Summary

  • Sharon AI’s SEC-filed agreement is for up to US$356m, not evidence that US$356m has already been drawn. It comprises a US$150m Facility A and a US$206m Facility B.
  • The 9.95% press-release headline is not the whole coupon schedule: A starts at 7.25% and steps up on dates the filing redacts; B is fixed at 9.95%. Both are monthly cash-pay bullet facilities due 42 months after first utilisation.
  • The filing links customer acceptance and an escrowed deposit to an expected prepayment of A. If A is repaid, the parent guarantee is expected to release and the financing to become non-recourse to Sharon AI, with management and IP-licence exceptions. The customer, deposit, drawdowns and fee amounts remain undisclosed.

The headline rate is not the financing map

Sharon AI announced an inaugural US$356m GPU-backed facility on 1 October at a fixed 9.95%, excluding fees. That is a useful headline for a capital raise; the SEC-filed agreement summary shows why it is not enough to price the risk. The total is a commitment “up to” US$356m, divided between two borrower facilities: up to US$150m in A and up to US$206m in B.

The two tranches do not begin on the same rate. Facility A pays 7.25% initially, then 8.25%, 9.25% and 9.95%; the step dates are redacted. Facility B carries 9.95% throughout. Interest is cash-pay monthly. The company’s press release compresses this into the 9.95% headline, while the signed schedule makes the early cost of A different and its later path uncertain to outside readers.

The difference can be illustrated without pretending to know the actual bill. If the full commitments were drawn and stayed outstanding for a year at A’s first rate, simple coupon arithmetic would be about US$31.4m before fees, or roughly 8.81% of US$356m. Applying 9.95% to the entire commitment would imply about US$35.4m. Neither number is actual interest or a forecast: the facility may be partly drawn, A’s rate changes, principal can be prepaid and the fee letters are not public. The calculation only shows why the two-tranche schedule matters.

The agreement calls both A and B bullet term loans, each repayable in one lump sum 42 months after its first utilisation, subject to mandatory prepayment. A is therefore not merely the “cheap” half of a flat-rate loan. It has a different time path, and the filing gives it a customer-linked exit mechanism.

Acceptance becomes a financing event

The SEC filing identifies two Australian subsidiaries as the borrowers and says proceeds can fund or refinance servers, GPUs, CPUs, networking, storage, data-centre infrastructure, debt-service reserves and transaction costs for a GPU-compute customer contract. It does not disclose the customer’s name or the amount currently outstanding.

More unusually, the company’s filing says it expects Facility A to be mandatorily prepaid from a customer deposit held in escrow when the customer accepts Service Order 2. The order and deposit amount are redacted. That makes acceptance more than a service milestone in the public description: it is expected to release a pool of cash toward the first debt tranche.

The distinction is important. A deposit applied to principal is not the same thing as recurring revenue from operating GPUs, and it says nothing by itself about whether the remaining B tranche can be serviced. Nor does a signed facility prove that the customer has accepted anything. The SEC-filed press release says the project supports contracted compute, but the contract names, acceptance conditions, utilisation amounts and fee schedule are not public.

There is also no basis for assigning this loan to Sharon AI’s separately announced US$373m five-year cloud contract. The financing documents refer to a confidential customer master agreement and several service orders; the public filings do not identify the counterparty or connect them to that other announcement. Similar GPU language is not proof of the same contract.

The guarantee changes only after A exits

The borrowers’ immediate holding companies guarantee the facilities. Sharon AI itself gives a limited, releasable guarantee for both tranches. The agreement also grants senior security over substantially all obligor assets, shares in the borrowers and immediate holding companies, project bank accounts and material project contracts, subject to stated protections.

The company says that once Facility A is repaid, specified parent-level defaults and undertakings cease and the financing becomes non-recourse to Sharon AI, except through management and intellectual-property licence arrangements to be entered into by a separate subsidiary. That is a staged transition, not a declaration that the group has no continuing role or that all risk disappears at acceptance. The expected sequence is acceptance, release of the escrowed deposit toward A, repayment of A, then a narrower parent recourse perimeter. Each link remains conditional.

Facility B survives that first exit. It cannot be voluntarily prepaid until A has been repaid in full, and certain B prepayments during the first 18 months carry a make-whole. For creditors, sequencing can preserve the seniority and expected return of the second tranche. For the borrower, it limits the ability to clear B first even if another source of cash becomes available. The lenders’ security and contract rights remain attached to project-level borrowers after the parent guarantee changes.

A balloon still waits at the end

The 42-month bullet maturity keeps a second clock running after acceptance. The agreement has mandatory prepayment provisions, including cash-flow sweeps after the Customer Acceptance Date and specified proceeds from asset sales, insurance or customer termination payments. It also requires a quarterly gross loan-to-value test, but the maximum levels are redacted. Neither the SEC summary nor the release shows a draw schedule, the funded balance, reserves held, the acceptance date or how much the deposit could repay.

That makes the financing more nuanced than either “US$356m at 9.95%” or “contracted GPUs have financed themselves.” The public record shows a lender-designed bridge from customer acceptance to repayment of A and a possible release of parent support. It does not show the bridge has been crossed, how much principal remains on the other side, or whether operating collections can meet B’s monthly interest and eventual bullet payment.

The next useful disclosures are therefore tranche-level, not another aggregate-capacity headline: amounts committed, drawn and outstanding; the dates of A’s rate steps; fees and reserves; the accepted service order and deposit applied; the amount of A repaid; the date and exact scope of any guarantee release; covenant headroom; and B’s outstanding balance and service source. Customer identity can remain confidential while these financing states are reported.

The market signal is not that GPU-backed debt has become cheap. It is that customer acceptance is being used as a contract-defined pivot between two different forms of credit exposure. Until utilisation, acceptance, prepayment and recourse release are reported, the facility is a detailed financing design with its most important transition still ahead.

Sources