Mira Murati left OpenAI in late 2024, just as the company was approaching its IPO peak. As CTO, she departed with a group of core engineers.
Many expected her to chase another GPT-style language model. Instead, she built Tinker.
What Thinking Machines Lab actually does
In short: Tinker is an API that lets developers apply "capability patches" to any open-source large language model.
Want Llama 4 to be better at writing Python? Feed Tinker a batch of labeled data, and it uses reinforcement learning to fine-tune the model into a "Python expert" version. No need to maintain your own GPU cluster or build a training pipeline — just API calls.
The product concept resembles OpenAI's fine-tuning API, but the underlying approach is different — Tinker's core is reinforcement learning (RL)-based fine-tuning, not standard supervised fine-tuning. RL fine-tuning is harder to control than SFT, but it delivers significantly better results on certain tasks. DeepSeek's reasoning improvements came from this same path.
Thinking Machines' chief scientist is John Schulman — one of the authors of the PPO algorithm and a principal architect of OpenAI's early RL training framework. In short, they brought in the person who knows RL best.
The other two: CTO Soumith Chintala, creator of PyTorch; and CEO Mira Murati herself. Together, they form a rare combination: model framework + RL theory + product intuition.
The Google deal
On April 22, at Google Cloud Next in Las Vegas, Thinking Machines Lab and Google Cloud announced a new contract.
Scale: billions of dollars (TechCrunch exclusively reported the deal, with a source describing it as "single-digit billions"; the exact amount is undisclosed).
Contract details:
- Hardware: Nvidia GB300 NVL72 systems — large liquid-cooled cabinets with 72 GPUs each, delivering 2x faster training and inference than the previous generation
- Networking: Google's custom Jupiter network architecture, designed for GPU-to-GPU weight transfer
- Software services: Kubernetes Engine, Spanner database, Cloud Storage (with Anywhere Cache)
- Virtual machines: A4X Max instances, each with four Nvidia Blackwell Ultra GPUs, 72 CPU cores, and 37 TB of memory
Why does RL fine-tuning need this level of compute? Founding researcher Myle Ott explained the core bottleneck:
"During each training iteration, weights must be synchronized back and forth across thousands of GPUs thousands of times. If interconnect bandwidth isn't sufficient, training efficiency collapses."
This isn't ordinary inference deployment — it's compute-intensive training. Myle Ott's assessment:
"Google Cloud let us get up and running at record speed, with reliability that fully meets our requirements."
Background details
Thinking Machines was founded about 14 months ago, in early 2025.
Funding: In July 2025, it closed a $2 billion seed round at a $12 billion valuation. A seed round of $2 billion is extremely rare across the entire tech industry, reflecting investor confidence that goes beyond conventional judgment.
Team: 130+ people. But it faces a challenge — Meta has already poached 7 founding members from Thinking Machines, including the lead engineer on Tinker. This scale of talent loss is significant in the AI industry.
The Nvidia deal: In March, Thinking Machines reached another agreement with Nvidia — purchasing over 1 gigawatt-equivalent of Vera Rubin GPU systems, with Nvidia also making a direct investment. On compute, they are placing bets on both Google and Nvidia simultaneously.
The contract is non-exclusive: They can continue using other cloud providers. But securing early access to Google's GB300 batch shows Google is actively competing for this lab.
Incidentally, at this Google Cloud Next, Google also announced new contracts with Anthropic and Meta — locking in several top labs at once, a clear move by Google Cloud to capture the AI computing market.
Why this path is interesting
The AI industry now has two diverging tracks:
One is continuing to stack general capabilities — training larger models, chasing big benchmark scores.
The other is vertical fine-tuning — taking existing open-source models, strengthening them for specific tasks, and delivering them to enterprises. Thinking Machines is pursuing the latter.
The business logic: Most enterprises don't actually need GPT-5. They need "a model that can accurately process our contract documents" or "an AI assistant that passes our internal code review standards." Customization is more compelling than generalization, especially in industries with high data security requirements.
If Tinker can make RL fine-tuning work smoothly, it could become a true infrastructure layer in the enterprise AI stack — not flashy, but hard to bypass.
A 14-month-old company with a multibillion-dollar Google compute contract, and a chief scientist who is a central figure in RL. What they release next is worth watching closely.
Sources: Exclusive: Google deepens Thinking Machines Lab ties with new multibillion-dollar deal (TechCrunch); Google Cloud inks AI infrastructure deal with Thinking Machines (SiliconAngle); CocoLoop; Thinking Machines Signs Multi-Billion Google GB300 Deal (Implicator.ai)