TL;DR
Thinking Machines Lab, Mira Murati’s frontier-AI startup, signed a multi-billion-dollar infrastructure deal with Google Cloud. This is the first time TML has picked a hyperscaler as a partner. The agreement is worth “single-digit billions,” runs on Google’s Nvidia GB300-powered instances, and is explicitly non-exclusive. It lands five weeks after TML’s gigawatt-scale Nvidia partnership in March, and while the company is reportedly in talks for a ~$50B valuation round. The spend on Google is likely the first visible slice of where the fresh money would go.
What Got Announced Today
Rebecca Bellan at TechCrunch broke the story this morning: Thinking Machines Lab and Google Cloud have signed a multi-year agreement in the “single-digit billions” range. Google said it will provide TML with access to its Nvidia GB300-backed AI infrastructure for training, serving, and reinforcement-learning workloads.
Myle Ott, a founding researcher at TML, went on record with one sentence: “Google Cloud got us running at record speed with the reliability we demand.” That’s the closest either side came to a number. Neither Google nor TML disclosed the exact commitment, term length, or whether the spend is upfront or consumption-based.
Two details are bigger than the headline dollar figure:
- This is TML’s first cloud services provider deal. Up to now, TML had been dealing directly with Nvidia for silicon (the gigawatt partnership announced March 10) and with investors for cash. Google is the first hyperscaler to show up on the supplier list.
- It’s non-exclusive. TML can and will use other clouds in parallel. That’s a materially different deal than, say, Anthropic’s long Amazon tie-up or OpenAI’s historical Azure dependency. Google gets volume, not lock-in.
What Google Is Actually Selling Them
The GB300 bit is worth unpacking, because a lot of readers will see “Nvidia chips through Google” and assume it’s the same Vera Rubin deal TML did with Nvidia in March. That’s a different deal on different silicon.
GB300 (product name GB300 NVL72) is Blackwell Ultra, the current generation Nvidia started shipping in volume in Q1 2026. It’s a liquid-cooled rack that ties 72 Blackwell Ultra GPUs to 36 Grace CPUs, purpose-built for inference and reasoning workloads. A single NVL72 lands as about 37 TB of combined fast memory (20 TB GPU + 17 TB CPU LPDDR5X) and tens of petaflops of FP4 compute.
Vera Rubin, the 1 GW deployment TML committed to with Nvidia back in March, is the next generation. It doesn’t start landing in customer data centers until Q3 2026 at the earliest, and the gigawatt deployment with TML is phased over multiple years starting in 2027.
The two deals serve different purposes. Nvidia 1 GW is a long-dated capacity bet for the frontier models TML hasn’t shipped yet. Google Cloud on GB300 is for running Tinker and doing the reinforcement-learning experiments today, on hardware that already exists in racks and already has cooling plumbed in.
Tinker is the workload that made this urgent
Tinker, TML’s fine-tuning API, went live in October 2025. It exposes four primitives (forward_backward, optim_step, sample, save_state) and supports about 40 open-weight models across Qwen, Llama, DeepSeek, GPT-OSS, Nemotron and Moonshot Kimi, from tiny 1B base checkpoints up to Qwen’s 397B. Customers list UC Berkeley, Princeton, Stanford, and Redwood Research.
Tinker’s pricing is in tokens (prefill $0.03–$2.00 per 1M, sampling $0.09–$10, training $0.09–$12 per 1M, storage $0.10/GB-month). That’s a consumption business, which means TML needs elastic compute, the kind you rent from a cloud, not from a colo provider. And Tinker workloads lean heavily on RL, where you alternate long sampling runs with small weight updates. GB300’s rack-scale NVLink is tuned for exactly that traffic pattern.
The shape of the deal makes Google specifically the Tinker backbone. The Nvidia Vera Rubin deployment is for whatever frontier model TML is training behind the scenes.
Why Google Won This
I’ve been watching the hyperscaler AI-lab poker game for a while, and Google’s pitch on this one wasn’t a mystery. Three things were in their favor:
- GB300 at scale, now. Google already runs one of the largest fleets of Blackwell Ultra in the industry and has them available as first-party SKUs. AWS has GB300 too, but the allocation for new labs has been tight after Anthropic’s expansions. Azure has been OpenAI-dominated. If you’re a lab that needs thousands of GB300s next month, Google is the realistic door.
- Neutral ground. Google Cloud has made a point of landing AI-lab deals that cross competitive lines. In April, Anthropic expanded its compute deal with Google and Broadcom despite sharing equity with Amazon, the same lab that just passed OpenAI on ARR. For TML (friendly with OpenAI alumni but also competitive with them), that posture helps. Google doesn’t demand your alumni shut up about their old employer.
- TPU optionality. The deal announced today is GPU-heavy, but Google is the only cloud with a first-party training accelerator (TPU) that doesn’t come from Nvidia. TML has not committed to TPUs, but the door is open. Over a five-year horizon, having that option priced in is worth something.
Amazon, Microsoft and Oracle presumably pitched. None of them landed.
Paying For The Compute
Where is TML getting the money for all this? The seed was $2B at a $12B valuation in July 2025, with Nvidia, AMD, Cisco, and Jane Street on the cap table alongside Andreessen Horowitz. Bloomberg reported in March that TML is in talks for a follow-on round that would price the company at roughly $50B, though nothing has closed publicly. Meanwhile the Nvidia partnership in March carried its own “significant investment” tag. This is all happening inside a year where startups raised $300B in Q1 alone, with frontier AI labs taking most of the oxygen.
