You can tell these guys know nothing about LLMs or how they’re provided. I love how they show OpenRouter’s graph of token usage as if it speaks for usage across the board. DeepSeek looks like the king because people who use Anthropic and OpenAI use them on either a direct basis or AWS Bedrock …
And the bar chart for token costs, really? As if that’s information? Their sources are the API docs ffs. If they had at least modeled something to estimate token costs that would be interesting, but showing the public prices and calling it research is dumb.
I love how they show OpenRouter’s graph of token usage as if it speaks for usage across the board.
Based on the light-gray shade on your comment, I'm guessing most HN'ers don't understand how questionable that particular graph makes the rest of the presentation look.
It's like publishing a physics textbook and including an appendix on astrology.
This seems healthy to me. A market where returns are less dependent on seven mega-cap names is probably more stable, not less. If earnings growth broadens out and capital starts flowing back toward quality and free cash flow outside the obvious AI winners, that should reduce concentration risk and make the whole market less fragile.
Historically stocks that had a good run then tended to underperform:
> […] Since 1926, the median ten-year return on individual U.S. stocks relative to the broad equity market is –7.9%, underperforming by 0.82% per year. For stocks that have been among the top 20% performers over the previous five years, the median ten-year market-adjusted return falls to –17.8%, underperforming by 1.94% per year. Since the end of World War II, the median ten-year market-adjusted return of recent winners has been negative for 93% of the time. The case for diversifying concentrated positions in individual stocks, particularly in recent market winners, is even stronger than most investors realize.
> Historically stocks that had a good run then tended to underperform
This is more of a mathematical axiom than a financial effect, because you're defining "underperform/overperform" with respect to an average that contains them.
It's interesting, but the usual disclaimers apply with factor research
- Does it replicate internationally?
- Is it explained by another phenomenon such as the beta anomaly or small cap premium. Implication: large caps with high beta are already known to underperform, so this this isn't a new regularity.
On slide 6, they list the Mag7 stocks (not defined in a foreword), but on slide 8 they list the free cash flow (FCF) of a somewhat different set of companies. Why not stick to the Mag7 FCF only? It muddies the waters.
True. This slide also struck me the most, but for the data it shows. Free cash flow for all the hyperscalers basically evaporated in the last couple of months, with Oracle in the minus. What does this mean?
This feels like telecom "massive demand, bad returns".
If a good enough model can be swapped in every few months, the value moves away from the model and toward cheap inference. That is great for users, but not always great for returns on huge capex.
I like that the invisible hand of market is slapping the Mag-7 for capex which is the only way to discipline them. Investors are waking up to say: hey, you are spending all your profits on data-centers, where is the return for me ? But, it surprises me that there are vast pools of capital which we collectively call the "market" that makes these calculations, or maybe a simpler causal explanation is the missing stock repurchase bid. At some point, one of the hyperscalers (msft ?) will break from the pack and announce reductions to capex and increase stock repurchases to stem the decline.
> The once high-flying "Magnificent Seven" are looking more like the Dreadful Seven.
> The why: Wall Street is growing increasingly impatient with Big Tech's astronomical capital expenditures on artificial intelligence, projected to balloon 70% to exceed $700 billion this year.
What has been the best way to determine return on the AI specific capex for hyperscalers?
I would naively expect Microsoft’s to be the highest since they are probably mainly just selling access to their capex through cloud since they aren’t seriously pursuing frontier AI, I’d imagine Google to be in the middle (selling TPUs, general cloud GPUs, Gemini, revenue lift on ads from better AI) but also spending heavily on infra to compete with OAI/ Anthropic, and then Meta to be on the low end since they are likely getting serious revenue lift from AI but not monetizing their models by API.
I am invested in some of the companies that are downstream of the capital expenditures of Big Tech (e.g., COHR), so I have nothing to complain about.
I am really struggling to see what's the investment thesis behind Google valuation increasing 2x in response to AI, though. Assuming no magical AGI singularity, by the end of the day, they're still selling the same services, but the services have gotten more expensive for them to provide. Everyone was already using Google Search, but now, provisioning AI summaries on top of requires more compute. Everyone was already using Google Docs and Meet, but now, AI features cost Google more. Etc, etc.
The only place where they stand to make money is selling AI compute to enterprises. But with the current supply-chain challenges, the margins there are probably getting thinner.
Google is a good bet because if AI continues to boom, they're in a good position (frontier lab, vertically integrated). If the bubble bursts and the frontier labs fail, they might do even better.
>Assuming no magical AGI singularity, by the end of the day, they're still selling the same services, but the services have gotten more expensive for them to provide.
