Alphabet plans to borrow another $20-25 Billion, on top of the $50 billion in bonds and $85 billion in equity it sold earlier in 2026.
By Wolf Richter for WOLF STREET.
Alphabet, one of the biggest hyperscalers building out the AI infrastructure, wants to borrow more money, again, this time between $20 billion and $25 billion in a bond offering of up to 10 parts, with maturities ranging from 2 years to 40 years, according to Bloomberg, citing sources.
But if the yield is high enough, investors will come. Alphabet has received $115 billion of orders for this offering. The managers of the bond sale are Goldman Sachs, JPMorgan Chase, Morgan Stanley, Bank of America, Citigroup, and Wells Fargo.
Alphabet’s shares dropped 1.0% on the news.
And these bonds compete with Treasury securities for investor attention. To take on the additional risk, investors in these Alphabet bonds get a higher yield than on Treasuries, and it draws demand from Treasuries. And so when this news emerged, the 10-year Treasury yield rose by about 6 basis points to 4.67%.
Earlier this year, Alphabet had already sold about $50 billion in bonds, including 100-year bonds, in various currencies. Plus in June, it raised $85 billion by selling stock and mandatory convertible preferred. Share buybacks were scaled back last year and went to zero this year.
Alphabet spent $45 billion in Q2 on capital expenditures, mostly for AI infrastructure such as data centers, doubling from a year ago, thereby creating negative cashflow of $6 billion in the quarter.
And there is much more cash-burn to come: During its Q2 quarterly earnings call, Alphabet raised its full-year 2026 capital expenditure guidance to a range of $195 billion to $205 billion (from the $180 billion to $190 billion guidance in the prior earnings call). And it said that capital spending in 2027 would increase significantly from those levels.
So far this year, and not including this newest bond offering, just four companies – Alphabet, Amazon, Meta, and Oracle – sold $194 billion of bonds to fund the AI cash-burn. Alphabet’s deal would push the year-to-date total up to $219 billion. This does not include the funds raised via equity sales.
The AI infrastructure giants – Alphabet, Amazon, Microsoft, Meta, Oracle, and SpaceX combined – are expected to plow between $800 billion and $900 billion in cash into capital expenditures in 2026, primarily on AI infrastructure, and that amount keeps getting ratcheted higher, and we’re seeing some of it in the exponential surge of construction spending on data centers.
These companies used to be asset-light and used to produce enormous amounts of cash flow from their endlessly scalable services offerings, that allowed them to engage in large-scale share buybacks, which was so appealing to Big Tech investors.
But now they’re turning into asset-heavy companies, similar to manufacturers, with massive amounts of money tied up in mega-facilities that they funded in part by loading up on long-term debt. Their income will be hobbled by large amounts of expenses from interest, depreciation, and operating costs, while share buybacks – except for Microsoft – have vanished or have flipped to the opposite: share issuance. And that’s a complete change in business model.
In terms of the US economy, this investment boom and cash burn — Corporate America plowing hundreds of billions of dollars of their cash and investors’ cash quarter-after-quarter into the economy, instead of into share buybacks — is strongly stimulating across many sectors, and is becoming one of the drivers of inflation.
Enjoy reading WOLF STREET and want to support it? You can donate. I appreciate it immensely. Click on the mug to find out how:
jobs are being quoted “go away” pricing. things are booming out there. china price looks a lot like 3 years ago domestic price. these guys (data center/mil/defense) are paying *any* price to get things done, and they are maxing out capacity. service cost, ppi, etc. shows this, as do margins. roaring ’20s.
yep, I hear the same things from my flatbed trucker friends. Rates are CRAZY high (good for truckers, not so much everyone else). (part of which is the illegals getting pulled off the road and the other part is the the data center is driving it all).
Nobody that has been in trucking for more than a year is doing great unless it is big LTL companies.
Spot rates are way up but spot rates for 2-3 years were barely covering fuel costs. We were in the great freight recession. Even still, volume isn’t great, just a lot of trucks went bankrupt and off the road. The purging of illegals is just window dressing. It makes for big headlines but it’s not pulling much capacity off the road in the grand scheme of things. Especially when “Canadian” drivers are illegally running loads in violation of cabotage laws.
Very informative.
Wow, hundred year bonds for AI ‘factory’ construction. For interest sake (pun intended) I looked up great events 100 years ago, and thought about today’s trajectories and the changes I’ve seen in my 70 years.
Insanity, was my conclusion. Desperation. I think we need other other advances to pursue. Other priorities. If AI is going to be the driving lifeblood of all things, count me out. Just an opinion.
I expect a bubble burst.
I definitely moved my guess to sooner. Good info Wolf.
Make that many bubbles…not just monetized ones.
Meh, just borrow it out of Apple’s petty cash…
Does seem like Berkshire and Apple will own it all after the inevitable crash…
Apple feels very much like Sears in 1991 to me.
