Lambda and Sharon AI price GPU-backed debt 317 basis points apart

Lambda borrowed $1 billion against GPUs at a 6.78% coupon with an investment-grade rating attached, while Sharon AI borrowed $365 million against GPUs at 9.95% through a special purpose vehicle, putting the combined raises at $1.356 billion and the two coupons 317 basis points apart.

Both facilities are described as GPU-backed, which is what makes the gap worth attention rather than a shrug. Two loans secured by the same category of equipment, priced three points apart in the same week, are not being priced on the equipment. The market is charging for the borrower, for the legal structure, for the terms attached to each, or for some blend of the three, and that is the argument every financing plan in the data center buildout now has to absorb.

The pair is useful because it arrived together: coupons struck weeks apart, with different lenders and different rate backdrops, are hard to hold side by side, and two GPU-backed raises reported in the same week, secured by the same class of asset, give lenders and sponsors a comparison they can act on when the next term sheet is drafted. That does not settle what either borrower is worth, but it does remove the argument that the two numbers came from different markets.

The Lambda facility carries an investment-grade rating, the element of the pair that is hardest to reproduce, and an investment-grade assessment of paper secured by GPUs reads as a judgment about the credit behind the machines; it came attached to a $1 billion borrowing at 6.78%. The coverage does not name the agency that assigned the rating, describe the terms behind it, or say whether the Sharon AI debt carries a rating of its own.

The structures differ as well, at least in the labels the reporting uses—a $1 billion facility with a rating for Lambda, $365 million of SPV debt for Sharon AI—and special purpose vehicle debt conventionally means borrowing against a ring-fenced pool of assets, which the reporting does not describe. Contracted compute revenue, a tenant commitment, a sponsor guarantee and a reserve account are the kinds of terms that move a coupon in project finance, and none of them is published here.

The buyer base is another likely contributor: rated paper at $1 billion is easier to distribute than an unrated $365 million loan, and the narrower group of lenders that holds the smaller deal expects to be paid for taking illiquidity as well as credit. The reporting names no lenders on either facility, so that explanation cannot be tested here, though it probably accounts for some of the 317 basis points.

Put the coupons in dollars and the difference stops being abstract: at 9.95% versus 6.78%, the extra 317 basis points on $365 million works out to roughly $11.6 million a year. That is what the gap costs the smaller borrower every year it is priced this way, and it is the number a developer weighing a rated facility against an SPV structure has to run through its own model.

The supply of borrowers is not in question: nonresidential construction spending rose in August for a fifth consecutive month, and the chief economist at the Associated Builders and Contractors expects the momentum to stay confined to data centers and power, with material and labor costs reemerging as a constraint on what gets built. Rising input costs imply a larger cheque for each hall, and a larger cheque usually means more debt raised against the same revenue plan, so every building in that pipeline arrives with a financing need attached to it.

The week's filings sketch the queue: Skybox and Prologis filed for an 846,300 square foot single-story building at the Hutto PowerCampus that would be the fourth and largest on the site, with construction set for January 2027 and $550 million of Skybox capital behind it; Intro Group has earmarked $270 million for the first phase of Egypt's Kemet Data Center in the Suez Canal Economic Zone, an 80MW project across two phases that is partly solar and has no named tenant or disclosed interconnection terms; and Xeal's Laitent says it can reach more than 200MW of permitted electrical infrastructure across 1,600 US sites and intends to deploy 100,000 Nvidia GPUs to run inference on capacity its EV chargers are not using.

The two halves of an AI buildout are financed differently, and this week's coverage shows both: property and power get built against land, interconnection and leases, which is the shape of Skybox's $550 million and Intro's $270 million, while compute gets bought against chips and contracts, which is the shape of the two GPU facilities. When the equipment half of that stack begins pricing like corporate credit, the blended cost of capital for the whole project moves with it, because the same project budget pays both bills.

Someone absorbs the extra 317 basis points, and it will not be the lender: on a single-phase project the difference lands in the developer's return, and if the market settles at the wide end it lands in the rent a tenant eventually agrees to, or in the number of phases that get built. The reporting does not say who the tenant is for either GPU facility, so where the cost finally lands cannot be traced in these two deals.

Once lenders have two prices to work from, a borrower with a sponsor, contracted revenue and a fleet large enough to rate can aim at the low end, while a borrower without them is offered the high end, and the distance between the two ends is now visible to everyone in the room. Precedents shape term sheets quickly, because lenders generally prefer to price off the last comparable deal rather than underwrite an asset class from scratch.

None of the borrowers in this week's pipeline looks like Lambda: they are single-site, single-phase vehicles, often in one market, which is the profile that prices toward the wide end of a range once lenders start differentiating. Two prints in one week also make a thin basis for a market, because a rate becomes a benchmark after comparable deals price repeatedly, and the reporting offers no prior GPU-backed facilities to compare against and no sign of how either deal was distributed. That leaves a spread between two borrowers, and a curve for the asset class is some deals away.

The disclosure sets a limit on how much the spread can be made to say: the reporting gives a size, a coupon and a rating for Lambda, and a size and a coupon for Sharon AI, but no tenor, no advance rate against the hardware, no covenant package, no lender lineup and no amortization schedule. A shorter tenor alone would push a coupon higher, and so would a loan sized more generously against the same equipment, so the 317 basis points is consistent with a credit story and with a terms story at the same time.

If the two lenders are working from different assumptions about how long the collateral holds its value, that alone would move a coupon, and the reporting publishes no depreciation schedule, no refresh assumption and no residual-value estimate for either facility; for asset-backed lending those assumptions usually matter more than the sticker price of the hardware.

The concentration in the construction data cuts the same way: if momentum in nonresidential spending is confined to data centers and power, the lenders financing it are building books that move together, and a market that prices each compute borrower on its own credit will distinguish the credits that can absorb a slowdown from those that cannot. That is an argument for the discipline the spread implies, and it is also the reason a second and third print matter more than this one.

A third GPU-backed facility pricing near 6.78% would suggest the rating and the scale are what the market is buying, and would hand the buildout a lower reference for the cost of compute capital; a third print near 9.95% would suggest the Lambda coupon is the outlier a particular borrower earned, leaving the rest of the queue to fund closer to high-yield levels. The economics of the same square footage of compute look very different under the two.

Intro has budgeted $270 million for Kemet's first phase, Skybox is set to break ground at Hutto in January 2027, and Xeal says it has more than 200MW of permitted electrical infrastructure waiting for chips. Debt for all of it will be measured against the two coupons from this week, and the next facility to price will show which of them is the number the market actually believes.