Mira Murati's Thinking Machines Lab Gets Multibillion Google Cloud Deal

Mira Murati's artificial intelligence startup Thinking Machines Lab has secured a multibillion-dollar infrastructure deal with Google Cloud, announced April 22, 2026. The agreement — described by sources as "single-digit billions" — gives Thinking Machines Lab access to compute systems powered by Nvidia's latest GB300 chips. The deal is non-exclusive, meaning the startup can also use other cloud providers, but the scale of the commitment to Google Cloud infrastructure signals that Thinking Machines Lab is building foundational models, not just fine-tuning existing ones.

Murati was OpenAI's Chief Technology Officer for six years, the architect of GPT-4's deployment, and the person who temporarily replaced Sam Altman as OpenAI CEO during the November 2023 board crisis. She left OpenAI in September 2024 and founded Thinking Machines Lab in February 2025. The company raised a $2 billion seed round in October 2025 at a $12 billion valuation — one of the largest seed rounds in history — and launched its first product, an AI assistant called Tinker, the same month.

Why This Deal Is Larger Than It Looks

Single-digit billions is a wide range — it could be $2 billion or $8 billion. At the lower end, it represents a substantial three-year infrastructure commitment for a startup at Thinking Machines Lab's stage. At the upper end, it is comparable in scale to the Google Cloud commitments made by Anthropic ($2 billion initial, scaled to over $4 billion subsequently) and approaches the scale of Meta's recent Google Cloud GPU booking.

What the deal actually buys: access to Google Cloud's AI Hypercomputer clusters running Nvidia GB300 chips. GB300 is Nvidia's current-generation Blackwell-architecture accelerator — the successor to H100/H200. Access to GB300 clusters through Google Cloud means Thinking Machines Lab can run training workloads at a scale that requires the Anthropic-tier GPU cluster access that Google typically reserves for its largest AI partners.

The non-exclusivity clause is commercially significant. Thinking Machines Lab retains the ability to use AWS, Azure, or self-managed infrastructure for specific workloads. This gives it negotiating leverage with Google Cloud (and with Nvidia directly) and keeps its options open if a competitor offers better terms for a specific workload type.

Who Mira Murati Is Building Against

Thinking Machines Lab enters a market where the frontier model positions are already staked:

OpenAI: GPT-4o and upcoming GPT-5 (SPUD) remain the most widely deployed frontier models. Murati was the engineering architect behind GPT-4 — she understands that codebase better than anyone who stayed at OpenAI.

Anthropic: Claude Sonnet 4.6 and Claude Opus 4.7 occupy the technical AI assistant market for developers and enterprises. Anthropic has $7+ billion in Google Cloud commitments and 1+ million TPU chips from Google.

xAI: Elon Musk's Grok 3 and Grok 3.5 have outperformed GPT-4o on several benchmarks and have a captive distribution channel through 600 million X users.

Google DeepMind: Gemini 3 Ultra at 2 million context window and Gemini 4 previewed at Google I/O 2026 represent Google's internal frontier. Thinking Machines Lab is simultaneously a Google Cloud customer and a potential competitor to Google's own models.

What Murati is building toward: Tinker, the first product, is a general-purpose AI assistant — not vertically specialised. The Google Cloud deal at multibillion scale suggests the next phase involves training a foundational model, not just serving a product built on existing models. If Thinking Machines Lab releases a frontier model in 2026 or early 2027, it enters the top-tier competitive set directly.

The Talent Signal

Thinking Machines Lab has recruited from the same talent pool that built the frontier models at OpenAI, Google, and DeepMind. When Murati left OpenAI, several of her closest collaborators followed. The company's engineering team has direct experience with:

• GPT-4 training and RLHF methodology
• Large-scale distributed training systems at 10,000+ GPU scale
• Safety evaluation frameworks for frontier models
• Production inference serving at millions of users

This is not a team learning how to train frontier models — it is a team that trained them before. The GB300 access from Google Cloud gives them the compute to match that experience.

What Developers Should Watch

A new frontier model is likely in 2026-2027: Multibillion-dollar GPU commitments at a startup are not for inference serving — they are for training. When Thinking Machines Lab releases a foundational model, it will enter benchmark comparisons immediately. Developers choosing a primary AI API should watch for this entry.

Google Cloud deepens its AI partner moat: Google now has Anthropic, Meta, OpenAI, and Thinking Machines Lab as major TPU/GPU customers. Every frontier lab training on Google Cloud is both a customer and a validation signal. The more frontier labs depend on Google Cloud for training, the stronger Google's position in the inevitable future where it sells inference serving to the same labs' customers.

The post-OpenAI talent diaspora matters: The wave of senior AI researchers who left OpenAI, DeepMind, and Anthropic in 2024-2025 is now building funded, compute-backed alternatives. Thinking Machines Lab, Reflection AI, SSI, and others represent genuine frontier competition within 18-24 months.

Key Takeaways

• Thinking Machines Lab secures single-digit billion Google Cloud deal April 22, 2026: access to Nvidia GB300 clusters; non-exclusive; signals foundational model training, not just inference serving for Tinker product
• Mira Murati credentials: OpenAI CTO for 6 years, GPT-4 architect, acting CEO during November 2023 board crisis; founded TML February 2025, $2B seed at $12B valuation in October 2025
• Competitive positioning: enters market against OpenAI, Anthropic, xAI, and Google DeepMind; has direct GPT-4 training expertise; Google Cloud deal comparable in scale to Anthropic's initial Google commitment
• Google Cloud strategic pattern: now has Anthropic, Meta, OpenAI, Thinking Machines Lab as GPU/TPU customers; builds structural dependency among frontier labs even as it competes with them via Gemini
• Watch for: foundational model launch from TML in 2026-2027; GB300 training at multibillion-dollar scale implies frontier model training, not product-layer fine-tuning

For the Google Cloud AI infrastructure context, read Google TPU 8t and 8i at Cloud Next 2026: The Inference War Starts Now. For the broader AI competitive landscape, read Claude Opus 4.7 Developer Backlash: What Went Wrong and What to Do. For the AI funding context, read Cursor Raises $2B at $2.5B Valuation: AI Coding Tool Funding Explained.