Morgan Stanley Rises to Top of Wall Street's AI Debt Deals
Claire Weston
Morgan Stanley rode a wave of AI-infrastructure debt innovation to $2.3 billion in capital-markets fees in H1 — up 64% year-on-year, vaulting to global No. 2 — a sign that AI financing is reshaping the investment-bank league table.
How did Morgan Stanley jump to second place?
LSEG data show the bank's H1 debt and equity capital-markets fees hit $2.3 billion, up from $1.4 billion a year earlier — a roughly 64% increase.
That lifted its global ranking from fourth to second, behind only JPMorgan and ahead of Goldman Sachs.
This means → the leap was not driven by traditional business growth but by an entirely new revenue stream: AI-infrastructure financing.
How big are these deals?
A $3.2 billion Google-backed bond for data-center developer TeraWulf; a $27 billion debt package for Meta and Blue Owl's Hyperion data-center project; a $35 billion chip-financing advisory for Broadcom.
Co-head of investment banking Mo Assomull noted a step-change in scale: deals have moved from "$1 billion, $2 billion, $5 billion" to "$10 billion, $20 billion and beyond."
In plain terms = single transactions have gone from "large" to "mega" — traditional financing frameworks can no longer contain them.
What exactly is a "construction bond"?
Leverage-finance co-head William Graham designed the TeraWulf bond structure: it blends the liquidity of public-market bonds with the covenant protections of project loans, wrapped in a credit backstop from Google.
In plain terms = TeraWulf borrows the money, but Google's balance sheet stands behind it — investors see lower risk, so the cost of capital drops. TeraWulf raised $3.2 billion at a 7.75% yield.
A "lockbox" mechanism — lease payments routed directly to bondholders, plus additional collateral — pulled insurers, asset managers, and pension funds into data-center financing for the first time.
Can GPUs serve as loan collateral?
In May, Morgan Stanley and MUFG arranged a $3.1 billion loan for cloud provider CoreWeave to buy and install Nvidia GPUs.
It was the first GPU financing done as a broadly syndicated term loan — GPU financing meaning the chips themselves serve as the underlying asset — and drew nearly $20 billion in investor demand.
This means → the collateral has shifted from "the data-center building" down to "the chips inside it," pushing capital one layer deeper into the hardware stack.
Why don't all banks want this business?
Some banking executives have said explicitly they do not want to be the top player in data-center financing — partly because data centers are drawing community opposition across the U.S.
JPMorgan CFO Jeremy Barnum said this week the bank reviewed certain data-center financing terms and walked away.
This reflects a tension: AI-infrastructure financing is ballooning in scale, but the risk boundary remains blurry. Morgan Stanley itself projects AI buildout will consume $10 trillion in capital over the coming years — who bears the tail risk is the question this nascent market must answer.
Content is for reference only, not financial advice.