Per Bloomberg, roughly 70% of Microsoft’s AI revenue came from OpenAI alone last year. Per Barclay’s estimates, 75% of Amazon’s AI revenue is coming from just OpenAI and Anthropic. Per UBS estimates, 75% of Google’s AI revenue is coming from OpenAI and Anthropic.
Nearly 3/4 of AI revenue coming from 1 or 2 companies is wild, crazy, irresponsible stuff.
Why exactly? Most major sectors of the economy have 2-5 major players.
In the AI space, the major players are selling a vision of huge amounts of demand. But this much concentration in AI sales indicates the opposite: that there is no demand outside of the OpenAI and Anthropic, who are effectively being propped up by the hyperscalers, chip manufacturers and private debt.
If there is no demand outside of the circular financing that is going on, and if any existing end user “demand” is only propped up by the big guys heavily subsidizing the price of the LLMs, the bubble will likely pop, or at the very least deflate rapidly.
I kinda don’t understand where this mindframe comes other than you wanting it to be so. Anthropic has grown from $9B ARR Dec 2025 to $74B ARR in July 2026, literally the fastest rate of revenue growth in history.
I contribute $10/HR of AI revenue every hour I’m working. Departments within businesses are fighting each other for more tokens. It’s 100x worse than when we used to fight for tech resources to get our projects out of the development queue.
There is a nearly infinite demand for intelligence. It may be the case that revenue isn’t now covering creation and operating costs, but you’re living in a fantasy world if you think there isn’t demand for AI.
William, you are only fighting over tokens because they are being subsidized for you. Once the big players actually charge you what is being spent , your company will cut back on token spend, just like Uber is doing, just like Walmart is doing, and so on etc.
Note that I am not an bearish on the technology of AI itself, but the industry has grown to proportion that are not commensurate with the demand that exists. Once the bubble deflates, the move by companies who really actually need it will be to have their own on-prem compute. But that will be a fraction of what is being promised/committed now by the hyperscalers.
Even beyond the demand question, the capex is so enormous that the actual revenue that will be needed to justify the capex (both existing and promised) dwarfs any actual realistic money that is in customers pockets right now.
@Jeziel
I’m fighting for tokens because they allow me to work so much more productively. I would pay 10x the current cost and tell the CEO to cut headcount elsewhere to pay for it. Give me $200k of tokens over a $200k loaded salary of a software developer any day of the week. That’s the comparison being made here and it’s based on value for money.
This all gets very weird at the macroeconomy level, as at the end of the day we’re selling a product/service (food manufacturing and delivery) to relatively well off consumers, which are disproportionately work-from-home software developers. This aligns with my understanding of the real underlying cause of the great depression being vast efficiency gains being made on the business side of the ledger that ended up killing the spending power of their customers. Tremendous population growth in the 20th century was able to generate new demand and ultimately rebalance, but obviously demographics are trending the other way at the moment.
No, as you will someday come to understand, the cause of the Great Depression was the Bubble that preceded it.
And the interlocking trusts who had similar circular financing/ownership. But you’ll come out ok, unlike those people you encouraged your CEO to fire so you can buy more tokens.
Wolf, do you have any data on the spread for these bonds against Treasuries? That is something I would like to keep an eye on in case they start blowing out.
So sources cited by Bloomberg said that price talk started at 155 basis points above Treasuries for the long-term bonds. I have not heard of a final deal on the spread or rates. Eventually, the results will be announced. The preliminary SEC filing was full of blanks.
We’ll be at $40T in debt by the end of August, spending more than $1.247T annualized in interest expense.
Data center growth will continue to gobble up power. SMRs won’t arrive in volume to push down the cost until 2035, if they even arrive in volume at all. This means high power bills. Keep in mind inflation always follows the cost of oil. Well, I think we can definitely say it now follows electricity rate, maybe just not quite as volatile MoM by certainly YoY.
There doesn’t appear to be an off ramp for inflation outside of a recession.
There’s no way to tell what the next 3-5 years will look like once we hit AGI by 2027 or 2028 at the absolute latest. We’ll all be guessing by that point.
November will be the most consequential election in my lifetime. If the Dems / DSA crowd takes Congress, TACO is gone. That in of itself will make things get pretty crazy politically. Then, the biggest question of all is do the Dems / DSA go after JD Vance? If they’re successful with that mind boggling power grab, then that’s the moment the military has to decide if they’re going to step in to stop what will obviously be a coup d’etat.
Strap on your hard hats, people, and buckle up.
AGI by 2028 “at the latest”? The Democrats doing a literal coup?
Pass the joint buddy you’ve got the good stuff.
Many SMEs in terms of AI believe AGI will arrive in the 2027/28 timeframe, so this isn’t something I’m making up out of thin air. Far from it.
