keeda

From the outside, Microsoft seems to actually be one of the better places, and is making at least token efforts to reduce headcount in more amicable ways, as exemplified the buyout offer.

But in general, the strategy seems to be overload people until they burn out and quit, because hey, AI will replace them anyway.

Not quite, what I'm saying is, a single reference or source may not be comprehensive or accurate or reliable, so it's best to get a better idea by putting together multiple. We still won't know everything but we can guess much better.

> So who from "outside" throw money "into the system" ?? Certainly it not coming from selling inference.

But that's what some of the sources show, like the blogs from various VCs and payment processors (which report actual revenues exploring) along with multiple industry and government surveys (which show widespread increasing demand across the board.)

Now these are not enough to provide an estimate of how much money though, especially given the circular funding, which is where the Exponential View report helps with some numbers on the deduplicated revenue flowing through.

And note, these are independent (and often competing) parties, which makes it more unlikely they are all colluding. This is what I meant by having to piece together multiple sources of information.

> But making it artificially scarcity and cashing on it is not good heart doing and nowhere close to goodwill.

I would agree, but I don't see any signs of artificial scarcity? There is a ton of evidence that the scarcity is real, the main ones being the hyperscalers themselves who are putting triple digit billions of their money where their mouth is... that mouth being firmly attached to Nvidia's rear ;-)

For 1) the report cannot determine an exact split between VC-backed startup vs other revenue because the startups are private. What it focuses on is deduping the revenues, so the headline number of $110B - 175B annually is the outside money flowing into the system, and not circular flows.

However that outside money includes revenues, VC-backed cash, and training costs, which cannot be disentangled by this methodology. But note:

a) That does not imply circularity, as the apollo.com link suggests.

b) We also can't assume this is mostly from burning investor cash rather than actual going concerns. Large VCs like A16Z, YC and Menlo Ventures, without disclosing margins, report seeing dramatic revenue growth in their startups, e.g.

https://a16z.com/revenue-benchmarks-ai-apps/

https://www.cnbc.com/2025/03/15/y-combinator-startups-are-fa...

So while often VCs are also investors in the frontier labs, the flow of money here is explicitly not "hyperscalers -> startups -> frontier labs -> hyperscalers," but rather "VCs + revenue -> startups -> frontier labs -> hyperscalers." And VCs report that revenue is the fastest growing in history, which is corroborated by all the other links measure AI adoption growth.

Furthermore, this $110 - 175B number also gels with the $177B number that the a16z link estimates.

For 2) the slides are much better for this (and the caveats are important) but the simplest I can put it is, at this point the marginal cost to serve tokens is far below and falling faster than token prices. Each new generation of hardware drops the cost to serve even further. As above, this rhymes with other sources including Anthropic's S-1.

Regarding your point about the 2008 GFC, the difference here is that their individual incentives are not at odds with the global good, because at this point it appears to be bad for everybody! In the GFC there was no real pain until the crash; here the pain is all around from the start. The hyperscalers are actively torching their extremely profitable cash flows (along with a lot of goodwill) and taking on heavy debt, despite being punished by the stock market, to fund this buildout that apparently everybody hates but can't stop using.

Firstly, I would just ask, from a game theoretical perspective, why would these companies, all bitter competitors, most with very high fiscal discipline and absolute firehoses of double-digit billions of free cash flow per quarter pre-AI, actively do something that would be so destructive to themselves and the world economy?

Secondly, there are a ton of sources indicating significant growth and revenue coming from outside the AI industry. A few references:

1. The State of the AI Economy is the most thorough evaluation of deduplicated AI revenues I've found: https://www.exponentialview.co/p/the-state-of-the-ai-economy -- see especially the linked PDF report. I'm not a financial analyst, but they seem to cover all the bases. If nothing else, you should read this, it paints a very nuanced picture.

2. Enterprise surveys:

- https://knowledge.wharton.upenn.edu/special-report/2026-ai-a... (Has links to reports from prior years)

- https://menlovc.com/perspective/2026-the-state-of-enterprise... (Enterprise AI deployment spend up to 177B, 5x YoY)

- https://a16z.com/state-of-markets-ii/

- https://www.deloitte.com/us/en/what-we-do/capabilities/appli...

3. Government surveys:

- https://www.federalreserve.gov/econres/notes/feds-notes/moni...

- https://www.genaiadoptiontracker.com/?tab=genAIwork&view=ove... - Toggle "Overall" and check the drop down for the variety of industries using GenAI; apparently 40% of "Agriculture, Forestry, Fishing, and Hunting" workers use GenAI. Also relevant: one tab over shows how much work hours they use AI for -- still only at ~5%.

- https://ec.europa.eu/eurostat/statistics-explained/index.php...

- https://www.oecd.org/en/about/news/announcements/2026/01/ai-...

4. Payment processor data:

- https://stripe.com/guides/indexing-the-ai-economy

- https://ramp.com/data/ai-index (also here: https://fred.stlouisfed.org/blog/2026/10/which-businesses-ar...)

This link mentions a bunch of the above and more: https://www.ryzlabs.com/guides/enterprise-ai-adoption-statis...

