AI’s bottleneck isn’t power or chips, but funding

Original author: Long Yue

Source: Wall Street Journal

Money is becoming the most scarce chip in this high-stakes gamble on AI.

Over the past two years, global capital markets have firmly believed that computing power equals truth, but power shortages can be expanded, and chip supply can catch up. Recently, CoreWeave plummeted while tech giants’ CDS surged, suggesting the entire industry is beginning to recognize a harsh reality: the ultimate physical limit to AI’s rapid advancement is actually the availability of capital and the cost of financing.

When capital is no longer infinite and no longer cheap, AI has retreated from a technological myth back to the most basic arithmetic of return on investment (ROI).

Goldman Sachs' latest estimate shows that next year, the largest cloud giants will spend $1.2 trillion on capital expenditures, while their operating cash flow during the same period will be only $1.1 trillion, forcing at least 35% of construction spending to rely on bond issuance, exceeding $400 billion. Meanwhile, the yield on 10-year U.S. Treasuries recently surged above 5.3%, as the AI industry’s annual demand for hundreds of billions in bond financing directly competes with the U.S. government’s nearly $2 trillion fiscal deficit for limited funds in the long-term bond market.

Goldman Sachs warns that credit tightening and rising capital costs pose a risk of delays in computing infrastructure development, and the semiconductor supply chain may face a sharp drop in orders between 2027 and 2028, as demand fails to translate into actual purchases.

CoreWeave, the core asset in hashpower leasing and a barometer of AI sentiment, was the first to expose a crack in the funding chain.

On Wednesday, October 14, CoreWeave plunged over 15% in a single day, then fell nearly 6% after hours. Its $2.75 billion unsecured bonds, issued in April this year, are trading below 80 cents on the dollar, firmly in distressed asset territory, with yields to maturity surging past 15%. According to Dow Jones Market Data, its five-year credit default swap (CDS) spread has jumped to 881 basis points, implying a market-implied probability of default within five years exceeding 50%.

Even though CEO Michael Intrator issued an internal memo urgently emphasizing that the company had $104.2 billion in unfilled orders and over $25 billion in new customer commitments for the third quarter, the market chose to vote with its feet.

Orders are future expectations, but buying GPUs, renting data centers, and pulling high-voltage power lines require real money today. Gil Luria, Head of Technology Research at D.A. Davidson, bluntly stated:

They have lost the ability to borrow at a 9% interest rate. If CoreWeave borrows at 9% but can only generate a 5% return, no amount of scaling makes sense: they are destroying value with every step.

Panic quickly spilled over. On that day, Oracle fell more than 7%, Dell dropped over 10%, and data center stocks such as Applied Digital and IREN plunged nearly 10%. Goldman Sachs’ TMT sector index showed that over the past five days, the U.S. memory sector declined by 8.5%, the AI data center sector fell by 7.9%, the semiconductor sector dropped by 7.6%, while the S&P 500 index declined by 0.17%. Musk left a telling comment on social media: “Yikes.”

The head of the chip giant also no longer hides concerns about funding. At the DBS Bank summit in Singapore, Broadcom CEO Hock Tan and Intel CEO Patrick Gelsinger jointly labeled capital as the industry’s biggest bottleneck. Hock Tan stated publicly:

"I think the biggest constraint is capital." "Many investors don't even know what they're investing in."

Elon Musk then explicitly agreed on social media: "Don't cry at the casino." Intel CEO Lisa Su added:

If there is any challenge ahead, it is capital and the cost of capital; on the other hand, in the long term, people need to see ROI.

If only small and medium-sized AI companies were facing funding shortages, the storm would not yet shake the broader market. The core issue is that even the world’s largest and most profitable cloud giants have had their self-sustaining capabilities entirely consumed by AI.

Goldman Sachs' equity strategy team disclosed a set of figures in its latest flow report: global hyperscale cloud providers' total capital expenditures are projected to rise to $1.2 trillion by 2027, while the combined expected operating cash flows of these giants amount to only $1.1 trillion.

This is a rare cash flow inversion in the history of tech giants. Alphabet recorded its first-ever quarterly negative free cash flow 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.

The cash earned wasn't enough to cover expenses, and the remaining gap was filled entirely by debt.

Goldman Sachs’ credit strategy team estimates that at least 35% of next year’s capital expenditures by cloud giants will need to be funded through external bond issuance, meaning over $400 billion in massive corporate tech bonds will enter the public markets. Meanwhile, in equity financing, 45% of the approximately $430 billion in equity raises on U.S. markets this year flowed to AI-themed companies.

