As a keen student of both AI systems and capital markets, I've watched the shift from traditional software economics to the token economy with fascination. What once looked like a straightforward SaaS evolution has become something far more fundamental: a new operating system for monetizing intelligence itself. The implications for investors are profound and many are still pricing AI companies through an outdated subscription lens.
In classic SaaS, you paid a monthly fee for access. Revenue was predictable, gross margins were high once the product was built, and success showed up in net revenue retention and low churn. Generative AI upended that. The dominant model is now token-based pricing—paying for every chunk of computation consumed. A token is roughly three-quarters of a word, but the real unit of value is intelligence delivered per dollar spent. Input tokens (your prompt and context) and output tokens (the model's response) are priced differently, with outputs commanding a premium because they require more heavy lifting.
This usage-based approach aligns incentives beautifully in theory: the more value a model creates, the more it gets used and the more it earns. In practice, it introduces volatility, optimization games, and a much tighter coupling between technical efficiency and financial outcomes.
Watching the Major Players Price IntelligenceOpenAI set the early standard with its tiered GPT family—budget-friendly models for everyday work sitting alongside premium reasoning engines. Their consumer subscriptions (Plus, Pro) often deliver outsized token value compared to raw API rates, functioning as both user acquisition and usage laboratories. Anthropic has leaned into a premium positioning, charging more for what many see as stronger reasoning and coding performance, backed by large context windows. Google’s Gemini strategy frequently looks like the value player, especially with lighter Flash variants and competitive long-context pricing. xAI has positioned Grok models as efficient contenders, emphasizing capability at more accessible rates.
Then there are the open-source and smaller providers Llama derivatives, Mistral, DeepSeek, and others driving effective costs dramatically lower for those willing to manage their own infrastructure. The market is rapidly segmenting: frontier performance for the highest-value tasks, efficient mid-tier models for volume, and open weights for customization and cost control.
Smart companies are layering on optimizations prompt caching, batching, speculative decoding, quantization that quietly determine who wins on unit economics. The public pricing you see is only part of the story; the real game is what happens after discounts, commitments, and inference tricks.
How This Flips the Investing Playbook
Traditional SaaS investing rewarded land-and-expand via sticky workflows and predictable MRR. Token economics rewards something closer to a utility-plus-platform hybrid. Usage can explode as models improve or as agents start running multi-step loops—creating 10x or 100x spend from a single customer. That’s the upside. The flip side is “success risk”: your best customers can also become your most expensive to serve if you haven’t optimized inference.Investors should retrain their eyes. Instead of obsessing solely over ARR growth and gross margin percentage, study:
Token velocity and average consumption per user or per workflow
Effective price per token after all optimizations
Inference efficiency (tokens per GPU-hour or per dollar of compute)
Caching hit rates and context utilization
The spread between flagship and efficient models in the portfolio
The companies that thrive will treat declining prices as inevitable and compete on intelligence-per-dollar while relentlessly driving down their own costs. Vertical integration (custom silicon, data center control, proprietary inference stacks) becomes a meaningful moat here. Pure model licensing without efficiency advantages risks becoming commoditized fast.
Efficiency Winners Are Emerging
The most compelling setups I’m seeing combine a few traits:Strong price-performance ratios that encourage adoption rather than sticker shock
Rapid iteration toward cheaper, specialized models instead of forcing every task onto the most expensive frontier version
Hybrid revenue (consumer subscriptions for stable base and brand, API for upside, enterprise commitments for predictability)
Deep focus on inference optimization—mixture-of-experts architectures, distillation, caching layers—that keep margins healthy even as volumes scale
Less efficient signals include over-reliance on raw frontier pricing without good alternatives, opaque unit economics, or context windows marketed as features without corresponding usage guardrails. High context sounds impressive until customers get token shock and start tightening prompts or switching providers.
What Forward-Looking Investors Should Study NowWatch how management teams talk about their cost curves. The best ones obsess over tokens-per-dollar on both sides of the transaction. Look for product roadmaps that naturally drive usage (better agents, memory layers, tool use) while offering clear on-ramps for cost-conscious customers. Diversification across consumer, developer, and enterprise segments provides a more resilient base than pure API plays.The token economy is still young. We’re moving from “how many seats did we sell?” to “how much intelligence did we deliver, and at what efficiency?” This is closer to cloud infrastructure economics than traditional software, but with software-like margins once scale and optimization kick in.For investors, this demands a more technical lens. Model architecture choices, inference strategy, and customer usage patterns are becoming as important as go-to-market motion. The winners won’t just build smarter models—they’ll build smarter economies around them. Those who master the token flywheel, turning capability improvements directly into scalable, high-margin usage, are the ones that will define the next decade of AI returns. The market is teaching us this lesson in real time. The question is whether we’re observant enough to listen.
Comments