The Gigawatt Lab: How Mira Murati's Thinking Machines Is Rewriting the Rules of AI Competition

A $12 billion seed-stage startup commits to one gigawatt of NVIDIA compute — while its co-founders keep walking out the door

Executive Summary

NVIDIA and Mira Murati's Thinking Machines Lab have announced a multiyear strategic partnership deploying at least 1 gigawatt of Vera Rubin systems — a compute commitment rivaling entire national AI strategies — while the 13-month-old startup has already lost four co-founders, three of whom returned to OpenAI.
The deal crystallizes a new structural reality in AI: the emergence of "gigawatt labs" where capital intensity has become so extreme that only a handful of entities can compete at the frontier, creating a compute oligopoly that may prove more durable than any model advantage.
For investors, the Thinking Machines paradox — sky-high valuation, massive infrastructure bets, yet chronic talent instability — offers a microcosm of whether the current AI investment cycle is building durable enterprises or inflating the most expensive bubble in technology history.

Chapter 1: The Partnership That Rewrites Scale

On March 10, 2026, NVIDIA and Thinking Machines Lab announced what amounts to one of the largest compute partnerships in AI history. The deal commits Thinking Machines to deploying at least one gigawatt of NVIDIA's forthcoming Vera Rubin systems, with deployment targeted for early 2027. NVIDIA is also making a "significant investment" in the startup, though neither party disclosed the amount.

To understand what one gigawatt of AI compute means, consider the scale. A gigawatt is roughly the output of a large nuclear power plant. It could power approximately 750,000 homes. In the context of AI data centers, one gigawatt of Vera Rubin systems represents perhaps 50,000 to 100,000 GPUs running simultaneously — a cluster that would rank among the most powerful computing installations ever built. Jensen Huang himself has projected that companies could spend $3 trillion to $4 trillion on AI infrastructure by the end of the decade, and deals like this are the building blocks of that projection.

The partnership also includes a commitment to design training and serving systems optimized for NVIDIA architectures, and to broaden access to frontier AI for enterprises, research institutions, and the scientific community. In Murati's words: "NVIDIA's technology is the foundation on which the entire field is built. This partnership accelerates our capacity to build AI that people can shape and make their own."

For context, OpenAI allegedly signed a historic $300 billion compute deal with Oracle in September 2025. While we don't know the dollar value of the Thinking Machines-NVIDIA deal, a one-gigawatt Vera Rubin deployment would likely cost billions of dollars annually in hardware alone, before accounting for energy, cooling, and operational expenses.

Chapter 2: The Murati Paradox — $12 Billion and a Revolving Door

Thinking Machines Lab's story is, in many ways, the story of the AI industry's contradictions made manifest.

Mira Murati co-founded the company in February 2025, shortly after departing OpenAI where she had served as Chief Technology Officer — and, briefly and dramatically, as interim CEO during Sam Altman's November 2023 ouster. Her departure from OpenAI was seen as one of the most significant talent exits in the AI industry, and her new venture attracted immediate investor interest.

By July 2025, Thinking Machines Lab had raised over $2 billion from a constellation of top-tier investors including Andreessen Horowitz, Accel, and — notably — both NVIDIA and rival chipmaker AMD's venture arm. The company's valuation at this seed stage exceeded $12 billion, making it one of the most richly valued startups at such an early stage in Silicon Valley history.

Yet even as the capital poured in, the talent began to drain out. In October 2025, co-founder Andrew Tulloch left for Meta. Then, in January 2026, three additional co-founders — Barret Zoph, Luke Metz, and Sam Schoenholz — departed to return to OpenAI, the very company Murati had left to start Thinking Machines.

This is the Murati Paradox: a startup valued at $12 billion, backed by the most powerful chipmaker on Earth, with a billion-dollar war chest and a gigawatt compute deal — and yet unable to retain half its founding team through its first year of existence. The company's first and only public product, an API called Tinker released in October 2025, allows researchers and developers to fine-tune AI models, but details about its broader technical roadmap remain scarce.

The revolving door raises uncomfortable questions. Is the talent instability a sign of fundamental disagreements about direction? Is it simply the gravitational pull of OpenAI's enormous resources and market position? Or does it reflect a deeper structural problem — that in the AI industry of 2026, even $12 billion and a gigawatt of compute may not be enough to compete at the frontier?

Chapter 3: The Rise of the Gigawatt Lab

The NVIDIA-Thinking Machines deal is not an isolated event. It represents the emergence of a new category in the AI ecosystem: the gigawatt lab.

