The elite Silicon Valley echo chamber is finally cracking. For years, the narrative was simple: whoever builds the biggest cluster of Nvidia chips wins. If you had $100 billion and a direct line to Microsoft or AWS, you were the future of artificial intelligence. If you didn't, you were just background noise.
Then came the reality check.
Mira Murati, the former chief technology officer of OpenAI, spent years at the very center of that brute-force scaling philosophy. She helped shepherd GPT-4 and the early iterations of OpenAI's reasoning models into existence. But when she left to build her new venture, Thinking Machines, she didn't just copy the OpenAI playbook. Instead, her team is quietly drawing inspiration from the open-source breakthroughs coming out of China.
It turns out that the most interesting ideas in AI right now aren't coming from multi-billion-dollar compute clusters in Iowa. They're coming from labs in Hangzhou and Beijing that figured out how to do more with less.
What Silicon Valley is learning from Beijing
To understand why Thinking Machines is taking this route, you have to look at how the global AI race shifted.
For a long time, US labs assumed that Chinese companies were perpetually two years behind because of export controls and chip sanctions. But scarcity breeds creativity. Blocked from buying the latest US chips in unlimited quantities, Chinese researchers focused intensely on algorithmic efficiency.
They didn't just try to build bigger models. They figured out how to make existing architectures run circles around larger competitors.
When labs like DeepSeek released their reasoning models, it sent shockwaves through the industry. They proved that you could match the reasoning capabilities of closed, expensive US models like OpenAI's o1 for a tiny fraction of the training cost. They did this by rethinking reinforcement learning.
Instead of relying on massive human feedback datasets, which are expensive and slow to build, they let the models teach themselves through structured incentive systems.
Murati's startup is paying close attention. Building a startup from scratch in 2026 means you cannot compete on raw GPU volume alone. Even with massive venture backing, a new startup cannot easily build a computing infrastructure that rivals Microsoft or Google overnight. You have to be smarter. You have to adopt the efficiency methods that Chinese developers pioneered.
Inside the Thinking Machines strategy
So, what does this actually look like in practice?
Thinking Machines is focusing heavily on reasoning-focused AI. These are systems that don't just spit out the next most likely word. They stop, think, generate internal monologues, correct their own mistakes, and try different paths before giving you an answer.
But instead of training these models from absolute scratch on trillions of tokens—an exercise that costs hundreds of millions of dollars—the team is taking a page out of the Chinese open-source playbook.
Smart initialization over brute force
Instead of wasting compute on teaching a model basic grammar and facts, smart startups are starting with incredibly strong open-source base models. Companies like Alibaba with their Qwen series have provided some of the best foundational bases in the world. By taking a highly capable base model and focusing all your resources on the post-training phase, you save massive amounts of capital.
Advanced reinforcement learning
The real magic happens during post-training. Chinese researchers popularized techniques like Group Relative Policy Optimization (GRPO). Traditional reinforcement learning requires a separate "critic" model to grade the main model's outputs. This process eats up valuable GPU memory.
GRPO gets around this by comparing a group of outputs against each other and calculating scores based on the average. It reduces the memory footprint significantly. For a lean startup like Thinking Machines, using these math shortcuts is the only way to build frontier-class models without a tech giant's balance sheet.
Distillation and synthetic data
Another major strategy is using larger models to train smaller, more efficient ones. By using frontier models to generate high-quality reasoning paths, you can train a smaller model to mimic that high-level thinking. It is a highly structured, deliberate training style that prioritizes quality over raw data quantity.
The end of the compute monopoly
For the last three years, the venture capital pitch was boring. "We have the best team, give us $5 billion for chips."
That era is over.
The strategy behind Thinking Machines proves that the industry is entering a new phase where architectural cleverness matters more than raw scale. If you can achieve 99% of the performance of a giant cluster for 5% of the cost, the business math changes completely.
This shift is democratizing the tech space. It means smaller, agile teams can actually build models that challenge the monopolies. They don't need to build their own power plants or sign exclusive deals with cloud providers. They just need to write better algorithms.
We are seeing a massive talent migration because of this. Top researchers want to work on elegant software solutions, not just act as system administrators for massive data centers. They want to work on reinforcement learning algorithms that actually teach models how to think.
How to adapt your own AI roadmap
If you are running a engineering team or building an AI-native product, you should stop waiting around for the next massive closed-source model release. The lessons that Thinking Machines is drawing from the Chinese ecosystem are directly applicable to any tech team today.
- Prioritize post-training over pre-training: Do not try to train a foundational model. Find the best open-source base that fits your hardware constraints and pour your resources into fine-tuning and reinforcement learning.
- Implement reasoning traces: If your application requires complex decision-making, design your prompts and pipelines to force the model to generate its reasoning steps before outputting the final answer. This simple change drastically improves accuracy.
- Keep your infrastructure flexible: Do not lock yourself into a single proprietary API. The efficiency gains in the open-source world are moving so fast that you need the ability to swap your underlying model overnight without rewriting your entire application stack.
The path forward isn't about throwing more money at the problem. It is about being incredibly deliberate with the compute you have. Mira Murati knows this, Chinese developers proved it, and the market is shifting to reflect it.
