Build a Portfolio of Intelligence

Why AI strategy is about designing the right mix of models—not choosing a single vendor

One of the questions I hear most often from executives is: Should we use OpenAI, Anthropic, or Google?

It is a reasonable question. But it’s far too small and misses the point entirely. Choosing an AI provider might be an important procurement decision, but it’s not an AI strategy. The AI landscape now contains many different types of models. They perform different jobs, operate at different levels of capability, come with different ownership structures, and make very different trade-offs between cost, speed, privacy and performance. The companies that create the most value from AI will not simply select a favorite model. They will build a portfolio of intelligence. Including home grown models that yield unique competitive advantage.

To make sense of that portfolio, leaders need to understand three dimensions:

• Model family: What kind of intelligence does it provide?

• Ownership: Who controls it, and how much can you change it?

• Capability: How much intelligence does the task really require?

These categories aren’t clean technical boxes—models span across all of them. But they provide leaders with a far more useful mental model than a shortlist of chatbot vendors. Let’s look at each one in turn.

Six families of intelligence

Models are designed to fulfill different functions.

General-purpose models

General-purpose models are the broad thinkers of the AI world. They can converse, analyze documents, reason through problems, write code, create plans, use tools and support a wide variety of knowledge-work tasks. GPT, Claude and Gemini are the best-known examples. Leading models are now multimodal, meaning that they can work across combinations of text, images, audio and video rather than being limited to language alone.

Reasoning is increasingly a capability within this family rather than a separate species of model. A general-purpose model might respond immediately to a simple question, then shift into a deeper reasoning mode when asked to analyze a strategy, debug complex software, or solve a difficult scientific problem. These are the models most leaders encounter first. But they are only one part of the landscape.

Media-generation models

Media models create images, video, speech, music and other forms of synthetic content.

Models such as Nano Banana, ChatGPT Images, and Midjourney create images, Google Veo, Seedance, and Runway generate video, and ElevenLabs, Suno, and Google Lyria produce music and audio. Other models specialize in image editing, voice synthesis, animation or three-dimensional assets.

These models are transforming marketing, product design, entertainment, training, simulation, content creation, and communications. A company might use a general-purpose model to develop a campaign concept, a media model to produce the visual assets, and a predictive model to determine which version is most likely to perform well.

Predictive models

Predictive models are the quiet workhorses of enterprise AI. They don’t write essays or generate cinematic video. They predict what is likely to happen next, and in some cases what has already happened.

The predictive family includes models for demand forecasting, fraud detection, churn prediction, credit risk, preventative maintenance, classification, anomaly detection, ranking, and recommendation. The recommendation engines behind digital commerce, streaming services, and social feeds belong here. They typically generate possible choices, score them and then re-rank them to determine what a user is most likely to want next.

Predictive AI has been generating enormous business value for over 15 years. Generative AI, relatively the new kid on the block, has not replaced it. In many cases, the two work together: a predictive model identifies which customer is likely to leave, while a general-purpose model helps a service agent understand why and decide what to do about it. Starbuck’s created an AI model called DeepBrew. The model uses predictive models to hyper personalize offers for an audience of one, then a generative model to further personalize and generate unique language designed to resonate with the customer and elevate conversion rates.

World models

World models learn how environments behave and how the world works. Rather than simply responding to a prompt, they represent objects, movement, cause and effect, and how a situation might evolve over time. They generate or simulate possible future states of an environment. For example, Google DeepMind’s Genie creates interactive environments from descriptions. NVIDIA’s Cosmos models are designed to generate physics-aware worlds and future states for robotics, autonomous vehicles and other forms of physical AI.

World models will be important for training robots, improving the reliability of autonomous vehicles, simulating factories, designing products, and exploring scenarios that would be too expensive, dangerous or slow to test in reality. Expect to hear a lot more about world models in the coming 12 months.

Physical AI models

Physical AI models perceive the real world and take action within it. Many are described as vision-language-action (VLA) models. They combine visual perception, language instructions, and physical control over actions. So, a robot can see an object, understand a request, plan a sequence of movements, and then execute them.

Leading humanoid robotics company, Figure, built a proprietary model named Helix that turns visual information and instructions into motor commands so its robots can perform complex physical tasks.

World models and physical models are closely connected, but they perform different roles. World models simulate what might happen while physical models decide what to do. A debate is currently raging in the robotics world over whether VLAs or world models will ultimately power robots of the future.

