Hacker News

points by dyslexit 3 days ago | 0 comments

To be more specific:

- 161.2B w/ Broadcom: 2026-2031

- 120.5B w/ SpaceX: 2026-2029

- 114.5B w/ Google: 2026-2033

- 110.0B w/ Amazon: 2026-2036

- 44.9B w/ Microsoft: 2026-2033

- 44.6B w/ Nscale: 2026-2033

- 41.3B w/ Lambda: 2026-2034

which is around $63.7 billion / year if spread over 10 years, but with most of it due earlier than that.

[–] smallmancontrov 3 days ago

Isn't the SpaceX contract cancelable with 90 day notice? How many of the others are too?

[–] dyslexit 3 days ago

You're right, SpaceX will be cancelable within 90 days after March 2027.

SpaceX was the only one mentioned as cancelable. These are listed under "Infrastructure Service Arrangements" on page 170.

[–] throwaway2037 3 days ago

That is wild. What if there is a huge break-through in LLM models where they can reduce compute by 90%? All of those data centers would suddenly crater (many customers cancel) and go bankrupt. I feel it could go either way.

[–] Silhouette 3 days ago

Surely the thing the AI giants fear the most in this respect must be the possibility that competitive open models and locally available hardware reach the point of viability for most people. If that happens then where is the moat for the giants?

If progress with frontier models doesn't continue at the same pace then the investments and funny money deals might start to look very fragile.

If progress does continue then that progress might be towards smaller models more than quantum steps up in quality or capability that would increase the risk of the scenario above.

All of this is set against a backdrop of public hostility and political problems that I can only imagine increasing in the near term as well.

This feels like an unusually high stakes game of poker. I don't think I'd want to bet on Big AI right now if I had a few billion lying around. Then again - maybe that's why I don't have a few billion lying around!

[–] BlobberSnobber 3 days ago

Well, they are ensuring that hardware prices stay extremely high by buying all the flash memory at a premium, so at least they've got that aspect kinda under control?

[–] Silhouette 2 days ago

I've made that same point myself in the past. It's more of an accidental moat though - and not one they can defend for very long if they're no longer getting value out of those components themselves.

I saw an interesting observation the other day that Jensen Huang has played this game the best because his company is the one actually selling real products and making real money in return. If all the AI giants are buying up huge amounts of hardware speculatively because they assume they will need those data centres in the future and then the bottom falls out of the market then it's the AI companies left holding the baby. And if we do get to a point where local models are competitive then the OpenAIs and Anthropics of the industry might face real difficulties but it would just be a different market for Nvidia to also be selling real hardware and still in exchange for real money.

[–] mike_hearn 3 days ago

That's already happened. It's one reason prices have dropped so much lately (n.b. in the quarter that's not in their S-1, I think!) They've been able to improve model efficiency a lot through better algorithms and training models to be less verbose.

[–] vlovich123 2 days ago

Training costs haven’t plummeted like they have for inference afaik and the efficiencies are just translated into running larger and longer training. Basically Jevon’s paradox in action just like on the inference side.

[–] edmundsauto 1 day ago

I am projecting but I think these companies are looking at their output (“intelligence”) as electricity. There’s never too much electricity, really - it just goes into lower margin uses.

For intelligence, I bet if inference got 90% more efficient they would just increase the “thinking” by 90%.

[–] acchow 2 days ago

We’ve already seen 90% reduction in intelligence costs multiple times over the last few years. Each time, token consumption rises 20-100x

[–] pierreb-aiva 2 days ago

Jevon’s paradox would apply. If I can use order(s) of magnitude more AI for the same price, I would absolutely do that. It would also unlock many use cases that are currently gated by model pricing, which is largely dictated by the price of compute + frontier intelligence markup.

And thats just the inference side of things. For training, the big players like Anthropic would still be able to benefit from cheaper compute, by generating more RL rollouts, running more RL environments, and thus still providing frontier intelligence that would justify the markup.

