CoreWeave's Capital Flywheel Comes to a Halt

Author: Liu Yangnan, Jiazi Guangnian

"I never want to do things that don't make money," said Brian Venturo, co-founder of CoreWeave, in an interview.

The three founders have also practiced this principle. Since the IPO in March 2025, the three founders of CoreWeave have collectively cashed out $3.3 billion, all making it onto Bloomberg's Global 500 Billionaires Index. Among this, $2.4 billion was completed in 2026, and the three still collectively hold about 20% of the company's shares.

As of the time of writing, CoreWeave's stock price has evaporated nearly two-thirds from its peak in the summer of 2025, and the founders' sell-off seems more like a timely stop-loss. Although a company spokesperson responded that these sell-offs were "pre-arranged, fully disclosed, and completely in line with standard market practices," the accelerated exit of core members inevitably shakes external confidence in the company's development given CoreWeave's current situation.

The market has already issued warnings. Recently, CoreWeave's unsecured bonds issued on June 18 of this year fell below 80 cents, becoming "junk bonds." According to the Financial Times, based on pricing in the credit default swap market, the implied probability of CoreWeave defaulting on its debt in the next five years exceeds 60%.

The shadow of a "double whammy" on stocks and bonds looms over what was once an AI star.

Fast forward three years, at that time, CoreWeave was at the throat of the AI computing power supply chain, representing a host of new cloud computing firms. Giants like Nvidia, Microsoft, Meta, and OpenAI were all active around CoreWeave, hoping to secure more GPU supplies.

No one expected that this collective enthusiasm that propelled CoreWeave to its peak would bring it crashing down in 2026.

1. The "Pandora's Box" of the East Coast

CoreWeave was established with the bold style typical of the cryptocurrency world.

Founded in 2017 as Atlantic Crypto, it transformed into cloud computing power and rebranded as CoreWeave during the crypto winter in 2019, raising only $1.2 million in its seed round that February; in 2021, it secured $50 million in convertible bonds from Magnetar, with a coupon rate of 10%, at that time being referred to as "the largest private HPC cloud in North America."

In 2022, CoreWeave continued to ramp up, investing about $100 million to purchase H100 GPUs despite the company's annual revenue being only $15.8 million, a massive procurement that even caught the attention of Jensen Huang.

However, Silicon Valley investors did not believe CoreWeave could compete with capital-rich cloud computing giants like Amazon, Microsoft, and Google, which were not only CoreWeave's main competitors but could also be significant customers, leading many investors to think they were crazy.

According to industry expectations at the time, expanding to hundreds of thousands or even millions of GPUs by 2024-2025 would require hundreds of billions of dollars. Given the pace of financing at that time, CoreWeave's equity financing speed and scale could not keep up with the pace of card purchases and capital expenditures for renting data centers, which would continuously dilute the equity of the founders and early shareholders.

Thus, the founders turned their eyes to the East Coast.

However, as the company was in a loss-making state and had no real estate or land as traditional collateral, the door to traditional corporate credit loans was essentially closed. The only significant asset on the balance sheet was the GPUs.

In August 2023, CoreWeave officially opened Pandora's Box.

They secured a $2.3 billion delayed draw term loan (DDTL), co-led by Magnetar and Blackstone Tactical Opportunities Fund, with participation from Coatue, DigitalBridge, BlackRock, PIMCO, and Carlyle.

According to estimates by Wing VC, based on a typical financing rate of about 70% for GPU collateral, this loan was backed by H100 servers worth over $3 billion, equivalent to about 70,000 to 100,000 GPUs.

After the transaction was disclosed, CoreWeave quickly became the biggest innovator at the intersection of AI and finance.

Jasvinder Khaira, Senior Managing Director at Blackstone, stated in an official press release that "CoreWeave is clearly a leader in driving innovation," and explicitly characterized this transaction as leveraging Blackstone's "scale and flexibility to provide innovative financial solutions to market leaders." Media attention also focused on the enormous wealth effect, calling it "one of the largest private financings in American corporate history."