Well, they're hoping to sell new services on expensive $200/month subscriptions.
The hope is that agents have more value than traditional software, because they do the work for you instead of just enabling you to do the work.
Google has a compelling story for many AI scenarios: it has lots of outs. It's the only frontier lab for which that's true. A massive bubble bursting wouldn't be existential for Google; it would be quite painful, but survivable, and even offers some potential upside (picking up assets and researchers from the wrecked, mangled corpses of other frontier labs on the cheap).
GCP has wild YoY earnings growths and it has accelerated since cloud AI became a thing. AI is finally bringing Google a real second line of business besides ads.
Markets are obviously, rationally, not happy about the overspending.
But what gives me pause is that a some of the mag 7 (think Meta) could change their mind on AI build-out tomorrow, and 1-year from now have the same amazing free cash flow they always did.
I don't think most people quite grasp the sheer size of the debt obligations Meta and specially Oracle have gotten into. Meta has done some clever tricks to keep it off the main books, but it IS there. If no AGI, its gonna drag them down for a decade.
This is probably a hot take and I am by no means a financial expert and this is probably quite wrong. I personally think that attempting to value these companies using the same methodology as the history of all American companies is fundamentally wrong. Sure, some mom and pop small local regional business that overperformed is probably more likely to underperform. But when it comes to big tech companies, these companies are operating a data and capital flywheel that doesn't easily just slow down. I mean, you look at the history of these computer tech companies, especially software companies. They really haven't slowed down. Like, look at Microsoft. It's just been growing from the very beginning, pretty much.
I respect your humility, and agree that you have an insight.
what I am curious about is how you would think about valuing these companies? Let's take it for a given that they are data and capital flywheels, and that this flywheel driven growth will continue for the next 100 years.
And say the valuation based on this measure says the company should be worth X today and X_y = X(growth rate per year)(y years) in y years (yes, I know we should use exponents not multiplication, but that's hard to render in this text editor).
and the price today is P.
What is the ratio of P to X? What is the ratio of P to X_y?
If P today is greater than X_y in 5 years, do you still buy it?
While I agree with other comments indicating that the headline is drawing on a small period of data, other data in the deck is pretty compelling.
Page 25 "The number of data centers in the US" gives an interesting insight as to the magnitude of the data center boom. 60% more data centers are being planned or are under construction. This might actually be underselling it in dollar amount, as I believe the average data center size under construction is larger than the average data center already constructed.
Page 27 "Cyclically adjusted P/E ratio near all-time highs" is certainly concerning and points to a near term correction.
as soon as you start calling a group of stock tickers the “magnificent 7” they’re destined to underperform, as that’s a feelings based assessment and many investors will continue to buy the feelings long past the value being fair.
“Do they make money? I don’t know but I know they’re magnificent!”
I think page 4 is a little disingenuous, I'm pretty sure you could pick lots of windows and show the mag 7 deviating negatively, but then later trends positive. I believe this whole thing will pop, but I'm not quite convinced it is just yet.
Page 8 for Oracle's free cash flow trending negative is really quite impactful. The Bloomberg AI bubble diagram [1] shows how this could really blow up. If Oracle falls they could take the whole market with them. We're just waiting on the mag 7 or any closely linked companies to fail to raise investment.
Page 16 really outlines how insane these evaluations are. I think most countries see it, hence aggressive selloffs of US bonds [2]. But everybody is just too insanely heavily exposed to it all now, it's going to wipe out everything. It's going to be a very awful time when heavily debt strapped countries can't issue debt anymore.
I think what we're going to see is some insane moves to keep these companies afloat longer in some desperate attempt to delay the pop, which will just make a bigger bubble. I could see Nvidia for example issuing bonds in excess of $100bn soon if the market has appetite for it [3].
58 comments
[ 4.8 ms ] story [ 54.1 ms ] threadAnd the bar chart for token costs, really? As if that’s information? Their sources are the API docs ffs. If they had at least modeled something to estimate token costs that would be interesting, but showing the public prices and calling it research is dumb.
Based on the light-gray shade on your comment, I'm guessing most HN'ers don't understand how questionable that particular graph makes the rest of the presentation look.
It's like publishing a physics textbook and including an appendix on astrology.
> […] Since 1926, the median ten-year return on individual U.S. stocks relative to the broad equity market is –7.9%, underperforming by 0.82% per year. For stocks that have been among the top 20% performers over the previous five years, the median ten-year market-adjusted return falls to –17.8%, underperforming by 1.94% per year. Since the end of World War II, the median ten-year market-adjusted return of recent winners has been negative for 93% of the time. The case for diversifying concentrated positions in individual stocks, particularly in recent market winners, is even stronger than most investors realize.