Granted, I had forgotten that it takes 2/3 of the Senate to remove a president / VP, so Trump getting removed is certainly within the realm of possibility but Vance probably not so much.
Why would the Dems even want to remove Vance? Trump ceding to him one way or another would be a gift to the Dems given how unlikeable he is to basically everyone. Perfect foil to run against.
Taco is gone – Such Unserious people – Ugh
I think you need to put down the Tom Clancy novels and go back and touch up on how this thing America works.
No inflation off ramp but a recession and AGI guessing by that point – LOL
This is great stuff.
Carry On…
I forgot that the Senate threshold is 2/3, so TACO is a possibility while Vance is not likely. I agree the later is not likely, but if Trump is removed, this will be a very big deal for a lot of people.
Again, AGI arrival in 2027/28 is what all sorts of AI SMEs are saying, so this isn’t me guessing. It’s the experts making their very well informed predictions.
A full impeachment and removal of Trump after the midterms is mathematically and politically highly unlikely.
Even if Democrats flip the House of Representatives and vote to impeach him by a simple majority, they face an almost insurmountable obstacle in the Senate.
The U.S. Constitution requires a strict two-thirds supermajority vote (67 senators) to convict and remove a president. Even in the absolute best-case scenario for Democrats in the midterms, they would still need to convince 15 to 17 Republican senators to vote to convict Trump —an outcome with no historical precedent.
Just a bunch of political theater – AGAIN…
Removing a Vice President requires a two-thirds majority in the Senate. No way the Dems hold that kind of supermajority after Nov.
I would love to know why you think the election results one way or another are going to have a huge effect on our fiscal trajectory.
NYC?
What about it?
I’m not saying either party is more likely to create a bad fiscal trajectory at least for now with the filibuster in place. But I think any reasonable person can say that once the filibuster is eliminated by the Dem / DSA crowd this will lead to higher deficits than we would have had with the GOP maintaining control over one of the three branches. That’s a reasonable conjecture to make.
You clearly don’t know your history. Huge deficit increases have occurred under Republican presidents, not Democrats, because Democrats will raise taxes on the billionaires and multi-millionaires. By the way, that’s what every real patriotic American wants to have happen now anyways. And all this idiocy regarding DSA is just that. There may be between 5 but certainly no more than 10 DSA Representatives elected to congress, and they will have zero power. The other foolishness of the DSA scare is typical political theater from the mendacious GOP.
“There doesn’t appear to be an off ramp for inflation outside of a recession.”
Well that’s always the case. The thing most closely correlated with inflation is economic growth, which is why economists always point out that deflation is a much more scary scenario than inflation.
They point that out because they represent the banks and governments for whom borrowing money cheaply is beneficial. It’s not in good faith.
Most economists are academics working in universities for $150k a year and trying to cover the mortgage in Cambridge. The idea that they “represent banks” is laughable.
It’s just relatively simple to look at economic history and see that we don’t have a good playbook for how to handle deflation and that deflationary periods tend to be much longer (long depression, great depression, Japan lost decades) and thus do more harm socially.
Well if that’s the case, why do we have a Fed then?
I thought there job as to thread the FFR / M2 needle such that they’re able to bring down inflation?
Granted, they’ve failed for 63 months in a row now, but they’re supposed to be trying to make the necessary adjustments that bring down inflation without sending us into a recession.
Both things can be true.
Most inflation/deflation has to do with underlying factors related to productivity, resource allocation, natural events (weather, etc), and goodness knows what else.
We understand how economies work about as well as we understand how human health and bodies work. By analogy, the Fed is something like a doctor, making interventions here and there, sometimes successfully, sometimes doing harm, but most of the time not even being able to tell as we don’t have good means of measuring impacts or the counterfactual of what would have happened had they done something else or nothing.
Complaining that the Fed hasn’t perfectly hit it’s mandates is like complaining your obese uncle who smokes and has unique genetics that we barely understand doesn’t have perfect blood pressure. The doctor can prescribe statins and give advice, but the big factors are either mostly outside of their control or not well understood by anyone.
The panic of 1907 where the government had to rely on a private party to inject liquidity as a means of last resort and the long depression of 1873-1893 with it’s persistent deflation that couldn’t be turned around were primary catalysts for setting up the Fed (see the 2% INFLATION target), along with bimetallism debates and the political strife it caused. Which shows if nothing else that the Fed is an attempt to manage economic problems that emerge in economic systems without anything like a Fed. It’s not a bogeyman, it’s not benevolent, it’s not a puppet master, it’s a government agency tinkering on the margins.
“If the Dems / DSA crowd takes Congress, TACO is gone.”
There isn’t even a remote possibility that the “DSA crowd”, all dozen of them, would come to dominate democratic party politics. They will attract outsize media attention, like Bernie Sanders, but have similarly no impact on actually changing the govt as they aren’t in swing districts that become critical swing votes.
