Artificial Intelligence

[gfp]

Update this morning from Irrational Analysis - more a Semiconductor blog than a strictly AI blog, but always worth a read and highly entertaining.  

 

https://irrationalanalysis.substack.com/p/market-memo-intc-qcom-wolf-on-nvts?utm_source=post-email-title&publication_id=1509468&post_id=195937503&utm_campaign=email-post-title&isFreemail=true&r=oqhe5&triedRedirect=true&utm_medium=email

[MungerWunger]

 

[valueventures]

Without reading through 20+ pages on this thread, curious if many of you have material AI exposure throughout your portfolio. I continue to be astonished by the returns these AI stocks are putting up and am kicking myself for underestimating the impact. I realize this is a value investing board rather than a growth investing board. I'm not planning to chase growth or go FOMO, but am thinking about some broader re-positioning (I own a fair amount of big tech and am growing increasingly concerned about the increasing capex and potentially lower future resulting ROICs). 

[RichardGibbons]

5 hours ago, valueventures said:

Without reading through 20+ pages on this thread, curious if many of you have material AI exposure throughout your portfolio.

 

I have a dual strategy based on regrets from the dotcom bubble a quarter century ago--long term holds (LT) of companies that I think will do well, and speculations (S) where I think the odds of a quick double are high, but there's some chance of a big loss as well, and I'm quick to take gains. I also have a lot of names in my portfolio, so the speculations are typically just under 1%, while many of the long term holds are bigger, but because they've grown to be bigger.

 

For the sake of the question, I'm going to assume that AI includes not just AI stocks directly, but those that make the infrastructure as well.

 

GOOG (LT)
NVDA (LT)

PGY (LT)

IREN (S)

CRDO (S)

 

Since the start of the year, I've also had positions the following companies at various times, often losing shares when selling high implied volatility covered calls, all the plays profitable, and all the speculations making at least 50% in a matter of a few months.

 

UPST (LT)

ALAB (S)

BE (S)

POET (S)

 

My main screw ups were owning NBIS and LITE last year and being way too trigger-happy, roughly breaking-even on both.

 

I also just started running an automated strategy that often owns a chunk of TQQQ. That's not a deliberate decision related to AI prognostications, but would likely benefit if AI does well without destroying the world.

[LC]

On 4/29/2026 at 9:55 AM, frommi said:

In the first two minutes you already realize that these guys have no clue what they are talking about. 10 times more compute necessary for 2x improvement in model performance is not exponential, its getting dminishing returns for each amount of compute.

 

Could also be that they believe a 2x performance improvement is enough to drive 20x revenue growth? I.e. gets them past whatever tipping point they envision? 

[bargainman]

On 4/22/2026 at 8:56 AM, rogermunibond said:

 

 

 

Hmm I mean sure a lot of the data is garbage, but every model I've seen trained filters out all that crap data before the training begins.  I'm not sure what he's talking about as if the models are actually training on that?  Maybe they did when he worked at OpenAI?  Kind of a strange comment.  But perhaps it's because my work has primarily focused on SLM and on device so I don't see the greater slop?  Anyway Karparthy is more knowledgeable than I am, and this is just a snippet so perhaps it's out of context (pun intended)

[bargainman]

On 4/29/2026 at 8:55 AM, frommi said:

In the first two minutes you already realize that these guys have no clue what they are talking about. 10 times more compute necessary for 2x improvement in model performance is not exponential, its getting dminishing returns for each amount of compute.

 

No comment on the guys.  But that was in reference to training only.  I'm not sure what the newer numbers are but training is a smaller part of the entire usage lifecycle.  That's a fixed cost.  After the model is trained it can be used for inference, and future trainings.  Earlier numbers were 60-90 percent inference vs training.  But with agentic AI I'm going to guess it will skew even further in inference's favor.  Also note that 10x compute does not mean 10x the cost

 

"The Vera Rubin NVL72 rack delivers approximately 10x higher inference throughput per watt"

 

So the cost of 10x compute may be the same in a year as it is now for inference at least.  The training costs per compute will also come down for the same models, but of course there will be more models.

 

Agentic is really the difference.  I know of people blowing through $2k tokens in a month and begging for more.  They probably get paid 200k+ a year, so if the agentic token usage increases their productivity even in the low teens, it's a no brainer.  Now that won't happen in all industries...

 

 

[NnnnotSoSmart]

