Mozilla’s CTO thinks AI should be built like the internet

The internet became powerful because no single company had to bless every connection before value could move across it. That is the hidden force inside Mozilla’s argument for AI: not nostalgia for a freer web, but a warning about what happens when intelligence becomes a deployment layer owned through contracts, clouds, app stores, chips, and institutional defaults. The question is not whether AI should be “open” in the abstract. It is whether the next general-purpose infrastructure will let new actors compose with it, contest it, and localize it — or whether they will only rent permission from whoever controls the narrowest bottleneck.

The tension inside AI should be built like the internet

When Mozilla’s CTO argues that AI should be built more like the internet, the phrase sounds almost too clean. The internet is the rare technical metaphor that still carries moral authority: interoperable protocols, permissionless publishing, global reach, messy pluralism. It suggests a world where no single actor owns the interface between human intent and machine execution.

But AI is not arriving as a neutral protocol. It is arriving through products, subscriptions, enterprise procurement, sovereign cloud deals, foundation-model licenses, content pipelines, and regulatory negotiations. That difference matters. The internet separated transmission from the applications built on top of it. AI collapses more of the stack into a few operational chokepoints: training data, compute, model weights, inference hosting, identity, safety layers, distribution, and user interface.

This is why the internet analogy is useful only if it is treated as a stress test. If AI becomes infrastructure, openness cannot mean merely releasing a model file or publishing a benchmark. It has to mean that builders outside the dominant centers of capital can adapt the system without asking permission at every layer.

That pressure is visible in Mozilla’s CTO thinks AI should be built like the internet, which matters here less as a headline than as evidence of where deployment authority is hardening.

Why the easy reading is too small

The easy reading is that Mozilla is making an open-source argument against Big Tech concentration. That is true, but it is not enough. A simple “open versus closed” frame lets everyone claim the right side. A model can be open-weight while still depending on closed data pipelines, expensive hardware, privileged distribution, or hosted services that few competitors can replicate.

Meta’s release of an open agentic model shows the ambiguity. Open-weight releases can widen experimentation and help researchers inspect, fine-tune, and deploy systems outside proprietary APIs. They can also strengthen the incumbent’s gravitational field by making its architecture, tooling, and assumptions the default substrate for everyone else. Openness at one layer can become dependency at another.

That is the trap in the mainstream reading. It treats openness as a property of artifacts. In deployment, openness is a property of power. Who can run the system? Who can afford it at scale? Who can modify it for a language, legal regime, school system, hospital workflow, or national security constraint? Who absorbs liability when it fails?

This connects directly to the question I raised in Justice-Centered AI Has to Leave the Workshop: justice does not become real because a concept is available. It becomes real when institutions can operationalize it under pressure.

The control mechanism underneath the announcement

The real mechanism is not ideology. It is stack position.

AI power accumulates where decisions become defaults. The actor that controls the model may matter less than the actor that controls deployment: the cloud vendor that prices inference, the platform that places an assistant into daily workflow, the procurement office that selects the approved tool, the state agency that defines compliance, the chip supplier that determines what can be run locally, the browser or operating system that decides which agent gets surfaced first.

That is why the open-source AI report matters beyond its taxonomy. The fight is no longer just about whether code is visible. It is about whether the conditions around AI allow meaningful participation after visibility. A country, startup, newsroom, university, or public agency may be able to download weights and still lack the compute budget, technical staff, evaluation system, or legal confidence to deploy them responsibly.

The Global South lens sharpens this. Many countries already learned the cost of joining the digital economy through infrastructure they did not control: undersea cables, hyperscale clouds, payment rails, mobile operating systems, ad markets, app stores. AI risks repeating the pattern at a higher layer. Local firms may build applications, but the strategic margin sits elsewhere: inference pricing, model access, compliance tooling, identity, data localization, and uptime.

That is why the cloud-contract question is not administrative trivia. As I argued in The AI Budget Is Hiding Inside the Cloud Contract, the budget often reveals the power map before the press release does.

Who inherits the deployment constraint

Builders inherit it first. The more AI becomes infrastructural, the less the hard problem is prompt craft and the more it is durable operation: latency, monitoring, security, evaluation, user trust, data rights, and cost control. Open models help, but they do not remove the burden of making intelligence behave inside real institutions.

Operators inherit it next. Hospitals, schools, banks, courts, logistics firms, ministries, and telecoms will not adopt AI as a philosophy. They will adopt vendor packages that survive procurement, audits, staff training, and political scrutiny. That gives enormous advantage to whoever can translate AI capability into institutional safety. The control layer forms around the boring parts.

Investors inherit a different constraint. The obvious capital rush goes toward models, agents, and applications. The more durable question is which companies sit near enforcement points: identity, evaluation, observability, secure deployment, data governance, localization, and inference efficiency. Infrastructure wealth often flows to the layer that others must pass through.

States inherit the most difficult version. Sovereignty in AI will not be proven by announcing a national model. It will be proven by whether local institutions can run critical systems without total dependency on foreign platforms. This is the same pattern behind orbital connectivity: connectivity, compute, and intelligence are converging into sovereignty questions because they define who can keep operating when politics turns hostile.

The test for whether power actually moves

The decisive test is not whether more AI systems are described as open. It is whether more actors can make consequential changes without negotiating with the same small set of gatekeepers.

That means asking harder questions than the announcement cycle prefers. Can a regional health ministry adapt a model to local languages and clinical norms without sending sensitive data through foreign infrastructure? Can a startup in Nairobi, Jakarta, São Paulo, or Manila serve customers at reliable inference costs without being crushed by cloud dependency? Can a university lab audit a model’s behavior and deploy alternatives, not just publish critique? Can public institutions choose tools that reflect local law rather than platform policy disguised as inevitability?

Mozilla’s internet analogy is strongest when it forces this standard. The old web was not pure. It produced monopolies, surveillance markets, spam, and platform capture. But its early architecture left enough open surface for unexpected builders to enter before the gates hardened. AI may not get that grace period. The capital intensity is higher, the state interest is sharper, and the distribution channels are already consolidated.

So the question is not whether AI should be built like the internet. It is whether anyone with power is willing to give up enough control for that sentence to become more than branding. Open weights are a start. Open deployment is the fight.