As per TACO, Trump will be completely neutered, as congress will have no reason to allow him his executive free range in exchange for endorsement. It’s already nearly there as Trump approval ratings are worse than any president in modern history and given what we know today, the GOP is going to get murdered in the midterms.
If you are interested in facts, you should know that GW’s approval rating dropped as low as 22% during the great recession. Obama also got as low as 38%.
Right–I should clarify with “by this point in their second term” and that the only close contender was Nixon on the verge of resignation. On the current trajectory he will finish well below W Bush and that’s before any statistically validated economic downturn (see today’s jobs report?) or an escalation of the Iran situation. At a certain point, these things take on a life of their own and people don’t want to be the last one still wearing bell bottoms.
ROTFLMAO! Today, 33% of all Dems identify as being Socialist friendly which is as little as one mid-term election cycle to jumping up to 50% or above. When this happens, we can say the DSA crowd will have taken over. And it certainly seems like only a handful of Congressional Dems are willing to stand up to the DSA crowd. All this points to the clear fact that the DSA crowd is winning and taking over the left.
Also, the way they’re doing this is the right way. They’re not trying to create a new party. Rather, they’re just promoting their DSA people within the existing framework of the Dem party.
However, it looks like the GOP is about to get a dose of real fracturing, because Tucker, MTG, Nick etc are looking into creating a 3rd party which will clearly split the ticket in national elections.
There are exactly two DSA members in the house and zero in the Senate. A total of 30 DSA endorsed candidates (most not members) have won congressional primaries, almost entirely in uncompetitive districts. There are 0.12M DSA members vs over 45M registered democrats.
Doesn’t the fear-mongering and fantasy doomsdayism ever get tiring? Fourteen years ago your fantasies were that we were destined for Sharia Law within a year if we lost the “flight 93 election”. Grow up.
Nuts.
Dumbest take I have seen in a long time.
There aren’t even enough votes to remove trump from office. He could be impeached but anyone who didn’t fail civics class would know that it takes 2/3 of the Senate to remove a sitting president from office. Not happening even with a Democrat landslide in November.
It’s more likely a coup will come from MAGA. They’ve already had Jan 6th so the precedent is there. Luckily though Trump is ancient and in a mental decline, him making it to 2028 is very unlikely. Not that whatever comes after him; maga-remnants or corporate democrats, will be any better. The country is lost to the rich elites.
Only safe bet is to do what the rich do. Plow every last bit of money you can into investing.
Leftist “fever dream ” btw, the 1st coup was attempted in 2016 and, how’d that work out ?
Who was it that attempted a coup again?
Wolf, what happened to your readership?
I’ll check back in a couple of years.
One Hundred (100) year bonds: A hundred years ago we had tubes and were 20 years from the invention of the transistor; Charles Lindbergh had not yet flown across the Atlantic. Don’t these fools even know Moore’s law on doubling transistor density. So mind boggling can’t even wrap one’s head around it. Only potential speculative answer is pension funds forced into buying these bonds, by some mechanism or “deregulated” lack of safeguards, may shed light on this situation.
Moore’s law was an observation of transistor -quantity-, not density.
Transistor silicon is only a small part of the total expense required to deploy a data center; the land / building / electrical / cooling all have their own lifetimes that are indifferent to whatever chips get put in the server racks.
It may surprise you to learn that Google makes a little money outside the AI industry, which is probably why some lenders feel safe buying its 100-year bonds.
government spending drives inflation. Private sector investment does not. government can balance its budget by printing money. Private sector has to borrow from the debt market. Higher debt means higher interests, which crowds out other investment.
“government spending drives inflation. Private sector investment does not”
Places that don’t even have governments have inflation. Read some history.
Cash Man ” The road goes on forever the party never ends.
Wolf,
Great article, and it’s the part nobody is saying out loud.
Could you possibly do a follow-up article on circular financing? I think many readers will also find it enlightening.
Thanks!
Probably not. I get vertigo just thinking about it.
Look, this thing is going to end in tears for investors. But AI is amazing. The Dotcom Bubble gave us the internet as we know it today. And the Dotcom Bust then wiped out thousands of companies. And the Nasdaq collapsed by 78%. There were a few survivors though, and some of them are huge companies today, such as Amazon (which was considered on the verge of not making it during the Dotcom Bust). Something like this will play out again. Deja vu all over again. sorry.
Increasingly seems like the most logical outcome.
As I’ve said, I’m using AI all day every day in a tech/biz ops management role, and it genuinely both makes me 10x more efficient, but also unlocks things I couldn’t do in any time without going back to grad school several times and hoping I actually remember everything I learned. Most of my day is analytics/dashboarding, ops optimization tool-making, etc., so understand it doesn’t have the same direct impact on our warehouse forklift drivers, though in principle that’s the kind of role that could be literally replaced by autonomous pallet jacks in the medium term.
