Apollo: AI Isn't Delivering Returns, and a Trillion-Dollar Chip Debt Market Is Emerging

When AI Becomes a Cost Center: Apollo's Cold Water

Torsten Slocke, chief economist at investment management firm Apollo Global Management, has thrown a bucket of cold water on the red-hot AI narrative. His verdict is blunt: outside the US tech sector, companies see almost no financial return on AI, and AI is a cost center that damages profit margins. The infrastructure bills are real, the token consumption is real—only the revenue line hasn't moved.

This view deserves serious attention not because of its sharp wording, but because the person saying it manages real money. Slocke is not a podcast guest making provocative claims for traffic; his judgment rests on continuous tracking of corporate financial data. Venture-backed companies are consuming more tokens while their unit economics actually deteriorate—and everyone in the room chooses to call this situation "strategic investment" rather than loss.

There is a widely overlooked mismatch here: the tech giants selling GPUs are seeing revenue climb steadily, while traditional enterprises buying AI capabilities are fretting over the bills. When "the people selling shovels" make a fortune, did "the people digging for gold" actually strike gold? That is the core question Slocke is putting to the market.

The Trillion-Dollar Chip Debt: A Credit Market Collateralized by GPUs

If Slocke is questioning the demand side of AI, a deal disclosed by Bloomberg reveals the risks accumulating on the supply side. The $2.3 billion loan Blackstone provided to AI cloud provider CoreWeave is becoming the template for "GPU-collateralized debt": buy compute, rent it out, use rental income to service the debt, and seize the chips upon default.

The scale of this model being tracked points to the trillion-dollar level. The problem lies in the nature of the collateral—GPUs on server racks depreciate in roughly three years, and the "tenant" behind this debt may be a startup that no longer exists 18 months from now. The real estate analogy is intuitive, but data centers don't have stable tenants on 20-year leases, and whenever Nvidia releases a new-generation architecture, the resale value of older hardware shrinks dramatically.

The fragility of this structure lies in multiple mismatches: debt maturity mismatched with hardware lifespan, rental income mismatched with tenant survival, and collateral valuation mismatched with the pace of technological iteration. When pricing power in the compute rental market rests with a handful of chip manufacturers, how far GPU-anchored credit expansion can go is a question with no historical precedent to draw on.

How GPU-collateralized debt works: This type of financing is essentially asset-based lending, but unlike traditional equipment financing, its collateral is highly specialized compute hardware. Lenders typically size loans based on the GPU's purchase price, remaining depreciation schedule, and expected rental cash flow, and recover funds upon default by seizing, reselling, or re-leasing the chips. The key risk is that the secondary market liquidity for GPUs is far lower than for traditional collateral assets such as real estate or aircraft, and resale prices depend heavily on Nvidia's product iteration cadence. Once a new-generation architecture is released, the value of older cards can shrink dramatically within months, causing collateral coverage ratios to deteriorate rapidly.

Alpha School's Honesty: Half Its Students Didn't Hit the Target

The New York Times investigated Alpha School—a private school network that markets itself on the claim that "AI can double children's learning speed." Data published by the school shows that about 50% of K-8 students failed to reach this goal in math. In other words, half the children did not double their speed, and some of them may have even slowed down.

In an industry full of selective disclosure and carefully curated benchmarks, proactively publishing a 50% miss rate seems rare. This is genuine transparency, and it may also be a costly "humblebrag"—the difference lies in how you view the gap between marketing promises and empirical results.

More noteworthy is the follow-up move: Alpha School is expanding into public education partnerships, which means the "double the speed" claim is about to enter municipal procurement documents as a clause taxpayers can question. AI in education has always been the easiest sector to tell stories about, but also the hardest to verify with standardized tests. When a quantifiable promise is publicly tested and only half delivered, it will have knock-on effects on the valuation logic of the entire "AI + education" sector.

Building a Reasoning Model from Scratch: Not Satisfied with API Wrappers

On the engineering practice side, developer Rashka built a PyTorch reasoning course from scratch, covering inference scaling and the reinforcement learning loop, rather than stopping at the API wrapper level. The course's core idea is to help people understand the underlying mechanisms and then build things by hand in a "kitchen" they are familiar with.

The course also points out an easily overlooked trap: distillation can let small models "apprentice" the reasoning abilities of large models, but small models often sound confident while being logically hollow. This observation has direct implications for real-world deployment—when you run a distilled model in a resource-constrained environment, fluent output does not equal reliable reasoning.

