Who Will Win the LLM Wars

We’re wired to think in winner-take-all terms.

One company dominates a market, one sports team wins the championship, and one AI model rules them all.

But that’s not how the future of AI will unfold.

Instead of a single dominant LLM, we’ll use multiple models—some public, some private—routed dynamically based on security, expertise, and efficiency.

ChatGPT for general tasks, an EnterpriseGPT trained on your proprietary data for internal insights, and specialized LLMs for industry-specific knowledge.

This is how Mixture-of-Experts (MoE) models already work, and it’s how enterprises will structure AI going forward.

The future of AI isn’t about one model to rule them all—it’s about choosing the right model for the right job at the right time.

☕️ AI Tangle - Is OpenAI's and Microsoft's Relationship Growing Even More Dicey?

 🔮 AI Lesson - ChatGPT for Conference Survival

🎯 The AI Marketing Advantage - How Google’s AI Mode Impacts Search Ads

💡 AI CIO - A First Look At Rebels of Reason

 📚 AIOS - This is an evolving project. I started with a 14-day free Al email course to get smart on Al. But the next evolution will be a ChatGPT Super-user Course and a course on How to Build Al Agents.

The Age of LLM Routing: Right Model, Right Task

The next wave of AI adoption won’t be about picking OpenAI, Claude, or Gemini. It will be about LLM routing—the process of intelligently directing prompts to different models based on context, security, and cost efficiency.

Here’s how it will work:

1. General-Purpose Queries → ChatGPT, Claude, Gemini

   • These models will continue to serve as broadly capable assistants for everyday tasks: drafting content, summarizing reports, brainstorming, or answering knowledge-based queries.

2. Enterprise Data & Security-Sensitive Queries → EnterpriseGPT

   • Companies will deploy internally-hosted, fine-tuned LLMs trained on their own proprietary data. These models—running on internal infrastructure or cloud environments with strict access controls—will handle queries involving sensitive, compliance-heavy, or domain-specific data. This is where I think models like IBM’s open-source Granite fit into the picture.

3. Industry-Specific & Domain Experts → Specialized Models

   • Certain queries—medical, legal, engineering—will require highly specialized AI models. Instead of relying on general-purpose LLMs that may hallucinate or lack depth, businesses will tap into domain-trained AI models built for high-precision tasks. For example, Lexis Nexis Lexis+ AI is an integrated solution for legal drafting, research, and insights.

4. AI Orchestration via LLM Routers

   • The glue that holds this system together will be LLM routers—software middleware that directs prompts to the right model based on access control, expertise, and efficiency. Think of it as an AI load balancer that ensures each request is handled by the most suitable model, much like modern cloud applications distribute workloads across different compute resources.

This Mixture of Experts (MoE) design pattern is already being used in cutting-edge models like DeepSeek and Mistral, which dynamically activate different model pathways depending on the complexity of a query. Enterprises will take the same approach—not choosing one model, but intelligently leveraging multiple models for optimal results.

The End of “One AI to Rule Them All”

Historically, businesses have consolidated around a single software provider (think Microsoft for enterprise productivity, SAP for ERP). AI will not follow that model.

Here’s why the future is multi-model, not monolithic:

• Security & Compliance → A single, centralized model means single-point vulnerability. Enterprise data needs to be compartmentalized across models with varying security levels—some public, some on-prem. That way, organizations can control which data goes to which source.

• Cost Optimization → Running every query through a high-end, proprietary LLM is wasteful. Instead, AI workflows will prioritize efficiency—leveraging lightweight, task-specific models for simple queries while reserving high-cost inference for complex reasoning tasks.

• Regulatory Constraints → Governments and industry regulators are already pushing back against the idea of a single AI provider having total control over data flows.

• Competitive Differentiation → If every company is using the same LLM, there’s no competitive moat. Proprietary models—trained on unique company data—will be a key differentiator, just as proprietary databases have been for the last two decades.

The idea that one AI model will dominate everything is as outdated as the idea that one database would power the entire internet. The future is about AI ecosystems, not AI monopolies.

How Enterprises Should Prepare for a Multi-Model Future

If you’re leading an AI strategy in your company, here’s what you should be thinking about today:

1. Deploy an LLM Router as Your AI Gateway

Instead of having employees manually select between OpenAI, Claude, or an internal model, businesses should build or adopt an LLM routing layer that automatically determines the best model for each query.

Think of it as an AI load balancer, ensuring that:

• Public models handle low-risk, generic queries.

• Enterprise models process sensitive, internal data.

• Industry-specific models execute specialized tasks.

This router could be API-driven, seamlessly integrated into corporate workflows, and equipped with audit logs to track model usage across departments.

2. Invest in Private, Enterprise-Specific LLMs

Public LLMs are powerful, but they’re not enough for enterprise decision-making. The real advantage will come from fine-tuning models on proprietary data.

