Rebirth 08: Rise from copycat phones

Chapter 585 The astronomical price of a $30 billion GPU
In mid-October, the research committee within the Zhiyun Group approved a non-public research plan, allocating $8.5 billion for a non-public research project.

This non-public research project is the Brain 6 system project of the Neuromorphic Computing Project!

These funds will be used to further improve the Brain 6 system and solve many subsequent technical problems. Although the Brain 6 chip is available and the design of the Brain 6 server has been completed, the construction of the entire server cluster still faces many technical challenges.

In particular, this server cluster is a comprehensive computing system that combines a superconducting quantum computer, a traditional GPU graphics card, and a Smart Brain 6 server.

The research and development of such a comprehensive system that spans multiple computer types presents significant technical challenges.

Currently, only Zhiyun Group possesses this technology globally... although Zhiyun Group has never publicly disclosed it.

However, only Zhiyun Group builds its integrated computing center without using both superconducting quantum computers and traditional GPU servers; other companies don't do this... because they don't have quantum computers and don't need that technology.

Now, Zhiyun Group plans to incorporate the neuromorphic computing system into this comprehensive computing power system... This will inevitably encounter many technical challenges that need to be addressed one by one.

In addition, the development of the Brain 6 system itself, as well as the continuous development and improvement of its hardware and software, are quite expensive.

Finally, after initial discussions among Xu Shenxue, Fu Zhengyang, Lin Andong, and others, $8.5 billion was allocated to the first experimental integrated computing system that combines a superconducting quantum computer, a data center GPU system, and a neuromorphic computing system.

This is only the first batch of funding. If the subsequent work goes smoothly and successfully, that would be a good thing. If it doesn't go smoothly and additional funding is needed, then even more funding will be invested.

In theory, as long as there are no major technical problems that cannot be solved, Xu Shenxue will firmly promote the technological research and development and construction of this new type of comprehensive computing center!
It's not just about saving on electricity bills, but also about developing more advanced artificial intelligence technology.

In other fields, Xu Shenxue might consider the issue of expenses, but when it comes to the research and development of cutting-edge technologies, especially the core technologies of artificial intelligence, he has always been very willing to spend money.

The future is uncertain, but for now, artificial intelligence technology has become the core supporting business of Zhiyun Group.

Strictly speaking, artificial intelligence technology itself has not brought any revenue to Zhiyun Group... its direct revenue is very small.

However, many other businesses within the Zhiyun Group, especially those that generate substantial revenue, are built upon artificial intelligence technology.

The most typical examples are the two new core businesses: intelligent robots and virtual devices.

Without outstanding artificial intelligence technology, it would be impossible to create these two hardware devices that can be described as cutting-edge technologies, thereby driving the development of an entire industry chain.

Even in the traditional smart terminal field, Zhiyun Group has begun to apply artificial intelligence technology on a large scale. The S series mobile phones are directly connected to the Yun AI interface, which can provide users with various work and life suggestions, translation and other commonly used functions.

When promoting the S20 series phones, Zhiyun Group started describing them as AI phones, claiming that the S20 is the phone that understands you best on Earth...

The next step for Zhiyun Group's smart terminal business unit is to further deepen its cooperation with the Yun AI team, and further utilize local computing power to enable the built-in Yun AI to continue to understand users, learn on its own, and then understand users even better.

To put it simply, it's about creating a local and network-connected AI system for mobile phones... but this requires too much local computing power, so progress hasn't been significant.

Even with smartphones, Zhiyun Group has integrated artificial intelligence technology and made it a new selling point.

The same applies to other businesses; they basically use artificial intelligence technology wherever possible... Even the satellite internet communication network developed by Zhiyun Satellite integrates artificial intelligence technology.

The security operations department uses the group's internal artificial intelligence system to assist in its business operations...

The security team uses an artificial intelligence big data system to collect and monitor global public opinion, as well as big data on management and technical staff, in order to prevent intrusion by foreign commercial espionage agencies and to combat corruption.

Even those working in finance use artificial intelligence extensively to collect, organize, and analyze a wide range of information in order to provide more information for financial sector management to make decisions.

Incidentally, artificial intelligence technology is also very useful in the securities market... Among the major clients of Zhiyun Group's APO graphics cards, in addition to a large number of high-tech enterprises, a considerable number are financial institutions, especially various securities funds.

