The real question we should all be asking about AI

Michael Nagle/Bloomberg/Getty Images

The stock market is turning its back on Big Tech.

Google just had its worst day ever. Intel tumbled more than any day since the Nixon administration. Amazon and Meta both had significant declines this week, too. The Nasdaq tumbled and is just over a percentage point away from a correction (a 10% decline from its all-time high).

The common thread: Building the infrastructure to power AI is really, really expensive – and Big Tech profits aren’t yet justifying that massive expense. AI companies are increasingly doling out more cash for data centers than their businesses are able to generate.

But if the stock market is supposed be to delivering some kind of message, Big Tech isn’t hearing it.

Spending on AI has reached $1.6 trillion since 2023 and is expected to hit $1 trillion every single year for the next decade or so, according to JPMorgan. None of that changed this week.

That’s because the market that’s financing all of this – the bond market – is mostly shrugging its shoulders while the stock market is in the midst of a mild freakout. Lenders have been lining up to take on Big Tech’s debt to pay for the technology that’s reshaping the global economy.

That means the questions the market seems to be asking about AI – will Big Tech keep growing its spending? (it most certainly will) and will AI will turn a profit anytime soon? (doesn’t look like it) – are the wrong ones.

The real question we should be asking about AI is: How much longer will the bond market sustain the nonstop spending spree?

That will ultimately dictate whether AI turns into a success or a bust.

The way AI math works is like this: Spend a lot of money upfront on the infrastructure needed to support AI (data centers, servers, chips, power, labor, building materials, etc.). Then, once the data center is switched on, AI companies can monetize that investment … somewhere between three months to two years later, according to Goldman Sachs.

With profit arriving that far down the road, hyperscalers (the Big Tech guys like Google, Amazon, Microsoft and Meta) need to borrow a ton of money upfront and pay back investors down the road.

“A ton” doesn’t really do it justice. Those four hyperscalers alone have been collectively responsible for more than a third of all capital expenditures from the S&P 500 over the past 12 months, notes John Higgins, chief economic adviser at Capital Economics.

That’s why a third of Big Tech’s capital expenditures this year are expected to be financed with debt – something like $250 billion worth of bonds from hyperscalers alone, according to Goldman Sachs. That’s expected to peak next year at around 35% of hyperscalers’ spending.

Mark Felix/AFP/Getty Images

But bond market investors don’t have infinite money to spend. And there’s a ton of competition for lending right now – especially from US Treasuries as the government is trying to pay for Social Security, Medicare and an expensive war, all at the same time.

That raises an obvious question: Exactly how much money are investors willing to lend AI companies?

It’s more than you might think.

The hypothetical maximum amount of debt the market can absorb is unknowable, but Goldman Sachs did the math anyway and came up with $510 billion of available investment-grade (top-tier) debt for the biggest hyperscalers. That’s more than double what they’re currently expected to be borrowing by the end of the year.

In other words: The biggest AI companies have a lot of runway left to borrow.

The question isn’t how much; the question is how much longer .

The answer depends on two parts: Who is paying and who is receiving.

“We are not concerned about access to capital,” said Amanda Lynam, chief credit strategist for Goldman Sachs on a recent company podcast . “We are just expecting there to be a more nuanced conversation about where to issue and at what price.”

Private investment funds, venture funds and private equity are ready to do the heavy lifting to support the AI buildout, Lynam said, with trillions more dollars of “dry powder” to finance it.

But hyperscalers aren’t the only game in town. Other companies, including nanoclouds, are playing in the space too. They’re highly indebted companies that rent out computing to other businesses, and their credit situation is starting to look iffy. Their non-investment-grade bonds deliver high yields that investors willing to take on more risk are willing to gamble on, in exchange for potentially massive rewards..

That’s the market in which CoreWeave plays, with its bond yields surging to 15% this month as losses piled up. Elon Musk said the nanocloud industry, which includes CoreWeave, is due for a massive shakeout (he called it “guppy-level death.”)

Stephanie Keith/Getty Images

Every new technology has winners and losers, and Musk may be right about nanoclouds. We’ll see.

Hyperscalers, by contrast, look like winners. Their cloud revenue (a stat Wall Street uses to measure AI demand) grew 50% in the second quarter. They have huge contract backlogs and the capacity to absorb higher borrowing costs.

Those bottlenecks can support high prices, which helped hyperscaler revenue surge 19% over the past 12 months, according to Higgins.

And the advent of agentic AI has sent demand exploding: requests from agents consume 15 times as many tokens as human-initiated requests, according to OpenRouter.

“The outlook for these things doesn’t look obviously bleak, given that AI is gradually being used more and more in the economy,” Higgins said.

The concern is that the cost of borrowing could get unwieldy. And that could change the economics that are still working for AI.

Negative sentiment that’s pervasive in the stock market could start to bleed out into the bond market, sending bond yields higher and raising borrowing costs for AI companies.

The bond market works best when there’s an end in sight, Lyman noted. Every AI investor believes there’s a light at the end of the tunnel. We just have no idea how long this tunnel is.

The uncertainty about when companies will ultimately start turning a profit and making money off all their massive AI investments could give the bond market pause, she said.

In the meantime, the outlook for the technology and the big companies’ finances looks quite strong, said Stephanie Aliaga, a global market strategist with JPMorgan, in a recent note to investors.

Warning signs would include a downshift in businesses’ AI usage, cheaper pricing or fewer customer commitments.

We’re not seeing any of those yet.

Michael Nagle/Bloomberg/Getty Images

股市正在抛弃大型科技公司。

谷歌刚刚经历了有史以来最糟糕的一天。英特尔的跌幅超过了尼克松执政以来的任何一天。亚马逊和Meta本周也都大幅下跌。纳斯达克指数暴跌，距离回调仅差一个百分点（较历史高点下跌10%）。

共同点在于：构建人工智能基础设施的成本极其高昂——而大型科技公司的利润目前还不足以支撑这笔巨额支出。人工智能公司在数据中心上的投入越来越超过了其业务本身的盈利能力。

但如果股市本应传递某种信息，那么大型科技公司显然没有听到。

据摩根大通的数据显示，自2023年以来，人工智能领域的支出已达到1.6万亿美元，预计未来十年左右每年都将达到1万亿美元。本周，这一趋势并未改变。

这是因为为这一切提供融资的债券市场大多对此漠不关心，而股市却正经历着轻微的恐慌。贷款机构纷纷排队承担大型科技公司的债务，以支付那些正在重塑全球经济的技术研发费用。

这意味着市场似乎在问关于人工智能的问题——大型科技公司会继续增加投入吗？（它们肯定会）以及人工智能会在短期内盈利吗？（看起来不太可能）——都是错误的问题。

关于人工智能，我们真正应该问的问题是：债券市场还能承受多久这种持续不断的支出狂潮？

这将最终决定人工智能是走向成功还是走向失败。

人工智能的运作模式是这样的：前期投入大量资金用于支持人工智能所需的基础设施（数据中心、服务器、芯片、电力、劳动力、建筑材料等等）。然后，一旦数据中心投入使用，人工智能公司就可以从这些投资中获利……根据高盛的说法，这大约需要三到两年的时间。

由于利润要很久以后才能到账，超大规模数据中心运营商（像谷歌、亚马逊、微软和Meta这样的大型科技公司）需要预先借入大量资金，并在未来偿还投资者。

“巨量”这个词其实不足以形容其规模。凯投宏观首席经济顾问约翰·希金斯指出，仅这四家超大规模数据中心运营商在过去12个月中的资本支出就占到了标普500指数成分股公司总资本支出的三分之一以上。

因此，据高盛估计，今年大型科技公司三分之一的资本支出预计将通过债务融资——仅超大规模数据中心运营商就将发行价值约2500亿美元的债券。预计明年这一比例将达到峰值，约占超大规模数据中心运营商支出​​的35%。

Mark Felix/AFP/Getty Images

但债券市场投资者的资金并非无限。而且，目前的贷款竞争非常激烈——尤其是来自美国国债的竞争，因为政府正试图同时支付社会保障、医疗保险和一场耗资巨大的战争的费用。

这就引出了一个显而易见的问题：投资者究竟愿意向人工智能公司借多少钱？

比你想象的要多。

市场能够吸收的债务上限理论上无法得知，但高盛还是进行了估算，得出最大的超大规模数据中心运营商可获得的投资级（顶级）债务规模为5100亿美元。这比他们预计到年底的借款额高出一倍多。

换句话说：最大的人工智能公司还有很多资金可以借贷。

问题不在于金额多少，而在于还要持续多久。

答案取决于两个方面：谁付款，谁收款。

高盛首席信贷策略师阿曼达·莱纳姆在最近的一次公司播客节目中表示：“我们并不担心资金获取问题。我们只是预计，关于在哪里发行以及以什么价格发行债券，将会有一场更加细致的讨论。”

Lynam表示，私人投资基金、风险投资基金和私募股权基金已准备好为人工智能建设提供大力支持，他们还有数万亿美元的“可用资金”用于融资。

但超大规模数据中心并非唯一的竞争者。包括纳米云在内的其他公司也在涉足这一领域。这些公司负债累累，向其他企业出租计算资源，它们的信用状况开始变得岌岌可危。它们发行的非投资级债券收益率很高，愿意承担更高风险的投资者愿意冒险一试，以期获得潜在的巨额回报。

CoreWeave 就身处这样的市场，由于亏损不断累积，其债券收益率本月飙升至 15%。埃隆·马斯克表示，包括 CoreWeave 在内的纳米云行业即将迎来一场大规模的洗牌（他称之为“小鱼苗级别的死亡”）。

Stephanie Keith/Getty Images

每项新技术都有赢家和输家，马斯克关于纳米云的看法或许是对的。我们拭目以待。

相比之下，超大规模数据中心运营商似乎是赢家。它们的云收入（华尔街用这一指标来衡量人工智能需求）在第二季度增长了50%。它们拥有庞大的合同积压，并且有能力承受更高的借贷成本。

据希金斯称，这些瓶颈可以支撑高价，这使得超大规模数据中心运营商的收入在过去 12 个月里飙升了 19%。

而智能AI的出现使得需求呈爆炸式增长：根据OpenRouter的数据，来自智能体的请求消耗的令牌数量是人类发起的请求的15倍。

“鉴于人工智能正在经济领域得到越来越广泛的应用，这些事物的前景看起来并不明显黯淡，”希金斯说。

令人担忧的是，借贷成本可能会变得难以承受。而这可能会改变目前仍然适用于人工智能的经济模式。

股市中普遍存在的负面情绪可能会蔓延到债券市场，推高债券收益率，增加人工智能公司的借贷成本。

莱曼指出，债券市场在前景光明时运作最佳。每一位人工智能投资者都相信隧道尽头终有光明。只是我们根本不知道这条隧道究竟有多长。

她表示，企业何时才能最终开始盈利并从其庞大的人工智能投资中获利尚不明朗，这可能会让债券市场感到担忧。

与此同时，摩根大通全球市场策略师斯蒂芬妮·阿里亚加在最近给投资者的一份报告中表示，科技行业的前景和大公司的财务状况看起来相当强劲。

预警信号包括企业人工智能使用量下降、价格降低或客户承诺减少。

我们目前还没有看到任何此类情况。