Many customers cancelling would only be a risk if the compute expenditure was more decentralized. But its not

[–] solyin 3 days ago

The question is, how're those datacenters build outs coming along? two reasons for Anthropic to cancel those contracts. A) They don't need the capacity because they have capacity elsewhere, or b) they don't need the capacity because no one is using Claude. (I guess c) they don't need the capacity because theyre running Claude more efficiently is also technically possible.) For B to happen, they've got the demand in the form of contracts, and models don't really seem like they're shrinking, though there's a lot of non-public info, Astra > insize than sol, and Opus > insize than a Sonnet. So then we're looking at situation A), and the think to look at is SpaceX has the vertical integration to be able to plonk down datacenters and get them powered with literal jet/rocket engines/turbines that they make from scratch because they're a space rocket company when no on else can. So, personally, I don't think it's that big a risk to SpaceX's revenue. Or rather, is cancelable in 90 days, but smart money is going to see those cancelations coming from a lot longer away, if you're tracking datacenter construction. (Disclaimer, I picked up some SPCX; not financial advice, etc)

[–] chrisweekly 3 days ago

> "SpaceX has the vertical integration to be able to plonk down datacenters and get them powered with literal jet/rocket engines/turbines that they make from scratch"

Is that really a viable path? (Honest question)

[–] burnerRhodov3 3 days ago

At this point, you can get the chips you just can't get the power. So in typical elon fashion, he just "solves" the bottleneck. Long term, no, that is not the strat. Thus the pivot to space.

[–] vel0city 3 days ago

What happens when he sucks the natural gas pipes dry in the peak cold of a Texas winter and thousands of people die to keep xAI burning?

[–] burnerRhodov3 2 days ago

I don't know why I'm responding to this, because I suspect this isn't being argued in good faith, but let's actually look at the numbers.

According to the U.S. Energy Information Administration (EIA), in 2024:

- *Texas natural gas production:* 27.5 billion cubic feet per day. - *Texas natural gas consumption:* 14.6 billion cubic feet per day.

So Texas produces nearly twice as much natural gas as it consumes domestically.

Natural gas contains approximately 1,025 BTU per cubic foot. Assuming 50% efficiency in a natural gas power plant, and 1 kWh equals 3,412 BTU.

27.5 − 14.6 = *12.9 billion cubic feet per day.*

12.9 billion × 1,025 BTU × 50% efficiency ÷ 3,412 BTU/kWh ÷ 24 hours = *approximately 80.7 GW of continuous electrical generation.*

That's the equivalent of more than EIGHT 10-GW AI computing campuses operating around the clock.

Lastly, a significant amount of natural gas is produced alongside crude oil. When producers can't economically capture or transport that gas, some of it is burned off through a process called flaring. Those flames you sometimes see at oil and gas facilities are an example, although refinery flares also serve important safety functions.

Seems like you just like to say snarky things about Elon, but can't be asked to do any actual math.

[–] vel0city 2 days ago

That comment was at least 75% snark, I agree.

I do agree and understand we flare a lot of natural gas, waste because there's just not the economics to transfer it elsewhere. I definitely agree, it would be nice for that to be put to some economic use if we're pulling it out of the ground.

> Texas produces nearly twice as much natural gas as it consumes domestically.

And it exports that other half it doesn't consume domestically. It's not like half of what is produced just sits around forever. We've massively scaled up LNG exports. We export to surrounding states. But hey just act like we've got billions of cubic feet lying around doing nothing ready to go for datacenters.

> approximately 80.7 GW of continuous electrical generation.

Ooh wow. There's 420.8 GW of demand sitting in the ERCOT connection queue specifically for data centers today. All that "extra" and we'll cover like 20%. A good gas power plant will do better than 50%, but that's still not enough. And I imagine there will be even more datacenter demand coming soon. And that's just at the grid connection levels, not behind-the-meter generation like what we're talking about.

