CoreWeave Crash, AI Trade Cracks, Anthropic IPO Red Flags, Bill Gates Doomer

The AI infrastructure sector’s explosive demand, driven by GPU-hungry workloads, has fostered new financing models but also exposed companies like Coreweave to heightened debt and refinancing risks amid a volatile economic environment.

Financial Structure of Gpu Lending and "Box" Model

Coreweave’s business model hinges on a novel financial structure called the “box,” as explained by CEO Michael Intrator. When a customer such as Microsoft approaches Coreweave for compute resources, they sign a contract, which, along with the GPUs purchased (typically from Nvidia), related data center capacity, power contracts, and financing agreements, all get bundled into a discrete financial unit—the box. The box operates as a closed cash-flow “waterfall,” where customer payments first service data center operating costs, utilities, then debt interest and principal, with residuals flowing to Coreweave. Critically, revenue and liabilities stay isolated within each box, preventing financial contagion across the company’s broader operations.

Chamath Palihapitiya notes this OPCO/PropCo approach shields Coreweave from company-wide bankruptcy risk—if a single project falters, only that asset goes into receivership, and creditors can reclaim GPUs or other box-specific assets. Lenders’ risk exposure thus hinges more on the customer’s creditworthiness (for instance, Microsoft’s five-year commitment) than on Coreweave’s.

A key draw for financiers is the rapid payback: Intrator highlights that, by structuring deals for five years, financing for any individual box is typically retired within 2.5 years, a risk profile more conservative than that of even SpaceX’s one-year payback model. These short payback windows incentivize lenders, driving initial borrowing costs lower and reinforcing the self-contained, noncontagious nature of each box.

Rising Borrowing Costs and Margin Compression

Coreweave’s economics are becoming stressed by macroeconomic forces. As Chamath explains, the Federal Reserve’s hikes of the risk-free rate raise the borrowing cost for everyone, including Coreweave. Unsecured debt that once averaged about 8.5% now faces GPU-specific loan rates spiking up to 15%, as observed recently on bond markets. This erosion of margin is especially problematic given the company’s need to refinance roughly half of its total contracted expansion—about $35 billion of outstanding unfinanced commitments on top of its existing $35 billion debt—for future capacity buildout.

Furthermore, while Coreweave’s business benefits from fixed-price contracts (which guarantee long-term revenue with hyperscale “investment grade” customers, representing 75% of revenue), it prevents the company from capitalizing on historically high spot rates for compute. As Jason Calacanis and Chamath note, the spread between fixed contract prices and spot market compute can represent upwards of $20–30 billion in opportunity cost, causing a revenue/cost mismatch if borrowing rates climb but contract revenues do not.

This puts Coreweave near a precarious breakeven: According to Chamath, if new financing is only attainable at double-digit interest rates, Coreweave could be just $5 billion from violating debt covenants—exposing them to the risk of being unable to service all obligations.

Credit Market Signals and Default Risk Perception

Such acute financial stress manifests in market sentiment. Trading in Coreweave’s debt credit default swaps (CDS)—essentially an insurance premium against default—has spiked, now pricing at levels that imply a 16% annual and roughly 55% five-year default probability, a sign that the smartest market players view Coreweave’s situation as a toss-up between solvency and default over the medium term.

This anxiety feeds into the bond market, where Coreweave’s bond prices have dropped to about 80 cents on the dollar, raising implied coupon yields from the contractual 9.6% to nearly 15% for new debt. The selloff has also triggered a 20% drawdown in valuations for the “neo cloud” companies like Coreweave and peers. Adding to the challenge ...

Anthropic’s upcoming IPO is drawing intense scrutiny due to its extraordinary revenue growth, the company’s dominant market position alongside OpenAI, and questions about how much capital it should raise to ensure long-term resilience in a rapidly evolving AI landscape.

Exceptional Revenue Growth and Market Position

Anthropic reported $16.8 billion in quarterly revenue, translating to over $5 billion per month. Notably, its net revenue retention was an unprecedented 900%, indicating current customers are spending nine times more than they did last year. The company achieves positive free cash flow, with an 85% adjusted gross margin and an 11% adjusted operating margin, highlighting strong economic fundamentals and robust customer demand for its frontier AI models.