Specialty models

Specialty models trade breadth for depth. They’re built for a particular domain, scientific discipline, or tightly-defined tasks. For example, a specialty model might analyze medical images, discover new materials, interpret legal contracts, detect cyberattacks, or understand the language of financial markets.

DeepMind’s AlphaFold is a powerful example. This model was designed to predict the three-dimensional structures of proteins and has helped reveal many millions of them, accelerating medical research and our understanding of biology in the process.

Specialty models may use the same underlying technologies as general-purpose models. What distinguishes them is the depth of their training, data and optimization for a particular field. BloombergGPT is a large language model that was trained on Bloomberg’s extensive financial data to perform finance-specific language tasks.

In a world where general intelligence is becoming widely available, this kind of specialized intelligence may become one of the most important sources of competitive advantage. The most highly competitive companies of the future will develop speciality models based on proprietary data that creates unique value and a competitive moat. More on that in my next post.

Who owns the intelligence?

The second dimension concerns ownership and control. Consider four main types of models: Closed, open weight, open source and home grown.

These types aren’t completely mutually exclusive. A home-grown model, for example, may begin with an open-weight model. But they describe four meaningfully different ways an organization can access and develop intelligence. It’s important for leaders to understand all four, especially open weights.

Closed models

With a closed model, the provider controls the model, its weights, and the process used to train it. Customers typically access it through an application or API. ChatGPT, Claude, and Gemini are mainstream examples of this approach.

Closed models offer convenience and rapid access to frontier capability, but you are effectively renting intelligence. The provider handles training, infrastructure, maintenance and continuous improvement. The trade-off is that the customer has less control over the underlying model and is dependent on the provider’s pricing, policies, availability and product roadmap. And as we recently found out, you might have access to these models cut off on a government whim.

Open-weight models

With an open-weight model, you can download the trained model weights and run the model yourself.

This allows an organization to deploy it in its own cloud or data center, operate it locally, fine-tune it for a particular domain, distill it into a smaller model or otherwise adapt its behavior. Clients use open-weight models to control costs, build organizational resilience, and enable extensive customization to yield home grown models. More on that in a moment.

Examples of open-weight models include Google’s Gemma, Alibaba’s Qwen, and models from DeepSeek, Mistral, and Kimi. A relative newcomer is Chinese model maker, Z.ai with their impressive GLM-5.2 reasoning model that you can use for a fraction of the price of a competitive frontier model. At the time of writing, Chinese models offer impressive price/performance and capability that only lags U.S. frontier models by 3-6 months. But some companies still won’t touch them due to security risks and worry that models might perform code injection when writing software, opening up backdoors.

What you don’t necessarily receive with an open-weight model is everything used to create it: the complete training data set, data-processing pipeline, training history, and a full recipe. You get the finished brain, but you usually don’t get the full story of how it was built.

Open-source models

The term “open source” is frequently used loosely in AI marketing, so businesses should look beyond the label and examine what is actually available. With a genuinely open-source AI model you get the weights together with the code, training recipe, data information, and other components needed to study, modify and reproduce the system. The Open Source Initiative explicitly distinguishes this from simply releasing weights.

Ai2’s OLMo family is one of the clearest examples of a truly open-source model. OLMo releases include accessible training data, training code, recipes, checkpoints and evaluation materials—not just the final model. Thousands of open-source models are also available to download and use on sites like Hugging Face and GitHub. Many of these models are specialist models designed for particular tasks such as text-to-speech, anomaly detection, image generation, or computer vision.

Home-grown models

A home-grown model is built around an organization’s proprietary knowledge, data, decisions and experience. It might be trained from scratch, but that is becoming increasingly unusual for most companies. More commonly, an organization will begin with an open-weight model and fine-tune, distill or continue training it using its own information.

For example, a bank might build a model that understands its risk policies and historical fraud patterns. A manufacturer might create one that has learned from years of equipment telemetry and maintenance decisions. A pharmaceutical company might develop one around its compounds, experiments and research history.

Ultimately, what matters is the resulting unique intelligence you create, not where the base model originated. This is where value and long-term competitive advantage will be created. Most businesses already possess proprietary data. The strategic opportunity is to transform that data into proprietary intelligence that competitors cannot simply buy from another vendor.

So we’ve covered model purposes and origins. Now let’s talk size and capability.

How much intelligence does the task need?

The third dimension is capability. In my last post, I argued that not every task needs a genius. The rapid improvement of AI has encouraged companies to reach automatically for the smartest model available, even when the work doesn’t require it. That’s like hiring a Nobel Prize winner to sort the mail.