Consider the compute commitments that have been announced in the past 18 months:

Entity	Compute Commitment	Estimated Value	Primary Chip Partner
OpenAI-Oracle	Multi-gigawatt	~$300B	NVIDIA/Custom
Microsoft-OpenAI	Stargate Phase II	~$100B+	NVIDIA
Thinking Machines-NVIDIA	1 GW Vera Rubin	Billions (undisclosed)	NVIDIA
xAI (Musk)	Memphis + Mississippi	~$20B+	NVIDIA
Meta	Multiple data centers	~$65B (2026 capex)	NVIDIA/Custom
Google DeepMind	TPU v6+ clusters	~$75B (2026 capex)	Custom TPU

The combined AI infrastructure spending of the hyperscalers and frontier labs is approaching $650 billion in 2026 alone. What distinguishes the "gigawatt lab" from the hyperscaler is the concentration of compute for a single purpose: training and serving frontier AI models. A hyperscaler like AWS or Azure serves millions of customers across diverse workloads. A gigawatt lab channels essentially all of its compute toward a single research agenda.

This concentration creates a new competitive dynamic. The barrier to entry for frontier AI research is no longer just talent or algorithms — it is the physical infrastructure of power, cooling, and silicon. A one-gigawatt deployment requires not just the chips but also the electrical infrastructure to power them, the water or cooling systems to manage thermal loads, and the networking to connect them. These are multi-year construction projects, not software deployments.

The Iran war and the Strait of Hormuz blockade have added an unexpected dimension to this infrastructure race. As we documented in our analysis of "Silicon's Hunger," the disruption to energy supplies and the helium supply chain from Qatar has created physical constraints on data center operations, particularly in the Gulf region. The AWS data center drone attack in the UAE demonstrated that AI infrastructure is now a legitimate military target. The gigawatt lab is not just an economic bet — it is a geopolitical asset that must be secured, powered, and defended.

Chapter 4: NVIDIA's Kingmaker Strategy

NVIDIA's investment in Thinking Machines Lab is part of a deliberate pattern. The company has invested in nearly every major AI lab: OpenAI, Anthropic, Cohere, Mistral, and now Thinking Machines. Jensen Huang has transformed NVIDIA from a chip supplier into the kingmaker of the AI industry.

The strategy works on multiple levels. First, every investment creates a customer. A startup that takes NVIDIA's money is very likely to deploy NVIDIA's hardware, creating a self-reinforcing cycle of demand for GPUs. Second, the investments provide NVIDIA with early visibility into frontier research, allowing it to optimize its chip roadmap. Third, by investing in competing labs, NVIDIA ensures that the total demand for its chips grows regardless of which lab "wins" the AI race.

This is not without historical precedent. In the late 1990s, Intel Capital made strategic investments in hundreds of technology companies, creating an ecosystem that depended on Intel's x86 architecture. The result was a decades-long monopoly on personal computer processors. NVIDIA appears to be executing a similar playbook, but at a much larger scale and with higher stakes.

The risk, however, is that NVIDIA's centrality becomes its vulnerability. The U.S. government's Project Vault initiative, announced on March 10, includes provisions for critical mineral stockpiling that explicitly address the semiconductor supply chain. The EU's Industrial Accelerator Act imposes local content requirements on technology infrastructure. And the growing compute sovereignty movement — exemplified by Nscale's $2 billion raise for European AI infrastructure — suggests that not all governments are comfortable with an AI ecosystem centered on a single American chipmaker.

For Thinking Machines Lab specifically, the NVIDIA partnership is both an enabler and a constraint. One gigawatt of Vera Rubin systems provides extraordinary capability, but it also locks the company into NVIDIA's architectural roadmap. If a competitor develops a fundamentally different approach to AI compute — as Groq attempted with its LPU architecture before NVIDIA's acquisition — Thinking Machines would face enormous switching costs.

Chapter 5: Scenario Analysis — The Future of the Gigawatt Labs

Scenario A: Compute Consolidation (40%)

Thesis: The gigawatt labs succeed in building frontier models that justify their enormous capital expenditures, leading to a consolidation of the AI industry around 3-5 dominant players.

Evidence:

Historical precedent: The cloud computing industry consolidated from dozens of providers in 2010 to three dominant players (AWS, Azure, GCP) by 2020, driven by the capital intensity of data center infrastructure. AI labs face even higher barriers.
OpenAI's $300B Oracle deal and Thinking Machines' 1GW commitment suggest that only entities with access to tens of billions in capital can compete at the frontier.
NVIDIA's GTC 2026 announcements (Vera Rubin + LPX inference chip + NemoClaw platform) are designed to make its ecosystem even more dominant, raising switching costs for all players.

Trigger: One or more gigawatt labs demonstrates a clear capability breakthrough (e.g., AGI-adjacent reasoning, scientific discovery automation) that validates the infrastructure spending.

Timeline: 12-24 months for initial consolidation signals; 3-5 years for stable oligopoly.

