Money is becoming the scarcest chip in the high-stakes gamble that is AI.
Over the past two years, global capital markets firmly believed that compute is truth, but power shortages can be solved by expansion, and chip supply can eventually catch up. Recently, with CoreWeave plunging and tech giants' CDS surging, the entire industry chain seems to be waking up to a brutal fact: the ultimate physical limit of AI's breakneck advance is actually the supply of capital and the cost of financing.
When capital is no longer unlimited and no longer cheap, AI retreats from a technological myth back to the most basic arithmetic problem of return on investment (ROI).
Goldman Sachs' latest estimates show that hyperscale cloud giants' capital expenditure next year will reach $1.2 trillion, while their operating cash flow over the same period will be only $1.1 trillion. At least 35% of construction spending will be forced to rely on debt issuance, exceeding $400 billion. At the same time, the 10-year U.S. Treasury yield recently broke through 5.3%. The AI industry's hundreds of billions of dollars in annual bond issuance demand is directly competing with the United States' nearly $2 trillion fiscal deficit for limited funds in the long end of the bond market.
Goldman Sachs warns that credit tightening and rising capital costs create risks of delays in compute infrastructure construction, and the semiconductor supply chain may face a demand cliff in 2027 to 2028, where demand cannot be converted into actual purchases.
The trigger is lit: The math of borrowing money to build compute no longer adds up
CoreWeave, a core computing power leasing target and the "wind vane of AI sentiment," was the first to tear open the rift in the capital chain.
On Wednesday, October 14, CoreWeave plunged more than 15% in a single day and fell nearly another 6% after hours. Its $2.75 billion unsecured bonds issued in April this year traded below 80 cents on the dollar, deep in distressed territory, with yields to maturity surging past 15%. Dow Jones market data showed its five-year credit default swap (CDS) jumped to 881 basis points, and the market-implied probability of default within five years has exceeded 50%.
Even though CEO Michael Intrator urgently issued an internal memo emphasizing that the company holds $104.2 billion in undelivered orders and more than $25 billion in new customer commitments in the third quarter, the market chose to vote with its feet.
Orders are expectations for the future, but buying GPUs, renting data centers, and stringing high-voltage power grids require real money today. Gil Luria, head of technology research at D.A. Davidson, put it bluntly:
"They have lost the ability to borrow at a 9% interest rate. If CoreWeave borrows at a 9% cost and can only earn a 5% return, then no matter how large the scale expands, it is meaningless: they are destroying value with every step."
Panic spread rapidly. That day, Oracle fell more than 7%, Dell fell more than 10%, and data center names such as Applied Digital and IREN plunged nearly 10%. Goldman Sachs' TMT index showed that over the past five days, U.S. memory stocks fell 8.5%, AI data center stocks fell 7.9%, and semiconductor stocks fell 7.6%, while the S&P 500 fell 0.17%. Musk left a meaningful sigh on social media: "yikes."
The heads of chip giants also no longer hid their concerns about the capital chain. Broadcom CEO Hock Tan and Intel CEO Lip-Bu Tan appeared together at a DBS Bank summit in Singapore and directly labeled capital as the industry's biggest bottleneck. Hock Tan publicly said:
"I think the biggest constraint is capital." "Many investors simply don't know what they are investing in."
Musk then bluntly agreed on social media: "If you're in a casino, don't cry." Intel CEO Lip-Bu Tan added:
"If there is any challenge ahead, it is capital and the cost of capital; on the other hand, in the long run, people need to see ROI."
Cash flow inversion: Even the most profitable giants need to borrow $400 billion
If only small and medium-sized AI companies lacked money, the storm would not be enough to shake the broader market. The core problem is that even the world's most cash-generating hyperscale cloud giants have had their self-funding capacity devoured by AI.
Goldman Sachs' equity strategy team disclosed a set of figures in its latest flow report: total global hyperscaler capital expenditure will climb to $1.2 trillion in 2027, while these giants' expected operating cash flow will total only $1.1 trillion.
This is a rare cash flow inversion in the history of tech giants. Alphabet posted negative free cash flow for a single quarter for the first time in company history in the second quarter of this year, with a net outflow of $5.9 billion; Amazon's free cash flow turned negative by $7.6 billion over the past 12 months.
Cash earned is not enough to burn, and the remaining gap is entirely filled by debt.
Goldman Sachs credit strategists estimate that at least 35% of cloud giants' capital expenditure next year will need to be solved through external bond issuance, meaning the public market will see more than $400 billion in massive tech company debt. On the equity financing side, of the roughly $430 billion in U.S. stock follow-on offerings this year, 45% flowed to AI-related companies.
The Goldman Sachs research report pointed out the fragility of this pass-the-parcel game:
"As long as rates rise and the gap between debt costs and equity costs narrows, the path for companies to sustain valuation expansion through bond issuance and buybacks will be severely suppressed."