That said, I also get pushback from finance of the capex spend for AI. Per stats, I’m at a roughly $10/hr expense rate, against my ~$75/hr loaded salary. From my perspective, that’s a bargain given the 10x lift, but from their perspective it’s just incremental cost. Which is true, so long as my efficiency doesn’t increase revenue through growing demand or increase profit margins by lowering costs.
This is the crux of AI as a business.
Can it increase aggregate demand? Questionable IMO. It feels like we’re well beyond the limits of incremental value through expanding consumption in the West. There are only so many products and services that can be consumed within human waking time and without running into “paradox of choice” style scenarios. You quickly start running into zero-sum scenarios where every new sale is based around poaching someone else’s customer at a cost. Great for the consumer, but not for business profitability. The steps we’d need to take to change this are societal, like reducing the workweek to allow more time for consumption etc. Essentially become France.
Can it lower costs? 100% and it really is the case that labor costs are the obvious thing it can replace. Our company has an 8-person finance + accounting team. I bet this could easily be reduced to 2 or even one person with an embrace of AI, plus the replacement of various single-purpose SAAS subscriptions. Of course, the issue then is that you’ve put people out of work and they aren’t increasing demand unless we radically change how we distribute in society.
AI is really like the industrial revolution in that it’s going to force societal changes that are huge and impossible to predict. The internet is going to feel like the age of whale oil vs electrification–an interesting blip in history. Being an hourly worker in 2100 is going to be like being a farmer in the 1800s.
It was from last year, but are you familiar with the study from METR that showed that many people thought they were way more productive with AI but in fact were less?
I don’t know you or your work, but just putting it out there. For some people, I’m sure they are more productive. Others are not, and use it to be lazy and produce crap work.
What happens in/to a society with very very high unemployment and no opportunities? It probably is one not worth living in and likely dangerous as well. Wealth has to be spread around, or it will be taken away. Eventually.
No hourly paid jobs? Here is a salary job openings forecast, security details. Or, the military…. using debt to pay for it.
Elon Musk travels with a private security detail of up to 20 security personnel, including armed bodyguards and a medical professional. His team—who use the code name “Voyager” for him—operates similarly to a mini Secret Service, and select members have been deputized as special U.S. Marshals.
I would not trade one minute of my life for extreme wealth.
@tsonder
It’s a fair question, and I’m sure for a huge chunk of the workforce it’s not a huge improvement, as they are already far from the productivity threshold. Those that haven’t been pushed by responsibility or laziness to become more efficient before will have their jobs eliminated–the accounting team I mentioned in the post, for example, who aren’t even familiar enough with excel formulas that were released over a decade ago that would already make their day-to-day work much faster and less error-prone.
I work in operations/biz management for a food manufacturing and delivery startup (think Factor or Chef Unity), having worked in operations/biz/tech jobs at International Paper, Amazon, Wayfair, UPS, Microsoft and some other startups. I have a masters in Supply Chain management from MIT, a masters in analytics from Georgia Tech, and an MBA from Notre Dame. I’m the kind of person with over $2M net worth at 44 earned entirely within the last 15 years through salary and compounding investments. Being employed remotely in three full time jobs during COVID and netting $400k annually for a couple years helped.
Three examples from the last week. In how AI has saved me dozens if not hundreds of hours.
1) We cook food in industrial kitchens and pay for thousands of cook hours every week across our facilities. We cook over 250k meals every week across 40 skus. As you can imagine, the scheduling of that cooking is very complex, to account for how many people to bring in on any given shift to do the cooking given our variable volume, you have to ensure certain items are done before others to allow for shipping timelines or use in downstream meals, the industrial cooking equipment has capacity constraints, cleaning schedules vary weekly based on allergens used in meals, and certain items can’t be completed too early due to spoilage risk.
It’s an uncomfortably manual process now, mostly in the hands of the kitchen managers and visibly inefficient. To do this scheduling properly, you need to develop a constrained optimization model (if you’ve ever used solver in Excel, this is the best known consumer-level example).
We’ve already spent hundreds of manager hours on meetings simply scoping this project over the past year, and the engineering team was hopefully targeting a Q1 2027 release of an MVP candidate after hundreds of hours of expensive engineer labor. I started vibe-coding a version using Claude three weeks ago and presented it to the CEO today. We’re rolling it out to ops Sept 07. I logged my time and it was 37 hours and 12 of meetings with ops staff to develop a 10,000 line python script. I know enough python and have enough of a background in optimization to get a sense of what the AI is doing, but could never wrote that level of validated code myself without going back and getting another graduate level degree in CS.
Of course, many companies have teams that manage this sort of optimization software–for example, my friend at American Airlines works as part of a team of 600 that manages their scheduling software, and he knows the reaper is stalking.
