the scale the kernel demanded. More importantly, the community could no longer lose its core collaboration system through one vendor’s decision.
That is credible exit in practice. A fallback matters only when people can actually use it.
AI is a harder case
The analogy breaks at the compute layer.
An internet protocol can be implemented on widely available machines. A frontier model can require capital, accelerators, energy, data and specialist talent available to very few organisations. Stanford’s 2026 AI Index reported that industry produced more than 90 percent of notable frontier models in 2025. Public company filings show infrastructure spending at a scale that smaller labs cannot match, although those capital budgets cover more than AI.
Openness therefore needs four separate tests.
Weights and licenses. Downloadable weights can enable local deployment, inspection and customisation. They do not necessarily include training code, sufficient data information or permission for every use. “Open-weight” is the accurate term unless the full release satisfies an explicit definition such as the Open Source Initiative’s Open Source AI Definition.
Interfaces. The Model Context Protocol connects AI applications to tools and context. Agent2Agent addresses communication between independent agents. Their open specifications can reduce integration friction. They cover different scopes, remain young, and do not guarantee that switching providers will preserve behaviour or safety.
Evaluations. Public test cases and reproducible methods make comparison possible. AI behaviour is probabilistic and sensitive to context, so an evaluation cannot provide the same deterministic guarantee as a networking conformance test. Open evidence helps. It does not turn judgment into a protocol.
Governance and credible exit. Foundation hosting and published decision rules can reduce single-vendor control. The stronger test is practical: are there multiple implementations, portable artifacts, conformance tests and enough contributors to sustain the alternative? Can a user leave without rebuilding the system around the model?
Open weights alone cannot answer those questions. Neither can an open protocol sitting above a closed identity, compute or distribution layer.
What the history actually offers
The internet’s open foundations were not inevitable victories. Government funding, academic work, community maintenance, commercial investment, regulation and standards institutions reinforced one another.
AI will need a similarly mixed system. Public compute can widen access. Open-weight models can create model-level exit. Open protocols can reduce integration friction. Transparent evaluations can make claims easier to test. Commercial firms can fund the engineering that turns all four into infrastructure people trust.
The conditions are measurable. Multiple independent implementations exist. Switching works in practice. Contributors can participate without one firm’s permission. A business can earn money without owning the underlying standard.
Open infrastructure kept the internet contestable because it made alternatives real. AI gets the same benefit only when openness reaches beyond the download button.