Currently, how to break the AI "bubble"? - Safe and reliable encrypted digital currency trading platform

Recently, stocks of AI-related companies have been heavily impacted due to various cyclical financing rounds in the US,

The high-level investments with no visible returns and the lack of global market liquidity have suffered severe setbacks.

Major CSP manufacturers are shifting to depreciation,

Cash flow,

Capital expenditure as a percentage of revenue has already been relatively high, as I mentioned mid-last year,

But at that time, the market ignored it,

Continuing to aggressively push AI,

Now it’s being hotly discussed again,

The core reason is exactly the same as back then,

Still no profit or blockbuster applications matching capital expenditure,

In simple terms, the market’s bubble is believed to be caused by excessively high investment costs.

But at present, everyone can’t really stop investing in AI,

Instead of futilely predicting when capital expenditure will break or blockbuster applications will emerge,

It’s better to study current industry development trends to solve the “bubble” problem in AI.

How to solve it?

The answer is cost reduction,

Any technological development trend must involve a continuous decrease in unit costs,

Otherwise, it cannot be widely adopted.

From brick phones to small transceivers to smartphones,

From the internet to mobile internet, this has been the path.

So where is AI’s biggest dilemma now: high capital expenditure and power shortages,

Correspondingly, it’s GPU costs being too high,

Excessive power consumption,

How to solve this?

There are two paths:

1. Reduce the system costs of compute cards

(1) Self-developed chips,

A typical example is Google,

And Google, like Apple, has abundant cash,

An ecosystem closed loop,

A hexagonal warrior combining software and hardware,

Naturally becoming the last investment of the legendary Mr. Ba.

From the chart below, it’s clear that Google’s TPU7 configuration is quite capable,

Two units roughly comparable to B300,

But in terms of price,

Buying a B300 probably can produce four TPU7 units.

Therefore, the TPU shipment guidance for 2026 has reached four million units,

Almost matching the predicted Rubin shipment for 2026.

The industry impact is that,

If general-purpose GPUs and self-developed cards reach a long-term market ratio of 1:1,

It means that with the same capital expenditure,

You can achieve about 1.5 times or more in computing power and storage utilization.

The same logic applies to supporting products like optical modules,

More cards mean increased demand for supporting components.

(This paragraph was written by me on September 7,

Just before this wave of storage explosion,

You can review the content from that time for details.)

(2) Splitting a traditional GPU into inference and deep training configurations,

Using high-low memory configurations,

NVIDIA’s Rubin+CPX combination and the Ascend series chips’ 950pr+950dr are examples,

The logic part can use the same chips,

But costs are reduced by adopting different memory configurations.

For example, 288G HBM4 costs an astonishing 30,000+ RMB,

But 288G GDDR7 might only be around 15,000 RMB,

Lightweight text preprocessing can significantly lower inference card costs,

Deploying more cards for inference.

The Ascend 950pr, a self-developed high-bandwidth inference card solution based on non-traditional HBM, is an innovation from China,

It has not been officially released and is not suitable for public discussion,

But industry insiders probably have an idea of what it is.

2. Reduce server power consumption

This week’s news already mentioned,

To reduce AI server power consumption,

NVIDIA recently decided to replace the memory used in servers from the usual DDR5 (fifth-generation double data rate synchronous dynamic random-access memory),

With LPDDR5X — also used in many flagship smartphones.

As for domestic AI chip power saving,

Like Qualcomm,

They are much more aggressive than NVIDIA,

They are using LPDDR for CPU storage to save power,

While we are directly targeting GPU storage,

No more spoilers,

Let’s just say, you know.

In addition, switching server cooling from air cooling to liquid cooling,

This goes without saying,

Cooling is also a key factor in energy consumption.

In summary,

Many structural changes will occur within the industry, and these links still hold the key to future opportunities.

Storage,

This week, the situation was also quite intense,

One issue is liquidity,

The other is the so-called large-scale expansion of memory capacity.

First, Hynix and Samsung’s new capacities are only M15X and P4 in Pyeongtaek,

Others are just old process capacities converted from their existing DDR4/LPDDR5 lines,

Not really new capacity.

The headline “XX expansion by 8 times” is more of a word game,

First, the advanced 1C process has a small base,

Second, over 50% of it is just capacity upgrades with a process improvement,

The actual total increase is relatively limited.

And this expansion will at least be implemented after the second half of 2026,

Some even as futures in 2027-28,

Installation can be adjusted according to market demand.

Maintaining DRAM prices at a relatively high level,

Expanding sales to meet demand,

Is what memory manufacturers need to do,

To achieve long-term high margins similar to HBM,

With both volume and price increasing.

Since DRAM is dominated by only three companies — Samsung, Micron, and Hynix,

It’s said that two Korean companies have some collusion behind the scenes.

Additionally,

It’s important to note that,

The spot prices you see are different from the contract prices supplied by manufacturers,

The contract prices are much lower than spot prices,

Even if spot prices halve, the manufacturer’s prices are still about 50% lower than spot prices.

Expanding from HBM to general storage,

Has greatly increased the proportion of AI business,

The “moment of optical module” where storage manufacturers’ volume and prices rise together has arrived.

Bank of America also issued a report describing the current awkwardness,

There is significant disagreement among buyers,

Research analysts are optimistic about storage,

But fund managers worry about becoming bagholders,

In simple terms, knowledgeable insiders see no problem,

But those managing the money are worried,

And are only willing to jump in after the secondary market bottoms out.

**$TAO **$DAI $AIA