Morgan Stanley and Goldman Sachs are not in those meetings to sell equity. They are in those meetings to sell a rating.
Two banks. Two AI companies. One ask, as reported by the Financial Times: an investment-grade credit rating shortly after IPO. Not a valuation. Not a roadshow narrative. A debt rating โ the only number that lets pension funds and insurance companies touch the paper at all.
The agencies have not moved.
Then there is the figure that should be pinned above every crypto credit desk on earth. Up to $105 billion in credit support from NVIDIA behind OpenAI's Ohio data center, structured to terminate the moment the borrower's rating comes in satisfactory.
Read that clause twice. The guarantee is a bridge to the rating. Once the rating exists, the bridge is gone.
I have audited structures shaped like this before. Not in AI. In crypto. In 2021 I watched lending desks mark phantom collateral against real liabilities, and the tell was always identical โ a support mechanism with a termination trigger that fires precisely when the support is no longer needed, except it always still is. The paperwork was worse. The architecture was identical.
Speed is the only currency that does not inflate. So let us move.
Context
Morgan Stanley and Goldman Sachs are running the process for both OpenAI and Anthropic. The prize is a corporate credit rating inside the investment-grade band. The practical consequences are mechanical, not emotional. Bonds sold at tighter coupons. Index eligibility. Access to the pension and insurance bid that is structurally prohibited from holding speculative-grade credit.
An investment-grade rating is not a compliment. It is a permission slip. It says a committee modeled expected loss and arrived at a number low enough that a fiduciary can defend the position in a deposition. That is the entire product.
The secondary prize is partner relief. Enterprise buyers, cloud vendors, and compute suppliers all carry counterparty risk on their own books. A rating does not change the cash flow. It changes how every counterparty accounts for the cash flow, which changes how much of it they are willing to commit.
The catch is the number underneath. FT analysts still classify both companies as distinctly speculative grade. Neither has demonstrated the ability to generate stable positive free cash flow. Not projected โ demonstrated. That word is doing an enormous amount of work in every sentence that follows.
The historical frame is unforgiving. Meta, Netflix, and Tesla all waited more than a decade after listing for investment-grade status. Netflix needed years to convert subscriber growth into the kind of predictable cash generation a ratings committee will underwrite. Meta needed a balance sheet capable of absorbing a nine-figure write-down on a pivot it later abandoned.
Then there is the counterexample. SpaceX listed this year and moved into investment grade quickly โ reportedly the first of the mega-cap tech cohort to do it on that timeline. Starlink carries contracted, recurring, price-inelastic revenue with government anchor tenants. That is a credit profile. It is not a growth story wearing a credit profile as a costume.
Two data points, pointing in opposite directions. The only question that matters is whether OpenAI and Anthropic are the second SpaceX or the third Netflix.
And there is a structural detail the equity bulls keep skipping. A rating is not an event, it is a surveillance relationship. Agencies re-rate annually and on material developments. An investment-grade grade granted early, on the strength of a vendor guarantee that later terminates, converts into downgrade risk every single quarter thereafter. Companies do not just want the rating. They want to hold it.
Core
Now the forensic part, and this is where crypto readers should stop scrolling.
Vendor financing is not equity. NVIDIA extending up to $105 billion in credit support for a data center that will be filled with NVIDIA silicon is a receivable, not a gift. Revenue gets recognized on the hardware. The risk stays on the vendor's balance sheet until the guarantee unwinds. It is not one check, either. Facilities like this draw down in tranches tied to delivery milestones, which means the collateral behind the receivable is the same hardware that generated the receivable. In 2021 and 2022 the crypto lending complex ran exactly this shape โ headline commitments that looked like confidence and functioned like leverage. The difference here is documentation quality, not structural elegance.
The termination trigger is the entire story. A backstop that ends when the rating is achieved means the rating is being manufactured, at least in part, by the backstop. Rating committees know this. That is precisely why they have not moved. The job is to model expected loss on a standalone basis. Strip the NVIDIA guarantee out and you are modeling a company with brutal depreciation on specialized compute, consumption-priced revenue, and a capital expenditure curve drawn by someone who has never had to service a coupon through a downturn.
The spread differential matters more than the headline. Investment grade versus speculative grade is not a cosmetic upgrade โ it is routinely a 200 to 350 basis point gap, plus a hard exclusion from the largest pools of capital on earth. On a multi-hundred-billion-dollar capital program, that gap compounds into a number that dwarfs most of the AI industry's current revenue.
