I’ve been in the tech game for a while now. I’ve seen bubbles burst and empires rise from the ashes. I’ve had two successful exits, RemoteTeam and MovieLaLa, and I’ve written checks for over 200 startups, including some of the biggest names in AI like Anthropic, OpenAI, Scale AI, and Hugging Face. But nothing has been as wild as the current AI boom. And at the heart of it all is a concept that most people, even those in the industry, don’t fully grasp: token economics.
I’m not talking about crypto tokens. I’m talking about the tokens that power Large Language Models (LLMs). These are the basic units of text that models process. And how you manage them, how you price them, and how you optimize them can literally make or break your AI product. I went down the rabbit hole, analyzing over 100 different LLMs, and what I found was shocking. Here are the 11 most critical lessons I learned about pricing, efficiency, and the future of tokenization.
1. Your Tokenizer is a Silent Killer
Most founders I talk to don't even know what tokenizer their model uses. They just pick a model off the shelf and start building. That's a huge mistake. The tokenizer is the first thing that touches your data. It breaks your text down into tokens, and different tokenizers do it differently. A bad tokenizer can inflate your token count by 20-30%, which means you're paying 20-30% more for every single API call. It's like a silent tax on your entire business. We once had a startup we invested in that was burning through their cash way too fast. We couldn't figure out why. It turned out their tokenizer was terrible. We switched them to a more efficient one and their costs dropped by a third overnight. That’s how much of a difference it can make.
2. The Great Multilingual Rip-Off
If you're building a multilingual product, you need to be extra careful. Many of the most popular models are trained primarily on English. When you feed them other languages, the token count explodes. I've seen cases where Spanish text takes up twice as many tokens as the English equivalent. And for languages like Chinese or Japanese, it can be even worse. This isn't just a technical issue; it's a business issue. If you're charging your users based on usage, you're effectively penalizing them for using their native language. It's a terrible user experience, and it's a great way to lose customers to a competitor who has a better handle on their token economics.
3. The Hidden Cost of Context
Everyone is excited about the massive context windows of the new models. 1 million tokens! 10 million tokens! It sounds amazing. But what they don't tell you is that the cost of that context isn't linear. The more context you use, the more you pay per token. And it's not just the cost. The more context you use, the slower the model gets. I've seen models that are lightning fast with a small amount of context grind to a halt when you feed them a massive document. You need to be smart about how you use context. Don't just stuff everything in there. Be selective. Be strategic. Your wallet and your users will thank you.
4. The Illusion of
Choice
Every week, it seems like there are a dozen new LLMs hitting the market. It’s overwhelming. And the truth is, most of them are just slight variations of each other. They use the same architecture, the same training data, and the same tokenizers. The only difference is the name and the marketing hype. I call this the illusion of choice. You think you have all these options, but you’re really just choosing between different flavors of the same ice cream. Don’t get distracted by the noise. Find a model that works for you and stick with it. The real innovation isn’t happening at the model level; it’s happening at the application level.
5. The Streaming Trap
Streaming tokens is all the rage right now. It creates a better user experience, right? The user sees the output as it’s being generated, instead of waiting for the whole thing to finish. But here’s the catch: streaming can be more expensive. When you stream, you’re making a lot of small API calls instead of one big one. And each of those calls has a certain amount of overhead. It may not seem like much, but it adds up. I’ve seen startups that have doubled their API costs just by implementing streaming. So before you jump on the streaming bandwagon, do the math. Make sure the user experience benefits are worth the extra cost.
6. The Quantization Gold Rush
Quantization is the process of reducing the precision of the model’s weights. It’s a great way to make your model smaller and faster. But it’s not a free lunch. When you quantize a model, you lose a certain amount of accuracy. For some applications, that’s a perfectly acceptable tradeoff. But for others, it’s a dealbreaker. I’ve seen companies that have quantized their models so much that they’ve become completely useless. They’re fast, but they’re also dumb. The key is to find the right balance. Don’t just quantize for the sake of quantization. Do it because it makes sense for your specific use case.
7. The Fine-Tuning Fallacy
Fine-tuning is another one of those things that everyone thinks they need to do. They have this idea that they can take a base model and fine-tune it on their own data to create a super-powered, domain-specific model. And sometimes, that works. But more often than not, it’s a waste of time and money. Fine-tuning is hard. It’s expensive. And it’s very easy to mess up. I’ve seen more failed fine-tuning projects than I can count. Before you go down the fine-tuning rabbit hole, ask yourself if you really need it. Can you get the same results with prompt engineering? Can you use a retrieval-augmented generation (RAG) approach instead? Fine-tuning should be a last resort, not a first step.
8. The MoE Magic
Mixture of Experts (MoE) is one of the most exciting developments in the LLM space right now. The basic idea is that instead of having one massive model, you have a bunch of smaller, specialized models (the “experts”). And you have a routing network that decides which expert to use for each token. It’s a brilliant idea, and it’s incredibly efficient. MoE models can be much larger than traditional models, but they’re also much cheaper to run. I’m a huge believer in the MoE approach. I think it’s the future of LLMs. And I’m putting my money where my mouth is. I’ve invested in several startups that are building MoE models, and I’m incredibly bullish on their prospects.
9. The Multimodal Revolution
For the past few years, LLMs have been all about text. But that’s starting to change. The new generation of models are multimodal. They can understand and generate not just text, but also images, audio, and even video. This is a huge deal. And it’s going to have a massive impact on token economics. How do you tokenize an image? How do you price a video? These are the questions that we’re all grappling with right now. There are no easy answers. But one thing is for sure: the multimodal revolution is coming, and it’s going to change everything.
10. The Open Source vs. Closed Source Debate
This is one of the most heated debates in the AI community right now. On one side, you have the open source purists who believe that all models should be free and open for everyone to use. On the other side, you have the closed source advocates who argue that you need to keep your models proprietary in order to have a competitive advantage. I’m a pragmatist. I see the value in both approaches. Open source is great for innovation and experimentation. But closed source is where you’re going to find the most powerful and capable models. The key is to use the right tool for the job. Don’t be a zealot. Be a strategist.
11. The Future is Hybrid
So what’s the winning strategy? It’s not about choosing one model over another. It’s not about going all-in on open source or closed source. The winning strategy is a hybrid approach. It’s about using a mix of different models, each one optimized for a specific task. It’s about being smart about your token economics. It’s about being strategic about your use of context. It’s about building a system that is flexible, efficient, and scalable. That’s how you win in the age of AI. And that’s what I’m helping my portfolio companies do every single day.
This is not just about saving a few bucks on your API bill. This is about building a sustainable business. The companies that master token economics will be the ones that survive and thrive in the long run. The ones that don't will be left behind. It's as simple as that. So, take a hard look at your token strategy. Are you being as efficient as you can be? Are you leaving money on the table? The answers to these questions could determine the future of your company.
Frequently Asked Questions
How were these items selected?
Each item on this list comes from direct experience, either from building my own companies or from patterns I've observed across the 200+ startups I've invested in. I prioritize practical, actionable items over theoretical concepts.
Are these recommendations still relevant in 2026?
Absolutely. While specific tools and tactics change, the underlying principles remain consistent. I update my thinking regularly based on what I'm seeing in the market and across my portfolio companies.
Can I implement all of these at once?
I'd strongly recommend against it. Pick the 2-3 items that resonate most with your current situation and focus there. Trying to do everything simultaneously is a recipe for doing nothing well.
Which item on this list has the highest impact?
It depends on your stage and context, but in my experience, the items near the top of the list tend to have the broadest applicability. That said, sometimes the less obvious items create the biggest breakthroughs for specific situations.