2) Along the same lines, yesterday as part of developing training data for that optimization model, I’m reviewing the data we have on throughput rates for making meal components. Imagine thousands of entries going back years in time with a start and and time along with quantity for when we cooked beans and rice. I want to know what rates I should use for planning, but want to eliminate outlier values where there was an issue (equipment down, employees jerking around, whatever).
Before I would have met with someone in kitchen management, someone else on my data team, tossed around ideas, scheduled follow-ups, googled outlier theory, maybe looked into my old school notes, basically many, many hours of paid labor and probably a week or two to get to a theoretical solutions, then tried to figure out how to implement that method in either excel or python. This time I explained the problem in about the level of detail I just did here to Claude and pointed it to the spreadsheet with the data. In 45 seconds it came back with three options briefly explained, all of which I was familiar with from my studies, but would have had to brush up on a decade on. I told it which I would prefer, and it implemented it in both excel and python 30 seconds later, literally writing the formula in the cell and the lines of code in the script.
We have a belief that more recent data is more accurate and realistic than older data. I told it as such and again it suggested a time decay model for a weighted average and implemented it within literally two minutes total.
You could say, of course these are “coding” problems, but it’s really just basic business work and having your own partner who has PhD level knowledge of multiple domain areas, always communicates clearly, and can work at lightning speed at any time of day or night.
3) The engineering team deployed an AI took that crawls every night through the roughly 1,000 Google sheets docs we use across the company, looks for anything “suspicious” that might be an error, and sends a list describing the issues to sheet owners. The first run it identified 2,800 errors. We’ve gone through 1,700 of these errors in the past month and so far 77% were legitimate errors, with a surprisingly large share being “fat finger” mistyping errors in formulas completely invalidating the calculations being made, or simply poor formula writing practices.
Obviously these are “low hanging fruit” type stuff that will reduce over time, but this will free up tremendous time to get to more marginally valuable tasks that have been in the backlog complete. In the back of our mind, we’re all worried about what happens when the backlog runs dry and we have to reduce our current staffing levels, but for right now it’s absolutely exhilarating to have a superhuman working with you.
@Paul S
“What happens in/to a society with very very high unemployment and no opportunities? It probably is one not worth living in and likely dangerous as well. Wealth has to be spread around, or it will be taken away.”
I understand the concern, but I think you’re really overstating how important paid work is for most people. I am working in a very comfortable job for very high wages doing relatively interesting work, and if I won the lottery I wouldn’t even call in to say goodbye. Goodness knows that’s the case for most workers who are spending their day cooking in our kitchens, standing around the gap folding clothes, or out on the road at 3am driving through north Dakota for the 600th time to deliver a load. Workforce participation has been falling at an accelerating rate for years now, as have hours worked globally. Every white collar person I know is trying to engineer early retirement through FIRE-type strategies.
We shouldn’t forget that working for pay has only been the norm as an economic arrangement for about a hundred years and during this transition from mostly agricultural labor to mostly manufacturing/service, people made similar dire arguments about how not being in touch with the land or working on someone else’s schedule or being at the mercy of capital was death for the human soul and society.
I personally find it fascinating how many people you’ve heard bemoan the loss of second industrial revolution manufacturing in the US. Have they never listened to a Bruce Springsteen song or been inside a factory? It sucks. It’s loud, bad for your body, boring, etc. Many people stuck there worked double shifts in hopes that their kids would go on to college and have opportunities to avoid that life, because again, it’s a poor way for a human being to reach a state of flourishing.
As for where this all ends, as much as I find him an infuriating, fraudulent sychophant, I think Musk is onto something with his prediction of material abundance alongside something like socialistic redistribution as the future. Of course, we’ve had that as a utopian fantasy for over a century and plenty has been written about whether that would be good if it even happened, but I personally think most people would be pretty cool trying out the universe of Brave New World despite it being written as a warning.
Interesting detailed explanation of how you are using ai. But why do you think management is pushing back so hard if it is making you 10X more productive?
How is your productivity affecting the overall business?
Have they just not caught on yet?
And how does that accounting team keep their jobs?
Agree, some appear incapable of accepting a good/great idea can be seriously over-valued.
The problem is that no one is even remotely in a position to evaluate what the technology’s ultimate costs will be or what it’s ultimate outcome will be. We’re SOOO early that we can’t imagine the impact this will have on society.
Most people dismissed Karl Benz when he strapped an internal combustion engine to a wagon. Who would give up their beautiful horse with a name and ate free grass for an infernal auto was unbelievably expensive and relied on an exotic fuel and broke down constantly and needed tyres from some tree in Africa and was loud, etc.
Those people, and frankly Benz himself, didn’t foresee the invention of the assembly line, the development of industrial asphalt manufacturing and application, synthetic rubber, electric traffic lights, the supermarket, population growth that would require urbanization and the impossibility of removing the amounts of horse manure produced, etc.