This loan set a precedent and rapidly pushed GPUs from computing resources to another extreme of financialization.

Prior to this, GPUs had never been used as loan collateral. They were more often seen as depreciating consumables that were replaced every few years, with banks typically preferring stable assets like real estate.

But at this unique juncture in 2023, GPUs were layered with several halos.

First, they were sufficiently scarce; at that time, H100s were in high demand, and the potential of the second-hand market was considerable. Theoretically, even if orders defaulted, CoreWeave could sell the cards to recover funds.

Second, there was the trust backing of giants; most of CoreWeave's clients were top-tier companies like Microsoft and OpenAI, which had significant industry credibility, signing multi-year take-or-pay contracts. In other words, the collateral for CoreWeave was essentially the cash flow locked in those yet-to-be-delivered contracts.

Subsequently, other new cloud vendors with similar positioning to CoreWeave followed this model, including Lambda, Crusoe, and Fluidstack, albeit with varying amounts.

GPUs became a beautiful shell, encapsulating the entire East and West Coast's expectations for the ultimate outcome of AI. This intense expectation pushed CoreWeave to the altar and then brought it crashing down heavily.

2. Where Will the Wheel of Fate Turn?

At the beginning of 2023, the global market was still filled with the shock and excitement brought by ChatGPT, and the giants took the lead, with the wheel of CoreWeave's fate also beginning to turn.

Microsoft was the first giant to extend an olive branch to CoreWeave.

At that time, Microsoft and OpenAI were in a close relationship, and Azure was at a critical juncture of transforming from a traditional cloud vendor to an AI cloud, almost hesitant to show any doubt in major decisions. In 2023, Microsoft executives admitted in a conference call that Azure's GPU servers could not meet OpenAI's needs and required external procurement.

Thus, Microsoft signed a multi-year cloud infrastructure agreement with CoreWeave, potentially worth billions of dollars, becoming CoreWeave's largest customer. The company's disclosed prospectus showed that Microsoft ultimately accounted for 62% of revenue in 2024, serving as the cornerstone of the company's income. In 2025, CoreWeave still derived 67% of its revenue from Microsoft.

In April 2023, Nvidia doubled down, investing $100 million in the B round.

Perhaps with the backing of Microsoft and Nvidia, CoreWeave was able to smoothly secure a $2.3 billion loan from Magnetar and Blackstone, laying the groundwork for the subsequent arms race in production capacity.

In the era of GPU scarcity, obtaining a chip allocation from Nvidia was in itself a moat for the business model. The media even described CoreWeave as "Nvidia's favorite startup," inadvertently pushing CoreWeave to a high position.

In May 2024, Blackstone and Magnetar further increased their investment, jointly offering CoreWeave a $7.5 billion DDTL 2.0, setting one of the largest single records in the private credit market at that time. In the same month, Coatue led an $1.1 billion C round, raising CoreWeave's valuation to $19 billion.

In November 2024, The Information reported that Microsoft planned to invest about $10 billion in CoreWeave by the end of 2030. This was a number that had never been officially confirmed, but its timing coincided precisely with the window of CoreWeave's valuation surge.

In the same month, a $650 million secondary share transaction led by Jane Street, Magnetar, Fidelity, and Macquarie was completed, with Cisco also appearing on the buyer's list, pushing CoreWeave's valuation up to $23 billion.

Then came the good news of the IPO. Reuters cited insiders stating that CoreWeave aimed for a valuation of over $35 billion in its 2025 IPO, planning to raise over $3 billion.

On the eve of the official IPO, OpenAI added the final touch. A week after submitting the prospectus, CoreWeave announced a five-year contract worth $11.9 billion with OpenAI, while OpenAI secured $350 million in equity through the IPO private placement.

Although OpenAI was not the largest contributor to CoreWeave's revenue, it accounted for a significant proportion of the company's future long-term contract income. According to Forbes' estimates, in the next seven years, OpenAI's cooperative orders would account for about one-third of CoreWeave's total signed revenue, with a contract amount exceeding $22 billion.

In March 2025, the giants from the East and West Coasts embraced CoreWeave as it entered the IPO door, truly ascending to the altar.