This kind of from-scratch content is valuable because much current AI application development stays at the level of prompt engineering and API orchestration, with little understanding of models' internal mechanisms. Once you encounter scenarios that require custom reasoning paths, cost control, or debugging anomalous outputs, the gap in foundational knowledge becomes exposed.

Inference scaling: A technical approach that invests more compute during the model's inference phase to improve output quality, corresponding to scaling during the training phase. Typical methods include chain-of-thought sampling, multi-path search and voting, and letting the model conduct longer internal reasoning before giving a final answer. Its core assumption is that a model's capability depends not only on parameter scale but also on how long it is allowed to "think" during inference. This direction has direct implications for cost structure—inference scaling means each request consumes more tokens and compute, which is one of the technical backgrounds behind Slocke's point that "token consumption is real, but the revenue line hasn't moved."

The limits of knowledge distillation: Distillation refers to using a large model's outputs as training signals so that a small model imitates its behavior, thereby achieving approximate capability at lower cost. The problem is that distillation transfers the surface patterns of outputs, not the internal computational process required to produce them. As a result, a small model may learn the large model's "confident tone" and "reasoning format" without truly mastering the logical chain within. This gap is especially pronounced on tasks requiring multi-step reasoning, numerical computation, or factual consistency. For deployers, this means a distilled model's benchmark scores may overstate its real reliability, and additional verification mechanisms are needed to identify "fluent but hollow" outputs.

Jack Clark's Timeline: Self-Improving AI Moves Up to 2027

Anthropic co-founder Jack Clark has moved up his prediction for "self-improving AI" to 2027, citing the fact that models can now design autonomous experiments without human instruction. The gap between assistant and replacement is narrowing, and the trajectory has become steeper over the past six months.

The striking thing about this prediction is the tension it creates with the other stories: if models can truly achieve self-improvement within two years, then the trillion-dollar debt collateralized by GPUs that depreciate over three years, the still-unverified effects of AI in education, and the zero return on AI investment outside the tech sector will all need to be reassessed within a completely different time frame. What would a self-improving model optimize for? If its optimization objectives include "shutting down uneconomical compute," what happens to hardware-anchored credit structures?

A Judgment at the Intersection of Several Threads

Looking at this week's pieces of information together, they point to different sides of the same question. Slocke says AI has delivered no returns outside the tech sector; Blackstone is issuing trillion-level loans collateralized by hardware that depreciates in three years; Alpha School's promised doubling of speed has only been half delivered; Rashka reminds us that distilled models may be confident yet hollow; and Jack Clark has moved the self-improvement timeline forward.

These sources are independent of one another, yet they cross-corroborate on one point: the verification of AI's value lags behind capital investment. It should be stated honestly that Slocke's view represents the bearish camp, and there are also voices in the market arguing that AI returns simply haven't yet shown up in financial statements; Alpha School's 50% miss rate is data self-disclosed by a single institution, with a limited sample scope. As for Jack Clark's 2027 prediction, it is a forward-looking judgment from a single source and should be treated as a signal requiring continued observation rather than established fact.

For practitioners, the practical implications of these threads are: at the infrastructure level, the depreciation rate of compute assets and the matching of debt structures deserve vigilance; at the application level, quantifiable effect verification matters more than narrative; and at the technical level, the ability to understand models' underlying mechanisms is becoming the dividing line between "callers" and "builders."

One-line Summary

Five independent threads converge on one point: the verification of AI's value is lagging behind capital investment.

Paragraph Summary

This article brings together five independent pieces of AI news from the week: Apollo chief economist Slocke points out that companies outside the tech sector see no financial return on AI, making it a cost center that damages profit margins; Blackstone's $2.3 billion loan to CoreWeave reveals credit risk collateralized by GPUs that depreciate over three years, potentially reaching trillions of dollars; Alpha School proactively discloses that about 50% of its K-8 students failed to meet the "double learning speed" goal; developer Rashka's PyTorch course emphasizes understanding inference scaling and reinforcement learning from scratch, and warns that distilled models may be confident yet logically hollow; and Anthropic's Jack Clark has moved up his prediction for self-improving AI to 2027. The author argues that these threads cross-corroborate on the point that "value verification lags behind capital investment," while cautioning that bearish views, single-institution data, and forward-looking predictions all warrant careful treatment.