Anthropic, together with OpenAI, currently dominates the foundation model market. According to semi analysis discussed by David Sacks, Anthropic and OpenAI are set to take up 60% of the new global compute capacity brought online over the next year. Sacks emphasizes this as a classic “compute duopoly,” where both control the lion’s share of hardware and software infrastructure. In terms of revenue and market share, they are essentially the only two foundation model companies producing significant revenue, building substantial entry barriers and competitive moats within the industry. While new models from Meta, Google, and others are improving, Anthropic and OpenAI remain out in front on both capability and business traction.

Proposed Valuation and Capital Structure Concerns

There is vigorous debate among market participants about Anthropic’s appropriate IPO valuation and the optimal amount of capital to raise. Some propose a $2 trillion valuation with a $50–100 billion IPO raise, while others suggest a more resilient approach: raising $200–250 billion at a lower $1 trillion valuation to maximize long-term security.

Chamath Palihapitiya argues that raising $200 to $250 billion is necessary for Anthropic to secure its own compute infrastructure, insulate against future market corrections, and avoid down rounds should the spot market for compute become volatile. He reasons that a large war chest will allow Anthropic to finance five to ten gigawatts of computing power, thus supporting ambitious growth for the next five to ten years and providing a buffer against risks from supply chain disruptions, competitive pressure, or broader market downturns. Palihapitiya prefers prioritizing financial insulation over optimizing headline valuation, believing this reduces risk concentration and ensures operational continuity.

Echoing this logic, Ben Horowitz notes that running out of cash is one of the few existential risks for any technology company, particularly one investing heavily in infrastructure and R&D. He points out that founder equity stakes are small and the group is reportedly not driven by personal financial optimization, but by building a sustainable business with enduring competitive advantages. Consequently, securing compute capacity through debt-financed data centers—rather than exposing the company to unfavorable market returns or raising at a down roun ...

A group of technology investors and commentators debate Anthropic’s public safety warnings, regulatory behavior, and technical choices. They argue that Anthropic exaggerates threats, especially around AI and IoT, as both market strategy and regulatory maneuvering, while introducing risks of misalignment through its technical approaches.

Iot Worms and Distraction From Threats

Anthropic Warns Rogue Ai Could Self-Replicate Across Iot Devices, Posing an Existential Threat

Anthropic recently published a blog warning that rogue small language models could self-replicate as worms across IoT (Internet of Things) devices, such as smart TVs, doorbell cameras, and robot vacuums. The company claims these AI-powered worms could rapidly spread, setting up inference engines on compromised devices and forming physical swarms that would be hard to disable, raising the specter of a looming existential threat just “months away.”

No Self-Replicating Ai Worms Exist; Traditional Worms Infect Millions of Iot Devices Through Vulnerabilities Like Default Credentials and Open Ports, Making the Ai Threat Seem Redundant and Dramatic

David Sacks challenges Anthropic’s claims, stating that no AI-powered self-replicating worms currently exist. He notes that traditional worms have already infected millions of IoT devices using classic hacking techniques, such as exploiting default passwords and open ports. For example, the Mirai botnet infected numerous devices with default credentials in 2016, and the more recent Kimwolf worm exploited two million Android TV boxes via open ADB ports. Jason Calacanis adds that these vulnerabilities persist because users rarely change device passwords from insecure defaults. Sacks and Calacanis argue that Anthropic’s AI worm scenario is redundant, as existing malware already demonstrates the plausibility and scale of such attacks without AI.

Constraints Like Slow Token Generation and Embedding Models in Payloads Make Ai-powered Worms Less Effective Than Conventional Malware

Calacanis and Sacks further point out technical limitations: embedding an AI model in the worm drastically slows its propagation due to the slow speed of model inference (e.g., generating less than one token per second), making AI-powered worms less practical than classic malware in real-world use. They argue that making these dramatic AI worm claims is a sensationalist distraction from actual threats.

Pattern of Regulatory Advocacy and Competitor Targeting

Anthropic Reported "Industrial-Scale Distillation Attacks," Sought Government Restrictions on Open-Weight Models, and Announced Technical Defenses Achieving the Same Protective Effect

The panel notes that Anthropic has a recurring pattern of escalating technical or security issues to government authorities in pursuit of regulatory intervention. Previously, Anthropic warned of “industrial-scale distillation attacks,” in which competitors use Anthropic’s own models to train their own, allegedly stealing intellectual property. Anthropic lobbied for government restrictions on open-weight models, claiming that such models were a conduit for IP theft. However, the company later announced it had implemented technical measures that reduced distillation attacks by 90-99%, the very approach critics recommended instead of legislative action.

