- My Forums
- Tiger Rant
- LSU Recruiting
- SEC Rant
- Saints Talk
- Pelicans Talk
- More Sports Board
- Fantasy Sports
- Golf Board
- Soccer Board
- O-T Lounge
- Tech Board
- Home/Garden Board
- Outdoor Board
- Health/Fitness Board
- Movie/TV Board
- Book Board
- Music Board
- Political Talk
- Money Talk
- Fark Board
- Gaming Board
- Travel Board
- Food/Drink Board
- Ticket Exchange
- TD Help Board
Customize My Forums- View All Forums
- Show Left Links
- Topic Sort Options
- Trending Topics
- Recent Topics
- Active Topics
Started By
Message
Compute Demand
Posted on 6/5/26 at 3:33 pm
Posted on 6/5/26 at 3:33 pm
Get some dry powder ready because once we get through this potential correction induced by IPo liquidity fears, possible rate hike, midterm worries and testing the new Fed chair, things may really run.
The evidence of compute constraint is overwhelming. Google — spending $175-185 billion building its own data centers and running its own TPU program through Broadcom — just signed an $11 billion annual deal to rent Nvidia GPUs from SpaceX because it still can’t meet demand. Anthropic rented the other half of the same SpaceX facility because Amazon’s $4 billion commitment isn’t enough. Microsoft CFO Amy Hood said on the record — “We’ve been short for many quarters. I thought we were going to catch up — we are not. Demand is increasing across many places.” AWS free cash flow collapsed from $26 billion to $1.2 billion in a single quarter as a deliberate bet on infrastructure that Andy Jassy says is already committed before it’s even built.
Every major AI player is simultaneously screaming compute shortage while paying whatever it takes to get supply. That’s not a bubble narrative. That’s a supply crisis hiding behind a stock market selloff.
The one legitimate pushback is cost discipline — as agentic AI bills arrive, enterprises will optimize token usage and tighten workflows. But that argument has a historical problem. Per-token costs are already falling 50-90% annually, and every time efficiency improves, volume growth has overwhelmed the savings. Enterprises don’t deploy less when costs fall — they deploy more. The internet didn’t slow down when bandwidth got cheaper.
And the most important point — the compute constraint isn’t going away. It’s about to intensify dramatically. Gartner’s 2026 analysis confirms agentic AI consumes 5 to 30 times more tokens per task than a standard chatbot. Goldman Sachs projects total token consumption multiplying 24 times by 2030. Enterprises moving from chatbot pilots to production-scale agentic deployments are discovering the multiplier only after their bills arrive. The AI buildout narrative isn’t slowing. It’s just getting started.
The evidence of compute constraint is overwhelming. Google — spending $175-185 billion building its own data centers and running its own TPU program through Broadcom — just signed an $11 billion annual deal to rent Nvidia GPUs from SpaceX because it still can’t meet demand. Anthropic rented the other half of the same SpaceX facility because Amazon’s $4 billion commitment isn’t enough. Microsoft CFO Amy Hood said on the record — “We’ve been short for many quarters. I thought we were going to catch up — we are not. Demand is increasing across many places.” AWS free cash flow collapsed from $26 billion to $1.2 billion in a single quarter as a deliberate bet on infrastructure that Andy Jassy says is already committed before it’s even built.
Every major AI player is simultaneously screaming compute shortage while paying whatever it takes to get supply. That’s not a bubble narrative. That’s a supply crisis hiding behind a stock market selloff.
The one legitimate pushback is cost discipline — as agentic AI bills arrive, enterprises will optimize token usage and tighten workflows. But that argument has a historical problem. Per-token costs are already falling 50-90% annually, and every time efficiency improves, volume growth has overwhelmed the savings. Enterprises don’t deploy less when costs fall — they deploy more. The internet didn’t slow down when bandwidth got cheaper.
And the most important point — the compute constraint isn’t going away. It’s about to intensify dramatically. Gartner’s 2026 analysis confirms agentic AI consumes 5 to 30 times more tokens per task than a standard chatbot. Goldman Sachs projects total token consumption multiplying 24 times by 2030. Enterprises moving from chatbot pilots to production-scale agentic deployments are discovering the multiplier only after their bills arrive. The AI buildout narrative isn’t slowing. It’s just getting started.
4
Posted on 6/5/26 at 4:06 pm to Boomer Rick
I’ve seen two recurring counterpoints that seem extremely practical and cause for concern.