Goldman Sachs' research report highlights the fragility of this game of musical chairs:

As interest rates rise and the gap between debt costs and equity costs narrows, the path for companies to sustain valuation growth through bond issuance and buybacks will be severely constrained.

The greatest misfortune of borrowing to build computing power is encountering a rare post-war global sovereign debt inflation.

For the fiscal year ending September 30, 2026, the U.S. federal budget deficit rose to $1.993 trillion, with annual spending reaching $7.4 trillion, and net interest payments consuming $1.1 trillion in tax revenue.

To cover massive deficits, the Treasury is flooding the bond market with historically unprecedented supply. Recently, the 10-year U.S. Treasury yield surged above the 5.3% mark, the 2-year yield jumped over 7 basis points in a single day, and the auction yield for 30-year Treasuries hit the highest level since August 2000.

The AI bond rush has forced competition with the U.S. Treasury for the same pool of existing U.S. dollars in deep waters.

Goldman Sachs’ swap trading head, Ilir Xholi, noted that between September 22 and October 7, the dollar 10y10y forward swap rate surged by 47 basis points, while the 2-year rate, which reflects market expectations for the Fed’s short-term policy, remained unchanged. Xholi wrote in the report:

This movement has completely detached from the Fed's interest rate path; fundamentally, the market is demanding a very stringent risk premium compensation when buying long-term assets.

The massive bond issuances by AI hyperscale cloud providers and semiconductor companies in September directly competed with U.S. Treasuries in the long-duration market for capital.

Meanwhile, global capital is flowing back home. Japan’s 10-year government bond yield has surpassed 3%, and foreign institutions have significantly reduced their positions in U.S. and European bonds. The global liquidity pool continues to shrink, systematically raising the cost of capital.

The benchmark rate has risen to a high of 5.3%, while the cushion for downstream credit spreads has vanished entirely.

Goldman Sachs credit strategist Sara Grut warned that in today’s U.S. investment-grade corporate bonds, the credit spread as a share of the total nominal yield has 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 over the past few months, credit spreads and government bond yields have begun to move higher together, exposing credit assets to dual headwinds.

On the high-risk funding side, this capital allocation has become even more stringent. Rich Privorotsky, head of the Goldman Sachs trading desk in London, noted that the spread for CCC-rated junk bonds has surged to around 1,250 basis points, widening by over 200 basis points in just one month.

Credit derivative prices for AI core leaders have all triggered alerts:

• Broadcom's 5-year CDS rises to 136 basis points, hitting a record high;
• Oracle's 5-year CDS has surged past 261 basis points, hitting a record high, while its stock price continues to decline.
• CoreWeave's unsecured bond spread has broken through 1,000 basis points directly;
• The $5 billion leveraged loan offered by Norwegian data center operator Volta has seen its combined financing rate approach 11%.

Privorotsky noted in the customer report:

Attention must be paid to the spread of credit stress into B-rated and leveraged loan markets. If credit stress continues to expand, its damaging effects will extend from AI to small and medium-sized enterprises, physical consumption, and the entire macroeconomy.

JPMorgan Chase CEO Jamie Dimon previously made the same assessment: when the cost of capital remains high, the private sector must borrow to buy GPUs, and the public sector must borrow to cover deficits—capital allocation will no longer be driven by narratives, but rather ruthlessly rationed by price—namely, elevated interest rates.

Cash flow disruptions and high financing barriers are being transmitted to the end procurement chain.

ASML's Q3 earnings report, released on Wednesday, was a case in point: despite significantly exceeding expectations, raising the lower end of its 2026 guidance, and management emphasizing future capacity shortages, the stock still dropped as much as 3% after the report.

The core logic behind the market sell-off is that ASML did not raise its earnings guidance for 2027 and beyond. Goldman Sachs TMT experts noted:

The credit crunch has caused numerous AI infrastructure projects to be delayed, making this factor the stock market's biggest concern.

Privorotsky thus identified an "air pocket" in the semiconductor supply chain:

The real risk today is that high capital costs and real delivery bottlenecks are preventing end-demand from translating into actual capital expenditures.

If you are a supply chain intermediary caught between strong end-demand and the reality of inability to deliver, at some point you will face the risk of sudden order cancellations. The semiconductor supply chain is deeply concerned about this, particularly whether it could escalate into a crisis in 2027 or 2028.

Privorotsky concluded his report with the most accurate assessment of the current AI boom:

We still have strong faith in the underlying potential of this technology, but we must increasingly acknowledge with clarity: the ultimate bottleneck limiting this industrial revolution is no longer on the demand side, but rather whether we have the will and the ability to secure sufficiently cheap capital for these super-sized supplies.