Artificial Intelligence

[beerbaron]

I have been using Kimi K2 thinking for a day. Its a great product IMO on par with GPT5. I think this is one of those thing that destroys about 500B of market cap on the western side. 

 

At this point LLM are clearly a commodity. I don't see how OpenAI keeps the consumer in at 20$ a month. If people switch between streaming platforms they will swith LLM. Very little moat on the consumer side.

 

I recommend everyone to try kimi. Its free.

[Dalal.Holdings]

No worries, I am sure these firms will account for depreciation in a reasonable way, just like when they tend to "adjust" away stock based comp...

 

 

[Malmqky]

Just now, Dalal.Holdings said:

No worries, I am sure these firms will account for depreciation in a reasonable way, just like when they tend to "adjust" away stock based comp...

 

 

 

Considering Google has 100% utilization for 6-7 year old TPUs, perhaps their depreciation useful life is actually accurate? Very skeptical of the rest.

[Milu]

28 minutes ago, Dalal.Holdings said:

No worries, I am sure these firms will account for depreciation in a reasonable way, just like when they tend to "adjust" away stock based comp...

 

 

Burry seems to be obsessed with calling bubbles and crashes, having that big success in 2008 was a huge moment for him but he's been prognosticating on other bubbles 4 or 5 times since then, he'll end up being right at some point and I'm sure the incorrect calls will be forgotten about.

[Milu]

1 minute ago, Milu said:

Burry seems to be obsessed with calling bubbles and crashes, having that big success in 2008 was a huge moment for him but he's been prognosticating on other bubbles 4 or 5 times since then, he'll end up being right at some point and I'm sure the incorrect calls will be forgotten about.

 

[Dalal.Holdings]

1 hour ago, Malmqky said:

 

Considering Google has 100% utilization for 6-7 year old TPUs, perhaps their depreciation useful life is actually accurate? Very skeptical of the rest.

 

The question is how much recurring capex is needed to buy new chips when they hit the market for these firms to simply maintain their competitive position (maintenance capex)

[Dalal.Holdings]

1 hour ago, Milu said:

Burry seems to be obsessed with calling bubbles and crashes, having that big success in 2008 was a huge moment for him but he's been prognosticating on other bubbles 4 or 5 times since then, he'll end up being right at some point and I'm sure the incorrect calls will be forgotten about.

 

Unlike those other times, Burry is actually digging into accounting issues (which is how he nailed 2008). I think that the accounting at these firms does not lie.

 

The capital intensity building at these firms is not a good sign IMO that they continue to be very high return on capital type businesses into the future.

[Milu]

18 minutes ago, Dalal.Holdings said:

 

Unlike those other times, Burry is actually digging into accounting issues (which is how he nailed 2008). I think that the accounting at these firms does not lie.

 

The capital intensity building at these firms is not a good sign IMO that they continue to be very high return on capital type businesses into the future.

Not fully convinced, in 2008 he was digging through the 400 page prospectuses on various CDS tranches which was quite great sleuthing and he got rewarded for it. The correct depreciation life and growth vs maintenance cap ex for these hyperscalers is quite a well known issue that many investors are well aware of and are investing accordingly.

 

I see a lot of comments lately as if this is some big gotcha discovery that none of the holders of MSFT, AMZN, GOOGL, META haven't already considered many times.

[Eldad]

15 minutes ago, Milu said:

Not fully convinced, in 2008 he was digging through the 400 page prospectuses on various CDS tranches which was quite great sleuthing and he got rewarded for it. The correct depreciation life and growth vs maintenance cap ex for these hyperscalers is quite a well known issue that many investors are well aware of and are investing accordingly.

 

I see a lot of comments lately as if this is some big gotcha discovery that none of the holders of MSFT, AMZN, GOOGL, META haven't already considered many times.

I don’t know that 2008 was really all that hard to see. It was that everyone was emotionally to point that they refused to see it. 

[Milu]

23 minutes ago, Eldad said:

I don’t know that 2008 was really all that hard to see. It was that everyone was emotionally to point that they refused to see it. 

Possibly, although many people tend to 'see' bubbles all the time, but in reality there wasn't a bubble so they were seeing nothing, then every now and then stock prices do fall in a large way and the people who 'saw' the bubble in that year are proclaimed the geniuses.

