NVIDIA did a sale and leaseback with one of its own customers. The most valuable chip maker in the world sold roughly 18,000 GPU servers to a company called Lambda, then reportedly signed a $1.5 billion deal to rent them back over four years. That was 2025. By 2026, NVIDIA had made Lambda a launch partner for its newest silicon, handing it the latest hardware to deploy before almost anyone else.

Once you understand why that sequence makes sense, you understand where the real leverage in AI actually sits.

Start with the scarce resource. Training a frontier model means coordinating thousands of high end chips for months without interruption. NVIDIA's H100 and H200 chips cost $25,000 to $40,000 apiece, and NVIDIA still controls roughly 75 to 80 percent of the AI chip market. Supply has not caught up with demand in any durable way since this cycle began. Lead times on the newest systems still run into months for buyers without a direct line to the source. The bottleneck in AI is not talent and it is not capital. It is access to compute. And access to compute is a relationship business before it is anything else.

The hyperscalers were not built for this. AWS, Google Cloud, and Azure were architected for web services and databases, and it shows. GPUs were bolted on as an expensive add-on rather than designed in. Provisioning is slow, availability is unpredictable during peak demand, and the pricing reflects a cost structure built for a different era. Founders feel that friction every time they try to scale a run.

A small set of GPU-native providers built their entire business around closing that gap. Lambda is the clearest example. Twin brothers Michael and Stephen Balaban founded the company in 2012. It found its footing selling GPU workstations and servers to researchers, then pivoted to a full GPU cloud platform in 2017. The strategy has held since. Go deep on one thing, NVIDIA-optimized infrastructure for AI workloads, and let everyone else fight over general purpose cloud. The focus shows in the numbers. Lambda generated $425 million in revenue in 2024, up 70 percent from the year before, and by 2025 that had climbed to an estimated $760 million on an annualized basis. In November 2025 the company raised a Series E of more than $1.5 billion led by TWG Global, at a $5.9 billion post-money valuation.

Now go back to the relationship with NVIDIA, because it is the whole story. It runs in three steps. NVIDIA first put money directly into Lambda's Series D in early 2025, alongside ARK Invest and Andrej Karpathy. Months later it signed the $1.5 billion leaseback, routing 18,000 of its own GPU servers through Lambda and renting the capacity back.

Then it went further. At NVIDIA's GTC conference in March 2026, Lambda became a launch partner for the new Vera CPU platform and NVIDIA STX, and now runs one of the largest deployments of NVIDIA's Quantum-X800 networking anywhere.

Invested, then deployed at scale, then handed the newest silicon first. That is not a vendor relationship. That is a chip maker treating one infrastructure provider as preferred. In a market defined by who can get hardware and how fast, that is the moat.

The ambition matches the relationship. Lambda has stated a target of more than one million NVIDIA GPUs and three gigawatts of liquid-cooled capacity by 2030, and it has already signed and committed over 320 megawatts toward it. It is not the only customer NVIDIA trusts at scale. Lambda also signed a multibillion-dollar, multi-year agreement with Microsoft in late 2025 to deploy tens of thousands of GPUs. The pattern is consistent. The companies that actually build at frontier scale route through Lambda.

The competitive set clarifies the thesis rather than complicating it. CoreWeave is the public reference point. It went public in March 2025 and now carries a market cap around $53 billion on roughly $6 billion in trailing revenue. RunPod runs the opposite playbook, aggregating GPU supply including consumer-grade hardware to undercut on price, which serves burst workloads and fails for anything mission critical. Lambda sits between the two. Owned infrastructure and a direct NVIDIA relationship like CoreWeave, with its last primary round priced at $5.9 billion in November 2025.

The company is building for a public market debut. In May 2026 Lambda brought in Michel Combes as chief executive, a veteran of Sprint, SoftBank International, and Alcatel-Lucent who spent his career deploying capital in infrastructure-heavy businesses. Co-founder Stephen Balaban moved to chief technology officer, and former AT&T executive John Donovan joined as board chairman. This is what a company does when it is preparing to scale capital formation and operate at a different order of magnitude. Combes and Donovan both describe the AI data center buildout in the same terms as the telecom and wireless expansions they ran for three decades. Lambda is reportedly targeting a public offering in the second half of 2026.

The risks are real and worth naming. Lambda depends on a single supplier, and any shift in NVIDIA's allocation or pricing flows straight to its cost base. The same relationship that is the moat is also the concentration. Hyperscalers can bundle compute into existing contracts and absorb losses to win share. Custom silicon from Google and Amazon chips away at the edges. A clear-eyed view holds all of that and still lands on the same conclusion.

The interesting position in AI is the layer underneath, not only the platform everyone is already discussing. It is the part of the stack without which none of it runs. Compute is that layer. The providers with direct hardware access sit closer to the center of the value chain than most people credit. Few are as close to the source as Lambda.

Bashar Aboudaoud
Managing Member, UpRound