I’m not betting against AI. I think it’s here to stay. I’m betting that the money spent on it got ahead of what the technology can actually do today.
I've essentially been using multi-agent systems non-stop for the past three years. I have worked at a number of businesses, industry being finance, niche being marketing. So I've been able to come across and solve real business issues...
The economics behind the buildout
People have overestimated the demand for AI and data centers.
The reasoning: CoreWeave and Oracle are the most leveraged bets on the AI buildout, because they borrowed heavily to build data centers. Nvidia sells the shovels in this gold rush.
For neoclouds—the specialist cloud providers selling access to AI computing—the financial question is how much their equipment can earn over its useful life, after operating and financing costs. GPU depreciation matters alongside borrowing rates because the hardware can lose economic value while the obligation to repay its financing remains.
GPU depreciation and the cost of debt
CoreWeave’s 2025 annual report assigns technology equipment a six-year useful life for accounting purposes. That is an estimate, not a guarantee of resale value or six years of equally profitable GPU rentals. If newer hardware or cheaper alternatives push rental revenue down faster than expected, the original investment becomes harder to recover. Accounting depreciation itself is a noncash expense; the concern is the equipment’s ability to generate cash.
Source: CoreWeave annual report
The borrowing costs also vary by financing arrangement. CoreWeave disclosed floating-rate debt at SOFR plus 2.25 percentage points under its March 2026 facility, and SOFR plus 5.5 percentage points under its August facility. It also issued senior notes carrying a 9.75% coupon in April. Those are specific debt terms, not one company-wide borrowing rate—and they should not be confused with a bond’s changing market yield.
Source: CoreWeave second-quarter filing
There is also a timing issue. CoreWeave says the August facility has an approximately five-year maturity, while the underlying customer contracts average approximately three years. The company presents that as lender confidence in continued GPU demand. My concern is what happens to renewal pricing and equipment economics during the remaining loan term.
Source: CoreWeave financing announcement
What the bond market is pricing
In an October 28, 2026 report, Morningstar said the yield on CoreWeave’s 9.625% senior notes due in 2032 had risen above 15% earlier that month, while their price fell below 80 cents on the dollar. These are dated secondary-market observations. The notes’ contractual coupon remains 9.625%; their market yield reflects the price investors pay. These notes are separate from the April issuance carrying a 9.75% coupon. Higher market yields can make future borrowing and refinancing more expensive.
Source: Morningstar, October 28, 2026 · CoreWeave June 18, 2026 issuance filing
The same report put CoreWeave’s five-year credit-default-swap spread at approximately 1,100 basis points, compared with roughly 500 basis points in June. That equals 11 percentage points of quoted credit-protection spread; it is neither a loan rate nor an 11% probability of default. Morningstar also cited its analysts’ projection that total debt could exceed $100 billion by 2030. That is a forecast, rather than debt already outstanding. The report said analysts were not seeing broader credit-market contagion at that time.
Source: Morningstar, October 28, 2026
Match the hardware to the job
Inevitably, as GPUs are so expensive, people will look into ways to make the most of their CPUs, I don't think people understand this enough. Deepseek was insane because when it came out it was the best coder on the market (or close) and didn't require too many resources. You could run it on older equipment.
DeepSeek’s full R1 model has 671 billion parameters, while its smaller distilled versions range from 1.5 billion to 70 billion, so hardware requirements depend on which version and configuration you run.
Source: DeepSeek R1 model documentation
CPU inference and more efficient models make that discussion concrete. Tools such as llama.cpp support CPU execution, quantization to reduce memory requirements, and hybrid CPU/GPU inference. That does not establish that CPUs can replace GPUs for every workload, or that total GPU demand must fall. It does show why hardware requirements need to be evaluated against the actual job.
Source: llama.cpp documentation
Where useful AI becomes a business
I think the only way to profitability for frontier labs is to sell to enterprise clients only. Opus 5.5 is too powerful for your everyday person's needs. Honestly GPT-4 is powerful enough for the average casual AI user.
RunPod illustrates the practical side of GPU cloud computing: rent capacity for a defined workload. Its Pods provide dedicated GPU instances billed by the second, while Serverless flex workers can scale to zero when idle. Those are concrete services for running applications. Their usefulness, however, does not by itself establish that every neocloud’s valuation or debt structure makes sense.
Source: RunPod GPU Pods · RunPod Serverless
The best startups will be the ones to have figured out how to create a seamless "computer-human" experience.
PLEASE understand the fundamentals of the companies you invest in. Don't just blindly follow. Have nuance. You can love the tech but also realize when people are speaking a bit delusionally.