However, the ticket to the altar was discounted. The IPO was originally planned to be issued at $47-$55, raising a maximum of $2.7 billion, but ultimately Nvidia subscribed to $250 million at a price of $40, anchoring the IPO price at $40 and reducing the issuance scale to 37.5 million shares. This was the largest tech IPO in the U.S. since 2021 and was an IPO that did not fall flat thanks to the support of giants.

Perhaps due to the fact that the loan category pioneered by CoreWeave was entirely in the private market, with no public quotes and the main debt ungraded, this unprecedented loan financing model had not attracted much public discussion.

But after the company went public, when this company that had always been in the shadows was placed under public scrutiny, the heavy halo above CoreWeave transformed into a target for short sellers.

The self-reinforcing capital flywheel beneath CoreWeave was built on a complex collective imagination, wrapped in aspirations for AI, worship of giants, and obsession with wealth. These emotions made this flywheel both sharp and fragile; when the weakest link showed cracks, it could transmit through layers until a complete collapse occurred.

Before that, CoreWeave must seize the last window of opportunity.

3. The Building is About to Collapse, Fixing the Fence After the Sheep are Lost

Doubts began to surface even before the IPO.

In March 2025, the Financial Times quoted an executive from an investment firm holding a tech short position, who believed that CoreWeave was essentially financing its capital-intensive hard asset business with relatively short-term notes... and using depreciating assets (Nvidia GPUs) as collateral. "This is not ideal."

However, the spotlight of the IPO overshadowed the scattered doubts. After going public, CoreWeave and investors still experienced a brief honeymoon period. The Q1 financial report showed that the company's revenue grew by 420% year-on-year, claiming to be "the world's first cloud vendor to complete GB200 commercial deployment."

Subsequently, in May 2025, CoreWeave issued $2 billion in unsecured bonds, with a coupon of 9.25%, maturing in 2030, and a Moody's rating of B1, four notches below investment grade. Despite the low rating, on the day of the bond issuance, CoreWeave's stock price still rose by 19%. At that time, investors were more focused on revenue growth and computing power deployment progress, seemingly insensitive to the credit risks reflected in the bond rating.

Until September 2025, short-selling firm Kerrisdale Capital released a report that systematically dismantled CoreWeave's financing cycle, stating that its core model involves using purchased GPUs as collateral to borrow money from non-bank institutions, which is then immediately used to buy more GPUs from NVIDIA, creating a continuous loop. The report directly characterizes CoreWeave as a debt-driven GPU rental business with no competitive moat.

On the day the report was released, CoreWeave coincidentally announced a new collaboration agreement with NVIDIA, which significantly mitigated the impact of the report on its stock price. However, a look at the financial statements reveals that CoreWeave's financial data had already raised red flags.

The company's total liabilities were still $5.1 billion at the end of 2023, expected to rise to $46 billion by the end of 2025, and reach $72 billion by the end of June 2026. Among these, interest-bearing debt amounts to approximately $35.6 billion, while the bulk consists of operating leases for rented data center facilities and electricity.

A more direct signal comes from liquidity. As of the end of June 2026, the company had $5.5 billion in cash, but had $20.9 billion in current liabilities due within a year, resulting in a cash coverage ratio of less than 30%. Meanwhile, the company's capital expenditures have not slowed down. In 2026, the full-year guidance was raised to $35-39 billion after Q2, up from the initial guidance of $30-35 billion at the beginning of the year. Regardless of the metric used, this is at least more than double the $14.9 billion expected for 2025.

As revenue doubles, so do interest and losses, and the gap can only be filled with more money.

A significant portion of this financing comes from money borrowed by CoreWeave from investors. Since its IPO, CoreWeave has issued five unsecured bonds and three convertible bonds, totaling over $10 billion.

Unsecured bonds are those that are not backed by any specific assets. In the event of a default, holders can only queue with other unsecured creditors for repayment and cannot prioritize taking GPUs or data centers. Due to the high risk, these unsecured bonds all carry a coupon rate of over 8.5%. Meanwhile, the average yield in the U.S. high-yield bond market is about 7%, and U.S. Treasury bonds are around 4%. This means that the market requires CoreWeave to pay a borrowing cost higher than that of ordinary "junk bonds."

