freedom

opinion are my own. respectful discussion only. AI. Machine Learning.

[freedom @xbhuslapwk]

A Look at the AI Bubble and GPU and Memory Chip hype:

I presented my thesis on the oversupply of AI chips by breaking the issue into four aspects. Here I presented my thesis with strong conviction and backed it with Live statistical evidences.

1. AI Model Adoption in the Enterprise Market
OpenRouter is widely considered one of the most useful sites for measuring API usage across AI models. It tracks live usage across major models, including DeepSeek, Claude, ChatGPT, Grok, and others. Because it tracks API usage across the market, we can use it as a metric to estimate each model’s realistic penetration in the enterprise market.
According to OpenRouter (https://t.co/vj6xTCXcJO), the top AI models are overwhelmingly dominated by Chinese AI companies. In the top 10 ranking, only Claude Opus 4.7 and Opus 4.8 barely make it into the chart. OpenAI and Anthropic have repeatedly claimed superiority and growing demand in the enterprise market, but apparently this is not reflected in the real data. Chinese models are now dominating the enterprise market because they are cheap and efficient. Therefore, the thesis about U.S. AI companies penetrating the enterprise market appears to fall apart when compared with real statistical data.

2. Economics of AI Data Centers
The AI data center business is brutal. It requires heavy upfront capital investment to acquire land and build electricity grids and water pipelines. In the old days, when data centers mainly ran on CPUs, the business was still considered a better real estate investment than many other forms of real estate because it operated somewhat like an equipment-leasing business, supported by long-term, relatively high-margin, value-added rental contracts.
In the AI era, however, the use case for GPUs is much more limited compared with CPUs. GPUs are primarily matrix-computation accelerators and perform poorly on command-driven execution, where CPUs excel. Moreover, innovation in AI chips has further worsened the situation because the upgrade cycle for GPU equipment is far more frequent, as the technology is still relatively new.
https://t.co/IdY1XESNHC

3. A Realistic Look at Addressable TAM and Revenue Streams
Currently, the main revenue streams of leading AI models rely heavily on recurring subscriptions and per-token charges. It has been reported that per-token usage has decelerated significantly due to unpredictable costs and unmeasurable ROI (https://t.co/mnpZv8pjIj). If a user spends a significant number of tokens on the wrong problem or a bad prompt, those wasted tokens cannot be reimbursed. It is clear that the recurring subscription model seems to be the most logical and predictable part of the business.
Assuming an adoption rate of 10% of the global population (800 millon people) and a monthly charge of USD 20, the total TAM would be around USD 192 billion. At a monthly charge of USD 100, which is absurd, the TAM would be around USD 960 billion. I think the plausible and realistic total TAM is around USD 300-400 billion across the consumer and enterprise markets, at least within the next 2-3 years.
By comparison, the total capital expenditure committed by MAG 7 cloud vendors for FY26 is around USD 750 billion, and they project that it will rise further in the next fiscal year. The financial figures themselves do not stand up. It does not make economic sense to justify such a level of investment to a 400 billion TAM market.

4. Oversupply and Chip Hype
ChatGPT was released in November 2022. Since then, models have evolved somewhat moderately. The context window has expanded quite a bit, and model accuracy has improved moderately. But if we ask ourselves whether the improvement from ChatGPT to Claude Fab 5 is impressive enough to justify an annual bill of nearly USD 1 trillion, the answer is undoutable NO. Can these numbers justify what are, comparatively, only moderate improvements in AI models? The answer is undoutable NO.
Back in early 2023, there was already a large influx of users on ChatGPT. The site occasionally broke down, and users were generally satisfied with the model with limited computational resources at the time. The next milestone was the DeepSeek moment, when its site also occasionally broke down and model was fine with limited computational resources. It was rumored that DeepSeek had only limited resources and was still capable of producing a superior model at the time. It is reported that DeepSeek used 2,048 NVIDIA H800 GPUs over a period of roughly 55 days. The total cost was less than 80-millions USD, inclusive of equipment and data centre operation costs. We quickly realized in that moment that the constraint was not necessarily a lack of training resources, but rather algorithmic efficiency. Despite being the smartest CEO on earth, somehow they fell for the lie and fraudulent spam orchestrated by semiconductor leaders and their mouthpieces like SemiAnalysis. And believed DeepSeek spent over 1-billion USD to 2-billion USD on training which is a more than 10x boast.
Then came the leak of Claude’s internal codebase. What we saw in that codebase leakage was piles of instruction sets built to tackle individual problems, with explicit instructions and pre-coded problem solvers. I suspect that what makes Anthropic today’s dominant leader in advanced AI models is not superior or unlimited computational power, but probably its internal lead in algorithmic efficiency and preprocessing.

