Closed AI vs Open Weights: Who Should Hold the Keys?

Imagine two financial systems.

One is built around a single, highly regulated global bank. If something goes wrong, you know exactly who is responsible.

The other is a global marketplace where thousands of financial institutions compete relentlessly. Innovation is faster. Prices are lower. New ideas emerge constantly. But responsibility is dispersed across countless participants.

Neither system is obviously superior.

Artificial intelligence is beginning to resemble exactly this choice.

The debate is often framed as Closed AI versus Open Weights, as though the choice were between disciplined professionals and untrained amateurs.

It isn’t.

Some of the most capable open-weight models in the world come from serious research institutions, national AI programmes and world-class engineering teams. Skill is not the dividing line.

The real dividing line is accountability.

Closed systems generally bundle capability and accountability within a single commercial provider. Open-weight systems often separate capability from accountability, allowing responsibility to migrate downstream to whoever deploys the model.

That distinction, more than intelligence itself, may determine where economic value ultimately accrues.

Markets have always placed a premium on accountability. Investors routinely assign higher valuations to businesses whose risks can be measured, insured and legally enforced. AI may prove no different.

First, a definitional correction

Most models commonly described as “open source” are more accurately described as open-weight models. Users typically receive the trained parameters, but not necessarily the complete training data, its provenance, or every component required to reproduce the model from scratch.

This distinction matters because the traditional open-source ideal—inspect it, audit it and reproduce it—is only partially available. Open-weight models can be run, inspected and fine-tuned, but they cannot always be fully reproduced or independently audited from first principles.

Investors who treat all “open” models as one category may therefore be pricing something they have not fully defined.

The Emperor and the Marketplace

The case for concentration is real.

Large enterprises want a counterparty. When something goes wrong, there must be a company to call, a contract to enforce and, where appropriate, an indemnity to claim. That has genuine commercial value and helps explain why many regulated industries continue to favour managed AI providers despite the higher cost.

The case against is equally real.

If governments, banks, hospitals and universities all rely upon the same provider, that organisation begins influencing how humanity accesses knowledge itself. Not necessarily through malice, but through defaults—what the model emphasises, what it declines to answer and how it frames uncertainty. That represents a remarkable concentration of soft power, even when exercised responsibly.

The marketplace offers a different model.

Thousands of participants compete.

Most fail.

Some succeed spectacularly.

Competition lowers costs, accelerates innovation and reduces dependence on any single organisation.

Open weights make that possible.

What they do not necessarily do is concentrate accountability in a single organisation. Responsibility may instead be shared among model developers, cloud providers, systems integrators and the organisations deploying the technology.

This is not to say open-weight deployment escapes accountability. A bank that fine-tunes an open-weight model and deploys it to customers remains fully accountable to its regulator, regardless of where the model originated. Accountability has not disappeared; it has moved.

The difference is that responsibility shifts downstream to whoever deploys the model. When that organisation is a regulated bank, a global consultancy or a large technology company with robust governance, capital and audit functions, this model can work extremely well. When it is a small organisation with limited oversight and financial resources, the same allocation of responsibility becomes more challenging.

The investment question is therefore not whether accountability exists. It is where accountability resides, who has the balance sheet to support it and who ultimately captures the economic value of providing it.

Neither system is complete.

Concentration provides order.

Distribution provides resilience.

History suggests that durable institutions rarely choose one over the other. They find ways to balance both.

AI Is Becoming Society’s Judgement Layer

The comparison to electricity or the Internet understates what is happening.

Electricity gave us power.

The Internet gave us information.

AI increasingly gives us judgement.

Which shares to buy.

Which medicine to prescribe.

How to draft a legal agreement.

How to educate a child.

Eventually, perhaps, how to negotiate treaties or design scientific experiments.

Throughout history, societies have wrapped judgement inside institutions. We created courts, medical boards, auditors and central banks. Not because the people inside those institutions were infallible, but because judgement without accountability is where systems begin to fail.

AI is the first technology attempting to scale judgement before we have fully built the institutions around it.

Investors Are Asking the Wrong Question

One model scores 92%.

Another scores 94%.

This is rather like arguing whether Ferrari is marginally faster than McLaren while ignoring which business actually generates superior returns on capital.

The more interesting question is not which model is smartest.

It is which organisations society will trust to make decisions at scale.

Banks care about governance.

Hospitals care about safety.

Governments care about accountability.

Trust is therefore not simply a moral virtue.

It may become a commercial moat.

Investors understand this well.

A bank’s greatest asset is not its vault. It is the confidence that depositors will return tomorrow.

Insurance provides another useful lesson. It separates risk from the underlying asset and prices that risk independently.

AI may eventually evolve in much the same way.

If governance, liability and indemnification become standalone services, trust may no longer belong primarily to the model developer. It may instead belong to whoever assumes and manages the risk.

That possibility would fundamentally change where economic value accrues, and it is one reason I hold this investment thesis with humility rather than certainty.

What trust may look like on a balance sheet:

• Long-term enterprise contracts and high renewal rates.

• Willingness to provide contractual indemnification.

• Penetration of highly regulated industries such as finance, healthcare and defence.

• High switching costs created by workflow integration rather than benchmark performance.

• Regulatory certifications and approvals that competitors cannot easily replicate.

What Would Prove Me Wrong?

Every investment thesis deserves to be tested.

Mine is no exception.

Commoditisation. If open-weight models narrow the capability gap while costing a fraction as much, many enterprises may choose cheaper models and build governance internally through insurance, compliance and internal controls.

History reminds us that superior technology does not always produce superior investment returns. Linux became one of the world’s most successful operating systems without creating a trillion-dollar Linux company. Open standards helped build much of the Internet, while many of the largest fortunes were created elsewhere in the ecosystem.

Liability law. If legal responsibility ultimately settles on the deployer rather than the model provider, much of today’s accountability premium may migrate away from frontier model developers.

The infrastructure layer. The most uncomfortable possibility is that this entire debate is settled beneath the model layer. If compute, memory, networking and energy infrastructure capture most of the economics regardless of which model wins on trust, then trust remains an important governance question but a weaker investment thesis.

My own positioning continues to favour the infrastructure ecosystem precisely because it does not require me to predict which model provider ultimately prevails.

Where This Leaves Me

Open weights are likely to continue driving innovation and reducing costs.

Closed systems are likely to remain attractive for organisations that prefer accountability bundled into a single commercial relationship.

Both approaches may coexist for many years because they solve different economic problems.

The investment question is therefore not whether AI should be open or closed.

It is where accountability ultimately resides, who has the balance sheet to support it, and who captures the profit pool created by trust.

That, more than benchmark scores alone, may determine where the greatest long-term investment opportunities emerge.