* https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4541122
This is more of a mathematical axiom than a financial effect, because you're defining "underperform/overperform" with respect to an average that contains them.
- Does it replicate internationally?
- Is it explained by another phenomenon such as the beta anomaly or small cap premium. Implication: large caps with high beta are already known to underperform, so this this isn't a new regularity.
If a good enough model can be swapped in every few months, the value moves away from the model and toward cheap inference. That is great for users, but not always great for returns on huge capex.
buy high, sell...whenever?
I just can't wait to get back to when innovation meant financialization, can you?
Building things? Real Jobs for real electricians? Real buildings? Real compute? Why would anybody want that!
https://finance.yahoo.com/markets/article/magnificent-7-stoc...
> The once high-flying "Magnificent Seven" are looking more like the Dreadful Seven.
> The why: Wall Street is growing increasingly impatient with Big Tech's astronomical capital expenditures on artificial intelligence, projected to balloon 70% to exceed $700 billion this year.
I would naively expect Microsoft’s to be the highest since they are probably mainly just selling access to their capex through cloud since they aren’t seriously pursuing frontier AI, I’d imagine Google to be in the middle (selling TPUs, general cloud GPUs, Gemini, revenue lift on ads from better AI) but also spending heavily on infra to compete with OAI/ Anthropic, and then Meta to be on the low end since they are likely getting serious revenue lift from AI but not monetizing their models by API.
Amazon $200B
MS $190B
Alphabet $175B-$185B
Meta $115B-135B
I am really struggling to see what's the investment thesis behind Google valuation increasing 2x in response to AI, though. Assuming no magical AGI singularity, by the end of the day, they're still selling the same services, but the services have gotten more expensive for them to provide. Everyone was already using Google Search, but now, provisioning AI summaries on top of requires more compute. Everyone was already using Google Docs and Meet, but now, AI features cost Google more. Etc, etc.
The only place where they stand to make money is selling AI compute to enterprises. But with the current supply-chain challenges, the margins there are probably getting thinner.
Well, they're hoping to sell new services on expensive $200/month subscriptions.
The hope is that agents have more value than traditional software, because they do the work for you instead of just enabling you to do the work.
But what gives me pause is that a some of the mag 7 (think Meta) could change their mind on AI build-out tomorrow, and 1-year from now have the same amazing free cash flow they always did.
what I am curious about is how you would think about valuing these companies? Let's take it for a given that they are data and capital flywheels, and that this flywheel driven growth will continue for the next 100 years.
And say the valuation based on this measure says the company should be worth X today and X_y = X(growth rate per year)(y years) in y years (yes, I know we should use exponents not multiplication, but that's hard to render in this text editor).
and the price today is P.
What is the ratio of P to X? What is the ratio of P to X_y?
If P today is greater than X_y in 5 years, do you still buy it?
Page 25 "The number of data centers in the US" gives an interesting insight as to the magnitude of the data center boom. 60% more data centers are being planned or are under construction. This might actually be underselling it in dollar amount, as I believe the average data center size under construction is larger than the average data center already constructed.
Page 27 "Cyclically adjusted P/E ratio near all-time highs" is certainly concerning and points to a near term correction.
-signed a bitter somebody that had to buy a new SD card for my camera last week.
They are sitting on their war chest, not spending money on scaling anything anywhere.
“Do they make money? I don’t know but I know they’re magnificent!”
Page 8 for Oracle's free cash flow trending negative is really quite impactful. The Bloomberg AI bubble diagram [1] shows how this could really blow up. If Oracle falls they could take the whole market with them. We're just waiting on the mag 7 or any closely linked companies to fail to raise investment.
Page 16 really outlines how insane these evaluations are. I think most countries see it, hence aggressive selloffs of US bonds [2]. But everybody is just too insanely heavily exposed to it all now, it's going to wipe out everything. It's going to be a very awful time when heavily debt strapped countries can't issue debt anymore.
I think what we're going to see is some insane moves to keep these companies afloat longer in some desperate attempt to delay the pop, which will just make a bigger bubble. I could see Nvidia for example issuing bonds in excess of $100bn soon if the market has appetite for it [3].
[1] https://archive.is/0bYLS
[2] https://sg.finance.yahoo.com/news/china-japan-uae-india-sell...
[3] https://uk.finance.yahoo.com/news/nvidia-raises-over-21-5bn-...