Heck, in less than ten years most dismissed electric cars as impractical and now we have to impose 100% tariffs to protect the big 3 from getting their lunch eaten.
Yeah sure you can do some sort of short term accounting to argue Anthropic or whomever is overvalued by the market and there will be some sort of washout recession. That happened many times in the auto and aerospace industries, and as wolf has mentioned, the dot-com bubble, but this is almost immaterial in helping predict whether AI will end up like the auto industry of the 1910s or the nuclear and space industries of the 1950s/1960s.
Much will be determined by future technological advancements that we simply cannot predict. I do know that the smartest and most powerful people who are also deeply involved in its development seem to be very much “on the bandwagon” of it’s going to change the world, which is interesting if not telling.
AI is amazing. From an investment perspective it might be better to play the pick’s & shovels, commodities and some of the strategic 493 “other” (non-Mag-7) S&P 500 stocks that will be able to lever AI to reduce costs/boost profits/destroy competition. The models are so powerful already there seems to be a LAG of understanding and ability to leverage them in the non-tech companies. Those that catch up quickly and integrate into their core business are going to print money …. IMO
Is it amazing if users have to pay full price for what it actually costs? This is a serious question.
Well it costs either almost nothing, or massive $$$
This is the issue, you can buy hardware and run a pretty good local LLM and you know your costs and ROI and capabilities. Its even getting constant upgrades with new models and optimisations arriving all the time.
Or you can buy off a hyper-scaler and have no clue on anything, because all the numbers are obfuscated and as company valuations rise the investors need more ROI, and so as a business you may be expected to foot those costs to meet investors silly expectations.
Just because someone is over-excited and willing to pay $X for the stock which means they need Y ROI, which sets your price as a user, is backwards.
The market is flooded with models and compute and improves day on day. Some point soon consumers will set the value and valuations will have to adjust accordingly.
Yes, if you consider the fully loaded costs of highly educated workers it’s dramatically cheaper on an output basis, even at today’s prices. But today’s prices are absolutely going to decline rapidly. Remember that the core of AI is three things
1) huge amounts of data, largely already gathered by scraping the history of the internet. Sure you need to keep feeding it the latest news, but we’re taking about filling potholes, not rebuilding the interstate system from scratch.
2) the math of the fundamental neural-network underpinnings and transformer math. Of course this will continue to be refined and perhaps step changes will happen, but so far you can just keep pouring more resources (data+compute) into existing algorithms and get large incremental improvements. Of course the near term question is whether this approach will ultimately lead to AGI that can improve itself and get to 101% human intelligence and reach “take off”. It may not, and it’s “just” a technology on par with the printing press, electrification, etc. It may and we’re suddenly faced with genuinely metaphysical questions.
3) the equipment to push the data through the algorithms. Nvidia got very lucky that it’s video game cards just happened to be able to efficiently do matrix multiplication much more efficiently than traditional x86 chips, but we’ve not yet even seen chips designed specifically for AI workloads. Google has already said in SEC-auditable statements that its beta in-house designed AI chips are 6x-10x more efficient-per-watt than Nvidia’s latest, and coming in mass production early 2028. Obviously AMD and Intel and every other hardware player has every resource at their disposal on similar efforts. The energy crunch is going to be a very near-term issue as most of these projections are essentially assuming that the 13 mpg of a model T Nvidia won’t be more than doubled by a VW Beetle in 30 years and doubled again by a Prius.
In terms of “picks and shovels” I’m thinking about electricity futures.
What is the energy source for such electricity generation? That is the problem, not the need for more electricity….the capability. There probably won’t be anymore from Canada for a long long time.
One example: “Manitoba announced it will not renew 500 megawatts of expiring hydroelectricity export contracts to the United States, choosing instead to redirect the power toward domestic and Canadian nation-building projects amid shifting trade pressures and local energy demands.”
Then the lights might go out in your sacred BC when BC Hydro can no longer import electricity from the US. It became a net importer of electricity with the US for a while not too long ago when the reservoirs ran low.
The US and Canadian grids are connected. A new interchange was just completed. When grids are connected, they become more resilient, and BC Hydro was a huge beneficiary of that.
100% on the lag issue. Part of the issue is that many workers are already so behind on even basic computer productivity skills–columnar excel formulas, building something themselves rather than just doing some step-wise process in some idiosyncratic SAAS suite, etc–that they are getting left further behind, because at this point AI is mostly doing player piano on the computer rather than making the computer invisible.
And frankly, it makes those of us that can do even basic software so much more effective that I don’t want to spend my time teaching a coworker how to do xlookup. Their productivity is much less important to me now in my own and the business’ success as I’d rather just disintermediate around them.
It’s amazing that investors are lining up for CapEx rates that are practically doubling year over year while revenue growth is around 20%. Google also has negative free cash flow for Q2. A
There are some articles indicating that at least 50% of the revenue increase are from Anthropic and OpenAI, two companies that are burning through cash including investments from the mag seven. Lastly the WSJ reported that the balance sheets and free cash flow is hiding the massive stock based compensation for AI gurus because it’s not realized until it’s vested. I call shenanigans!
