There are several reasons why US investors are growing nervous. The primary reason is debt, but general suspicions about OpenAI’s buildout and the perception that China has caught up aren’t particularly helping the optimistic case for US AI stocks.

Let’s cover the latter two first, because there’s much more to be said on the first one.

China is becoming a real threat to US AI supremacy. Besides the surprising growing allergy toward American chips (more on that later), they are “waging war” on American companies in two ways: by releasing research on the hardware and software efficiency frontier.

In other words, they are sending two “horrific” thoughts to US investors:

This is primarily based on two numbers: 82% and $4.6 million.

The first one refers to Alibaba’s Aegaeon token-level scheduler, which we recently covered, that has enabled Alibaba to reduce GPU inference usage by 82% compared to previously considered frontier inference strategies.

Put simply, for the same inference workload to be served, the system managed to service that load with 213 GPUs instead of the previous number of 1,192.

This immediately raises the question: how many of these efficiency breakthroughs are US Hyperscalers simply not factoring into their spending projections?

The fears are understandable. It’s clear that efficiency is taking a backseat in US buildouts and research, and money is being spent without regard for efficiency breakthroughs. Instead, companies are literally spending every single dollar they can on this.

This is a very risky bet for such an “unprofessional” way of spending money, as:

You don’t have to take my word on the "uncontrolled spending” matter; many of the leaders who spend this money believe so too, with examples like Sam Altman (OpenAI CEO), Sundar Pichai (Google CEO), Satya Nadella (Microsoft CEO), or Mark Zuckerberg (Meta’s CEO). The latter was particularly clear on this, stating he was happy risking “a couple of hundred billion” rather than being late to the AI party.

Therefore, China's release of work showing that we can constrain GPU usage by almost 90% is not something investors are happy to see coming from the land of the dragon. But if that has people nervous, Moonshot’s Kimi K2 Thinking model is as alarming as things get.

Is it better than GPT-5 Pro or Gemini 2.5 Deep Think? No, but that is not a fair comparison because these two are ‘best-of-N’ systems, where several models are run in parallel and the best one is kept, or deploy humongous amounts of compute per problem, which naturally leads to the best performance.

Thus, both of these techniques obscure the accurate apples-to-apples comparison between GPT-5 or Gemini 2.5 and Kimi K2 Thinking. Notably, Kimi K2 Thinking is one of those “DeepSeek moments” claimed to have been trained with a budget of under $5 million.

This is, once again, total rubbish, not because Moonshot’s claim is false, but because it’s misinterpreted. They were referring to the training run cost (the total cost to get you the model that eventually was deployed), but did not include all the very real overhead typical of AI training (dataset gathering, experiments, salaries, and others) that increase your total spending by a lot.

But two very real conclusions can be drawn and seem non-negotiable:

Both of which have investors wondering the same thing they asked back when NVIDIA crashed in March: Are we overspending?

Maybe, and OpenAI’s recent moves don’t suggest confidence at all.

Another topic top of mind for many is OpenAI’s recent controversies. For one, nobody actually understands how a company with a $13 billion run rate (recently updated to $20 billion by December, according to Sam) can spend $1.5 trillion in the next several years.

Who’s going to fund that?

We’ll answer that below, but besides obvious questions, their PR strategy is debatable, after its CFO’s clumsy comments basically begging the US to bail them out if their huge spending commitments overextend, even prompting politicians like Republican Governor Ron DeSantis jumping on this matter and earning considerable praise from several Democrat leaders and voters at a time when both sides can’t even agree on whether the sun rises in the mornings.

And while OpenAI is the face of uncertainty in US markets these days, it’s only a symptom of the real risk that keeps investors awake at night: debt.

The fact that debt is growing in importance as a key mechanism to fund AI CapEx is an understatement. The shift is visible in hard numbers.

At the center of all, as you may imagine, we have the Hyperscalers, which have started to rely more on debt. These companies have so much cash that they are investment-grade issuers, meaning that credit agencies believe they are unlikely to default on their loans.

The problem is the rapid growth of this number. US investment‑grade issuance from AI‑heavy Big Tech hit roughly $75 billion in September–October alone, including:

That’s way more than what they issued in the entire 2024 year. While all three were oversubscribed (more on why this is very important to the AI buildout later), JPMorgan now estimates AI‑linked issuers make up about 14% of the US IG index, overtaking banks as the dominant sector.

Oracle’s case is particularly remarkable. And worrying. The same month Oracle sold $18 billion of bonds, it also became the anchor tenant in a separate $18 billion project finance loan for a New Mexico data center campus tied to the “Stargate” initiative, thanks to a consortium of up to 20 different banks. That takes Oracle‑related new debt commitments to roughly $36 billion in six weeks.

But if there’s a company riding debt, it’s Meta; it’s the poster child for the trend. The Hyperion move (which is clearly set up as an off-balance-sheet structure, more on this later) shifts funding away from pure cash flows and toward lease-backed debt. Of course, markets took note: the bond news coincided with an 11% single-day stock drop as investors recalibrated their leverage and payback windows.

Alphabet’s case is the least concerning but still worth mentioning. It raised debt in dollars and euros this month, in addition to a May euro-denominated deal, while increasing 2025 capital expenditure guidance to $91–$93 billion. This is the clearest admission that organic free cash flow will not cover AI capex at the desired speed, even for cash‑rich firms, and the choice is to mix internal cash with market debt.

And what about the spreads? How expensive is this debt?

But what does this mean? This is a cause for monitoring, but we can’t yet panic.

Besides Oracle, the other companies have the cash to fund the AI spending. Hence, the reason they are tapping into the bond market is because, well, they can. It’s cheap capital, as investors are willing to accept interest rates that are not much higher than those requested by the US Government itself.

That said, even though it’s early to panic, corporate bond spreads for Hyperscalers are spiking back again to 78 basis points this November, to the point some analysts are citing this metric as something to “watch out” for.

But if corporate bonds are a nice way to gauge investor sentiment toward the AI bubble, what if I told you investors are betting hard that one very prominent public US AI company will default in the next few years?

And more controversial even, that we are repeating the same procedures that led to the 2008 financial crisis?