Thinking Machines Lab did raise $2 billion—but the financing closed in July 2025, not recently. Early coverage described roughly a $10 billion pre-money valuation; later reporting put the completed round at about $12 billion post-money. At the time, Mira Murati’s company had no public commercial product. By August 2026, it had released the open-weight Inkling and Inkling-Small models, operates the Tinker customization platform, and has announced a future one-gigawatt NVIDIA infrastructure deployment.
What happened in the $2 billion financing?
Thinking Machines Lab, founded and led by former OpenAI chief technology officer Mira Murati, publicly unveiled itself on February 18, 2025. It closed a $2 billion seed round in July 2025, led by Andreessen Horowitz. Reported participants included Accel, NVIDIA, AMD, Cisco, Jane Street and other investors. TechCrunch initially reported a valuation of approximately $10 billion before the new money was added. After the financing closed, TechCrunch reported an approximately $12 billion post-money valuation.
| Figure | What it means | Source and qualification |
|---|---|---|
| $2 billion | Seed financing closed in July 2025 | Reported by TechCrunch and other outlets |
| About $10 billion | Pre-money valuation | Early financing reports |
| About $12 billion | Post-money valuation after the round | Later reporting on the completed financing |
Those numbers are not inherently contradictory: a $10 billion pre-money valuation plus $2 billion of new capital produces approximately $12 billion post-money. WIRED characterized the financing as among Silicon Valley’s largest seed rounds, a description that depends on the comparison set rather than an official universal ranking. WIRED’s report also detailed the investor group and the concentration of high-profile AI talent.
The important context is that the company was pre-product when investors committed the money. The round demonstrated confidence in Murati, her recruiting ability and the strategic value of frontier-AI capacity; it did not demonstrate revenue, profitability or product-market fit.
Recommended Free Tools
#1 Best Overall
Who is Mira Murati?
- 2018: Murati joined OpenAI.
- 2022: She became OpenAI’s chief technology officer as ChatGPT emerged and the company expanded products such as DALL-E and Codex.
- September 2024: She left OpenAI.
- February 2025: She publicly launched Thinking Machines Lab as CEO and cofounder.
Murati held senior leadership responsibility and was involved in major OpenAI product and research efforts; it is more accurate to say she led or oversaw that work than to call her the sole creator of ChatGPT or those systems. TechCrunch’s launch profile describes her background and the initial team.
Who joined the startup?
The founding group included former OpenAI researchers and engineers such as chief scientist John Schulman, CTO Barret Zoph and chief architect Andrew Tulloch. The broader team drew from OpenAI, Meta AI, Mistral AI and other laboratories. Team pages and titles can change, so a launch roster should not be read as a guarantee that every listed person remains at the company.
That distinction became significant in January 2026, when reporting said cofounders Barret Zoph and Luke Metz were leaving to return to OpenAI. TechCrunch reported the departures. They show the recruiting and retention risk in a company whose early valuation was closely associated with a small group of elite researchers, but they do not by themselves establish a broader staffing or viability problem.
What was Thinking Machines trying to build?
At launch, the company described an AI research and product organization focused on systems that work across modalities, collaborate naturally with people, can be customized to users and organizations, and are more understandable than opaque, one-size-fits-all models. It also emphasized open science and practical applications.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Investor-facing reporting described a possible enterprise strategy in which models could be adapted around a customer’s business metrics or key performance indicators. That was a reported direction, not a confirmed product specification at the time. Axios and The Information covered that early thesis.
What has the company released since the funding?
Tinker: a managed model-customization API
Tinker is Thinking Machines’ developer-facing training API. Instead of operating an entire GPU cluster, a team can use managed infrastructure while controlling low-level operations such as forward_backward, optim_step, sample and save_state. The service is intended for fine-tuning and post-training workflows and provides SDKs, documentation, a cookbook and API-compatible interfaces.
The product was introduced on October 1, 2025. Its product page and documentation describe supported workflows and models. Pricing is usage-based; the pricing documentation listed checkpoint storage at $0.10 per GB-month when viewed in August 2026. Inkling configurations were shown with a limited-time 50% discount, so prices and availability should be checked before budgeting. Tinker is a poor fit for buyers requiring fixed-cost infrastructure, strict on-premises operation, direct GPU control or production-grade serverless inference beyond the service’s stated beta limitations.
Inkling: a large open-weight model
Thinking Machines released Inkling on July 15, 2026 with full weights under an Apache 2.0 license according to its model card. It accepts text, image and audio inputs and produces text outputs. The company presents it primarily as a customizable base model rather than claiming it is the strongest system on every benchmark.
Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchPC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Rank #3
- 975 billion total parameters, with 41 billion active parameters.
- Mixture-of-experts architecture.
- Context window of up to 1 million tokens, although a particular provider may expose a lower limit.