There are many more of such sources out there, Google AI overview is pretty good at surfacing them.

As another data point, these backlogs were at ~0.5T in 2022, before the AI boom started. Google for instance registered a 390% growth YoY in backlogs.

That’s not too far from the trillions being invested in the buildout.

I would say the irrationality of this buildout is exaggerated. It makes more sense to look at it as these companies making an existential bet on their futures.

> Tesla shares close at record high, boosted by 69% pop since Trump election victory

SPCX was not public back then, but I suspect it would have followed similar trends. In any case, Elon's still come out ahead. For now.

I guess we could look at it as going through a world-wide Code Red...

What I've found is that even the best models can write very good code "in the small", and the code is mostly correct, but they can be highly problematic in their design, architecture and abstractions they use. The most common problem is that there is a strong tendency to overcomplicate things, and often this can be traced (ironically!) to the agents not having enough context.

E.g. a very common failure mode is that agents are unaware of some invariant in the architecture (or the fact that they could potentially stipulate a new invariant) and write code to handle a dozen edge cases that would never occur. The code is correct, but there's a ton of unnecessary slop.

My suspicion (confirmed by friends and posts on HN) is that at most places, people don't bother reviewing the code and blindly ship all this slop. The reason things haven't imploded yet is a) the code is usually technically correct, and b) all that additional complexity is never exposed to humans -- it is just outsourced to even more agents to manage. Which explains Anthropic's exploding revenue, I suppose.

That said, it's not like things were better before AI. I once had to go Indiana Jones-ing into a humongous legacy Python codebase so full of magic that there were some code paths I could never figure out even with a debugger. I could only tell that if I got a row written to a specific table, some completely disconnected code path elsewhere would get triggered. I suspect what was happening was a function name was being constructed out of 3 different strings from 3 different sources, one of which was a column in the DB row that was getting written, and that function was being eval'd in some thread somewhere.

This was to fix a bug, which I managed to do without ever figuring this flow out. I still spent an inordinate amount of time trying to figure this out before I realized I was being too idealistic and closed the ticket and moved on. I bet an agent would have figured it out though.

It’s still early days, true, and courts may find legal reasoning to address this... But I fear this stems from a fundamental yet little-known flaw in copyright for code, based on a legislative hack, which can only be made worse by caselaw. I ranted about it a few days ago: https://news.ycombinator.com/item?id=50320875

> VerInf proved Llama 4 Maverick (the same 400B-parameter MoE model we prove) at ≈0.02 tok/s.

The world didn't end. It was a complete black box, but we followed the runbooks and kept things running, and only learnt enough of the code as needed to fix any issues or add any enhancements that popped up. Diving in, doing some software archaeology, and spot-changing relevant code became a skill we had to develop.

At scale and over time, I would guess that the percentage of code running on machines that is actually understood by anybody approaches 0. Yet the world does not end.

A 1000 of us regular devs scanning a piece of code is not as effective as a single review by, say, Tavis Ormandy. So "enough eyeballs" is only "enough" when one of those set happens to belong to a Tavis O. or the likes.

Not to mention, this assumes a 1000 of us even look at the code, which is statistically never true. Approximately 100% of us just install the software without looking at a single line of code. Hence the code auditability aspect of open-source software never really applied. And hence, "supply chain attacks."

Unfortunately, this misguided "law" was taken as gospel by many, and as you found out, still is.

NVDA is down a couple %, ORCL is mostly flat. It’s the neoclouds that have taken the most beating, down 10%+

I wonder if people know who the data centers are for…

It is actually, but the shady part is that this is not (just) to target ads to you, though that certainly happens… it is, as OP explained, so that they can better capture query intent AND your behavior (clicks, dwell time, subsequent searches, etc) that they then use to refine and rerank search results, which improves Google Search and keeps you using it, which becomes a virtuous flywheel that competitors can’t catch up to.

One way for competitors to break through is to buy traffic, but well Google blocked that too by using its monopoly profits to pay more than anybody else (Apple and Mozilla search deals.)

All of this came out in the Search Antitrust trial. Look up NavBoost for instance.

Sure 2.33T is less than 3.2T but these RPOs were $0.5T in 2022, and have been growing YoY at rates of anything from 84% (Microsoft) to 390% (Google.) Microsoft and Google alone are expected to recognize about $0.5T of that revenue over the next 12 - 24 months.

Is the CapEx and debt ("hidden" or otherwise) growing faster than the RPOs? Eyeballing the numbers, maybe? But generally service providers try to stay ahead of the demand curve rather than behind, so...

Is there concentration risk in the customers? Yes, OpenAI and Anthropic seem to be like $1T of these RPO's? That still leaves $1.3T of other demand.

Is there demand? I mean, RPOs come with a commitment, and have you seen the cloud revenue growth numbers recently??

As another comment here points out, is it really wise to assume the tech giants are NOT financially savvy enough to know what they're doing? Some of these, especially Amazon and Microsoft and yes even Alphabet, are well-known for being highly fiscally disciplined. This is a huge bet for them, even an existential bet, but maybe it's because this is an existential kinda situation.