Wall Street financial institutions are the second largest customer group for Zhuyun Group's computing chip business in the US market... They purchase tens of thousands of APO graphics cards every year.

They then either develop their own models or deeply customize the open-source GTAI system to create large models suitable for their own use in predicting and analyzing the stock market.

Artificial intelligence technology has penetrated into all aspects of Zhiyun Group's various fields, and even has been widely applied to major enterprises around the world... its influence is growing.

The fact that Zhiyun Group's computing chip business revenue reached more than $80 billion in the third quarter is the most obvious proof.

Under such circumstances, it is conceivable that Zhuyun Group attaches great importance to artificial intelligence technology... as it is the underlying technology supporting more than 90% of the group's business!
Spending tens or even hundreds of billions of dollars to improve artificial intelligence technology is very worthwhile and necessary for Zhiyun Group.

However, this money cannot be spent openly... because the Brain 6 system cannot be made public; it remains a classified project.

There are many classified projects within the Zhiyun Group, such as the superconducting quantum computer project.

The same applies to projects such as the research and development of room-temperature superconducting materials, some artificial intelligence projects, and many projects in cooperation with other institutions. These projects cannot be made public due to their confidentiality, but they still require money, so they obtain the corresponding research funding through non-public programs under the research committee.

Of course, this is just the internal structure; on the surface, these projects will be incorporated into some of the publicly known research and development.

For example, the research and development of room-temperature superconducting materials is directly linked to materials research... However, there is no need to disclose to the public what the specific research and development of materials is or what results have been achieved.

The superconducting quantum computer project, as well as the early AI series graphics card projects, were directly linked to the data center GPU project... Therefore, when investors look at the financial statements, they will find that Zhuyun Group has particularly large expenditures on data center GPU R&D and self-built data centers.

The development costs of data center GPUs, in particular, seem extremely high, which has led to some misjudgments from the outside world, who believe that the development costs of a graphics card like the APO6000 are extremely high!
In reality, the APO series graphics cards are mature projects that Zhiyun Group has been working on for many years. Now it's just routine technology advancement. Its R&D funding is not as high as outsiders imagine, at most a few billion US dollars a year. Then, adding the R&D of data center related technologies, such as data center communication technology, and other miscellaneous R&D investments in the data center GPU field, the annual investment is actually only around tens of billions of US dollars.

However, in the financial report, this figure was directly inflated to over 20 billion US dollars!

The extra 10 billion US dollars includes superconducting quantum computers and corresponding superconducting quantum computing centers, as well as some secret projects, such as defense computing chip projects.

This would scare many outsiders who don't know the inside story... It's just developing a GPU, and your Zhiyun Group spends more than 20 billion US dollars a year, are you crazy!
This makes many chip design companies very anxious when developing their own computing chips... they don't have that much money.

Of course, AMD also has its own perspective: Developing data center GPUs and related data center and communication technologies really requires that much money... and it's still not enough.

In June of this year, the United States passed a so-called plan to support local computing power chips, which aims to provide special support for American companies to research and develop and produce data center GPUs and related technologies.

As the only company in the United States currently developing data center GPUs, AMD has submitted a report stating that with the continuous advancement of process technology, the development cost of the next generation of data center GPUs based on 5-nanometer process technology will increase significantly, and it is estimated that developing the next generation of data center GPUs will cost more than 30 billion US dollars!

AMD's profits in other areas are no longer sufficient to support such a massive investment in data center GPU R&D.

Then they said that over the years, AMD has invested heavily in the data center GPU field, but the total revenue from the data center GPU business over the years is less than 10 billion US dollars, resulting in huge losses.

These limited revenues all come from various support orders from the United States.

For security reasons and to prevent potential backdoors in APO graphics cards, the US would never use them in security or sensitive areas... Even if Zhiyun Group were willing to sell them, they wouldn't dare use them...

For example, if they were to develop a theater AI system... they were worried that if they used APO graphics cards, missiles might end up landing on their heads.

Zhiyun Group has consistently maintained that APO graphics cards have no backdoors, and has even submitted various so-called technical funds for security audits, but Americans don't believe it... Heaven knows whether the data and information provided by Zhiyun Group are true or false.

