Mira Murati isn't interested in building another digital god. While OpenAI chases the ghost of AGI, her new venture, Thinking Machines Lab, just inked a deal with Nvidia that shifts the entire focus of the industry. We aren't talking about a casual hardware purchase. It's a multi-year, multi-billion-dollar commitment to deploy one gigawatt of power via Nvidia’s upcoming Vera Rubin architecture.
If you're wondering what a gigawatt looks like in the silicon world, it’s enough to juice roughly 750,000 homes. In dollars? Experts at Reuters and the Financial Times estimate the price tag north of $50 billion. That’s a staggering amount of capital for a startup that’s barely a year old. But Jensen Huang isn't just selling chips here. He’s buying into a vision where AI is something you shape, not something that dictates to you.
The end of one size fits all AI
Most people think the finish line for AI is a single, massive model that knows everything. Murati disagrees. Her strategy with Thinking Machines is built on the idea that the "frontier" isn't about raw size anymore. It’s about adaptability.
Think about it. Why would a specialized medical research lab use the same generic model as a teenager writing a TikTok script? They shouldn't. Last October, Thinking Machines dropped its first product, Tinker. It’s an API that automates the fine-tuning of open-weights models like Llama 3.2 and Qwen. It basically takes the "black magic" out of customizing AI.
This Nvidia deal provides the massive infrastructure needed to scale that customization. Murati’s team—which includes OpenAI heavyweights like John Schulman—wants to bridge the gap between "smart but generic" and "expert and specific." They’re building systems that work alongside humans, augmenting expertise rather than trying to replace it.
Why Jensen Huang is doubling down
Nvidia's participation goes beyond being a vendor. They’ve made a significant, undisclosed investment in Thinking Machines, following their lead in the startup’s $2 billion seed round. Why? Because the Vera Rubin chips, set to deploy in early 2027, need a use case that justifies their existence.
Vera Rubin is the successor to the Blackwell line. It’s designed for massive scale. By securing Thinking Machines as a primary partner, Nvidia ensures that the next generation of "customizable AI" runs exclusively on their stack.
It’s a smart move for both sides. Murati gets the "firepower" to compete with Google and Anthropic. Jensen gets a flagship partner who proves his most expensive chips are necessary for the next leap in productivity. Honestly, it’s a match made in silicon heaven.
Drama in the lab
It hasn't been all smooth sailing. Every high-stakes startup has its cracks. In January, Murati fired co-founder Barret Zoph over a reported relationship with an employee. In a twist that surprised nobody, Zoph and several other staffers immediately headed back to OpenAI.
Then you have Andrew Tulloch, another co-founder, who bailed in October to return to Meta. People love to speculate about "founder flight," but the reality is simpler. Building at this speed is brutal. It burns people out. Despite the leadership churn, the $12 billion valuation holds firm because the market believes in the tech, not just the names on the door.
What this means for your business
If you’ve been waiting for AI to become "useful" for your specific industry, this is the signal. Thinking Machines isn't building a chatbot. They’re building a factory for specialized intelligence.
The deal with Nvidia focuses on "training and serving systems" that allow enterprises and research institutions to build their own frontier models. This is the move away from the centralized power of the "Big Three" (OpenAI, Google, Anthropic). It’s about giving you the keys to the car.
- Stop waiting for a perfect model. Start looking at how you can fine-tune existing open-source models for your specific data.
- Watch the Vera Rubin launch. If your company relies on high-end compute, the shift to 1GW-scale facilities will change your cloud costs and capabilities by 2027.
- Audit your data now. Customizable AI is only as good as the data you feed it. If your internal documentation is a mess, no amount of Nvidia chips will fix it.
Don't get distracted by the billionaire drama or the flashy chip names. The real story is the decentralization of power. AI is becoming a tool you own and shape, not just a service you rent.
