Skip to main content
  • Home
  • Library
  • Read
    • Popular
    • Charts
    • Browse
    • Categories
    • Wikis
    • Glossary
  • Request
  • Activity
  • Leaderboard
  • Help
Our Apps
  • DaoSearch
  • Narratorr
Social
  • Discord
  • Patreon
The AI Wave Tycoon's Rise
Chapter 184

Forget it. Not a chance

2,977 words 10 min read 2 views 1mo ago Source ​
Skip to the end of the chapter
Prev
Next

After Thursday's power struggles, Zhou Han didn't cause any more trouble.

Or rather, the necessary probing has been done, and the outstretched claws have been withdrawn.

On Friday morning, State Grid Investment held an internal meeting.

At 11:47 a. m., as soon as the meeting ended, Zhou Han immediately called Han Luyi's cell phone.

"Xiao Han, congratulations." Zhou Han's voice was filled with laughter, her tone much warmer than yesterday. "You passed."

Han Luyi stood in front of the floor-to-ceiling window of the hotel room, looking down at the traffic below, and smiled, "Thank you, Sister Zhou."

Zhou Han said, "At 2:30 this afternoon, in the same conference room as yesterday, the legal department, the investment committee secretary, and the risk control team will all be there. Make sure you have the official seal ready. If Lawyer Gu has no further objections, we'll sign the agreement today."

At the signing ceremony, the atmosphere was noticeably more relaxed compared to the negotiations the previous day.

Zhou Han looked at him and said with a smile, "Xiao Han, let's confirm one last time. State Grid Investment has invested 300 million yuan in this round, acquiring 10% of Source Intelligence Technology's shares after the capital increase, and will appoint one director. The list of major company matters is subject to today's version of the shareholders' agreement."

Han Luyi nodded. "I can confirm."

"Power support and Data Center construction coordination are outlined in the strategic cooperation memorandum, but their specific implementation will depend on local approvals, power availability, and subsequent construction plans."

"Confirm."

"There's one more thing," Zhou Han said, looking at him and slowing her tone slightly, "Because of the state-owned shareholders, if Source Intelligence Technology subsequently involves overseas listing, VIE structure, or the transfer of core assets abroad, it will need to go through the approval process again. You're aware of that, right?"

Han Luyi picked up his pen and smiled.

"Of course I know."

Zhou Han carefully observed his expression. "What, don't you think it's a pity?"

"It's not a pity."

Han Luyi took off his pen cap and placed it on the table.

"Source Intelligence Technology was not originally founded with the goal of ringing the bell on Nasdaq."

Several people in the conference room looked up and glanced at Han Luyi.

Han Luyi calmly stated, "AI Large Model research is fundamentally different from other fields. It certainly requires money, and a lot of money, but it's not a field solely driven by money."

"Computing Power, data, engineering systems, industry scenarios, and security boundaries—all of these are indispensable."

"If it were just about telling a story, raising US dollar funds, creating impressive financial statements, and ultimately going public to raise money, then I wouldn't have needed to come to Beijing, nor would I have needed to take the money State Grid Investment."

He paused for a moment after saying this.

"What we're doing is creating a model base and implementing it in real-world scenarios, not just packaging for an IPO."

Gu Siyue sat down next to Han Luyi and turned to look at him.

Several copies of the agreement circulated between the two parties, one name after another was written at the end of the agreement, the legal department confirmed the signature page, and the secretary checked the number of copies.

The stamp fell with a dull thud, as if bringing a perfect end to the tug-of-war of the past few days.

The signing was completed at 4:20 PM.

Ultimately, Source Intelligence Technology secured the power support and 300 million yuan in cash promised by State Grid Investment, giving up only 10% of its shares and a board seat.

As the largest shareholder, Han Luyi retained absolute control over Source Intelligence Technology.

After the signing ceremony, Han Luyi saw Zhou Han off, and she warmly shook Han Luyi hand goodbye, saying, "Little Han, it was a pleasure working with you."

"Sister Zhou, it's a pleasure working with you."