• Train private LLMs using internal knowledge bases, contracts, customer interactions, and proprietary research.

• Deploy them on private cloud environments or on-prem infrastructure for maximum security.

• Use vector databases to enrich responses with proprietary context.

3. Optimize for Cost and Performance

Running every query through GPT-4.5 is expensive. Strategic AI usage means:

• Routing low-complexity requests (e.g., FAQ-style questions) to cheaper, distilled models.

• Using open-source LLMs (like Mistral or LLaMA 3) for cost-effective on-prem deployments.

• Reserving premium models for complex, high-value reasoning tasks.

By strategically tiering AI usage, enterprises can maximize efficiency while keeping costs under control.

4. Avoid Vendor Lock-In—Think Modular, Not Monolithic

The AI landscape is shifting fast. Locking your entire business into one AI provider today could be a strategic misstep.

• Use APIs and modular architectures to stay adaptable.

• Leverage open source models where possible to maintain flexibility.

• Keep an exit strategy in place—contracts should allow for easy migration if better models emerge.

5. Build AI Governance Early

As AI adoption scales, governance becomes critical. Companies must define:

• Who gets access to which models? (Define security tiers for LLM routing.)

• What data can be processed by public vs. private models? (Data compliance policies)

• How do we ensure auditability of AI-generated outputs? (Logging and oversight). This will require a strategy to do model evaluation (this is how we monitor how these models do what they do, similar to the way we do observability in cloud deployments).

Without governance, AI at scale becomes a black box liability. Companies that implement structured AI policies now will have a clear advantage as regulation tightens.

The AI Race Isn’t About “Winning” – It’s About Evolving

The idea of one AI model dominating everything is as outdated as thinking one database could power the entire internet.

AI is not a zero-sum game. The real winners will be those who build adaptive AI ecosystems—leveraging multiple models intelligently, optimizing for cost and efficiency, and integrating AI deeply into proprietary workflows.

The future of enterprise AI is not about choosing a single model. It’s about choosing the right model for the right task at the right time.

And that’s not a race with one finish line. It’s an evolving strategy—one that the smartest enterprises are already building today.

BoltAI - Switch between top AI services and local models. All from a single native app on your Mac.

POE - Poe is a platform that lets users interact with AI chatbots. It's a place to explore and experiment with different AI models, from simple inquiries to complex problem-solving.

LiteLLM - LiteLLM simplifies model access, spending tracking, and fallbacks across 100+ LLMs.

Kong AI Gateway - The AI Gateway provides a normalized API layer allowing clients to consume multiple AI services from the same client code base.

Extracting Text from Images Without OCR

Every spend all day in a meeting jotting notes on a whiteboard? Then you take a picture and send it around?

But then all you have is a picture to refer to; you can't update the dry-erase board from your desk.

What do I do when I need to pull text from an image on a whiteboard, slide, or infographic? I use ChatGPT to help—no special OCR (optical character recognition) software is required.

I take a picture with my phone and then upload it to ChatGPT.

Then I use a variation of this prompt.

Analyze the attached image and extract all visible text. If it's a slide or infographic, preserve headings and bullet points. If it's a whiteboard, list key points separately from equations or diagrams.

Three Useful Examples

1. Extracting Text from a Whiteboard Session
   Take a picture of your whiteboard with your phone and upload it to ChatGPT through the mobile app.

Extract all text from this whiteboard image. Separate key discussion points, action items, and any diagrams that contain text. Summarize bullet points for clarity.

Use Case: Capturing brainstorming sessions or meeting notes without manual transcription.

2. Extracting Text from a Presentation Slide

Ever find a meticulously created flow chart in a presentation and you’d love to extract the concepts? Or a picture with stats that you’d like to copy into a spreadsheet? Ever spend all day in a meeting jotting notes on a whiteboard? Here’s a shortcut. Screenshot it and then upload it to ChatGPT. Here’s an example prompt for that.

Pull the text from this slide image. Maintain structure by organizing it into a title, main bullet points, and speaker notes if present. If there’s a list of stats create a table and make it downloadable in CSV format. 

Use Case: Quickly converting slide content into editable text for documentation or reference.

3. Extracting Data from an Infographic

Ever seen a killer chart but it’s an image and you want the included data? Upload the image via ChatGPT and use this prompt.

Extract all text from this infographic and organize it into a table with columns for 'Section Title' and 'Key Data Points.'

Use Case: Transforming complex visual data into structured, reusable text for reports or presentations.

This approach goes beyond basic OCR by preserving formatting, context, and structure, making extracted text immediately useful.

As a special thank you for being part of our community, we're excited to give away 10 tickets to the All Things Open AI conference happening on March 17-18 in Raleigh, North Carolina.

If you're interested in AI and want to connect with like-minded individuals, this is your chance!

To grab your pass, use the code ”AllThingsAIxTheAIE” when registering. Don’t miss out on this opportunity to learn, network, and be part of something special.

I appreciate your support.