Compute Became Collateral
Every few decades an asset moves from being bought with equity to being bought with debt, and the industry built on it is never the same afterwards. Ships were once financed by merchants pooling their own money. Then lenders worked out what a hull was worth secondhand and how quickly it could be sold, and shipping became a lending business. Commercial property crossed the same line a century ago. Aircraft followed in the 1970s, fiber and cell towers in the 2000s. In each case the same hierarchy formed. The asset with the deepest queue of buyers borrows cheapest, and the best asset always has the deepest queue.
GPUs crossed that line this month, and Lambda, our newest portfolio company, is the reason I can put a date on it.
On August 13, Lambda borrowed $926 million against its chips and its customer contracts. Lending against GPUs is not new. CoreWeave raised $2.3 billion that way in August 2023 from Magnetar and Blackstone, and the sector now carries more than $20 billion of debt secured by NVIDIA hardware. What changed last week is the grade. A rating agency judged Lambda's debt investment grade, which by the company's account no private neocloud had achieved before. For three years this paper was private credit, priced by funds that specialise in risk. Now a far wider pool of buyers can own it, insurance companies and pension funds among them.
Three days earlier, NVIDIA explained why that became possible. On August 10 it announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to channel more than $500 billion of outside capital into AI compute. Jensen Huang summarised it in five words: in AI, compute is revenue.
Read what the release actually claims. NVIDIA compute is described as fungible and transferable across customers and operators, with the lowest token cost, the highest revenue and the longest life of any option available. That is not a product pitch dressed up for investors. It is a credit argument, and NVIDIA is telling the largest lenders on earth that its chips are the best collateral in the category precisely because they are the best chips in the category. CUDA is why the claim holds, because software that keeps five year old hardware productive is what protects resale value.
That is the strategy, and it is a good one. NVIDIA is converting a manufacturing lead into a financing lead. Every operator running NVIDIA borrows against an asset the market has agreed is liquid. Everyone else borrows against something a credit committee has to think about.
Lambda sits closer to that asset than anyone. It has been an NVIDIA launch partner for new silicon, and last year NVIDIA bought eighteen thousand GPU servers from Lambda and leased them straight back for $1.5 billion. I wrote about that deal here in July and called it evidence that access to compute is a relationship business. I would put it more precisely now. NVIDIA was testing whether its own hardware could carry a financing structure, using the partner it knew best, before taking the same model to Blackstone and KKR at three hundred times the size.
The relationship was real. What is new is where it shows up. Being an NVIDIA launch partner used to mean getting chips before rivals did. It now also means lenders look at Lambda's hardware and see something with a deep resale market behind it, so they charge less to finance it. Take one percentage point as an illustration. On a billion dollars that is ten million a year, every year, and Lambda intends to borrow multiples of a billion between here and 2030.
Look at how Lambda has borrowed since. A $275 million credit line in August 2025. Raised to $1 billion this May with J.P. Morgan leading, with more demand than the company needed. Then last week's $926 million, borrowed for longer, at a lower risk premium, from a wider pool of lenders. Each raise bigger, cheaper and further dated than the one before. Lenders do not move in that direction on enthusiasm.
They move on contracts. Lambda's rating rests on revenue a customer has already committed to pay, which here is the multibillion-dollar agreement it signed with Microsoft last November to deploy tens of thousands of NVIDIA GPUs. The Microsoft deal and the credit rating are the same fact seen twice. Microsoft locks in the revenue, and the agency confirms it is solid enough to lend billions against.
The equity market has drawn the same conclusion. This week Lambda was reported to be in talks for up to $3 billion ahead of a listing, at a valuation as high as $12 billion, roughly double where it raised last November, with several term sheets already in hand. Revenue is expected to pass $1.5 billion this year against a little over $520 million in the twelve months to September. A buildout on this scale runs on both kinds of capital, equity to carry the risk and debt to carry the assets underneath it. What changed this month is the debt, and the equity now reportedly being lined up on top of it points at a buildout larger than the one the company was describing a year ago.
This is also likely the last of it. A company raising three billion dollars at twelve billion with several term sheets in hand is not planning many more private rounds. Morgan Stanley, J.P. Morgan and Citi are engaged, and a listing is planned for 2027.
NVIDIA spent this year building a market where its own chips can be borrowed against at scale, and Lambda is the first private operator to sell investment grade paper into it. That is the news. Nothing is filed yet. The first real look comes with the S-1, the registration statement every company files with the SEC before listing, where the assumptions behind all of this get written down in public for the first time. All investors, us included, are keen to learn where it prices.
Bashar Aboudaoud
Managing Member, UpRound
Q3 2026 Live Webinar
Live Webinar: Thursday, September 17
You leave with three things. Where private valuations actually sit right now. Which sectors are repricing fastest. How I am reading prices into 2027.
20 minutes, then live Q&A. 20 slots only for UpRound readers.