I've also lived through continued cost increases of natural gas in Texas. What does adding an additional 420GW of demand for just datacenters do to our rates? How much more am I going to spend to heat my home this winter? How much more will I spend on my electric bill next summer?

I've also lived through the collapse of natural gas supplies in 2021. I wonder if these datacenters will turn off their turbines or let people freeze it something similar happened again?

Seems like you make a big point about doing the math but only doing half of it. Keep going, see where it takes you.

[–] solyin 2 days ago

Thank you for doing the math. North Dakota "solved" flaring by making it illegal and fined companies for flaring just a bit more than it would take companies to capture the flared off gas, thus making it economical to capture it.

[–] me-vs-cat 3 days ago

We'll be forgiven for expecting the new overlords will take over in the bang of boiling oceans, as we miss the significant whimper of a cold Texan winter.

[–] jordanb 2 days ago

I honestly think that the whole Memphis/Colossus debacle is what lead to the datacenter backlash. Before then I'd just hear rumblings from people complaining about bitcoin miners, but data centers in general were not seen as an odious land use.

Then Elon goes to Memphis, buys a warehouse and fires up 30 jet turbines in the parking lot, and you start getting reports about the incessant noise and being able to smell it from 5 miles away.

It seems like there's a fundamental problem with realizing other people have agency and if you fuck them hard enough they will fight back.

[–] wbl 3 days ago

SpaceX does not make turbines that can run for tens of hours producing megawatts of power. A turbopump is a different beast.

[–] burnerRhodov3 3 days ago

Not the whole turbine, but they are manufacturing turbine blades in bastrop texas. The turbines are available right away, but the blades are 12-18 month lead time, so Elon built a foundry. From what i heard they've only gotten a couple proto's out of the foundry/ no idea the grain quality. But they are doing it.

[–] ethagknight 3 days ago

all Anthropic contracts are cancelable if the intent is to stick public shareholders with a bag

[–] lumost 3 days ago

I'd imagine all of the vendors have an incentive to keep their main customer a float, while there might be penalties etc. for missed targets - it doesn't make sense for any of these vendors to ride the contract into chapter 11.

[–] runako 3 days ago

This is so important in analyzing these commitments. Anthropic's commitments will make them a Top 10 customer at many of their vendors. (Even at Google/Amazon, how many other customers do they have with $100B+ committed?)

Those vendors would very much like to retain as much of that business as possible. If the space goes sideways, I would very much expect some kind of renegotiation/work out.

[–] Mistletoe 3 days ago

This is always the intent for all recent IPOs I can remember.

[–] nostrademons 3 days ago

On a side note, these are pretty hefty portions of the total revenue of these companies. 161B with Broadcom over 5 years is about 32B/year; that's fully half of their 2025 revenue. 120B with SpaceX over 3y is 40B a year, more than double their 2025 revenue and almost triple 2024. 114B with Google over 7 years is about 3% of their total revenue and about 1/4 of all Cloud revenue. 110B with Amazon is about 9% of their AWS revenue.

If Anthropic's revenue story doesn't materialize and the AI bubble bursts, these companies are in a world of hurt. It's probably not fatal for Google or Amazon but would be a major drag on the stock price (perhaps why Google has dropped from 400 to 315 over the last couple months as the AI story soured), and it could be company-ending for Broadcom or SpaceX if they end up building out infrastructure to service these commitments.

I remember before I left Google, one of the TGIF questions (which are now AI syntheses of many peoples' individual opinions) was "What happens if AI is a bubble and it bursts?" His response was "If that happens, the whole industry is in trouble." I see now what he meant.

[–] brightball 3 days ago

SpaceX will just use the compute directly if they don't have buyers. They are already talking about the data center shortage as well because they can't get capacity elsewhere either. It's a weird problem to not be able to use your own data centers because you rented them out.

Between SpaceX, Starlink, Tesla, Neuralink, X, Grok, Grok Bot & Cursor there's not a demand shortage.