The compute shortage has lingered for years, yet data centers continue to be built but none ever seem to be completed. The colossus data centers that were actually built for Grok were quickly leased to OpenAI. At the same time, NVDA is selling a ton of GPUs which are apparently sitting in warehouses waiting for a data center. By the time they’re available will it already be time for the next gen chips?
The secondary factor that I think is the most compelling is electricity. Utilities and Nuclear got huge boosts from the power contracts last year. Those stocks seem to be on a protracted decline, which leads me to believe the market doesn’t really they’ll be executed, and all these data centers can’t move forward without a lot more power.
The compute shortage has lingered for years, yet data centers continue to be built but none ever seem to be completed. The colossus data centers that were actually built for Grok were quickly leased to OpenAI. At the same time, NVDA is selling a ton of GPUs which are apparently sitting in warehouses waiting for a data center. By the time they’re available will it already be time for the next gen chips?
The secondary factor that I think is the most compelling is electricity. Utilities and Nuclear got huge boosts from the power contracts last year. Those stocks seem to be on a protracted decline, which leads me to believe the market doesn’t really they’ll be executed, and all these data centers can’t move forward without a lot more power.
Posted on 6/5/26 at 4:21 pm to lsuconnman
quote:Yeah I’ve noticed this. If a major energy expansion is coming, nobody seems to be telling uranium or natural gas etc.
Those stocks seem to be on a protracted decline, which leads me to believe the market doesn’t really they’ll be executed, and all these data centers can’t move forward without a lot more power.
Posted on 6/5/26 at 4:22 pm to lsuconnman
Data centers are absolutely being completed. The Colossus facility SpaceX built is operational right now, generating $11 billion annually renting to Google and Anthropic. AWS, Google Cloud, and Azure are all reporting commissioned capacity with accelerating revenue to prove it. The timing mismatch between chip delivery and facility readiness is real but temporary — it’s a sequencing issue, not evidence that data centers aren’t completing.
On power — it’s a legitimate constraint worth watching. But go watch the Dwarkesh Patel and Dylan Patel interview from March 2026. Dylan runs SemiAnalysis and does primary supply chain research. His specific conclusion is that scaling power in the US will not be a problem — the hyperscalers have the capital and incentive to solve it through behind the meter generation, gas turbines, and nuclear.
What Dylan actually identifies as the most serious long term constraint isn’t power — it’s ASML. They manufacture the EUV lithography machines that produce advanced chips, roughly 70-100 exist globally by 2030, and each costs $300-400 million. That’s the binding constraint almost nobody is talking about.
On power — it’s a legitimate constraint worth watching. But go watch the Dwarkesh Patel and Dylan Patel interview from March 2026. Dylan runs SemiAnalysis and does primary supply chain research. His specific conclusion is that scaling power in the US will not be a problem — the hyperscalers have the capital and incentive to solve it through behind the meter generation, gas turbines, and nuclear.
What Dylan actually identifies as the most serious long term constraint isn’t power — it’s ASML. They manufacture the EUV lithography machines that produce advanced chips, roughly 70-100 exist globally by 2030, and each costs $300-400 million. That’s the binding constraint almost nobody is talking about.
Posted on 6/5/26 at 4:40 pm to Boomer Rick
quote:As long as there’s no backsliding on their commitments to carbon neutrality.
the hyperscalers have the capital and incentive to solve it through behind the meter generation, gas turbines, and nuclear.
Posted on 6/5/26 at 4:55 pm to Boomer Rick
quote:
The Colossus facility SpaceX built is operational right now, generating $11 billion annually renting to Google and Anthropic.
That’s also one of the huge SpaceX red flags. That revenue stream can be cancelled with 90-days notice. But it still begs the question, why did Elon build it if he isn’t going to use it?
Posted on 6/5/26 at 5:28 pm to lsuconnman
Just pointing out that Google is paying $920 million a month for compute capacity at xAI data centers
LINK
LINK
Posted on 6/5/26 at 6:23 pm to Jjdoc
I think 920 million/month answers the question why he built it.
Posted on 6/5/26 at 6:50 pm to Boomer Rick
I use systems like this for my line of work.
There will always be demand. Everything from engineering firms to universities will pay for more cpu-hours.
10 years ago it was like scrapping for rations on the supermike systems at LSU.
you had to submit a multipage proposal for like 50,000 cpu hours. (which is almost nothing)
The government machines are always running. ERDC, ARL, NRL and all the DOE machines. People are fighting for space. at least it was that way 5 years ago. Now im on a private cluster all to myself.