I'm NnnnotSoSmart regarding silicon but trying to learn...

~~~~~~~~~~~~~~~~~~~~

The EDA Primer: From RTL to Silicon

Semi Analysis

 

The semiconductor industry’s ability to keep building more powerful chips depends not on physics or lithography alone, but on EDA (Electronic Design Automation) software. These tools effectively translate human intent into manufacturable silicon. Without EDA, no chip designed after the mid-1980s would exist.

 

This primer is your guide to EDA in the semiconductor industry. In this first part, we will walk the entire journey from RTL (Register Transfer Level) code, the high-level hardware description language that engineers actually write, all the way to manufactured, packaged silicon. We will name the tools, explain the tradeoffs, and show why EDA is one of the most consequential and underappreciated sectors in technology.

 

In part 2, our EDA Market Primer dives deep into the business of EDA, profiling the major companies (Synopsys, Cadence, Siemens) and their revenue and business models. We provide comprehensive market analysis and monitoring the Chinese EDA effort, as well as IP licensing and outsourcing to design partners and the transition to Customer Owned Tooling (COT) with hyperscaler ASIC designs.

 

Part 3 then assesses how AI is disrupting the EDA industry, covering the full gamut from startups and engineer dashboards to agentic chip design flows from NVIDIA and the big three. The concept of using AI accelerators to create superhuman designs that go into future AI accelerators is the most exciting development that our industry has seen in decades. Stay tuned as we cover the incoming revolution in chip design.

 

https://newsletter.semianalysis.com/p/the-eda-primer-from-rtl-to-silicon

 

Edited May 12 by NnnnotSoSmart

[MungerWunger]

https://www.wheresyoured.at/where-are-all-the-data-centers/

[Fly]

12 hours ago, MungerWunger said:

https://www.wheresyoured.at/where-are-all-the-data-centers/

 

Sounds like a Big Short moment. Who are the most overextended names? Oracle? BE? MSFT?

[gfp]

So real, physical, constraints creating bottlenecks and slowing construction and full completions is being interpreted as bearish?  

 

I'm not hearing anything about huge data centers that get completed and sit un-utilized.  The article is basically arguing the opposite.  That they aren't able to build them and light them up fully as quickly as their PR has claimed.  

[rogermunibond]

Heard a story from a friend in corporate supply chain.  A large tech vendor that supplies this enterprise's PCs, laptops, servers etc unilaterally terminated their supply contract recently.

 

Pricing on CPUs, DRAM, etc was the issue.

 

 

[Paarslaars]

 

 

Interesting read on the AI supply chain bottlenecks.

[Paarslaars]

On 5/13/2026 at 5:13 PM, gfp said:

So real, physical, constraints creating bottlenecks and slowing construction and full completions is being interpreted as bearish?  

 

I'm not hearing anything about huge data centers that get completed and sit un-utilized.  The article is basically arguing the opposite.  That they aren't able to build them and light them up fully as quickly as their PR has claimed.  

Basically he is claiming that a shitload of the current forward earnings are based on upscaling datacenters, but if these are substantially behind schedule then all of these expected orders will be postponed.

 

And the current rally in semi is mainly based on forward earnings..

[NnnnotSoSmart]

Eric Jang – Building AlphaGo from scratch

AlphaGo is still the cleanest worked example of the primitives of intelligence: search, learning from experience, and self-play.

 

Eric Jang walks through how to build AlphaGo from scratch, but with modern AI tools.

 

Sometimes you understand the future better by stepping backward. AlphaGo is still the cleanest worked example of the primitives of intelligence: search, learning from experience, and self-play. You have to go back to 2017 to get insight into how the more general AIs of the future might learn.

 

Once he explained how AlphaGo works, it gave us the context to have a discussion about how RL works in LLMs and how it could work better – naive policy gradient RL has to figure out which of the 100k+ tokens in your trajectory actually got you the right answer, while AlphaGo’s MCTS suggests a strictly better action every single move, giving you a training target that sidesteps the credit assignment problem. The way humans learn is surely closer to the second.

 

https://www.dwarkesh.com/p/eric-jang

 

[NnnnotSoSmart]

The First Major-Exchange Compute Futures

 

"To finish the arc oil walked in the 1980s and natural gas in the 1990s, compute had to find a home at a major regulated derivatives exchange. Today it does."

 

Edited May 19 by NnnnotSoSmart

[rogermunibond]

US DOE Data Center infrastructure map as of 11/2025

 

[MungerWunger]

Wow, first SpaceX and now OpenAI!

 

OpenAI Prepares to File to Go Public in Coming Weeks

https://www.nytimes.com/2026/05/20/technology/openai-ipo.html

[treasurehunt]

Anthropic is expecting $10.9B in revenue and over $500M in operating profit in Q2 26. Stunning numbers!

 

 

[Spekulatius]

42 minutes ago, treasurehunt said:

Anthropic is expecting $10.9B in revenue and over $500M in operating profit in Q2 26. Stunning numbers!

 

 

No SBC? Rules are rules!

[treasurehunt]

21 minutes ago, Spekulatius said:

No SBC? Rules are rules!

 

Maybe it is included in Sales, Marketing & Partnership Costs. But I'd guess that the operating income vanishes if we look too closely. Still, Anthropic appears to be headed towards profitability.

[mattee2264]

Anthropic "profit" is because they are getting a discount on compute costs from SpaceX as they ramp up. 

[Spekulatius]

75% price cut for Deepseek.

https://finance.yahoo.com/sectors/technology/articles/china39s-deepseek-to-make-permanent-75-price-cut-on-flagship-v4pro-ai-model-133313442.html

[beerbaron]

6 hours ago, Spekulatius said:

75% price cut for Deepseek.

https://finance.yahoo.com/sectors/technology/articles/china39s-deepseek-to-make-permanent-75-price-cut-on-flagship-v4pro-ai-model-133313442.html

AMAZING. This is a model on par with top end frontiers from december 2025. 1/15th of the price of opus. Imagine if China pulls out a better coding frontier model just before the 3 big IPOs, that would cratter valuations.

 

How many over here know if there company allows dev. with chinese models? 

 

I

[Spekulatius]

Looks like China has a Palantir clone as well:

https://www.dw.com/en/new-chinese-surveillance-leaves-foreigners-nowhere-to-hide/a-77246713