I think about capability in three broad tiers: frontier, good enough, and embedded.

Frontier intelligence

Frontier models provide the highest levels of broad capability available at a given time. They’re appropriate for difficult reasoning, complex coding, scientific research, high-stakes analysis and unfamiliar problems where quality matters more than cost or speed.

But…frontier intelligence is generally far more computationally demanding. It costs more to serve, is therefore priced higher to use, and may take longer to respond. Users have been complaining about sluggish performance for Claude Fable 5 lately because Anthropic doesn’t have enough compute power to meet demand.

When it comes to frontier intelligence, it’s important to only use it where the additional intelligence creates additional value. Otherwise you’re wasting money and time.

Good-enough intelligence

Good-enough models aren’t bad models. They’re actually pretty good these days. These models meet the quality threshold for the work without providing expensive capability the task doesn’t need. A smaller open-weight model might be more than capable of classifying documents, routing service requests, extracting information from forms, drafting routine correspondence, or supporting a tightly defined agentic workflow. Why use a hefty large language model when a small language model (SLM) will do?

Good-enough models can be faster, cheaper, and easier to customize. They may also allow an organization to process much larger volumes of work within the same budget. The trick: Use the least expensive intelligence that reliably produces the outcome you need.

Embedded intelligence

Embedded models operate directly inside devices, products, vehicles, machinery and other local environments. They may run on phones, laptops, cameras, factory equipment, medical devices, or robots.

These models, such as Google Gemma 3n and the Microsoft Phi family, have been optimized for operation on everyday devices. They are typically distilled versions of larger, more powerful models, slimmed down to fit in the memory and storage footprints of edge devices.

Embedded intelligence can offer extremely low latency, work without a reliable internet connection, and keep sensitive data on the device. And small doesn’t mean stupid. Some of these models can pack decent intelligence, particularly on focused tasks. In many cases, a narrowly optimized embedded model is exactly what the application requires.

There is no universally best model

Every model choice involves trade-offs. More capable models may produce better results on difficult tasks, but cost more and respond more slowly. Smaller models may be cheaper and faster, but fail when faced with ambiguity or unfamiliar situations. Cloud models can provide access to enormous capability, while locally deployed models can offer greater privacy, control, and resilience.

These are the questions to ask and answer as you decide what models you need. What level of quality does the task require? How quickly must the model respond? How many times will it be used? Can sensitive data leave the organization? Does the model need to work offline? How much customization is required? What happens if the vendor changes its price, terms or model behavior?

Cost, latency, privacy, control, energy use, and capability are all architectural variables, and there’s no single model that optimizes all of them.

Build an intelligence portfolio

Consider a modern manufacturing operation. A predictive model forecasts equipment failure. A specialist vision model inspects products for microscopic defects. A world model simulates changes to the production line. A physical AI model controls a robot. A general-purpose model allows employees to interact with the entire system using natural language.

Some of those models may be closed. Others may be open weight. A few may be home grown because the company’s proprietary operating data creates a meaningful competitive advantage. Some work will require frontier intelligence. Much of it will use good-enough models. Time-critical decisions may rely on embedded intelligence operating directly on the factory floor.

The value comes not from selecting one model, but from combining many models intelligently. That requires a model portfolio, supported by routing and orchestration that sends each task to the appropriate form of intelligence.

So the hardest strategic question is therefore not: “Should we choose OpenAI, Anthropic or Google?” It’s deciding “which intelligence architecture will create the most value for our business?”

The future of enterprise AI won’t be built solely on the smartest models. It’ll be built by companies that know when to use frontier intelligence, when good enough is enough, when intelligence must be embedded, and where proprietary data can be turned into a home-grown competitive moat.

AI strategy is not vendor selection. It’s the design of a comprehensive intelligence portfolio that will accelerate your business and catapult you ahead of the competition.

Steve Brown is an AI futurist, global keynote speaker, and author. He advises Fortune 100 companies and global brands on AI strategy and transformation, helping them build AI-first organizations and stay ahead of the forces reshaping business, the economy, and society. He has delivered hundreds of high-impact keynotes across five continents, translating complex technologies into clear, practical action for leaders.

Steve’s latest book is “The AI Ultimatum: Preparing For a World of Intelligent Machines and Radical Transformation.” Get his book here, and learn more about Steve at www.stevebrown.ai.