news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

New Intelligence Report

Editor: Qiao Yang

【New Wisdom Introduction】Since last year, Altman has been flying around the world, looking for people to attract investment and discuss cooperation, and actively promoting his plan to realize "self-developed chips" within OpenAI. Now, this idea of ​​"decoupling" from Nvidia has finally made some practical progress.

OpenAI, which just released GPT-4o mini today, has new developments!

SemiAnalysis broke the news at the beginning of last month. He recruited top talents from Google's TPU team and vowed to get rid of dependence on Nvidia.

But not now. The estimated time of realization has been scheduled to the end of 2027.

Today, The Information further revealed that Ultraman is also in contact with Broadcom and plans to cooperate in the development of this self-developed chip.

Completely abandoning the current supplier Nvidia and starting to develop its own chips is actually a risky move.

NVIDIA’s current leading position in the industry is inseparable from its years of technological accumulation.

Meta, Microsoft, Google, Amazon, many technology giants have continued to invest for several years, but have not been able to squeeze out Nvidia's market share. Only the TPU developed by Google can achieve the same level of performance.

Even if the top team is used to "catch up" at this moment, it will take at least several years for the chips delivered to truly replace Nvidia's products.

Whether this attempt at "decoupling" will offend Nvidia, especially Huang himself, is anyone's guess.

But if we look at it from a positive perspective, things could have gone another way - OpenAI, because it has alternative solutions, has more bargaining chips in its deal with Nvidia, and can stockpile chips at a cheaper price.

Although the future is uncertain, it is not difficult for us to understand because Ultraman has always had this "radical" style of doing things.

He has long said OpenAI needs far more computing power than it currently has.

For example, this tweet published in February this year roughly expresses the hope to build more AI infrastructure and a more flexible supply chain system, including wafer production capacity, energy, data centers, etc.

Currently, OpenAI is renting Nvidia servers through Microsoft at a cost of about $2.5 billion per year.

In addition, according to a report by The Information in March, the two companies have been brewing an ambitious computing power plan: , which could cost up to $100 billion.

Although Ultraman's plans for chips and computing power are so complex and dazzling, these are not the ultimate goals.

What he ultimately wants to achieve is OpenAI’s grand mission - AGI or even ASI.

This kind of artificial intelligence can help humans colonize Mars or develop fusion energy, and it is definitely not on the same level as the "chatbot" GPT we see now.

Ultraman travels around the world for chips

Although OpenAI is not the only tech giant that is short of chips, the urgency of Ultraman's efforts is definitely among the highest as can be seen with the naked eye.

According to a senior manager of TSMC, starting from 2023 and continuing to this year, Ultraman approached various chip giants including TSMC to discuss the possibility of increasing production capacity.

However, most executives disagreed with the plan, and were unwilling to invest a large amount of capital and professional labor to increase production just for OpenAI. Doing so was too risky.

Last month, TSMC CEO CC Wei was asked about the prospects of the new factory and said that Ultraman's plan was "too radical to be believed."

But that was only the public claim. According to a senior TSMC manager, company executives privately promised Altman that they would be willing to expand chip production if he or OpenAI could commit to placing large orders for the new chips.

This arrangement is more feasible for TSMC because it can use existing factories instead of risking money and building dozens of new factories.

In addition to trying to stockpile more chips from suppliers, Ultraman is also simultaneously pushing forward the establishment of new data centers.

According to a senior executive in the industry, Altman plans to cooperate with external investors to "start a new business" and set up a new company specifically responsible for the construction of data centers, and then let OpenAI rent servers from this company.

People familiar with the matter said that Ultraman has sought approval from the U.S. Department of Commerce for the plan and is in talks with governments of other countries, such as in the Middle East, for cooperation.

Several governments, including the UAE, have said they want to develop AI-focused data centers to boost their economies and improve their countries’ standing in the tech industry while protecting local data from leaks and outside influence.

In addition, Altman is also considering raising funds from private equity firms and other investors.

For example, Softbank, a Japanese company that once invested $3 billion in perplexity, Altman even discussed an independent "Altman project" with them.

The Wall Street Journal reported a few months ago that Altman was negotiating with relevant investors, and the total amount of funds he hoped to raise was as high as 7 trillion US dollars!

This figure even exceeds the total national debt of some major economies.

According to the Securities Industry and Financial Markets Association, the total bond issuance of all US companies last year was $1.44 trillion. The total market value of Microsoft and Apple, the two most valuable companies in the United States, is only about $6 trillion.

Altman privately explained the composition of this astronomical figure: 7 trillion includes the total investment of the project in real estate, electricity and chip manufacturing over several years.

At present, we are not sure whether Ultraman has successfully attracted investment and whether it has started registering a company.

Altman himself rarely discusses specific details of his data center and chip plans publicly, but he has privately told industry executives that he will announce the plan this year.

According to people familiar with the matter, this plan caused Ultraman considerable trouble last year, and was even one of the triggers for the "recall storm" at the end of last year.

At the time, his idea was to keep the data center project independent of OpenAI, but the board was worried that it would distract him from OpenAI.

However, after some "palace fighting", this plan now has almost no resistance within OpenAI.

Earlier this year, Altman informed colleagues that OpenAI would take a stake in the new project, and that the new board had begun considering the proposal.

Challenge Nvidia?

According to The Information, a person involved in Altman's conversation said that OpenAI's chip team will be led by Richard Ho, and it is likely to choose an American company to jointly develop chips, with Broadcom being the most likely candidate.

After all, Broadcom had worked with Google to build TPUs and had already been in communication with the OpenAI chip team.

Richard Ho, the helmsman of the chip team, just joined OpenAI in November last year. He must be a talent specially recruited by Ultraman for the "chip manufacturing" plan.

Richard Ho graduated from the University of Manchester in the UK with a bachelor's degree in microelectronics engineering, and then went to Stanford to pursue a doctorate in computer science.

Since 2014, he has joined Google's hardware and semiconductor team as a senior engineer, participated in the TPU project, and was gradually promoted to chief engineer and senior engineering director until he left in 2022.

People familiar with the matter said that the OpenAI team does not seem to have started designing chips yet, and formal production will have to wait until at least 2026.

In addition, the R&D team is considering various memory component suppliers and possible chip packaging solutions in order to achieve optimal performance.

For example, one of the most important components of GPU, the HBM chip, directly affects the computing performance of the chip.

HBM is responsible for providing high-speed connections between processors and memory, speeding up data transmission and processing, thereby accelerating AI computing.

This type of memory chip is as hot-selling as GPUs and is always in short supply.

Altman discussed his chip plans earlier this year with Samsung and SK Hynix, both top HBM chip makers, according to people familiar with the matter.

Direct negotiations with upstream suppliers may have a significant impact on Nvidia's business, and it is difficult to say how Nvidia will react to this.

But the fact is that it’s not just OpenAI. Companies like Meta, Microsoft, Google, and Amazon are also investing heavily in developing their own AI accelerators.

These AI giants are not only Nvidia's largest customers, but are also becoming its most powerful competitors.

References:

https://www.theinformation.com/articles/openai-has-talked-to-broadcom-about-developing-new-ai-chip?rc=epv9gi

https://www.wsj.com/tech/ai/sam-altman-seeks-trillions-of-dollars-to-reshape-business-of-chips-and-ai-89ab3db0