[freedom @xbhuslapwk]

$RDDT up almost 14% from last Friday’s low.

[freedom @xbhuslapwk]

As always when market clams down and return sanity, the $RDDT post-earning loss will be wiped out completely, probably within the next week or two.

[freedom @xbhuslapwk]

@JonahLupton sounds legit

[freedom @xbhuslapwk]

@joecarlsonshow It wont ripping long.

[freedom @xbhuslapwk]

@HedgeyeComm clown

[freedom @xbhuslapwk]

@jpinsights Yes it is a concern. When a company sit on a pile of cash, they tends to make silly decisions.

[freedom @xbhuslapwk]

@outerbridgecap I think Reddit does not actively try to acquire new users. If your Karma is low, then it is quite difficult to make a post and even comment in some communities.

[freedom @xbhuslapwk]

@ReneSellmann Another Facebook in the making.

[freedom @xbhuslapwk]

@KobeissiLetter Its Monday time!

[freedom @xbhuslapwk]

@DeItaone It’s Monday guys! Negotiation start!

[freedom @xbhuslapwk]

@teechongyen This is a very concentrated portfolio. Risky bet.

[freedom @xbhuslapwk]

@Mylovanov London is a muslim city.

[freedom @xbhuslapwk]

Actual document: https://t.co/DtufWRgu4r

[Brian McCormick @bjmtweets]

$RDDT is likely to make significant gains from renegotiating its API deals

Google’s own internal evidence reportedly found that licensed data provides meaningful product quality benefits to Gemini, both during model pre-training and when Reddit content is retrieved to provide live AI answers in inferencing.

The disclosure appears in the DOJ plaintiffs’ public, redacted remedies filing from the Google Search antitrust case.

Reddit is also the #1 most cited domain for AI answers.

Semrush found that of 150,000 citations from AI it studied, Reddit leads all sources with a 40% citation frequency. Wikipedia was next at 26.3%.

Google is set to spend $200B in capex this year.

The industry is already looking at $1T in annualized capex spend.

Reddit is only making around $160M a year leasing its data to AI.

That means Reddit is only capturing about 0.016% of the annualized AI capex spend, despite being 40% of cited outputs.

If Reddit 10x its revenue from leasing out its data, it would still be only one tenth of a percent of the money being spent on AI.

The AI industry relies on Reddit for model quality and the loss of not having that data is significant.

Google as well as the rest of foundation models will be highly interested in signing a deal. The value Reddit can capture should increase, as the importance of AI, and spend here, has increased significantly over the past 2 years since the previous deal was signed.

[freedom @xbhuslapwk]

@bjmtweets @calistoker123 Thanks!

[freedom @xbhuslapwk]

@povdaily1 @TemptInvest There is no point for Google to own Reddit, rather just pay them.

[freedom @xbhuslapwk]

A bunch of clowns 🤡 acting like they really know how to run a business. These DAUs and WAUs are just useless numbers and provide no values. What really matters is regular customers who actually contribute to communities; these are engaged users with behaviour patterns: These numbers are keeping growing at lightning speed. $RDDT is relentlessly monetising on it. Those who know $RDDT should realise that it is difficult for a new user to post or comment on their platform because of low karma; the system treats them no different than bots.
Reddit should stop posting these useless numbers.

[Brian McCormick @bjmtweets]

$RDDT is close to $META P/E on an annualized basis now.

And Reddit is growing DAU and WAU much faster. Meta is only growing DAU 3% YoY.

Oh, and Reddit also has much more room for ARPU expansion.

My best guess is the DAU story is to shake out retail before more institutional accumulation. This business is too high margin and early in monetization stage.

Reddit is the highest quality business I’ve found with a NTM P/E under 20.

[freedom @xbhuslapwk]

@chinodgk @epictrades1 If that’s the case, everyone is so bullish and buying the dip, so why hasn’t the stock moved up? Clearly, the market is wrong. And people who say buy the dip ain’t really buying it because they lost all their money in the AI trade.

[freedom @xbhuslapwk]

Reddit is not valuable because it is another social network. It is valuable because it may be one of the last large-scale places where humans openly discuss real experiences, publicly on the internet. $RDDT

[freedom @xbhuslapwk]

Reddit $RDDT should create a dedicated data service to AI companies and charge for it, doing so reduce the overhead cost of in data cleansing and quality control. Once they do that, their data becomes ready-product and much more valuable. They can basically cut off these data scraping companies out of business. There’s still a lot they can do in their API pipeline. @Reddit $RDDT

[freedom @xbhuslapwk]

The market underestimating what the Reddit $RDDT really is.
What’s wrong?
The market still sees Reddit as a social media company competing with Meta.
But the reality is:
Reddit is a human-generated knowledge database that AI model, search engines, and advertisers need.