CoreWeave is not entirely unaware of its situation.

As early as July 2025, CoreWeave took some "remedial" actions. They announced an all-stock acquisition of Core Scientific for $9 billion, clearly stating in the official announcement that they aimed to consolidate data center ownership, gain access to Core Scientific's total power of 1.3GW (with 840MW already leased to CoreWeave and over 1GW of expansion potential), and eliminate $10 billion in future rental obligations, saving about $500 million annually.

However, this transaction was conducted entirely in stock rather than cash, as CoreWeave had no capacity for cash acquisitions at that time. On October 30, 2025, Core Scientific shareholders rejected the deal, citing undervaluation, and the acquisition was terminated.

CoreWeave seems to be on a train that is about to derail; they want to pull it back on track and convert the rapidly circulating cash on their books into more stable assets. But once the flywheel slows down, the looming debt maturity date awaits them, and they can only continue to accelerate to maintain a temporary balance.

In the first half of 2026, a significant amount of funds continued to flow into CoreWeave:

In January, NVIDIA invested another $2 billion; in March, an $8.5 billion DDTL 4.0 secured the first investment-grade GPU collateral rating (Moody's A3), causing the stock price to surge 12% on the announcement day; in April, Jane Street invested $1 billion, along with a $6 billion computing power contract, while Meta expanded its collaboration to $21 billion that same month; in June, CoreWeave was included in the Nasdaq 100 index; in August, the second-quarter report exceeded expectations, with backlogged orders reaching $104.2 billion, prompting an upward revision of the full-year revenue guidance.

However, CoreWeave's financing costs have also been climbing. In March of this year, they could still secure financing at SOFR+2.25%, which changed to SOFR+4.50% in May, and rose to SOFR+5.50% in August, more than doubling the spread within six months. The unsecured bond with a coupon rate of 9.625% issued on June 18 was CoreWeave's last instance of financing from the public bond market at face value and high interest.

By September, the bonds issued by CoreWeave were convertible bonds, also unsecured and at face value, but with a coupon rate of only 2.875%. This essentially represents a debt resolution strategy through equity dilution, but different people may interpret it differently—short sellers might see it as another link in a "Ponzi scheme," while bulls might view it as "exchanging time for space" for self-rescue.

However, looking beyond CoreWeave, the entire data center industry is experiencing a shaking opportunity window, undergoing a collective collapse of trust.

Credit risk data platform Credit Benchmark indicates that over the past six months, the credit conditions in the data center industry have deteriorated by 10% to 12%; Gartner's latest report predicts that by 2030, over 80% of Neocloud service providers will be forced to cease independent operations and exit the mainstream market, warning that expected AI revenues will not support all planned infrastructure expenditures to achieve profitable returns.

As internal and external troubles converge, CoreWeave must exert more effort to extricate itself from this round of "dual kill" of stocks and bonds.

4. Will CoreWeave Have a Bright Future?

Looking back at CoreWeave's rise and fall, it is evident that this is a journey that is hard to replicate. From the common development paths of AI companies, CoreWeave resembles an "anomaly" buoyed by favorable timing, location, and human factors.

Part of the momentum comes from the historical inertia of the founding team.

The three founders were all hedge fund traders before entering the AI space. Intrator founded the natural gas hedge fund Hudson Ridge, while Venturo and McBee have spent many years in energy trading.

From a natural gas hedge fund to Ethereum mining, and then to GPU cloud, CoreWeave's three transformations are essentially driven by the same intuition from traders: find an asset with limited supply and distorted pricing, use maximum leverage to build a position ahead of everyone else, and bet that the demand curve is steeper than everyone imagines.

