Alex Karp's recent AI critique landed because the concern is real — especially at banks.
When
"Customers want control over their compute, their models, their data stack and their alpha," said Karp.
"They're like, 'I am paying for tokens that create no value. [The proprietary model providers] are stealing the weights and alpha of my business, and they're creating a wealth tax that does not help the poor."
Banks have been weighing open-source and closed-source models for some time. Karp's "wealth tax" comment—which is really about cost, control and value capture—brings new urgency to the question as AI usage moves from pilots to production.
For bank executives, the strategic task is not to choose one model architecture. It is to build a tiered AI architecture that assigns the right model to the right job, with the right level of control.
Three practical questions follow:
- Which bank workloads demand frontier proprietary models because they require strong reasoning, broader knowledge or faster integration into enterprise platforms?
- Which workloads can be handled by a smaller, task-specific models deployed in a controlled environment?
- Which workflows do not require generative AI at all?
The answer will vary by institution. A global bank with deep engineering capacity, mature model governance, and significant AI infrastructure may make different choices from a regional or community bank. Asset size, risk appetite, talent depth, vendor strategy, regulatory posture and business urgency all matter.
When I heard Karp raise his concerns about closed source AI models, I thought the topic would be excellent for my debut commentary as senior Market Intelligence analyst for AI at American Banker. The concern he had was about these proprietary models from AI companies such as OpenAI and Anthropic. It makes sense to me. A couple of weeks before his CNBC appearance, I asked almost the same question at
During a conversation with the chief AI officer of a large regional bank, I asked why banks were not using locally deployed open-source models for more of their internal workflows. The logic seemed straightforward. If a bank can keep the model inside its own environment, it may reduce token costs, retain greater control over sensitive data, and avoid sending proprietary information through external systems.
His answer was immediate: who is accountable? If a bank downloads and deploys an open-source model, who is responsible if the model contains hidden vulnerabilities, leaks data,
That is why bank CAIOs can agree with Karp's concern without treating open source as the easy answer. Once AI enters production, the harder question is not cost. It is accountability.
The accountability gap
Banks do not evaluate AI systems only on technical performance. They also have to consider model risk, cybersecurity, data lineage, third-party risk, auditability, explainability and supervisory expectations. These are not secondary issues. They are the conditions for adoption.
The challenge is that generative AI and agentic AI are evolving faster than the supervisory framework around them.
For bank leaders, that creates an accountability gap. The technology may be promising. The economics may be attractive. The business case may be clear. But if something goes wrong, the bank still owns the outcome.
That is one reason proprietary providers remain attractive. They do not eliminate risk. But they often provide documentation, security reviews, contractual protections, implementation support and a more established enterprise-control environment. For many banks, those features are part of what they are paying for. In that sense, proprietary LLM fees are not just a technology expense. They may also function as a form of perceived risk reduction.
Open-source models will almost certainly play an important role in bank AI strategy. However, the economics must be evaluated carefully.
Moving from vendors' proprietary models to open-source models shifts part of the cost from usage fees to infrastructure, talent and operations. A bank may reduce token spending, but it may need to invest in computing resources, security controls, model monitoring, MLOps capabilities and specialized engineering talent.
Open-source models can also create documentation challenges. If training data, model lineage, and fine-tuning processes are not sufficiently transparent, risk teams may struggle to explain the model to internal reviewers, auditors or regulators.
The result is that open source is not a simple substitute for proprietary AI. It is a different operating model.
From AI impact to accountable AI
Over time, however, the balance may shift. Many Wall Street firms initially banned or sharply restricted ChatGPT after its launch, only to begin using generative AI more broadly once internal controls, approved tools and clearer use cases emerged. Open-source models may follow a similar path. As AI becomes more embedded in core bank workflows, regulatory expectations become clearer, and management teams gain more comfort with model oversight, banks are likely to make greater use of open-source models in selected areas where cost, data control and customization matter most.
This is also why the bank AI conversation needs to move beyond pilots and headlines.
Many banks are already seeing impact from AI. The harder challenge is turning that impact into outcomes that are measurable, scalable and accountable. Measurable means the bank can connect AI use to business value. Scalable means the use case can move beyond isolated experiments. Accountable means the bank knows who owns the model, the data, the controls, the costs and the consequences.
Karp's "wealth tax" framing may be imperfect, but it points to a real issue. As AI usage scales, banks will have to become more disciplined about where they pay for frontier capabilities, where they build internal capacity, and where they avoid unnecessary complexity.
The next question for bank AI leaders is workload mapping: which use cases justify frontier proprietary models, which can move to controlled open-source environments and which require stronger accountability structures before either path is safe.
The industry does not need another abstract debate about whether proprietary or open-source AI is better. Bank leaders need a framework for deciding which workloads belong where, what risks need to be managed, and how AI spending should be connected to measurable business value.
The next phase of bank AI adoption will not be defined by model choice alone. It will be defined by architecture, governance and accountability.
For AI leaders, that is the real question.