According to Epoch AI the cost of a data center is 60% the servers in them. The buildings and their various systems depreciate, perhaps on a 15-20 year schedule, but the servers themselves will be on the fastest depreciation schedule.
If we assume MACRS 5 year depreciation, this half-trillion in AI spending per year – on just the servers – is going to reappear on big tech’s income statements as massive depreciation expenses. Under a 5-year MACRS schedule, 20% gets written down in year one, 32% in year two, 19.2% in year 3, and so on.
Multiply those percentages by perhaps a half-trillion dollars in server-only spending, and it appears big tech is setting itself up to report negative or near-negative earnings for the next few years.
E.g. Alphabet reported $134B in earnings in 2025, but if they’re putting $200B per year into Capex and experiencing, let’s say, an average 20% per year in depreciation on their last couple of years’ capex, then I can see a couple years of spending at this pace turning into a (50B x 2 =) $100B per year depreciation expense. Thus, depreciation alone will erase most of their earnings for the next few years.
Wolf’s final points are the most important here. Big tech used to write copyright-protected software that could be replicated at virtually no variable cost and yet still sold at very high margins. Big tech used to produce their products very lean – often for the cost of payroll – but now the companies are transforming themselves into an asset-heavy, debt-loaded, commodity producers attempting to earn enough revenue to keep up with massive depreciation and energy expenses. They themselves seem to believe the only moat available in AI will be scale and the ability to keep up with the need to constantly spend more money on upgrades. Margins could be thin unless a monopoly or duopoly emerges.
This asset-heavy, high-input-cost business model looks more like a steel mill, a refinery, a trucking company, or a railroad than the software companies of five or thirty-five years ago. This detail is being missed because it is the software companies transforming themselves into a new industry.
It’s not the emergence of an exciting new tech industry; it’s the death of an old model of tech that drove up stock index margins, earnings growth, and efficiency for decades. It’s as if the fast-growing, high-margin, asset-light, low-debt tech companies all decided what they really wanted to do with their lives was manufacture trucks, or transport rocks, refine oil, operate mines, or forge steel.
So suddenly all these companies that once sold for >30x PE ratios now seem to deserve single-digit PE ratios, like we are used to seeing from companies like Nippon Steel, Ford, Dow Chemical, or Diana Shipping, where depreciating machines financed by debt do most of the work instead of people.
No company would voluntarily choose to trade down this way. The truth is the software companies were up against a wall. AI was going to commoditize the software/SAAS business anyway by writing code more cheaply than humans ever could, and software companies could either die or become low margin, asset heavy AI mills. They chose the latter path, but I do not think investors understand what has been lost.
In terms of sustainability, massive cash flows from the old software-writing business model is, in addition to new external investment, currently paying for the very expensive transition to becoming AI mills. But the transition is necessary because the old software business model is about to end. So the flow of cash available to reinvest into AI will falter unless cash flow from AI picks up.
Currently, we’re in a happy transitional period when cash flows from both software writing and AI subscriptions/tokens are boosting tech companies, allowing them to spend more on AI. But one of the following must be true: 1) AI will eat the software industry and thus is fairly priced, or 2) AI will NOT eat the software industry, and thus is a bubble. In either case, big tech loses something worth trillions, relative to what they have today! Either they lose their software businesses or their investments in AI fail to pan out.
An AI optimist would say cash flows from AI will pick up soon, and the companies that invested the heaviest will be able to name their price for still-too-scarce AI compute, guaranteeing cash flow and margins.
But there’s also a chance that the AI industry will need years to mature, or will go through a longer-than-expected “freeware” period like we’re experiencing now, until consolidation, price hikes, and enshittification occur. Any one of these possibilities could lead to a long period of bleeding in the tech industry, and that’s if investors don’t panic and flee.
But IT now constitutes almost 37% of the S&P500. If IT companies eventually turn into marginally profitable heavy industries with single-digit PE ratios, then how do the indices go?
Well, AI is useful to summarize thesis paper long comments!
Very good summary of where we are.
If nothing else, AI makes software development exponentially cheaper, and big tech has little choice but to try and lead the change in the short term. The people managing these companies all have MBAs and been forced to read the cases of Kodak ignoring the digital camera, Xerox ignoring the PC, and Nokia ignoring the smartphone.
That said, weird moats persist that are hard to understand. I personally would have predicted Apple would have lost its dominance with the iphone and Microsoft its with Office as the much cheaper products from Google are equivalent or better.
I think we’re about to enter a major economic transition where deflation is the dominant trend (see your comment about market values) that will disproportionately damage holders of certain kinds of wealth, but that the vast majority will lead much, much richer lives.