Like, before this you had extremely specialized models being painstakingly trained for an extremely specialized task using mountains of specialized data, painstakingly collected and labeled for correspondingly specific features, typically done by ML engineers (a specialized discipline) further specialized in that problem domain. And most users only have ever experienced all this abstractly through extremely narrow and specific functionalities like recommendations or ads.

On the other hand, any random person can hand these GenAI models any task that can be expressed as words or images or sound or some combination thereof and have a shockingly good chance of getting useful results. Yes they sometimes contain mistakes, but if you simply point them out, the models can often fix them!

I don’t know how much more general things can get than that.

It does get a bit uncomfortably close to “AGI”, which is a vague, heavily contentious term. But this sidesteps all that by not calling it AGI. The conventional “Generative AI” interpretation of GenAI thus becomes a convenient decoy for “General AI” ;-)

https://en.wikipedia.org/wiki/Robert_Metcalfe#Predicted_Inte...

Copyright very explicitly does NOT cover functional aspects, only creative expression. But the value of software is not in the expressive aspects of the code that generates it, the value is in the functionality it provides. And this is the exact thing that copyright does NOT protect!

Like, literally, we're protecting the fact that we used a for-loop instead of a while-loop to do a useful thing, not the useful thing itself. This is why rewriting enough code is sufficient to bypass all copyright protection, as the recent AI rewrites have shown.

This was always a huge legislative hack, rightfully controversial at the time it happened, but legislators didn't want to bother figuring out a proper IP regime for software -- maybe something between copyright and patents -- so they shoehorned code into copyright.

Google v Oracle was trivial to decide, but it got decided the wrong way. Because it became a trial around laws based on an ill-founded legislative hack and the decades of flawed case-law it spawned. Stepping back, it is very clear that Sun spent billions making a very useful thing and Google co-opted it for its own uses without recompense because they bet (correctly) that they could legally get away with it.

It is somewhat analogous to what the AI companies are doing with all the code and data out there. Funny that people are unhappy about THAT but cheered when Google did it to Sun / Oracle. I wonder what changed ;-)

My simplistic narrative, based mostly on the timing of market I observed, agrees with yours: the tech selloff was sparked by Google's earnings report. Mostly the fact that Google's Search cash cow seems to be getting cannibalized by its less profitable AI overviews, compounded by their cash flow turning negative due to their insane CapEx bill, which is only going to keep going up. And then Amazon reported a similar plight.

Hilariously, this even affected MSFT, which actually had a good quarter while being the only hyperscaler with positive cash flow, initially sending the stock to record highs. And then Google's earnings report hits and drags everything down. Satya must have been pretty annoyed.

Another way to look at it is, bureaucracy is an aspect of the corporate culture, and lasting security comes down to having the right culture. At least in that respect Microsoft is doing something right. Doubtlessly a lot of that bureaucracy is ineffective, but it's a signal that the cultural intention exists.

As a relevant counterexample, we've never heard of these kind of processes at Meta and so you get Cambridge Analytica.

My perspective is skewed by my experience being on a security-adjacent platform team for years. Our biggest challenge was not external threats, it was internal teams seeing us as friction on their productivity, to the extent that it regularly led to escalations.

So when Cambridge Analytica happened, I played it up for all it's worth; Meta may move fast because they didn't have our "bureuacratic team", but look where it got them! Unfortunately the $25B penalty happened just last week... I'd have gotten infinite political mileage out of that.

Of course in the short term we have to deal with pesky things like rent and food, for which we need jobs that are little more than filing TPS reports, which keep us from making progress towards our destiny…

But in the long term, after the economy implodes because AI starts filing all the TPS reports, and we rebuild it for a future where humans are actually allowed to reach for their full potential, we’re gonna need all the intelligence we can get!

Either way Anthropic will IPO soon and we’ll be one step closer to finding out.

One thought I had is that, bubble or not, this technology is likely going to upend economies of entire nations and impact all people, so it would be better if controlled by the people which is (in theory) represented by the government.

Looking at this as a backdoor bailout is the wrong perspective; it should be more like giving the people control over their economic future.

As is what constitutes the standard of “adequate.” Which BTW is the standard in ALL engineering (not just security or software) because we live in a world where economic pressures apply and costs must be managed. Trivial example: bridges only use an “adequate” amount of steel and concrete because they are not free. The word “adequate” fundamentally involves the concept of tolerances. There’s a reason we consider over-engineering a bad thing.

So which development approach would you say is better correlated with good security outcomes?

But they are also very highly incentivized to be more secure. They have millions of customers (hence a correspondingly higher expectation of trust) and billions in revenue from the trillions in economic activity happening on those platforms (hence a correspondingly larger target on their backs.) This means exponentially higher liabilities. And that probably matters more than any other factor.

I’m not even blaming Zuck for that, he didn’t really invent it, though he certainly popularized it — he just made an astute observation of the dynamics in software and turned it into a company motto.

Like, in the theorycrafting in my comment that I linked before, I didn't assume any change in userbase; just how lucrative (or not) each step in a chat-based journey could be versus a search-based one.

Personally I have doubts that AI search would ever be as lucrative as traditional search, which I expanded on in that comment I linked.