Scenario B: The Talent Exodus Unravels the Model (35%)

Thesis: The revolving door of talent proves fatal to several gigawatt labs, revealing that raw compute without stable research teams cannot produce frontier models.

Evidence:

Thinking Machines' loss of 4 co-founders in 13 months is not an isolated case. The AI industry is experiencing what researchers call "founder liquidity" — co-founders who vest quickly, take profits, and move on.
The 2000-01 dot-com precedent: Companies that raised massive capital but couldn't retain key engineers (Webvan, Pets.com, Kozmo) collapsed despite their funding advantages.
Anthropic's experience suggests stability matters: its relatively low turnover (until the recent Pentagon controversy) correlates with consistent model improvements.

Trigger: A major gigawatt lab fails to produce a competitive frontier model despite billions in spending, triggering investor reassessment.

Timeline: 6-18 months for the first high-profile disappointment.

Scenario C: The Energy Constraint Restructures the Race (25%)

Thesis: The Iran war's energy disruption and the broader global energy transition create physical constraints that prevent gigawatt labs from deploying their planned infrastructure on schedule.

Evidence:

Brent crude at $91-$119 in March 2026 has already increased data center operating costs by an estimated 20-40% for energy-dependent regions.
The AWS UAE drone attack demonstrated that data centers in conflict-adjacent zones face physical security risks.
xAI's Mississippi power plant regulatory battle shows that even in the U.S., securing gigawatt-scale power is a multi-year political and engineering challenge.
The nuclear energy summit in Paris (March 2026) explicitly linked AI data center demand to the nuclear renaissance, acknowledging that current grid capacity cannot support the planned compute buildout.

Trigger: A major data center project faces delays of 12+ months due to energy or permitting constraints, forcing a rethink of deployment timelines.

Timeline: Already emerging; full impact within 6-12 months.

Chapter 6: Investment Implications

The gigawatt lab phenomenon creates a distinctive investment landscape:

Clear winners:

NVIDIA (NVDA): The kingmaker benefits regardless of which lab succeeds. Every gigawatt deal is a multi-billion-dollar revenue commitment. With GTC 2026 revealing the Vera Rubin + LPX + NemoClaw stack, NVIDIA's platform lock-in deepens. Current forward P/E of ~30x looks expensive until you consider that $650B in 2026 AI capex is flowing disproportionately to one company.
Energy infrastructure: Companies that can deliver gigawatt-scale power to data centers — Constellation Energy, Vistra, Cameco (uranium) — are structural beneficiaries. The nuclear renaissance is directly tied to AI compute demand.
Cooling and electrical equipment: Eaton, Vertiv, Schneider Electric benefit from every data center buildout regardless of which AI lab occupies it.

High-risk, high-reward:

Frontier AI labs (private): The Thinking Machines valuation ($12B at seed) prices in enormous success. The talent exodus risk is unpriced. Investors in these rounds are essentially making binary bets.
Cloud providers without AI differentiation: Smaller cloud providers that can't match hyperscaler AI spending face disintermediation.

Potential losers:

SaaS companies: As NVIDIA's NemoClaw and similar agent platforms emerge, traditional SaaS companies face disruption from AI-native alternatives. This "SaaS-pocalypse" risk is accelerating.
Non-NVIDIA chip companies: Intel's continued struggles and AMD's secondary position suggest the compute oligopoly favors NVIDIA overwhelmingly. The Groq acquisition removed the most promising alternative architecture.

Conclusion

The NVIDIA-Thinking Machines Lab partnership is a milestone, but not in the way the press releases suggest. It is not simply an investment or a compute deal. It is a data point in the emergence of a new industrial structure — the gigawatt lab — that may define the AI industry for the next decade.

The paradox at the heart of this structure is profound. The AI industry has convinced investors that frontier models require gigawatt-scale infrastructure and billions of dollars in capital. Yet the actual production of intelligence — the research breakthroughs, the architectural innovations, the training insights — remains the work of relatively small teams of researchers. Thinking Machines Lab's four departed co-founders represented an irreplaceable concentration of expertise, and no amount of Vera Rubin GPUs can replicate what walked out the door with them.

This tension between capital intensity and human creativity is the defining question of the AI industry in 2026. The answer will determine whether the gigawatt labs become the Bells and GEs of the 21st century — durable industrial giants built on infrastructure moats — or the Webvans and Pets.coms of a new era, monuments to capital deployed without sustainable competitive advantage.

For now, the money keeps flowing, the deals keep getting larger, and the power plants keep being planned. One gigawatt at a time, the AI industry is building its future. Whether that future justifies the investment remains, as Murati might say, a question that people will need to shape and make their own.

Eco Stream