The enterprise angle is where the rating actually earns its keep. Procurement committees inside banks, insurers, and government agencies have counterparty thresholds written into policy. A rating moves a vendor across that threshold without a single new product shipping. That is the real commercial value, and it has nothing to do with model quality.
Now run the crowding-out math, because nobody else will. Pension and insurance balance sheets run finite duration budgets. Tokenized Treasury products, stablecoin reserve portfolios, on-chain credit funds โ all of them have spent three years trying to reach that same bid. They have been locked out, treated as unrated exposure, capped by allocation committees, and scolded by consultants.
When OpenAI and Anthropic issue, the marginal fixed-income dollar does not come from venture capital. It comes from the same treasury desks that would have bought the crypto-native version. Stablecoin issuers have quietly become the cleanest credit in our industry โ short duration, Treasury collateral, transparent reserve attestation. They do not get a corporate rating. They get a regulatory perimeter. That is the trade PayPal understood when it launched PYUSD: become a regulated counterparty before someone regulates you into one. OpenAI is running the identical play through a different door.
Volatility is the tax you pay for access. The access being purchased here is the right to be boring.
Layer2 already taught us how to read this. For two years the ecosystem sold decentralized sequencing on slide decks while production networks ran a single sequencer behind a multisig upgrade key. The architecture was never the promise; the architecture was the reality, and the promise was a marketing layer bolted on top. Compute has the same shape. Hash rate has been concentrating into a shrinking set of pools since the last halving as miner revenue collapsed, and nobody calls that decentralized anymore because the data stopped cooperating. Training compute is concentrating into fewer balance sheets than that. An investment-grade rating on a concentrated compute stack is not a decentralization story. It is a consolidation story with a credit enhancement stapled to it.
Three things belong on the screen before believing any IG timeline.
Coverage ratios. Investment grade is a debt-service metric, not a growth metric. EBIT has to exist before interest coverage can be computed. EBIT at scale requires gross margin discipline in inference pricing โ a market currently deflating faster than any model refresh cycle can offset.
Duration. Watch whether the first bond is five-year or ten-year paper. Five-year issuance tells you the syndicate does not trust the free cash flow curve past the current depreciation schedule. Duration is a confession.
The guarantee amendments. The $105 billion is not static. It will be amended, extended, or renegotiated, and every amendment is public information about who is actually carrying the risk in year four of a cycle that stopped being a surprise two years ago.
Contrarian
Everyone is watching the rating. The rating is the wrong variable.
The thing nobody is pricing is what happens the day after the guarantee terminates. Vendor and strategic backstops exist to carry an asset class to the point of institutional acceptance. Once acceptance is granted, the backstop is removed and the true cost of capital is revealed in public. If the rating was earned, the spread stays tight and nothing happens โ the most boring outcome, and the correct one. If the rating was rented, the spread reprices within two quarters and the equity absorbs the difference.
My prediction, on a twelve-to-twenty-four month window with a bear-market bias: at least one of these two companies receives a split rating โ investment grade from one agency, speculative from another. Split ratings are the fingerprint of a manufactured floor. They showed up in structured credit in 2007. They showed up in crypto lending in 2022, when one desk would rate a counterparty prime while another refused to extend it an unsecured line in the same quarter.
The second blind spot is worse. Crypto-native credit is about to be structurally outbid by paper that performs the same job with less transparency and more lawyers. If your thesis for on-chain credit is institutional inflows, that thesis has an expiration date and it is not the one on your calendar. Capital does not care about decentralization. Capital cares about recovery rates on default.
Takeaway
Watch three dates. The moment an agency officially comments on either filing โ silence has a cost, and right now the issuer is the one paying it. The next NVIDIA guarantee amendment, because that document will tell you more about real AI economics than any model launch this year. And the first coupon. Not the rating. The coupon.
If OpenAI and Anthropic get the rating early, expect the back half of 2026 to be a credit event for the entire alternative-asset complex, crypto included โ not because AI fails, but because AI succeeds at soaking up the fixed-income bid that was supposed to find its way on-chain.
And if they do not get it early, then SpaceX stays a sample size of one, and the real story is not about credit ratings at all. It is about how quickly an industry can build something enormous and still not be able to pay for it.
We do not price narratives. We price cash flows. Right now there is only one of those two things on the table.