If today’s Google deal is three years at, say, $3B in committed spend, that’s a meaningful chunk of whatever fresh round eventually clears. Most of the money AI labs are raising this cycle is going to chips, power, and cloud capacity, not headcount. At somewhere in the 100–150 range of employees (public trackers disagree on the exact number), TML’s salary bill is a rounding error next to its GPU spend.
| Lab | Known cloud partner | Disclosed size | Nvidia direct deal |
|---|---|---|---|
| OpenAI | Microsoft Azure (historical), Oracle | Tens of billions | — |
| Anthropic | AWS (primary), Google Cloud | $100B+ combined | — |
| xAI | Oracle, direct colo | Undisclosed | Direct buys |
| Thinking Machines Lab | Google Cloud (announced today) | “Single-digit billions” | 1 GW Vera Rubin (March 2026) |
| Mistral | Azure, Nvidia direct | Undisclosed | Direct |
The one thing worth noting on that table: TML is the only lab for which the first cloud deal is hitting the wire over a year after founding. OpenAI, Anthropic, and xAI all started with a cloud partnership. TML started with Nvidia direct and only now walked through the cloud door. The reversed ordering points at what’s actually scarce in 2026: silicon, while cloud managed services have become commodified.
What This Means for the AI Coding Space
Tinker is adjacent to the AI-coding stack even if it doesn’t ship an IDE. If a researcher at Anthropic, Mistral, or an open-weight shop wants to fine-tune a 120B base for a coding eval, Tinker on Google’s GB300 is now one of the cheapest paths that doesn’t require you to own hardware. That competes directly with Hugging Face’s training platform, Modal, and Together AI’s fine-tune API.
Cursor and Composer-style products sit one layer up; they consume finished models. But the models Cursor integrates tomorrow are getting trained on platforms like Tinker today. When Kimi K2.5 showed up in Cursor Composer 2 last week, its path from Moonshot’s lab to a Cursor customer’s tab-completion involved exactly this kind of elastic RL compute.
The second-order effect of a Google–TML deal, then, is that the supply of well-tuned open-weight coding models gets faster and cheaper over the next year. It’s steady, unglamorous pressure on Copilot’s and Cursor’s margins.
What to Watch Next
Three specific things are worth tracking:
- Does TPU show up in a follow-up? If Google can convert TML to TPUs for training, that’s a much bigger win than selling them Nvidia capacity. TPU adoption by a frontier lab that isn’t Anthropic would be a genuine market signal.
- The Nvidia 1 GW Vera Rubin ramp. That deployment is phased from 2027; the intermediate-term compute story is hyperscaler-based. If the Nvidia ramp slips, TML’s GPU plan leans harder on Google.
- Tinker enterprise pricing. If TML wants to service enterprise fine-tuning (not just academic labs), expect dedicated-capacity tiers announced this summer. That would be the commercial reason to have locked down GB300 inventory now.
FAQ
How much is the TML–Google Cloud deal worth? “Single-digit billions,” per TechCrunch’s reporting. Neither company disclosed the exact number, the term length, or the payment structure (upfront vs. consumption-based).
Is this the same as TML’s Nvidia deal from March? No. The March announcement was a direct gigawatt-scale deployment of Nvidia Vera Rubin systems, starting in 2027 and phased over multiple years. Today’s Google deal is for GB300 (Blackwell Ultra) capacity delivered through Google Cloud, available now. Different chip generation, different procurement path.
Why didn’t TML pick AWS or Azure? TML didn’t say publicly. The most plausible reasons: Azure’s capacity is tied up with OpenAI; AWS’s GB300 allocation is heavily committed to Anthropic; Google had both the chips and the willingness to take on a lab that’s competitive with its existing customers. Google has repeatedly landed AI-lab deals that cross Amazon’s and Microsoft’s lines.
Does this affect Tinker pricing? Not directly announced. But more reliable capacity should translate to tighter SLAs and possibly volume discounts for power users. If dedicated capacity tiers appear this summer, that would be the visible signal.
What does “non-exclusive” actually mean here? TML is free to use other clouds (AWS, Azure, Oracle) in parallel, and is free to run its own colo. That’s standard for research labs and the opposite of how OpenAI’s early Azure deal was structured. Google gets volume, not lock-in.
Sources
- Exclusive: Google deepens Thinking Machines Lab ties with new multi-billion-dollar deal (TechCrunch). Original reporting on today’s announcement by Rebecca Bellan.
- Thinking Machines Lab and NVIDIA Announce Long-Term Gigawatt-Scale Strategic Partnership. TML’s own writeup of the March 2026 Nvidia deal.
- Thinking Machines Lab inks massive compute deal with Nvidia (TechCrunch). Independent coverage of the Nvidia gigawatt partnership.
- Tinker, Thinking Machines Lab. Official product page with API primitives, pricing, and supported models.
- NVIDIA GB300 NVL72 product page. Specs of the chip generation the Google deal runs on.
- Anthropic ups compute deal with Google and Broadcom (TechCrunch). Reference for Google’s track record landing AI-lab deals that cross competitive lines.
Bottom Line
The headline is the dollar figure, but the actual story is the ordering. Most AI labs pick a cloud before they pick a chip vendor. Thinking Machines Lab did it in reverse: Nvidia first, hyperscaler second. That tells you what Mira Murati thinks is scarce in 2026, and why a “single-digit billions” check to Google still isn’t quite enough to close out TML’s compute plan. The Nvidia gigawatt deal remains the load-bearing piece, with Google as the bridge until 2027.