- Training on 45 trillion tokens across text, images, audio and video, according to the company.
- Weights distributed through Hugging Face and fine-tuning available through Tinker.
- Third-party deployment and inference support announced through providers and tools including Together AI, Fireworks, Modal, Databricks, Baseten, Hugging Face, SGLang, vLLM, TokenSpeed, Unsloth and llama.cpp.
“Open-weight” does not automatically mean every component of a project is open in the practical or legal sense. Users still need to review the Apache 2.0 terms, acceptable-use policy, model limitations, hardware requirements and each provider’s conditions.
Inkling-Small: a lower-compute option
Released July 30, 2026, Inkling-Small has 276 billion total parameters and 12 billion active parameters. It supports native reasoning over audio and images and offers up to a 1-million-token context window. The company positions it as requiring less compute and delivering lower latency than full Inkling. It is available through Tinker subject to the current rollout and pricing terms. The launch announcement provides the stated specifications.
How much hardware does Inkling require?
For teams considering self-hosting, the model card says the full BF16 Inkling checkpoint requires at least approximately 2 TB of aggregate VRAM. Its NVFP4 version requires at least approximately 600 GB. These are minimum aggregate-memory figures, not a promise of a particular throughput, latency or production configuration. Hosted providers can shift the hardware burden to their infrastructure but add service pricing, capacity and terms that vary by provider.
What does the NVIDIA partnership mean?
On March 10, 2026, Thinking Machines announced a multiyear partnership with NVIDIA. The announcement calls for deployment of at least one gigawatt of next-generation NVIDIA Vera Rubin systems, targeted to begin in early 2027, plus joint work on training and serving systems optimized for NVIDIA architectures and an undisclosed NVIDIA investment.
This is a future capacity target, not evidence that a one-gigawatt Vera Rubin supercomputer is already installed. The agreement suggests the company is pursuing frontier-scale model training and serving rather than only a lightweight application layer. It also creates exposure to capital spending, electricity and data-center requirements, hardware supply, operational complexity and concentration on one strategic supplier. Thinking Machines’ announcement sets out the timing and scope.
Why would investors fund a pre-product company?
- Founder credibility: Murati’s senior role at OpenAI gave investors a track record in frontier product development.
- Talent concentration: Recruiting recognized researchers can be strategically valuable when experienced model-building teams are scarce.
- Compute barriers: A $2 billion starting bankroll can secure chips, data, infrastructure and years of research time.
- Customization thesis: The company differentiated itself from the idea that one general model should serve every user identically.
- Strategic optionality: Models, hosted APIs, enterprise customization, research licensing and infrastructure partnerships all remain possible paths.
These factors explain the investment thesis implied by the plans and investor behavior. They are not proof that the eventual business will succeed.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What business model is emerging?
The visible strategy combines open-weight model distribution with paid customization and infrastructure services. Developers can download weights or use third-party inference, while organizations with proprietary data can pay for managed fine-tuning and post-training through Tinker. Enterprise integration, usage-based APIs and deployment partnerships could provide additional revenue paths.
That approach helps address a central open-model trade-off: releasing weights can accelerate adoption and ecosystem development, but it can make direct model licensing harder. Tinker may be the bridge between freely available model assets and recurring service revenue. There is no public basis to state current customer counts, revenue, profitability, enterprise contract sizes or how much of the original $2 billion remains.
Best Value
What remains unproven?
- How many paying customers Tinker has and whether usage is recurring.
- The economics of training and serving Inkling at frontier scale.
- Whether open weights plus hosted customization can produce venture-scale returns.
- How much of the announced NVIDIA capacity is committed capital versus a future capacity arrangement.
- Whether the company can retain enough research talent after cofounder departures.
- Whether technical performance, safety, reliability and distribution will justify the private valuation.
What the $2 billion headline does—and does not—tell you
For investors, the financing is a private-market valuation event, not a liquid public-market value. For general readers, “raised $2 billion” means capital committed by investors, not revenue, cash profit or guaranteed product success. For enterprise buyers, using Inkling through Tinker is different from self-hosting the weights: cost, governance, latency, context limits, hardware and provider terms can differ substantially.
The most accurate current description is therefore not simply “an OpenAI rival that raised $2 billion.” Thinking Machines is a Murati-led AI research and product company that raised extraordinary seed capital before launching publicly, then moved into open-weight models, managed customization and planned frontier infrastructure.
Bottom line
The July 2025 round was a remarkable vote of confidence in Murati, her team and the scarcity of frontier-AI talent. The meaningful test is now execution: adoption of Tinker, real-world customization results, sustainable training and inference economics, and delivery of the capabilities promised at frontier scale.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
Quick wins for a faster PC:
Scan for outdated or missing drivers - takes under a minuteDriver Scan →Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →
