Have The Hyper-Scalers Lost Their Minds?

For the first time, I’m starting to wonder if hyper-scaler management might’ve lost the plot in earnest.

By “in earnest” I mean I’m beginning to suspect a case of mild delirium might’ve taken hold in the C-suite among the big AI spenders. (Some of you will immediately quip that “mild” is an understatement of trillion-dollar proportions.)

It’s not that I expected capex plans to be cut this quarter. I obviously didn’t. No one did. Nor did I anticipate management would evidence a rekindled commitment to capital discipline despite clear signs of market angst around the trajectory of AI-related borrowing and spending.

There’s been no change to analyst estimates for 2027 hyper-scaler capex, which is seen at $1.2 trillion, and everyone knows the free cash flow situation’s going to get worse before it gets better.

The figure above shows you historical and projected FCF for the big-spending tech heavyweights.

So, again, it’s not that I’m surprised by anything we learned over two weeks of hyper-scaler reports. Indeed, we didn’t learn anything. That’s kinda the point. They’re still spending, planning to spend more and with FCF turning negative, funding that spending will entail debt or equity raises. (Both.)

But I wonder if management learned anything from the decidedly adverse price action seen in shares of Meta, Amazon and particularly Alphabet following earnings. Microsoft held up better, but note from the chart that FCF’s still positive there, and not expected to turn negative.

With that in mind, we’ve just seen two more hyper-scaler debt deals, and no sooner had they launched than Bloomberg was out previewing another one.

As the figure shows, Alphabet launched a seven-tranche EUR deal today, with maturities from two to 19 years (it’s the second time Alphabet tapped the EUR market this year) and Oracle a $10 billion, six-tranche deal with maturities from three to 20 years alongside a $10 billion stock offering.

That’s on top of Microsoft’s 10-tranche $25 billion deal which priced last week, and ahead of what Bloomberg says could be a $20 billion, cross-border placement from Meta. And don’t forget about the pending $40 billion SpaceX deal.

This is too much supply, and that’s showing up in the pricing. Alphabet’s paying more this time in the EUR market than it did in May and as Bloomberg noted in a separate piece, the company’s existing euro bonds “have widened as much as 60bps since they started trading.” That as investors in Alphabet’s massive equity offering are deeply underwater following last week’s rout in the stock.

Meanwhile, Anthropic amended its S-1 to include Q3 results. The updated filing also tallies the company’s commitments and lease agreements, which come to “approximately” $895.4 billion. I don’t know why “approximately” sounds funny there, but it does.

Maybe Claude can tell you why that’s not a scary sum, but… well, the only way $895.4 billion (“approximately” or exactly) isn’t a large number is if you’re Elon Musk on a day when SpaceX is up big.


 

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.

22 thoughts on “Have The Hyper-Scalers Lost Their Minds?”

  1. Recently, I read that Gemini, is hurting advertising earnings, and that Gemini is a pretty good AI that will be enough for a lot of users. The money those other big companies are pouring into building even more powerful AI might be wasted because for many (most?) people, Gemini answers their needs.
    Then today, in Tom Friedman’s column in the New York Times, he says China is moving forward very fast with “good enough” AI , and doesn’t need to spend billions to make AI just a little bit better. I am paraphrasing, but to me these things I am reading seem reasonable. Will there ever be free cash flow to pay back these big investments?

    1. Friedman’s article made me wonder too. Not only is China’s approach to create “good enough” (unlike the U.S. attempts at creating a God like can do anything and everything), it also is narrow domain solution focused which means much smaller models which means no data centers, locally run, open source (e.g. cheaper and one time cost for startup and occasional recurring costs for upgrades).

      Who’s going to pay for all those tokens when they can pay for a few Mac Minis and an open source install?

      What do I know? Maybe the military? They can spend big $ on a Trump class battleship housing a centralized AI command data center to fire off expensive rockets all to fend off cheaper armies of autonomous/locally networked systems of drones/robots?

  2. If they’re really convinced of the value of being the ones to find the holy grail, I can see logic and justification for going for all the cash they can gather while the gathering is still good (even if not so good as it might have been). A couple years of negative FCF isn’t likely to sink the big hyperscalers – if it only goes on for a couple or three….though I wouldn’t touch Oracle or some others. I’ve always thought companies with capital needs for a perceived home-run reward should grab all the cash they can when it’s available, never mind some dilution or loss of market cap – as long as they’re not going have a continuing burn rate that outstrips the ability to attain or restore positive cashflow. But what do I know….

  3. The bigger concern to me is if/how everything around the AI rapid evolution may affect core (making them legacy?) businesses – like the huge hit Intel took from weakening sales into the PC market. Some version of that is likely to come for AMD, perhaps MSFT as well; and what about Google’s search/advertising biz, and AMZN’s shopping domination. Then there’s AAPL; not sure what to make of them…

  4. Nov 2, 2026 · @Bill
    Bottom line
    The $895.4 billion looks less like a near-term cash problem than a very large long-term bet on growth. About $69.8 billion is due within 12 months, against $146.6 billion of liquidity, but most of the total is hard to exit.
    This report draws on the commitment, liquidity and quarterly results sections of Anthropic’s amended S-1, filed November 2, 2026, not the whole filing. The analysis was written by Claude, which Anthropic makes, so weigh it accordingly.
    What the total includes
    Executed leases that have not yet started make up about half of the total, and substantially all of it relates to compute and infrastructure.
    Component Amount ($B)
    Executed leases that have not started yet 451.2
    Leases and non-cancelable purchase commitments in the obligations table, as of Sept. 30, 2026 320.0
    Commitments signed after Sept. 30, 2026 100.2
    Short-term commitments ($17.4B + $6.6B) 24.0
    Total 895.4
    The largest named deals are below. The filing describes others as well, so this list is not the full total.
    Counterparty Arrangement Amount ($B)
    Broadcom Five-year equipment leases, excluding interest about 161.2
    Google Cloud hosting commitment, April 2026 to July 2033 about 121.3
    SpaceX Compute agreements running into 2029 up to 120.5
    Amazon Spending commitment, May 2026 to April 2036 at least 110.0
    Reasons for less concern
    The payments are spread over two decades, and the next 12 months look covered.
    • The payments are spread out: $447.5B is expected by the end of 2030, and nearly all the rest between 2031 and 2045.
    • The next year looks covered: $69.8B is due within 12 months, against $146.6B of liquidity before any IPO proceeds. Up to $77B in partner note financing is also available, subject to conditions.
    • Some of it can fall away: $213.8B is subject to termination rights or conditions, such as a counterparty securing financing. The SpaceX agreements can be cancelled on 90 days’ notice after their initial non-cancelable periods end in early 2027.
    • The business is growing fast: Q3 revenue was $17.0B, up from $1.4B a year earlier, and operating cash flow was +$6.6B over nine months.
    • The huge net loss is mostly non-cash: of the $23.3B Q3 net loss, $21.9B came from fair-value changes on convertible notes and similar instruments.
    Reasons it’s still scary
    The real exposure is above $1 trillion, and the costs are far more fixed than the revenue meant to pay for them.
    • The real total is higher: another $150B–$175B of estimated payments is left out because those deals have no fixed dollar amount. That brings the total to roughly $1.05T–$1.07T.
    • Most of it isn’t flexible: only $4.3B of the $69.8B due in the next year is contingent or terminable. Some power and data center deals run 20 years, with take-or-pay terms and uncapped fuel cost pass-throughs.
    • Long-term costs are matched against short-term revenue: 79% of nine-month revenue is consumption-based, and the filing itself warns that much of its revenue isn’t under long-term contract.
    • Margins are moving the wrong way: compute rose from 48% of revenue in Q2 to 58% in Q3, and operating income went from +$974M to a $567M loss. Excluding stock compensation and a non-cash charitable charge, the margin still fell from 19% to 15%. Price cuts also took about $600M off Q3 revenue.
    • Outside capital is doing the heavy lifting: financing brought in $110.7B over nine months, against $6.6B from operations. Anthropic also says it expects its growth rate to decline.
    • A default could spread: one under certain commercial agreements could trigger cross-defaults and accelerate substantially all of its debt.
    Does the math work?
    Covering the commitments through 2030 would take about $180B in annual revenue, roughly 2.7 times today’s run rate. These are rough calculations made for this report, not figures from the filing.
    • Scale: the total is about 13 times Q3 annualized revenue of $68B, and about 25 times trailing 12-month revenue of $35B.
    • Revenue needed: the $447.5B through 2030 averages about $105B a year. At Q3’s 58% compute ratio, covering that would take about $180B a year.
    That is achievable at the recent pace, but it leaves little room if growth slows or prices fall. The obligations stay fixed either way.

  5. “But I wonder if management learned anything from the decidedly adverse price action seen in shares of Meta, Amazon and particularly Alphabet following earnings. Microsoft held up better, but note from the chart that FCF’s still positive there, and not expected to turn negative.”

    An excellent question. I am beginning to suspect the market may actually reward the first AI hyper-scaler to pull-up on the reins meaningfully. Just take a look at Apple’s chart by comparison. Either way, the market’s recent turbulence seems to clearly suggest a fair number of investors are no longer comfortable subsidizing cutting-edge AI development without some greater clarity about how profitable that particular element will prove.

  6. Not that I know anything, but think about the people running and working at these companies. Do these people care about themselves or Google? They care about themselves. They went to a good school. Mapped out how they were going to get into big tech. Got in. And cashed in. Id argue there’s a non zero possibility that the plot has been lost for a few of these companies. Props to Microsoft even though they only have two products I like.

  7. I am going to upset somme with this analogy. ‘The problem’ is that hyperscalers have committed the worst error of any drug dealer. Be clear AI is intentionally a drug, intended to addict you to usage through psychological manipulation.

    Their error? Sampling and using the product. The downdraft, will feel like a hangover because that is what it will be.

    The fundamental error of the group think is that larger will result in better more sustainable profits. You think they are not distilling their own models and copying within a data center? The reason for the data center has grown from that of only producing better models based on learning from users to being a security firewall for their cheap distillation. The reason then that large scale llm’s are not better than distillation is because they are distillations cloned internally.

    In the meantime, the party continues. Keep passing the drinks. But keep one hand in your pocket as only then can you survive the hangover.

  8. What was Alphabet’s Neg FCF? Less than $10Bn for the qtr? For 100% growth in rev? And what incremental margins? In addition to the goodwill of giving your customers the product they obviously demand? Against what cost of capital? I am guessing those expected returns are well above said cost of capital. Remember stocks are the present value of ALL FUTURE cash flows.

    For too long “investors” lived in a world where growth opportunities were more difficult to find and in some cases (very few) “easy”. Sometimes one actually has to “invest” to get growth. Does anyone really think that AI will not add to all businesses competitiveness, potentially profits (and therefore cash flow), etc if adopted competently? Does anyone think AI demand 1, 2 4, 7, 10 years out does not exceed nominal GDP? And that there will not be operating leverage? AI has barely touched businesses to date. The opportunity to use AI in every business is extraordinary imo.

    Many companies will not invest wisely but so far for the “great” companies it seems to me to be actually very rational. Not sure I would want to be competing against the 3 largest but my guess is the 3 largest and possibly some others might just do very well on these “long term investments”. Neither myopia or rose colored glasses are good but rational financial analysis helps increase the odds of successful investing.

    A year from now let’s revisit whether the “market” was right or wrong about the concerns over the “Hyper-Scalers”. MSFT is not AMZN or GOOGL imo BUT the cash flow differences are in reality “rounding errors” in my mind. We shall see if there is differentiation or not over the next few years.

    There will be winners, there will be losers, we each (professional and individual) need to do the work, realize the risks, and do our best to maximize the risk adjusted returns. It is never easy.

  9. This morning, ChatGPT stopped allowing me to download images it created and ceased allowing me to export .csvs it made. It would still show me the results, but the files were restricted because whenever I made a request, the back-end inexplicably flipped my account from “full-access” to “read-only.” I went through the support resolution process and all the Reddit threads to no avail. This is a known problem, but OpenAI simply hasn’t addressed it. Here’s what ChatGPT had to say when I asked if this is the end of the road for me and it: https://heisenbergreport.com/wp-content/uploads/2026/11/ChatGPT.png

      1. What I’ve learned today is that if you curse enough during the “escalate to a human agent” process and if you post something embarrassing about ChatGPT on Twitter and tag OpenAI, you get a human response pretty quick. It seems to be fixed now, but the jury’s still out. I don’t really give a damn about the banner images (I mean, I love what it can do, but I can make my own banner images in Photoshop if I absolutely have to), but I do care about the data gathering because as mentioned here over the weekend, I’m thinking of ditching the Bloomie. If ChatGPT is going to break such that I can’t download the Excel sheets that I prompt it to create, then I need to keep the Bloomie. That’s where the frustration comes in.

        1. Vis-a-vis the potential Bloomie savings, I’ll adapt a famous Katt Williams quotable: “They had already told me how much I was gonna save, and I had already spent that money in my head.”

        2. Well, don’t be as crazy and overreact as myself. I kept a folder of my writings emailed in my Outlook account for many years and the whole folder disappeared. The rest of the email account was intact apparently un hacked.
          Do my writing in ink I photograph the handwritten pages. Then I let someone type it up for me in Google Docs and then send it over to my Apple account for a copy.
          When I used to paint pictures and sell them, the buyer owns the painting, but the artist still keeps the image copyright unless the buyer buys the copyright.
          So my first thought in your case, led me to wonder who owns your images. AI is gonna be a copyright shit show for years. The way the lawsuits played out for writers is really ridiculous.

10th Anniversary Boutique

Coming Soon