Iot Worm Blog Timing Suggests Coinbase's Model Beats Anthropic, Revealing Market Competition Fears

The timing of the IoT worm blog raised further suspicion. David Sacks points out that just before Anthropic’s worm blog, Coinbase revealed it had fine-tuned a small open-weight model for fraud detection, which was faster and more cost-effective than Anthropic’s flagship Opus model. This raised concerns that Anthropic’s safety warnings intentionally coincided with a competitor’s technical success, highlighting Anthropic’s fear of losing market share to more efficient, practical, and accessible open models.

Using Technical Challenges and Government Intervention to Create Regulatory Moats Instead Of Genuine Safety Advocacy

According to Sacks and Ben Horowitz, Anthropic’s advocacy pushes for legislative and regulatory action based on dramatic scenarios rather than technical necessity. Critics characterize this as regulatory capture—a strategy to build moats around Anthropic’s offerings by making open competition harder under the guise of safety. They note that while Anthropic dramatizes theoretical scenarios for regulators and the press, the company itself benefits from using the same internet-wide data that it claims to protect. In practice, when confronted with real-world security concerns in their own products (e.g., revealing exploits with Opus), Anthropic responds by blaming prompts rather than acknowledging product risks. The group believes this two-faced behavior is motivated by competitive positioning, not pure public safety concerns.

...

AI markets are undergoing rapid transformation, marked by price compression, commodified capabilities, and a marked shift toward smaller, specialized models. These developments are fundamentally altering competitive dynamics, job prospects, and the basis for durable business moats.

The Commodification of Ai Capabilities and Price Compression

Jason Calacanis highlights the ongoing commodification of AI-driven software, with prices plummeting as AI assistants make it easier and cheaper to create rival products. He cites examples such as $5 platform games ("hood warfare") and open-source text-to-speech (TTS) solutions now undercutting commercial products. Elon Musk’s orchestration approaches also now route users to the cheapest or most effective models, whether open source or proprietary.

This intense competition erodes traditional margins; software that once stood as a franchise can be quickly decompiled, replicated, and re-released by competitors. Calacanis and David Sacks agree that the pace and magnitude of price drops are unprecedented, with code now commodified and copied at a velocity never seen before. Ben Horowitz adds that the result is sharply lower prices—deflationary pressure for software—although the quantity of software being produced is soaring.

The deflationary dynamic redefines the economics of software production. It dramatically lowers barriers previously shaped by technical complexity, since AI assistants do most of the coding, and makes it difficult to build franchises for software not anchored by unique data or substantial specialization.

The Rise of Small Models and Efficient Specialization

Within enterprises, the preference is shifting toward smaller language models that are fine-tuned for specific purposes, rather than deploying costly and inefficient frontier models. David Sacks gives the example of Coinbase, which used a fine-tuned model to cut fraudulent transactions by 30%, outperforming Anthropic's best model, Opus 4.5, in both cost and accuracy.

These small models are highly efficient, using a fraction of the compute and energy required by larger counterparts, and they are excelling at specialized tasks. As compute constraints tighten, the economics favors these proprietary, fine-tuned solutions, which allow companies to capture customer value, reduce infrastructure costs, and maintain performance in resource-constrained environments. The shift is toward every workload having its own small, customized model.

Expansion of Software Creation and Counterintuitive Job Growth

Despite fears of automation, demand for developers is actually rising, even as coding assistants write up to 95% of the code. The remaining 5%—which still requires human expertise—represents a larger absolute amount of work as the total volume of software explodes. There is a need for technically skilled professionals to oversee, integrate, and maintain the growing ecosystem of AI-generated software.

The proliferation of coding assistants and agents enables a new category of bespoke software for niche applications that were previously uneconomical. The software market is now expanding beyond mass-market solutions to highly tailored, use-case-specific applications, resulting in ...

