TL;DR: Enterprise AI has three core layers: intelligence, the operating layer, and proprietary data. As intelligence becomes cheaper and models become good enough for most tasks, long-term value will shift toward the workflows, integrations, and feedback loops that make a business unique. The next phase of enterprise AI will be a race to control that operating layer, making sovereignty over proprietary data and workflows a critical source of alpha. Companies should use external AI platforms for generic work while keeping differentiated processes within company-controlled infrastructure, using local or open-weight models when performance and economics justify it.

First is intelligence: the models themselves, where the cost per unit of capability continues to fall.

Second is the operating layer: the systems, workflows, and logic through which the business runs. There is significant value in owning this layer.

Third is data. When proprietary data is connected to real workflows and continuous feedback loops, enterprise AI begins to compound.

The model still matters. But increasingly, the long-term value of enterprise AI resides in the operating layer and the proprietary data that make the model more relevant, personalized, and effective.

The first phase of AI adoption was centered around access to intelligence. Companies purchased subscriptions and tokens for ChatGPT, Claude, Gemini, Cursor, and other tools their teams wanted to explore. That phase was useful because it helped people understand what these systems could do, where they were effective, and what their limitations were. But access to AI is not the same thing as becoming an AI-native company.

Many companies are now recognizing the risks of assembling a fragmented collection of AI tools, scattered prompts, disconnected integrations, data silos, and manual workarounds. Enterprise software followed a similar path, often leaving companies with costly and disjointed technology stacks. There is little reason to recreate that problem with AI.

The goal should be interoperability within a vertically integrated, company-controlled system.

The next phase of enterprise AI will not be defined solely by individual AI vendors. It will be defined by company-specific platforms that understand the business, its workflows, its data, and the operating logic behind how it actually runs.

This is where the ontology and platform layers matter.

Every company needs a shared map of its customers, products, contracts, data, systems, permissions, workflows, and exceptions. But the ontology is only the foundation. The most valuable layer is what gets built on top of it: agents, orchestration, integrations, governance, observability, reporting, and execution. This is where AI moves from being a productivity tool to becoming part of the operating system of the business.

The strategies of OpenAI and Anthropic suggest that they understand a fundamental reality: intelligence is becoming increasingly abundant and less expensive. Models are reaching a level of performance that is “good enough” for most business tasks. Not every workflow requires the latest model with the most horsepower.

As intelligence becomes commoditized, the greater opportunity is to own the operating layer through which businesses apply it.

This helps explain the demand for forward-deployed engineers and the urgency with which leading AI companies are building those teams. These engineers do more than implement software. They embed platforms directly into a company’s workflows, adapting them to the exceptions, context, and operating realities that make each business distinct.

Once an AI platform is embedded deeply enough, its provider is no longer selling only intelligence. It gains visibility into how work gets done, where exceptions occur, and how decisions are made. Each interaction deepens the platform’s understanding of the operating environment through the distillation of data. Depending on the contractual protections and data policies in place, that information could also be used to train future models and products.

The market remains focused on model performance, but the long-term value lies in the operating layer, the data, and the feedback loop between them. A model sitting outside the business can be useful. A model embedded within the company’s workflows and informed by its proprietary data is far more powerful and valuable. That is vertical integration.

If an AI provider owns this layer inside your business, your data may be the only differentiated asset left. You are then trusting that provider to protect it.

That is where companies need to be careful. Once your workflows and data begin fueling someone else’s system, you have to ask two basic questions: What do I still own, and where does the value of my business reside?

This matters for every enterprises. Well-run companies are trying to improve margins, speed up execution, reduce manual work, compress workflows, improve the customer experience, and build enterprise value. That does not happen because a management team bought a few expensive AI subscriptions. It happens when AI is built around the company’s actual mandate, workflows, and data. The last part being the most critical.

So how do I ensure that my workflows and my data remain mine?

The growing capability of open-weight models may offer part of the answer.

Historically, open-weight models required a meaningful trade-off. They could be less expensive and provide companies with greater control, but their performance often declined on more complex tasks—particularly those requiring advanced reasoning. Frontier models maintained a clear advantage.

The release of GLM-5.2 suggests that this gap is beginning to narrow. It is one of the first open-weight models capable of competing more directly with frontier systems on meaningful performance measures. Its reasoning capabilities rival those of the latest frontier models at roughly half the cost.

Fable and GPT-5.6 remain among the strongest-performing models in the market, but the gap between leading closed models and their open-weight alternatives is beginning to narrow. Using Fable for most tasks is like commuting in a Formula 1 car. The capability is exceptional, but fully utilizing it requires a level of technical sophistication, creative problem-solving, and workflow complexity that most everyday business tasks do not demand. For most applications, it is simply overkill.

The relevant question is not simply which model performs best overall. It is which model delivers sufficient performance for a specific workflow at the right cost, with the appropriate level of control.

If AI providers are seeking to own the operating layer and integrate themselves into a company’s workflows, businesses should ask how far that access extends. Contracts and assurances matter, but trust should also be reinforced by verifiable controls. The regulatory and legal history of the broader technology sector gives companies legitimate reason to remain cautious. At sufficient scale, financial penalties can become little more than a cost of doing business.