Crushed by massive Treasury debt: The 10-year U.S. Treasury breaking 5.3% acts as a liquidity siphon
The greatest misfortune of borrowing to build compute is running into the rare postwar inflation in sovereign debt globally.
For fiscal year 2026 through the end of September, the U.S. federal fiscal deficit climbed to $1.993 trillion, with annual spending as high as $7.4 trillion, and net interest payments alone consumed $1.1 trillion of tax revenue.
To pay for the massive deficit, the Treasury Department is pressing historically huge supply onto the Treasury market. Recently, the 10-year U.S. Treasury yield once again broke through the 5.3% mark, the two-year U.S. Treasury yield jumped more than 7 basis points in a single day, and the winning yield at the 30-year U.S. Treasury auction set the highest record since August 2000.
AI's bond issuance wave is being forced to compete with the U.S. Treasury Department for the same stock of dollars in deep waters.
Ilir Xholi, head of swap trading at Goldman Sachs, pointed out that between September 22 and October 7, the USD 10y10y forward swap rate surged abnormally by 47 basis points, while the 2-year rate representing the Fed's short-term policy expectations did not move at all. Xholi wrote in a report:
"This move has completely detached from the Fed's rate path. In essence, the market is demanding extremely strict risk premium compensation when buying long-term assets."
"The massive September bond issuance by AI hyperscalers and semiconductor companies directly engaged in hand-to-hand combat with U.S. Treasuries for money in the long-duration market."
At the same time, global capital is still flowing back home. Japan's 10-year government bond yield broke 3%, and overseas institutions sharply reduced their long positions in U.S. and European bonds. The global capital pool's water level continues to shrink, and the cost of capital is being systematically pushed higher.
Credit buffer falls to zero: The entire industry faces hard capital rationing
With benchmark rates climbing to a high of 5.3%, the protective cushion of credit spreads downstream has completely vanished.
Goldman Sachs credit strategist Sara Grut warned that in current USD investment-grade corporate bonds, credit spreads as a share of total nominal yields have fallen below 15%, far below the historical average of 45%:
"In the past, when government bond yields rose, credit spreads often narrowed to absorb the shock. But in recent months, credit spreads and Treasury yields have begun rising in the same direction, and credit assets face dual headwinds."
At the high-risk financing end, this capital rationing is even harsher. Rich Privorotsky, head of Goldman Sachs' London trading desk, pointed out that CCC-rated spreads in the junk bond market have jumped to around 1,250 basis points, widening by more than 200 basis points in a month.
Credit derivative prices for AI's core leaders are flashing warnings across the board:
Broadcom's five-year CDS rose to 136 basis points, a record high;
Oracle's five-year CDS broke through 261 basis points, a record high, while its stock price kept falling.
CoreWeave's unsecured bond spreads directly pierced 1,000 basis points;
Norwegian data center operator Volta's $5 billion leveraged loan has a blended financing rate approaching 11%.
Privorotsky noted in a client report:
"Close attention should be paid to the spread of credit stress into single-B and leveraged loan markets. If credit pressure continues to spread, its destructive power will extend from AI to small and medium-sized enterprises, real consumer spending, and the entire macroeconomy."
JPMorgan CEO Jamie Dimon previously offered the same assessment: when capital costs remain high, the private sector wants to borrow to buy GPUs, and the public sector wants to borrow to cover deficits. Capital allocation will no longer tell stories, but will be brutally rationed by price—that is, by elevated interest rates.
Goldman Sachs warns: Semiconductor supply chain may face order gap risk in 2027-2028
The cutoff of funding and high financing thresholds are transmitting into the end-purchase chain.
ASML's third-quarter earnings released on Wednesday are a microcosm: although results greatly beat expectations and the lower end of its 2026 guidance was raised, and management emphasized that future capacity would be in short supply, the stock still fell as much as 3% after the earnings release.
The core logic behind the market selloff is that ASML did not raise its performance outlook for 2027 and beyond. Goldman Sachs TMT experts pointed out:
"Credit tightening has caused many AI infrastructure projects to delay delivery, and this variable has become the stock market's biggest worry."
Based on this, Privorotsky identified an "air pocket" in the semiconductor supply chain:
"The real risk now is that high capital costs and real delivery bottlenecks are preventing end demand from converting into actual capital expenditure.
If you are an intermediary in the supply chain caught between strong end demand and the reality of being unable to deliver, at some point you will face the risk of a sudden order gap. The semiconductor supply chain is deeply worried about this, especially whether it could turn into a crisis in 2027 or 2028."
At the end of his report, Privorotsky wrote what may be the most precise framing of the current AI boom:
"We remain optimistic about the prospects of this technology at the fundamental level, but we must increasingly recognize clearly: the ultimate bottleneck limiting this industrial revolution is no longer on the demand side. It lies in whether we are willing and able to raise enough cheap money for these super supplies."
Comments