 

Based on your postings I assume you see a bubble now which may or may not be correct. I'd be interested to understand when you first started determining we are currently in a bubble and have you made any changes to your investment approach or positions based on this discovery?

 

The best way to judge an accurate call with something like a bubble is what was the price of the asset (Bitcoin, S&P 500, Meta etc) when the person first start calling the bubble and what percentage below that price did the asset fall to when the bubble popped.

 

For me personally I feel that some stocks are possibly richly priced and many more are good value. Maybe by historical standards we are moderately overvalued, but I typically don't believe I have the ability to make any accurate macro calls, nor do I make much change to my investments so I just hold good companies through the ups and downs, and ideally buy more after large drawdowns.

[Eldad]

5 minutes ago, Milu said:

Possibly, although many people tend to 'see' bubbles all the time, but in reality there wasn't a bubble so they were seeing nothing, then every now and then stock prices do fall in a large way and the people who 'saw' the bubble in that year are proclaimed the geniuses.

 

Based on your postings I assume you see a bubble now which may or may not be correct. I'd be interested to understand when you first started determining we are currently in a bubble and have you made any changes to your investment approach or positions based on this discovery?

 

The best way to judge an accurate call with something like a bubble is what was the price of the asset (Bitcoin, S&P 500, Meta etc) when the person first start calling the bubble and what percentage below that price did the asset fall to when the bubble popped.

 

For me personally I feel that some stocks are possibly richly priced and many more are good value. Maybe by historical standards we are moderately overvalued, but I typically don't believe I have the ability to make any accurate macro calls, nor do I make much change to my investments so I just hold good companies through the ups and downs, and ideally buy more after large drawdowns.

I guess I can just sort of feel the mania currently. It’s not so much the quantitative factors that are flashing red (they are now but they have cried wolf many times in the post 2008 world). I saw a chart that this morning that showed that by far the best asset class this year is Nasdaq listed companies with no revenue.
 

Recently, it seems that all other asset classes must go down to feed the beast. In the past couple of weeks you  have had pretty huge index up days where almost every stock is down but the AI names. All of this is happening at the same time as the circular financing and the government funding talk. 
 

Then there are the irrational hopes for AI that are always “right around the corner” but never seem to materialize. Sam Altman etc. have to do a podcast or interview almost everyday now about no jobs, universal abundance blah blah blah. The FOMO is even more heated with this bubble because not only are they hyping their product, they are also actively trying to scare you into buying into it as your last and only hope. 
 

Just a hunch but it feels like the big one to me. But I’m not selling out or anything currently. 
 

[DegenerateGambler]

The problem with calling bubbles is timing and the surety of eventual collapse.  The problem with calling bubbles on these cashflow-positive big tech is that they can weather any kind of "collapse" scenario whereas in 2008 the banks and financial institutions had no way to survive.   Sure the big tech companies can overspend for a few years but that doesn't create the kind of bubble we saw in 2008 where literally everything was built on a pile of sand.  Meanwhile what can happen is his short position can get obliterated as people pile into the "AI trade" and then it "crashes" to a point still higher than where he started shorting them.  Also remember if the stock market really crashes the fed will come to the rescue giving you very little opportunity to exit your positions.   The Fed put is a real thing now, nothing is really allowed to fail, not even asset prices.   The "AI trade" can go on for quite a few years and then maybe the frothy, no-revenue companies will start failing but the big tech can just cut back spending an their operating cash flow will sky-rocket again.  I feel Burry bet too big on what is eventually an "adjustment" vs "crash" scenario.  The fed has incentive to keep printing money because US debt is so high - gotta inflate that debt away.  So where is the money gonna go?  Right now the only obvious answer is the "AI Trade".

[rogermunibond]

https://www.wsj.com/tech/ai/three-ai-megadeals-are-breaking-new-ground-on-wall-street-896e0023?

 

WSJ goes under the covers for three massive $30+ billion AI investment deals.  One for Meta, one for OAI, and one for xAI

[rogermunibond]

Quite an informative post on GPU life and depreciation estimates.

 

 

[Spekulatius]

AI has reached the SPV age. Deja vue they not a new construct either.

https://archive.ph/A40Fv

 

$30B off balance sheet for Meta with full operational control.

[Eldad]

Have you noticed that many people actually doing the work deride AI and complain about top down AI mandates? I have noticed in various software developer forums etc., the tone on AI is completely different than from the c-suite. 
 