"The due diligence team will be in Haicheng soon, and I'll arrange payment as soon as the due diligence is complete," Zhou Han said, then lowered her voice and added, "I look forward to your product being developed quickly and successfully implemented in the power sector."

Han Luyi smiled and said, "Then we'll have to trouble Sister Zhou a lot more later."

"It's not exactly troublesome," Zhou Han smiled. "If the project succeeds, it will benefit everyone."

That's a very straightforward statement.

The elevator doors closed, blocking Zhou Han's smile from view.

Facing the closed elevator doors, Han Luyi's smile slowly faded.

Gu Siyue walked to his side, noticed his expression, and asked, "What's wrong?"

"It's nothing," Han Luyi said, "I'm just a little tired."

Gu Siyue said calmly, "You'll get used to it."

Han Luyi closed his eyes. "Yeah, you'll get used to it."

After several clashes over the past few days, and after Zhou Han revealed her passion and fangs, Han Luyi realized something.

People like Zhou Han are neither the kind of benefactors people talk about, nor are they simply bad people.

She can smile and express her appreciation for you, but she can also test you ruthlessly during negotiations.

Her helping you doesn't mean she's your ally.

The fact that she's making things difficult for you doesn't mean she's your enemy.

She, they, were all just political animals.

The relationships between these people are complex and intertwined, so it's best to observe and listen more, and speak less.

Han Luyi soft sigh of relief.

On Friday night, Han Luyi sat on the plane bound for Hongqiao, reflecting on his experiences in Beijing over the past few days.

He came with a do-or-die mentality, but unexpectedly, he not only secured investment from Hongyuan, but the valuation was also higher than previously agreed.

More importantly, they simultaneously came into contact with both State Grid Investment and the Strategic Emerging Fund.

Han Luyi recalled a sentence He Yunshen had said to him privately at Hongyuan that day, "Capital is just an entry ticket, resources are the real currency."

He didn't understand it then, but now he does.

He Yunshen provided the money. Zhou Han provided the set and electricity.

If they can also secure a deal with the Strategic Emerging Fund, Source Intelligence Technology will have a much stronger hand than what a startup should have.

With these resources, Source Intelligence Technology is much closer to the next stage planned by Han Luyi —owning its own Data Center and independent training resources.

However, even with the help of his Sight, the scheming and intrigue with those old foxes left him feeling deeply exhausted.

He even felt a little admiration for him.

Those who lack Sight, living in such a vortex every day, still manage to stay clear-headed and move forward. How difficult that is.

Beside her, Gu Siyue was looking at her phone.

The cabin was dimly lit, and her profile was reflected in a cold white light by the screen.

"What are your plans when you get back?" Han Luyi asked.

Gu Siyue turned around and glanced at him.

"Boss Han, I've been on a business trip for three days to secure this contract. I have a mountain of cases on my plate, and I'm afraid I'll be working overtime all weekend."

"Yes, yes," Han Luyi smile, "Thank you for your hard work, Lawyer Gu."

Gu Siyue turned her head away, refusing to look at him. "This isn't work time."

Han Luyi, then realized what was going on.

It's not during working hours, so please don't call me Lawyer Gu.

He smiled, didn't say anything more, but closed his eyes and began to look forward to Glutinous Rice Ball's training progress.

After Zhao Wenyuan returned to Haicheng, he hadn't had time to follow up carefully.

But he knew the progress should be good.

Because the experience value of the Sight has increased from 10% to 15%.

This shows that Glutinous Rice Ball's training is bringing new feedback to his Sight.

This also proves that his judgment was correct.

The path of Large Model is indeed deeply connected to the Sight.

Han Luyi leaned back in his chair, feeling a little more relaxed.

My trip to Beijing is over; it's finally time to return to Haicheng.

However, at this moment, Zhao Wenyuan, sitting in his office, was not in a good mood.

It wasn't because of the pre-training with Glutinous Rice Ball.

On the contrary, Glutinous Rice Ball's pre-training progressed very smoothly.

The industry's pre-training strategies are already quite mature. After returning to Haicheng, Zhao Wenyuan read several papers and adjusted the specific training strategies and parameters using a small dataset. After confirming that there were no obvious problems with the loss curve, he officially launched the full-scale training on Dingsheng's cluster.