There will always be demand. Everything from engineering firms to universities will pay for more cpu-hours.
10 years ago it was like scrapping for rations on the supermike systems at LSU.
you had to submit a multipage proposal for like 50,000 cpu hours. (which is almost nothing)
The government machines are always running. ERDC, ARL, NRL and all the DOE machines. People are fighting for space. at least it was that way 5 years ago. Now im on a private cluster all to myself.
This post was edited on 6/5/26 at 6:51 pm
Posted on 6/5/26 at 7:44 pm to Boomer Rick
quote:
I think 920 million/month answers the question why he built it.
How much do you think it cost to build? Admittedly, I don’t know. However, I’ll be surprised if it isn’t a tax write off in two years.
Posted on 6/6/26 at 4:33 am to Boomer Rick
Thanks to all in this thread. Great info boiled down.
Posted on 6/6/26 at 8:45 am to UptownJoeBrown
Agree. Good thread. Some serious variables related to growth constraints, but no denying demand will keep increasing.
Posted on 6/6/26 at 9:37 am to beaverfever
quote:water too.
Yeah I’ve noticed this. If a major energy expansion is coming, nobody seems to be telling uranium or natural gas etc.
Posted on 6/6/26 at 2:38 pm to dstone12
I want to get into AIRJ at some point, but it’s difficult given the current state of their financials.
AirJoule (AIRJ) is an atmospheric water harvesting and dehumidification technology. It captures pure water from ambient air using proprietary, state-of-the-art sorbent materials and a vacuum-swing cycle of heat.
AirJoule (AIRJ) is an atmospheric water harvesting and dehumidification technology. It captures pure water from ambient air using proprietary, state-of-the-art sorbent materials and a vacuum-swing cycle of heat.
Posted on 6/7/26 at 1:38 pm to bayoubengals88
IDC projects over 1 billion actively deployed AI agents by 2029 — a 40x increase from 2025 levels — each consuming 5 to 30 times more compute per task than a traditional chatbot query.
The physical constraints on meeting that demand are largely immovable. ASML produces roughly 70-100 EUV lithography machines globally, each taking years to build. TSMC’s advanced node capacity and power grid connections carry multi-year lead times regardless of capital deployed.
The buildout doesn’t compress — it extends. Supply constraints acting as a governor on deployment speed turns a three to four year cycle into six to eight years minimum.
The physical constraints on meeting that demand are largely immovable. ASML produces roughly 70-100 EUV lithography machines globally, each taking years to build. TSMC’s advanced node capacity and power grid connections carry multi-year lead times regardless of capital deployed.
The buildout doesn’t compress — it extends. Supply constraints acting as a governor on deployment speed turns a three to four year cycle into six to eight years minimum.
Posted on 6/7/26 at 2:02 pm to Boomer Rick
So you’re buying ASML?
What about metrology testers like ONTO?
Nuclear? SMR?
What about metrology testers like ONTO?
Nuclear? SMR?
Posted on 6/7/26 at 2:10 pm to lsuconnman
quote:
The secondary factor that I think is the most compelling is electricity. Utilities and Nuclear got huge boosts from the power contracts last year. Those stocks seem to be on a protracted decline, which leads me to believe the market doesn’t really they’ll be executed, and all these data centers can’t move forward without a lot more power.
Market is paying for immediate availability. Not in utility NG directly, but power generation from turbines. Look at BW stock - they sell turbines for this. BE for data center energy. Many other examples. These are immediately available and don’t require government licensing like nuclear, which is still a huge unknown. Trump losing mid terms or MAGA losing in 2028 could hinder nuclear licensing as backlash to Trump admin pushing it forward
That said, I’m invested long term in some nuclear stonks. It’s a matter of when, not if, IMO
This post was edited on 6/7/26 at 2:11 pm
Posted on 6/7/26 at 2:23 pm to Boomer Rick
Boomer Rick,
Would you mind sharing your list of stocks to invest regarding this demand and AI?
Would you mind sharing your list of stocks to invest regarding this demand and AI?
Posted on 6/7/26 at 2:25 pm to bayoubengals88
There are a lot of ways to play it. My point is that the stronger companies will have more durable earnings than some believe. Take advantage of price discounts if they are offered in the coming months.
Posted on 6/7/26 at 2:56 pm to Boomer Rick
quote:
My point is that the stronger companies will have more durable earnings than some believe.
You waited four years to say THAT?
Spill the goods!!
Page 1 of 2
Popular
Back to top
Follow TigerDroppings for LSU Football News