After all, the APO6000 graphics card has hundreds of billions of transistors in its GPU core alone, making it impossible for outsiders to check them all individually… They can only accept whatever Zhiyun Group says. Otherwise, how could you possibly examine and inspect every single one of those hundreds of billions of transistors?

This is worrying, as there are concerns about backdoors, and there's no way to conduct audits. Therefore, to ensure security and prevent missiles from flying over their heads one day.

They are using AMD's data center GPUs in many systems, such as national defense and security systems and some civilian systems, which is also the main source of revenue for AMD's data center GPU business.

With the support of subscriptions and direct subsidies every year, AMD actually receives a considerable amount of money each year. Therefore, AMD has not lost much money in its data center business, but it is not profitable.

AMD is a private company and a publicly traded company. It's a loss-making business that it keeps running every year without making money, consuming a lot of funds and top technical talent!

Furthermore, the development costs of next-generation 5-nanometer process data center GPUs are higher. Not only are the GPU core development costs higher, but the research and development expenses in many other fields such as communication technology and data center technology are also rising.

AMD's internal estimate is that even at the lowest estimate, it would require tens of billions of dollars in R&D investment, and there's no guarantee they'll be able to develop it... This isn't just a GPU core, but an entire system.

Finally, at the hearing, AMD gave a rough figure: $30 billion!
Based on internal cost estimates, they doubled the price... Why double the price? Naturally, it's because AMD isn't the only one getting paid; there are a whole bunch of other people waiting to share the profits.

And this is just the starting price! AMD formally disclosed the technical risks when they participated in the hearing!
Although the project is relatively well-planned, there are still many uncontrollable technical risks!
Therefore, there is no guarantee that spending 30 billion US dollars will definitely result in success. There is also the possibility of encountering various technical problems, which could lead to project delays and increased costs.

It's very difficult to say how much it will ultimately increase.

After all, their goal is to create a top-tier data center GPU that can rival the APO6000 graphics card... Even if the ecosystem can't keep up, the hardware level has to be up to par.

However, even with its 5-nanometer deep customization process, top-tier 3D packaging technology, and a host of the latest cutting-edge technologies, the APO6000 hasn't been officially supplied to Zhuyun Group yet.

If AMD wanted to do something like this, it would be crazy to do it without hundreds of billions of dollars being poured in... Without external funding, AMD wouldn't be able to cover its losses even if it poured all the profits from its other businesses into it each year.

That's a massive sinkhole, the kind that could cost you money or even your life!

So, what AMD is saying is: the next generation of 5nm data center GPUs is feasible... but the ecosystem is uncertain, the technical performance is uncertain, and they're not even talking about computing power costs... there's nothing they can do about these things.

However, they can definitely produce the chip's technical specifications, such as the number of transistors and processing cores; the only difference might be higher power consumption. If they tell you there's a 5-nanometer data center GPU, they can definitely produce one.

But you have to put up the money first, whether it's direct subsidies or high-price support for orders, it's all fine, but don't expect me to invest my own money, I really can't afford to lose money like this!

This is how the $30 billion came about!
The biggest evidence that AMD's estimate of starting at $30 billion and not necessarily being able to achieve it even with $50 billion is the financial report of Zhiyun Group.

Last year, the group spent over US$23 billion on data center GPUs, including GPU core design, communication technology research and development, data center research and development, etc.

That's a whole year!

The Zhiyun Group spends over 20 billion US dollars a year, while my entire project package price is only 30 billion US dollars, which is already very cheap... First, use something that looks better and costs less to get the project started, and then find various excuses to raise the price. This is the basic operation of American companies.

Even if the project is delayed and costs double the expected amount, ending up costing $50 billion or even $60 billion, don't be too surprised, you congressmen... It's all quite normal. AMD is already preparing people for further price increases.

After all, once such a large project begins, money is being spent every day. If it is delayed by even a year, it will amount to at least tens of billions of dollars in expenditure!
Zhuyun Group spent tens of billions to develop the APO6000 graphics card, so why can't we spend tens of billions? That doesn't make sense!
AMD said a lot, but it all boiled down to one thing: send the money quickly, or they're not going to do it...