For highly proprietary workflows, one of the strongest ways to preserve control is to run open-weight models locally, build a company-controlled operating layer, and keep proprietary data within the company’s infrastructure.

That approach does not eliminate risk, but it reduces dependence on external providers and gives the company greater control over how its workflows and data are accessed, stored, and used.

What about security, particularly when many leading open-weight models are developed in China?

At roughly 40 billion parameters, concealing a meaningful backdoor is more difficult than many people assume. Because the model weights are open, the technical community can inspect and test them more closely, making suspicious behavior easier to identify.

That concern becomes more understandable as widely used models scale into the trillions of parameters, where the complexity and difficulty of inspection increase significantly. For now, however, I do not view it as a major concern when the model is run locally and appropriate security controls are in place.

With Thinking Labs and other U.S. companies investing in open models, I am confident that competitive U.S.-based alternatives will emerge before trillion-parameter models become the standard.

The fundamental goal of a business is to serve its customers with an excellent experience at an economically sustainable price. The more a company can reduce its cost to serve without compromising the customer experience, the more profitable and valuable it can become.

As the performance gap between frontier and open-weight models narrows, companies may no longer need to make the same trade-off between cost, capability, and control.

Where is it safe to use OpenAI and Anthropic?

If the workflow or data is generic, use them. If the workflow and data are proprietary, keep them in-house.

That means owning more of the workflow, integration, data, and decision-logic layers that make the business distinct. It also means remaining open to local or open-weight models when the economics, security requirements, and strategic value justify them.

The objective is not to be doctrinaire about open versus closed models. It is to understand where value is accruing, determine which parts of the technology stack the company should control, and protect the assets that preserve its differentiation.

The value of your business lies in maintaining sovereignty over your data and operating layer.

Make the market pay for your alpha. Do not give it away for free.

The information provided, including any accompanying materials and communications (collectively, the “Information”), is for informational purposes only and does not constitute investment advice or an offer to sell or solicit an offer to buy any securities, financial instruments, or investments. Plutus21 Holdings Inc. is a holding company whose subsidiaries and affiliates conduct separate and distinct businesses: Plutus21 Investment Management, L.L.C., Plutus21 Capital Management, L.L.C., Plutus21 Partners, Plutus21 Holdings Inc., and their affiliates are not responsible for any trading decisions, damages, or other losses resulting from the use of this Information. References to “Plutus21,” “we,” “us,” or “our” are used for convenience and do not imply that Plutus21 Holdings Inc. or any particular affiliate provides every service referenced in this communication. Any services are provided solely by the applicable Plutus21 affiliate and only pursuant to a separate written agreement and, where applicable, the relevant offering, advisory, subscription, or other governing documents.

This communication and any accompanying materials are provided solely for general informational and discussion purposes. They do not constitute legal, tax, accounting, investment, financial, or other professional advice; a recommendation; an offer to sell or solicitation of an offer to purchase any security, financial instrument, or investment product; or a commitment to provide consulting, investment-management, or other services. Recipients should consult their own professional advisers and independently evaluate the information in light of their particular circumstances.

Any opinions, projections, estimates, examples, case studies, or other statements contained herein reflect the views of the applicable author or affiliate as of the date presented, are subject to change without notice, and may involve assumptions and uncertainties. Plutus21 does not guarantee the accuracy, completeness, or continued availability of the information and undertakes no obligation to update it. Illustrative examples and case studies may not reflect actual results, and actual outcomes may differ materially.

To the extent this communication discusses securities, digital assets, investment strategies, or financial markets, such investments involve risk, including the possible loss of principal, volatility, and illiquidity. Past performance is not indicative of, and does not guarantee, future results. No assurance can be given that any investment strategy will achieve its objectives. Investment products and investment-management services, if any, are made available only through the applicable Plutus21 affiliate, to eligible persons, and in accordance with applicable law and definitive governing documents.

Plutus21, its affiliates, and their respective personnel may provide services to, invest in, hold positions in, or otherwise have interests relating to companies, assets, industries, or strategies discussed in this communication. Such interests may create actual or potential conflicts of interest, which will be addressed or disclosed as required by applicable law and the relevant governing documents.

Receipt of this communication, participation in preliminary discussions, or access to any materials does not create a client, consulting, investment-advisory, fiduciary, or other professional relationship with Plutus21 Holdings Inc. or any of its affiliates. Any such relationship will arise only through a definitive written agreement with the applicable affiliate. Nothing in this disclaimer limits any obligation arising under an existing agreement or applicable law.

Materials expressly identified as confidential or proprietary may not be reproduced, distributed, or disclosed without prior written consent, except to the recipient’s professional advisers who are subject to appropriate confidentiality obligations.

Additional legal and communications disclosures are available through the links provided below.
This Information is confidential and proprietary to Plutus21 Partners. It is intended solely for authorized recipients. Any unauthorized dissemination, distribution, or copying of this Information is strictly prohibited and may be unlawful.
Receiving this email does not create an advisory or fiduciary relationship between you and Plutus21.

Our full disclaimers can be found at: Full Disclaimers and Communications Disclaimers