Nvidia Powers Mira Murati’s Thinking Machines Lab with Multi‑Billion Chip Investment

Nvidia commits at least 1 gigawatt of GPUs to Thinking Machines Lab, marking a watershed AI investment

• At least one gigawatt of Nvidia chips will power Thinking Machines’ next‑generation models.
• The partnership includes joint design of training and serving systems using Nvidia’s latest architecture.
• Thinking Machines has grown from 30 to roughly 120 staff in a single year.
• Mira Murati described Nvidia’s tech as “the foundation on which the entire field is built.”

Why a gigawatt of GPUs matters for the future of AI research

NVIDIA—Nvidia’s pledge to supply a full gigawatt of cutting‑edge graphics processing units (GPUs) to Mira Murati’s Thinking Machines Lab is more than a financial infusion; it is a technical infrastructure upgrade that rivals the compute capacity of the world’s biggest cloud hyperscalers.

The deal, announced in early 2026, pairs Nvidia’s H100‑based accelerator stack with Thinking Machines’ flagship product, Tinker, a platform that helps researchers iterate on large language models while keeping humans in the loop.

Industry analysts see the partnership as a bellwether for how chipmakers will lock in emerging AI startups, ensuring a pipeline of demand for next‑generation silicon while giving startups the compute they need to compete for talent and breakthroughs.

Why Nvidia’s Gigawatt Commitment Matters for AI Compute

Scale of a gigawatt in the AI world

A gigawatt of GPU power translates to roughly 10,000 H100 GPUs, each capable of delivering 30 teraflops of FP16 performance. In practical terms, that amount of compute can train a model comparable in size to GPT‑4 in under a month, according to Nvidia’s own benchmark data (Bloomberg, 2024).

For comparison, the total GPU capacity of the public cloud’s AI‑focused offerings in 2023 was estimated at 850‑900 gigawatts, according to a joint report by the International Data Corporation (IDC). Thinking Machines, with a dedicated gigawatt, will sit in the top‑10 percentile of compute‑rich AI labs worldwide.

“A gigawatt is the kind of commitment you normally see from hyperscale data‑center operators, not a two‑year‑old startup,” noted analyst Karen Liu of Bloomberg. “It tells us Nvidia sees Thinking Machines as a strategic partner that can push the envelope on model efficiency and human‑centric AI.”

The partnership also includes co‑design of training pipelines. Nvidia’s DGX Cloud platform will be customized for Tinker’s workflow, allowing researchers to run distributed training across the full gigawatt without the typical latency penalties associated with multi‑cloud setups.

From a financial perspective, the gigawatt commitment likely represents a multi‑billion‑dollar investment when factoring in hardware, support, and software licensing. Nvidia’s FY2023 AI segment generated $10.2 billion in revenue, and the company has earmarked roughly 15 % of that for strategic partnerships, suggesting a possible $1.5‑$2 billion outlay for this deal (Bloomberg, 2024).

By locking in a fast‑growing customer early, Nvidia not only secures future sales but also gains access to novel research outcomes that can inform the next generation of its chips. The synergy between hardware and software is a hallmark of Nvidia’s AI strategy, first demonstrated with its CUDA ecosystem in the early 2000s.

As Thinking Machines scales, the gigawatt will likely be expanded, mirroring Nvidia’s historic practice of scaling commitments with partner growth. The next chapter explores how the lab’s origins set the stage for this ambitious collaboration.

The Birth of Thinking Machines Lab: From OpenAI CTO to Independent Startup

From OpenAI to a new AI frontier

Mira Murati left OpenAI in late 2023 after serving as chief technology officer, where she oversaw the rollout of GPT‑4. In early 2024 she gathered a core team of former OpenAI colleagues—including lead researcher Dr. Ananya Patel and systems architect Luis Gomez—to launch Thinking Machines Lab.

The startup’s first public offering, Tinker, debuted in November 2024. Tinker is a developer‑focused suite that streamlines the creation of models that interact with humans, emphasizing safety, interpretability, and real‑time feedback loops. According to the company’s blog, Tinker reduced the average training cycle for a 6‑billion‑parameter model from 48 hours to 28 hours on a 200‑GPU cluster.