Grok Bot's the fastest growing product in the history of the company and the moment you use it, it's clear why. Took me about 1-2 days to realize I'd have a very hard time going back to not having it available.

[–] blonder 2 days ago

A software product you can access via website or app being faster growing than cars, rockets, and satellites is not an indication of it being any good. It takes a lot less to convince a consumer to use it vs buying a car or starlink.

[–] brightball 2 days ago

Once people start using it, the usage is constantly increasing as well.

[–] Uke 3 days ago

And what is it, that makes it good?

[–] brightball 2 days ago

A lot of things. The UX for non-code related activities is excellent. They've removed the friction from setting things up by giving each agent their own machine which can give you remote desktop. The process of creating and working with each agent is as simple as naming it and then explaining it's job. You can keep feeding them information, ask them to train themselves, etc.

They can access anything you can, whether there's an MCP/CLI/API or not. I create speciality bots for each 3rd party tool that I use at this point. The bots can talk to each other, have group chats. You can give them access to the machines that you're running the software on and give specific bots permission to do things on certain machine (like an Obsidian bot that writes logs of daily bot activity on a specific machine with a Dropbox folder shared). I've got a Security Scan bot that downloads anything that a bot wants to download to it's personal machine at SpaceX and runs a slew of security scanners before the other bots can work with it.

I got locked out of one machine while I was travelling and I asked Grok Bot on another machine if it could get into that one over SSH, so it sent me a temporary authorization link via Tailscale, connected to the machine and restarted the stuck process for me.

The usability is off the charts. It's really like having a team of people working for you. I've run a conference for the last few years and was getting completely overloaded. Now I have 14 bots specialized just for the conference helping me do it better than ever and participating in architectural discussions to build a system that help long term.

If you find yourself trying to juggle a lot in your life and work, it's a gift.

[–] majky538 3 days ago

Also interested in use cases. I tried, but went back to "normal" life and apps.

It's not impossible integrate AI to projects/products directly.

[–] w2sdss 3 days ago

No the biggest problem will be investors will demand much higher dividends aka cash return..... how will alphabet continue to fund its projects with ever decreasing internal cash flows available after dividends? If the founders decide not to come good on this - the stock price will crash. If much of your employees compensation is stock based...? i have bad news for you

[–] monegator 3 days ago

one can truly dream

[–] demibabs 3 days ago

wait what, how is Broadcom worth so much with so little revenue?

[–] nostrademons 3 days ago

Their current stock price is pricing in these contracts, which have already been announced but have not been fulfilled (or booked as revenue) yet. So Broadcom's 2025 revenue was 63B; note that this is already up almost double from its 2023 revenue of 35B, largely on AI contracts. If Anthropic spends another $32B in 2026, that's in addition to their declared revenue; investors expect they'll be making $100B+ in 2026. Broadcom's operating margins are pretty good, so a lot of this flows through to the shareholders as earnings, and so investors are valuing it on the assumption that these contracts result in tangible cash flow.

If the AI companies go tits-up, that introduces counterparty risk, the revenue doesn't materialize as cash, and suddenly their earnings take a nose-dive. If they've staffed up significantly to design and sell these AI accelerators and suddenly there's no market, they could be in a lot of hurt.

[–] foobiekr 3 days ago

AI aura.

BRCM and Marvell are both the actual designers of many of the “we designed our own ai chip” chips that so many companies have been announcing.

Broadcom is also a major supplier of the networking chips that the hyperscalers use. Clusters are moving away from infiniband and nobody is particularly impressed by Spectrum.

[–] herodoturtle 3 days ago

I'm sure there are others here that are experts on this topic, but my 2 cents on this is that Broadcom specialise in custom / specialised AI chips, and their collaboration with Google to manufacture the TPUs is a pretty solid revenue moat over the next several years.

[–] tancop 3 days ago

AI focused investments and patent trolling.