I have seen many articles lately saying some workers feel like they have to kind of do the work twice because of mandated AI use requirements. The numbers in a PEW research report sort of seem to back this up. The higher the income bracket of those surveyed the more excitement and belief in AI. 
 

I see a Barron’s article today saying the new Verizon CEO says AGI in 2-4 years. Why would he feel the need to say something like that? Then I see META parts ways with Yann LeCun who seems to believe AGI isn’t possible. Is this the new Davos set mental virus? Have they replaced DEI and ESG as societal panaceas with this new obsession? 

[DegenerateGambler]

2 hours ago, Eldad said:

Have you noticed that many people actually doing the work deride AI and complain about top down AI mandates? I have noticed in various software developer forums etc., the tone on AI is completely different than from the c-suite. 
 

I have seen many articles lately saying some workers feel like they have to kind of do the work twice because of mandated AI use requirements. The numbers in a PEW research report sort of seem to back this up. The higher the income bracket of those surveyed the more excitement and belief in AI. 
 

I see a Barron’s article today saying the new Verizon CEO says AGI in 2-4 years. Why would he feel the need to say something like that? Then I see META parts ways with Yann LeCun who seems to believe AGI isn’t possible. Is this the new Davos set mental virus? Have they replaced DEI and ESG as societal panaceas with this new obsession? 

That's an interesting observation.  With regard to AGI - if we start from the base layer - atoms - would we definitely say that using semiconductors is the better way to create intelligence than using neurons?  I mean some of these semiconductors are getting close to 1-atom thin  and that is the physical limit without resorting to quantum computing.  And who is to say there are no quantum phenomenon going on inside our brains?  I'd say it's still early to say semiconductors is the better way to create intelligence from the basic building blocks of matter.   It may never surpass meat brains.

[LC]

On 11/11/2025 at 1:29 PM, rogermunibond said:

Quite an informative post on GPU life and depreciation estimates.

 

 

 

 

I would imagine these big training data centers are actively managing GPU temps and voltage. They can extend the life to 6 years if needed.

[Fly]

1 hour ago, LC said:

 

 

I would imagine these big training data centers are actively managing GPU temps and voltage. They can extend the life to 6 years if needed.

 

Wouldn't there be incentive to overclock these GPUs and try to train even more data? The best model wins it would seem. Move fast and burn GPUs? 

Edited November 13, 2025 by Fly

[beerbaron]

11 minutes ago, Fly said:

 

Wouldn't there be incentive to overclock these GPUs and try to train even more data? The best model wins it would seem. Move fast and burn GPUs? 

Impossible to overclock GPU in a large cluster, they would all have to be equally overclocked in order to see the gain. There is kinda of a sync period where between each propagation where data is more or less shared across the whole GPU cluster. Furthermore, GPUs are notoriously unreliable, losing a few GPU can significantly delay the training process, imagine if an overclock triples the early death of those GPU.

[nsx5200]

5 hours ago, DegenerateGambler said:

That's an interesting observation.  With regard to AGI - if we start from the base layer - atoms - would we definitely say that using semiconductors is the better way to create intelligence than using neurons?

 

Current 'state of the art' AI is still some form of feed forward network.  The training is fed back to update the weights, but in actual inference/implementation, it is expressed feed forward.  The organic brain is more organic, so it feeds forward, back, sideways, etc.  The TLDR is that structurally, organic brain is different than current computing AI.  There have been some success with current compute AI, but it is still no squishy brain, so IMHO, there will be fundamentally different problems that these different types of 'brains' can solve.

 

Even within the comparatively confined game like Go, which a statistic machine that has optimized moves up to the nth move in the future, Lee was able to gain some insight to beat AlphaGo once.  Compute brains are still fundamentally statistic optimizing machines at the core, with some randomness added to it, whereas organic brains are not, for better or for worse.  I think the trick is to recognize/learn the type of problems that is better suited for the type of brain to better solve the problem.  Right now, people think compute brain will equal organic brain, and IMHO, I just don't see that happening without a breakthrough in compute brain's structure.  If I'm right, then it'll take people some time to realize that.

Edited November 13, 2025 by nsx5200
Add condition for compute brain's future

[Tenerife]

100% agree with your statement. At best they're weighing machines, weighing from the past events. 