Within a year, the lab grew from roughly 30 engineers to about 120, a 300 % increase, according to people familiar with the matter. The rapid hiring spree reflects the broader talent inflation in AI, where median compensation for senior ML engineers jumped from $180k in 2022 to $260k in 2025 (Wall Street Journal, 2025).

Murati’s vision—“AI systems that work with humans rather than operating autonomously”—has guided product roadmaps and attracted talent eager to shape more collaborative AI. The lab’s culture emphasizes interdisciplinary collaboration, bringing together ethicists, cognitive scientists, and hardware engineers under one roof.

Industry observers, such as Dr. Samuel Ortega of the MIT Media Lab, argue that the lab’s human‑centric focus could carve a niche distinct from the compute‑only race dominated by firms like Anthropic and Google DeepMind. “If Thinking Machines can prove that human‑in‑the‑loop models are both safer and more cost‑effective, they’ll set a new benchmark for the industry,” Ortega said in a recent interview (MIT Media Lab, 2025).

The partnership with Nvidia therefore serves a dual purpose: it supplies the raw compute needed to train large, interactive models, and it validates Murati’s thesis that next‑generation AI will be built on a foundation of massive, yet efficient, hardware.

Looking ahead, the timeline of key milestones—from founding to the Nvidia deal—highlights the rapid acceleration of the lab’s ambitions.

Financial Stakes: Estimating the Scale of Nvidia’s Investment

Putting a dollar figure on the gigawatt

While the WSJ article did not disclose the exact monetary size of Nvidia’s investment, industry analysts can triangulate the figure using known hardware pricing and Nvidia’s historical partnership spend.

The list price for a single H100 GPU in 2025 was approximately $30,000. A gigawatt of compute, roughly 10,000 H100 units, would therefore cost $300 million in hardware alone. Adding software licensing, support contracts, and co‑design engineering—typically 30‑40 % of hardware cost—pushes the total to between $400 million and $500 million.

Bloomberg’s 2024 coverage of Nvidia’s AI ecosystem investments noted that the company allocates roughly 12 % of its annual AI revenue to strategic partnerships. With FY2023 AI revenue at $10.2 billion, that translates to a $1.2 billion partnership budget, making a $400‑$500 million commitment plausible for a high‑potential startup.

From Thinking Machines’ perspective, the infusion of capital is expected to fund not only hardware but also talent acquisition, research grants, and operational scaling. The company’s headcount grew from 30 to 120 in twelve months, implying a hiring budget of at least $150 million when accounting for market‑rate salaries and equity packages.

Financial analyst Laura Chen of Morgan Stanley highlighted that “strategic hardware deals often double as de‑risking mechanisms for chipmakers.” By embedding its technology early, Nvidia reduces the risk of losing a future heavyweight customer to a rival like AMD or Intel.

The financial stakes are further underscored by the potential upside: if Thinking Machines’ human‑centric models achieve higher efficiency, they could lower the total cost of ownership for AI workloads across industries, indirectly boosting Nvidia’s broader market share.

In the next chapter, we explore how the influx of capital translates into a competitive talent war, reshaping the AI labor market.

Talent War and Workforce Growth: The Human Capital Behind the Chips

Hiring at breakneck speed

Thinking Machines’ staff surge—from 30 to about 120 engineers in a single year—mirrors the broader AI talent crunch that has defined the mid‑2020s. According to a 2025 LinkedIn report, demand for machine‑learning engineers grew 62 % year‑over‑year, while supply lagged, pushing average salaries up by 35 %.

To attract top talent, Thinking Machines has offered equity stakes comparable to those at early‑stage unicorns. One insider, who asked to remain anonymous, said the company’s compensation packages now include a 0.5 % employee‑wide pool of restricted stock units, a figure that rivals offers at OpenAI and DeepMind.

Industry veteran and AI recruiter Maya Patel of Hired.com noted, “Startups that secure a hardware partner like Nvidia can promise researchers access to compute that most rivals can’t match, which is a powerful recruitment lever.”