[–] AlphaSite 3 days ago

Forward P/E.

[–] mirekrusin 3 days ago

from what you've written it's more like:

- 2026: $97.08B

- 2027: $97.08B

- 2028: $97.08B

- 2029: $97.08B

- 2030: $66.96B

- 2031: $66.96B

- 2032: $40.09B

- 2033: $40.09B

- 2034: $14.59B

- 2035: $10.00B

- 2036: $10.00B

Total: $637.0B

Considering everything that is cancellable, ie. hard commitment only, it's:

- 2026: $56.79B

- 2027: $56.79B

- 2028: $56.79B

- 2029: $56.79B

- 2030: $56.79B

- 2031: $56.79B

- 2032: $29.93B

- 2033: $29.93B

- 2034: $10.00B

- 2035: $10.00B

- 2036: $10.00B

Total: $431B

If they go bankrupt, there are some more cancellations available, also from the other side – if some providers don't deliver, it can be lower as well.

[–] dyslexit 3 days ago

That's helpful, but 2026 isn't cancelable, nor is at least half of 2027. Also, I think some of these numbers are off. Only SpaceX is cancelable but the above shows declines for years outside of the SpaceX commitment (assuming it does get cancelled, and in full).

I also left out AMD at $20B for 2027, so that would need to jump up.

Finally, the above is only counting the commitments they've already made. It's not very likely they'll only spend $10B in 2035.

[–] mirekrusin 3 days ago

It goes down because contracts expire.

AMD is missing, true but commitment schedule isn't known. Also missing estimated additional TPU purchase for 28-31, ~$125B-$150B.

- 2026: $97.08B = Broadcom $26.87B + SpaceX $30.13B + Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B

- 2027: $97.08B + AMD (TBD) = Broadcom $26.87B + SpaceX $30.13B + Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B + AMD (TBD)

- 2028: $128.33–134.58B + AMD (TBD) = Broadcom $26.87B + SpaceX $30.13B + Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B + additional TPUs $31.25–37.50B + AMD (TBD)

- 2029: $128.33–134.58B + AMD (TBD) = Broadcom $26.87B + SpaceX $30.13B + Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B + additional TPUs $31.25–37.50B + AMD (TBD)

- 2030: $98.21–104.46B + AMD (TBD) = Broadcom $26.87B + Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B + additional TPUs $31.25–37.50B + AMD (TBD)

- 2031: $98.21–104.46B + AMD (TBD) = Broadcom $26.87B + Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B + additional TPUs $31.25–37.50B + AMD (TBD)

- 2032: $40.09B = Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B

- 2033: $40.09B = Google $14.31B + Amazon $10B + Microsoft $5.61B + Nscale $5.58B + Lambda $4.59B

- 2034: $14.59B = Amazon $10B + Lambda $4.59B

- 2035: $10.00B = Amazon $10B

- 2036: $10.00B = Amazon $10B

Combined total: $782–807B+, including estimated additional TPU purchases and conditional AMD spending.

[–] Espressosaurus 3 days ago

If we assume none of it is front-loaded (which seems, uhhhh unlikely) we're looking at ~$100 billion a year for the first 4 years of commitment. That's significant.

[–] egeozcan 3 days ago

https://tinylittleoffice.com/write/#/v1/write/yZVWxTsMwEN39F...

Got the numbers on a spreadsheet. Of course this is a super simplistic version assuming equal yearly commitments.

[–] cavisne 3 days ago

It’s been rumored but this is a pretty huge TPU dependency. Assuming Broadcom + Google is TPU spend.

[–] youngtaff 3 days ago

There is the FT covers it…

> Its largest commitment is with Broadcom, which co-develops the TPU AI processors with Google and leases them to Anthropic in deals worth $161bn over five years. From 2028, Anthropic expects to spend up to $150bn more on TPUs.

https://www.ft.com/content/a6c43402-719e-4295-ae5f-541edb0ae...