The explosive growth in artificial intelligence (AI) infrastructure is marked by fundamental differences from previous technology bubbles, particularly in terms of demand, supply constraints, and the critical metrics for business health. Industry leaders such as Ben Horowitz, David Sacks, Jason Calacanis, and Chamath Palihapitiya highlight severe shortages in compute resources, persistent demand, and mounting energy bottlenecks, all of which underpin the market’s current dynamics and help clarify the risk profile around an "AI bubble."

Demand Remains Robust While Supply Remains Constrained

AI’s current wave is driven by demand that continues to accelerate, while the computing power needed remains scarce. David Sacks notes that both major model companies and startups report being compute constrained, not demand limited. Startups now approach their investors for help acquiring compute rather than capital, and some venture firms have started securing compute resources themselves to gain a competitive edge. Jason Calacanis highlights how some accelerators use access to excess compute to attract founders—evidence that demand is driving new forms of strategic differentiation.

Customers Agree: Compute Constraints Trump Demand Constraints

Industry consensus is clear: compute, not customer demand, is the binding constraint. Companies large and small struggle to obtain the graphical processing units (GPUs) required to train and deploy AI models. According to Ben Horowitz, GPU prices remain elevated even as cryptocurrency token prices fall, showing that AI demand is keeping compute hardware values high.

Older Gpus Rent At Premiums, Indicating Severe Supply Scarcity and Extended Depreciation, Supporting Compute Infrastructure Value

The supply-demand imbalance is also reflected in the longevity and value of GPU assets. Jason Calacanis explains that while GPUs were originally expected to have operational lifespans of three to five years, they are now being used well into year six, seven, or eight for inference workloads. This extension indicates not only scarcity but also a lengthier and more valuable depreciation curve, enhancing the long-term worth of compute infrastructure.

Building New Data Centers Faces Regulatory Challenges and Sourcing Power and Cooling, Creating Supply Constraints Unlike the Dot-com Bubble

Ben Horowitz notes that not only are GPU prices high, but the cost to secure data center space is rising because of labor shortages and regulatory hurdles. Political and logistical complications make it difficult to expand capacity rapidly, exacerbating the shortage. Building out infrastructure is a prolonged and expensive endeavor that is not easily remedied, setting the AI era apart from earlier bubbles.

Differentiation From Historical Tech Bubbles

Dot-com Bubble: Excess Bandwidth vs. Ai Infrastructure: Insufficient Supply vs. Explosive Demand

A sharp contrast exists between the AI sector and the dot-com bubble. In the late 1990s, bandwidth was briefly scarce and expensive, but by 2001, excess supply caused prices to collapse. Ben Horowitz recalls that those who borrowed money to lock in bandwidth contracts lost out as a result of overcapacity.

Internet Bubble Burst Due to Excess Capacity; Ai Cycle Shows Insatiable Demand, Companies Struggle to Secure Compute

Today, rather than fighting excess, the AI sector battles shortages—of hardware, infrastructure, and power. Demand has not subsided; in fact, it continues to surge with no sign of slowdown. Ben Horowitz and David Sacks argue this creates a robust tailwind for companies holding compute assets.

Current Gpu and Data Center Assets Sustain Value and Debt Servicing Remains Sustainable

Ben Horowitz and David Sacks emphasize that asset owners are in an increasingly strong position. Compute infrastructure retains or even grows in value, and companies are able to service their debts due to constant high demand and robust cash flow opportunities, unlike what investors faced in the dot-com bust.

Energy and Power as the New Bottleneck

U.S. Faces 1.4 Twh Electricity Shortage By 2050, Equivalent to six Californias' Consumption

Chamath Palihapitiya warns that AI’s energy needs are colliding with massive national shortages. By 2050, the U.S. will be short by about 1.4 terawatt-hours of electric capacity—even assuming all planned plants are built on time. This deficit is equivalent to six times the electricity consumed by California, showing just how acute the infrastructure shortage has become.

Power Scarcity as a Competitive Moat: Companies Owning Land, Power, and Data Center Shells Gain Unreplicable Advantages, Shifting Ai Infrastructure From Financial Engineering to Asset Control

Energy, land, and physical data center shells are becoming unique, unreplicable assets. Palihapitiya argues that owning these hard assets will become a formidable moat, shifting the basis of AI competition from pure venture finance to asset control. This makes the economics of AI infrastructure fundamentally different—not engineered through fin ...