The influx of capital from Nvidia also enables Thinking Machines to expand its research labs in Austin and Cambridge, cities that have become AI hubs due to tax incentives and university pipelines. The company’s new Austin campus, slated to open in Q3 2026, will house 60 engineers and feature a dedicated GPU cluster built on Nvidia’s DGX A100 platform.

From a macro perspective, the talent war has spurred a 15 % increase in AI‑related patents filed in the United States between 2023 and 2025, according to the United States Patent and Trademark Office (USPTO). This surge reflects the competitive pressure to innovate while securing scarce engineering talent.

Bar chart below visualizes the staffing growth of Thinking Machines relative to three of its peers—Anthropic, Cohere, and Stability AI—over the past 12 months.

As the race for engineers intensifies, the next chapter examines how Nvidia’s hardware commitment reshapes competitive dynamics across the AI ecosystem.

Strategic Implications: How This Deal Shapes the Competitive Landscape

Building an ecosystem advantage

Nvidia’s partnership with Thinking Machines is part of a broader strategy to embed its GPUs into the next generation of AI startups. In 2023, Nvidia announced a $25 billion AI ecosystem fund aimed at securing early‑stage customers, a move that has since resulted in over 200 formal collaborations (Bloomberg, 2024).

The deal also positions Nvidia against rivals such as AMD, which has launched its MI300X accelerator but has yet to secure a comparable number of startup partnerships. Analyst Raj Patel of Citi wrote, “Nvidia’s moat is no longer just hardware performance; it’s the network of developers, startups, and research labs that rely on its stack.”

From a market‑share perspective, the partnership could shift the balance of AI compute spending. IDC’s 2025 forecast predicts that by 2028, Nvidia will control 70 % of the high‑performance AI GPU market, up from 55 % in 2023. The influx of dedicated gigawatt‑scale customers like Thinking Machines accelerates that trajectory.

Donut chart below breaks down the sources of Nvidia’s AI revenue in FY2024: data center GPUs (62 %), automotive AI (18 %), professional visualization (12 %), and other (8 %). The partnership adds a new slice—strategic startup collaborations—that, while not yet quantified, is expected to grow into a measurable revenue stream.

Beyond revenue, the collaboration influences standards. Nvidia’s CUDA and cuDNN libraries are becoming de‑facto standards for AI research, and co‑design work with Thinking Machines could embed new primitives for human‑in‑the‑loop training, potentially shaping future AI frameworks.

The competitive ripple effect is already visible: rival startups have announced parallel hardware deals with AMD and Google’s TPU team, indicating a fragmentation of the AI compute market. The next chapter looks ahead to how these dynamics may evolve over the next five years.

What Does the Future Hold for Nvidia and Thinking Machines?

Projected growth and technology roadmaps

Looking five years ahead, both Nvidia and Thinking Machines are likely to double down on human‑centric AI. Nvidia’s roadmap, unveiled at GTC 2025, promises the Hopper‑next generation GPU with a projected 1.5× performance per watt over the H100, which could push the gigawatt threshold to 1.5 GW for a single partner.

Thinking Machines has hinted at a second‑generation Tinker platform that will integrate reinforcement‑learning‑from‑human‑feedback (RLHF) loops directly into the training stack, reducing the need for post‑hoc alignment work. If successful, this could halve the compute cost per model, effectively stretching Nvidia’s gigawatt investment further.

Line chart below tracks Nvidia’s AI‑related capital expenditures from FY2022 to FY2025, illustrating a steady upward trajectory that aligns with the company’s strategic emphasis on ecosystem partnerships.

External forecasts from Gartner suggest that AI‑driven enterprises will spend $1.2 trillion on compute by 2028, with GPUs accounting for 45 % of that spend. Nvidia’s early lock‑in with startups like Thinking Machines positions it to capture a disproportionate share of that growth.

However, risks remain. Supply‑chain constraints for advanced semiconductors could delay the rollout of next‑gen GPUs, while regulatory scrutiny over AI safety—exemplified by the EU’s AI Act—may increase compliance costs for both firms.

In summary, the Nvidia‑Thinking Machines partnership is a microcosm of the broader AI arms race: massive hardware commitments, talent battles, and strategic positioning. As the industry evolves, the gigawatt of GPUs today may become the baseline for tomorrow’s AI super‑labs.