But this time, they encountered a market of entirely different scale that cannot be liquidated at will. The depreciation cycle of GPUs is measured in years, data center leases are measured in decades, while debt interest is calculated quarterly— the original capital turnover cycle of the data center industry does not align with CoreWeave's light asset, high turnover model. However, at the 2023 juncture, the collective frenzy of the initial wave of technological enthusiasm and endorsements from giants left a lot of hot money eager to find a place. CoreWeave rapidly matured the financialization of GPUs during this window.

A user on X has categorized the financialization process of different assets into five stages:

• Company credit financing: Banks lend based on the overall credit of the company. Assets have not yet become independent financing units.
• Asset-backed financing: Assets themselves become collateral, but transactions remain bilateral and highly customized.
• Securitization and structuring: Asset cash flows are placed into SPVs and split into different tranches sold to capital market investors. Private credit and insurance capital begin to participate.
• Standardized price discovery: Spot markets and accepted benchmarks emerge, giving assets market-usable prices.
• Derivatives and risk transfer: Futures, insurance, and CDS allow risks to be separated from asset ownership. At this point, the asset class ultimately matures.

The user concluded: "It took about half a century for oil to transition from long-term contracts to WTI futures and a complete forward curve in 1983. Aircraft financing took about 40 years to evolve from bank loans and leasing companies to EETC securitization, professional appraisers, and residual value insurance. Computing power completed the first three stages in just three years. The fourth stage is beginning to take shape, and the fifth stage's products are already formed:

1. August 2023: CoreWeave completed its first H100 collateralized loan, amounting to $2.3 billion.
2. May 2024: Blackstone and Magnetar entered through a $7.5 billion private credit facility.
3. October 2025: Meta and Blue Owl completed Hyperion, the largest private credit transaction in history, amounting to $27 billion.
4. March 2026: CoreWeave DDTL 4.0 received the first investment-grade rating for GPU-backed financing, rated A3.
5. May 2026: CME and ICE announced GPU futures in the same week, testing the transition from the fourth to the fifth stage.

The first half of the financialization of computing power was packaged and expanded using the credit of large-scale cloud vendors. Large financial institutions have already built this part. As financialization progresses to the latter two stages, it essentially shifts the focus from credit to the assets themselves. The consumable nature of GPUs makes it difficult to obtain a full credit rating, and the weak links in this model naturally show cracks.

Fortunately, CoreWeave's founding team is also actively working to resolve the crisis while cashing out.

During the Q2 earnings call, management clearly identified "self-builds" and "deeper vertical integration" as core strategies, stating that they would bring stronger operational control and higher long-term profit margins.

Specific actions include advancing a 180MW self-built data center in Hammond, Indiana, and converting a laboratory/manufacturing building in New Jersey into a data center. The company's goal is to bring most of the contracted power capacity online through self-build methods, rather than relying entirely on long-term leases from data center partners.

CoreWeave is also continuously replacing debt costs to minimize interest expenses.

At the same time, co-founder Brannin McBee is leading efforts to actively market its GPU collateralized loans and cash flow-supported debt to insurance companies, aiming to package these assets into products that insurance companies can buy.

Insurance capital is large, long-term, and relatively patient. If CoreWeave can persuade insurance companies to buy its debt under investment-grade or near-investment-grade conditions, it can replace the high-yield debt market's short-term funds, which can quickly become expensive once market sentiment cools.

Whether CoreWeave can successfully navigate these challenges, its emergence has been like a giant stone that accidentally fell, bringing a huge freshness to the AI field, which has long been dominated by technology and a relatively singular discourse system.

The AI community has never lacked prophecies about the endgame. To this day, the development of AI evokes both anticipation and unease. Ryan Hoffman, head of Credit Benchmark's analytical products, stated: "No one can truly predict the future direction. We are in a highly uncertain period, and even the situation in the next three months cannot be predicted."

CoreWeave's current stock and bond crisis serves as a warning to everyone—whether capital's role in emerging technology industries is positive is a key variable. The flow of capital must resonate with the rhythm of the industry itself; otherwise, it could be a case of drinking poison to quench thirst.

For CoreWeave, the main task in the next three months is clear: to quickly digest the debt left by this round of capital frenzy before its financial situation spirals out of control and to build a healthier closed loop.

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.