AMD used Zhiyun Group's financial statements as evidence to demonstrate how much money was needed to build data centers, and then tried various ways to obtain funding support... The US side really had no choice but to grit their teeth and continue to provide support... Even though they knew that in the end, probably less than half of this money would actually be spent on research and development, but there was nothing they could do!
After all, we've been supporting them for so many years, and the sunk costs are considerable... Besides, they actually have a share of this money too!
Of the $30 billion in support funds allocated, only a large portion will ultimately reach AMD. Where does the rest of the money go?
Of course, everyone should share the money... The councilors also have families to support, and they are under a lot of economic pressure. If they don't share the money, how will they support their families?

Of course, the above are all secondary reasons, since they can make money no matter what project they support.

Most importantly, AMD's data center GPUs are truly indispensable in the security field...

They dare not use Zhiyun's APO graphics cards. Although other American companies also have data center GPU projects, such as Intel and Qualcomm, their data center GPUs are even worse.

They want to support these companies' data center GPU projects and then split the profits, but they can't even come up with a good excuse.

It's better to grit your teeth and continue supporting AMD. At least AMD can actually provide you with a data center GPU, even if the performance is lackluster and the ecosystem is even worse. But at least they can provide it, which is an easier excuse to find and explain to the outside world.

After spending so much money, at least they should be able to produce something. It's not like they've spent hundreds of billions of dollars over several years and ended up with nothing. It would be hard to explain to the public... Now a bunch of people are criticizing them. They allocate so much money to AMD every year, but after all these years, they still can't compete with Zhiyun Group.

Not to mention competing with APO graphics cards from Zhiyun Group, the gap is actually widening, and there is a lot of public criticism about this, with many feeling that AMD is just taking money without delivering results.

AMD is using Zhiyun Group's financial statements as evidence to demand money, but Zhiyun Group's financial statements themselves are inflated...

Whether AMD can finally produce its next-generation 5nm data center GPU, and how good its performance will be, is hard to say, but those involved will definitely make a fortune!
The reason why many items in the financial statements of Zhiyun Group are inflated is that Zhiyun Group has also included a large number of non-public items.

This neuromorphic computer project is no exception; its massive research and development was directly applied to the development of data center GPUs.

The funding for building the first generation of computing centers was also included in the overall computing center expenditures... Therefore, outsiders can see that Zhiyun Group's annual investment in data center construction is also very extravagant.

After finalizing the follow-up research and development plan for the Brain 6 system and the plan to build a new generation of comprehensive computing power center, Xu Shenxue gave instructions on the relevant matters and did not pay attention to them every day... In the end, this kind of thing cannot be done in a day or two.

Optimistically speaking, the first new integrated computing center will not enter the construction phase until next year at the earliest. Before that, various technical issues will need to be resolved, and it is optimistically estimated that it will not be operational until the year after next.

As for the pure AI6 system, such as a single server, we can see that now, but a single server is not very useful.

The unique neuromorphic computing system of Brain 6 emphasizes the number of neurons. The less than two hundred neurons on a single server are actually not very useful!

Even if you combine a dozen or so servers, resulting in hundreds of billions of neurons, their effect is still minimal.

According to Professor Lin Andong's technical approach, at least a hundred servers with about two trillion neurons need to be built to barely cooperate with existing superconducting quantum computers and data center GPUs, and finally bring out their unique advantages in the fields of low-precision computing and fuzzy computing.

Ultimately, only through computing centers composed of such integrated computer systems can we train new artificial intelligence technologies that are at least one level more advanced than existing ones.

A small quantity is useless, but combining so many Brain 6 chips is no easy task... Its difficulty is no less than that of the early stages of general-purpose computing supercomputers or GPU computing centers in previous years, and may even be more difficult.

Finally, there is an even more crucial question: how to effectively connect with these three different types of computing systems, work together, and jointly train artificial intelligence.

There are huge technical challenges involved here.

In other words, the technical team at Zhiyun Group was extremely confident, believing that their R&D personnel were all top geniuses, the elites in their respective fields. Coupled with ample funding and various resource support, they were confident that they could develop it within the expected timeframe.

If technicians from other companies wanted to develop this kind of system, they would probably be completely baffled... Integrating three completely different types of computer systems to work together is technically very difficult.

Currently, apart from the technical team at Zhiyun Group, which has two possible solutions, other companies basically have no chance of handling this.

However, as mentioned before, other companies don't need this technology either, since they don't have superconducting quantum computers with tens of thousands of controllable qubits, nor do they have mature and usable neuromorphic computing systems.

(End of this chapter)