Skip to content

Chapter 3: The AI Landscape — Models, Providers, and Aggregators

Companies spent $37 billion on generative AI in 2025—up from $11.5 billion in 2024. That's 3.2x growth in a single year.1 Meanwhile, the performance gap between the top model and the 10th-ranked model shrank from 11.9% to just 5.4%.2 The landscape is simultaneously exploding in investment and converging in capability.

Understanding the AI landscape isn't about keeping up with news—it's about recognizing which capabilities make the "intelligence is cheap" thesis actionable for your specific context. The 107x cost collapse we discussed in Chapter 1 doesn't matter if you can't navigate the providers, aggregators, and architectural decisions that determine how cheaply you can actually deploy.

At Yirifi, we learned this lesson early. We started with a single provider—simple, clean, fast to integrate. Then came the latency issues for real-time features. Then the rate limit surprises during a demo. Then the 40% price increase. Now we route requests through what we call our AI Tool Gateway: different models for different tasks, automatic failover when providers hiccup, cost optimization that happens without anyone thinking about it.

"We don't commit to one provider. Our AI Tool Gateway routes requests based on task type, cost constraints, and availability. Model routing isn't a nice-to-have—it's how you avoid vendor lock-in while getting the right model for each job."

Universal insight: The AI stack has four layers: Foundation Models → Providers → Aggregators → Applications. Decisions at each layer compound. The companies that win treat model selection as ongoing optimization, not a one-time decision.

Memorable close: "Routing is strategy. One provider is a bet."

What You'll Learn

  • The 4-Layer AI Stack: The mental model for navigating foundation models, providers, aggregators, and applications—plus how Perplexity's 38-person team achieves 91.3% optimal model selection with their "smallest viable model" routing.

  • Foundation Models Landscape: Claude's 97.8% security compliance for code, GPT's 85.4% multimodal dominance, Gemini's 2M token context windows, and DeepSeek's $0.07/million token disruption. The open vs. closed decision framework and the $50K threshold for hybrid approaches.

  • Provider Landscape: OpenAI's 99.3% uptime reality, Anthropic's 90% cache discount (with hidden costs), Google's 99.9% SLA commitments, and the four-step decision framework that actually drives provider selection in production.

  • The Aggregator Layer: OpenRouter's 5% markup tradeoff, LiteLLM's 3-17ms self-hosted alternative, and why RouteLLM achieves 85% cost reduction while maintaining 90-95% of GPT-4 quality.

  • The 6 Questions Before Choosing Any Model: From Stripe's sub-300ms fraud detection requirements to the 42% AI initiative abandonment rate—the framework that cuts through benchmark hype.

  • When to Fine-Tune: The decision hierarchy (prompt engineering → RAG → fine-tuning), why 73% of fine-tuning projects fail ROI, and how CFM achieved solutions 80x cheaper than large LLMs with LoRA.

  • Future-Proofing Your Stack: The abstraction trap that sank early LangChain, four patterns that survive model churn, and why the multi-model future requires architecture decisions today.

The Real Question

The previous chapter established the mindset. Now comes the toolbox.

For startups, use aggregators early. Flexibility matters more than marginal cost savings from direct contracts. You don't know which model will be best in six months—and neither does anyone else.

For established organizations, start with a provider but architect for routing from day one. Your first integration will change. Your compliance requirements will narrow options. Your volume will eventually justify custom terms. Build the abstraction layer before you need it, not during the crisis when your primary provider raises prices 40%.

The model landscape will keep shifting. What won't change: the companies that treat model selection as ongoing optimization rather than one-time decision will consistently outperform those locked into rigid architectures.

Let's find out.

References

Part Overview | Book Contents

  1. Menlo Ventures. "2025: The State of Generative AI in the Enterprise." ↩

  2. Stanford. AI Index Report 2025 ↩