[Milu]

Would encourage people who keep highlighting the depreciation estimates and whether 4 years, 5 years, 6 years is correct etc to listen to first 10 minutes of this.

 

Pay particular attention to where they highlight how google is still running their 10 year old TPU chips at 100%, so if anything their could be a case that many of these companies are possibly still being conservative about the depreciation estimates. 

 

[nsx5200]

27 minutes ago, Milu said:

Pay particular attention to where they highlight how google is still running their 10 year old TPU chips at 100%, so if anything their could be a case that many of these companies are possibly still being conservative about the depreciation estimates. 

When TPU came out, it was ahead of its time.  Depreciation is an accounting thing.  We know that things can be fully depreciated but still useful in generating revenue.  A lot of that depends on whether the semi fabs can continue Moore's law.  If somebody figures out how to make compute hardware cheaper quicker, than those depreciation will speed up.  Similarly, if they find out they're hitting some barrier that slows down the improvement of compute hardware, than those depreciation will slow down.  Those barriers may include access to critical minerals, destruction of existing fabs due to war, etc...  So the TLDR summary is that nobody really know, and 4-6 year guesstimate based on Moore's law and historical empirical data is probably a decent one, but still a guesstimate.  Whether these companies are fudging these grey area numbers to 'cook the book' is a separate discussion.

 

Just curious, does changing the depreciation schedule from 4-6 year really change your investment thesis that much?

[beerbaron]

3 hours ago, nsx5200 said:

When TPU came out, it was ahead of its time.  Depreciation is an accounting thing.  We know that things can be fully depreciated but still useful in generating revenue.  A lot of that depends on whether the semi fabs can continue Moore's law.  If somebody figures out how to make compute hardware cheaper quicker, than those depreciation will speed up.  Similarly, if they find out they're hitting some barrier that slows down the improvement of compute hardware, than those depreciation will slow down.  Those barriers may include access to critical minerals, destruction of existing fabs due to war, etc...  So the TLDR summary is that nobody really know, and 4-6 year guesstimate based on Moore's law and historical empirical data is probably a decent one, but still a guesstimate.  Whether these companies are fudging these grey area numbers to 'cook the book' is a separate discussion.

 

Just curious, does changing the depreciation schedule from 4-6 year really change your investment thesis that much?

 

Well, if you listen to Jensen's interview on BG2 he talks about how B100 VS H100 and puits some 2X number and much higher when optimized to the new hardware. According to him, it make absolutely no sense to buy H100 if you can buy B100. The operating cost of the H100 VS the B100 is much higher per TFLOPS. So one could argue that if there were not supply shortage those old GPU would have a negative value. Similar to when miner ASICs were profitable until an ew generation came up... they became fancy space heaters over night and made more NPV to thrown then in the trash than keeping operating them.

 

I would say if NVIDIA scales up production it basically destroys the amortization period. Furthermore, amortization is currently artificially long because of a supply issue and that if your are not a hypercaler with long terms commitments you are shit out of luck.

 

I asked ChatGPT to make me a comparative table comparing the same TFLOPS (reference TFLOPS is 100xH100 cluster) between architectures. 

 

TAKE WITH A GRAIN OF SALT AS I'M NOT AN NVidia ANALYST. But it gives and idea.

 

Cost Component	A100 Cluster (600 GPUs)	H100 Cluster (100 GPUs)	B100 Cluster (56 GPUs)
GPU CapEx	$9,600,000	$3,000,000	$1,820,000
Server / Node Infrastructure (CapEx)	$6,000,000	$1,000,000	$1,400,000
Total CapEx	$15,600,000	$4,000,000	$3,220,000
---	---	---	---
5-year Power Cost	$1,365,000	$398,500	$223,200
5-year Cooling / PUE Overhead	(included in power)	(included)	(included)
5-year Maintenance / Spares	$1,330,000	$450,000	$322,000
5-year Staff / Operations Cost	$5,500,000 – 6,000,000	$1,000,000 – 1,300,000	$800,000 – 1,100,000
Total OpEx (5 years)	$8,200,000 – 8,700,000	$1,850,000 – 2,150,000	$1,345,000 – 1,645,000
---	---	---	---
5-Year Total TCO	$23.8M – $24.3M	$5.85M – $6.15M	$4.56M – $4.87M

Edited November 15, 2025 by beerbaron