From that moment on, the process became a true test of patience and perseverance. Once a training session began, it often lasted for dozens of days without interruption.

Hundreds or thousands of high- Computing Power graphics cards form a cluster, and massive amounts of data are segmented, packaged, and fed into the model.

Every second, countless matrix calculations are transmitted between graphics cards.

Periodically, snapshots of intermediate results should be taken to prevent data loss due to occasional failures.

Snapshots shouldn't be taken too frequently, as this will slow down training efficiency. Nor should the intervals be too long, as long intervals could lead to the loss of several hours or even days of training if the cluster fails.

The funding requirements, engineering capabilities, cluster scheduling, and fault recovery involved are not something small companies can easily afford.

This is why Zhao Wenyuan has always said that Large Model is not something everyone can afford to do.

Without sufficient Computing Power, you can't even get to the poker table.

Without a strong enough engineering team, even if you get to the poker table, it's useless; you'll only be dragged down by your own training tasks.

In the case of Source Intelligence Technology and Dingsheng signing a cooperation agreement, Zhao Wenyuan received a specially configured account.

This account has internal permissions at Dingsheng Cloud and can access GPU clusters specifically designed for Model Training. Such GPU clusters are not services that ordinary customers can simply purchase by opening a webpage and topping up their account balance at Dingsheng Cloud or any other cloud service provider.

You can't buy it even if you have the money.

Therefore, Han Luyi was able to obtain Dingsheng's Computing Power by taking a huge shortcut.

Zhao Wenyuan acknowledged this, but he was also very unhappy because they had to guard against Dingsheng stealing the Annotated Data.

The Glutinous Rice Ball produced in this round of training can only be a "crippled version".

Zhao Wenyuan dared not put the most core, most valuable, and most representative data of Source Intelligence Technology's advantages into Dingsheng's cluster.

It's like a master chef finally getting access to a top-notch kitchen, but having to hide the most essential ingredients and cook with only half the recipe.

How could you possibly make a dish that glows?

However, these are not what bothers Zhao Wenyuan the most right now.

What truly annoyed him was the new task Han Luyi had given him before returning from Beijing: to research compatibility with domestically produced graphics cards.

When leaving the Beijing, Zhao Wenyuan excitedly boarded the plane.

It could even be described as being overwhelmed by excitement.

Domestic Computing Power, independent control, and freedom from the overseas GPU ecosystem make large-scale Model Training and inference possible on domestically produced graphics cards.

Just mentioning these keywords is enough to get tech enthusiasts' blood pumping and their hearts racing.

Zhao Wenyuan even created a document on the plane, listing several pages of plans.

<Roadmap for Domestic Computing Power Adaptation of Tangyuan Model>

As soon as Zhao Wenyuan got off the plane, he didn't even go home; he went straight to his office with his suitcase.

He poured himself a cup of coffee, placed it on the table, turned on his computer, rolled up his sleeves, and prepared to break free of the ecological shackles binding domestically produced graphics cards.

As a result, Zhao Wenyuan bumped his head against the wall.

No, it can't even be called a wall, it's practically a mountain.

The CUDA ecosystem dilemma, the nightmare that tormented Google back then, has reappeared in an even more terrifying form.

To the average person, a graphics card is just a graphics card. Foreign graphics cards are considered graphics cards, and so are domestic ones. The only difference is whether the performance is higher or lower.

But for those who build Large Model, a graphics card is not just a standalone piece of Hardware; it represents an entire ecosystem.

At the bottom layer is the driver, above the driver is the runtime, above the runtime is the programming model, and above the programming model are Operators, communication libraries, compilers, debuggers, and performance analysis tools.

Above that are training frameworks like PyTorch, TensorFlow, and JAX.

Large Model Training, in turn, builds upon these frameworks, utilizing various highly optimized Operators and distributed training capabilities.

Nvidia's CUDA ecosystem is like a super city that has been under construction for over a decade, with its infrastructure, such as water, electricity, fire protection, and hospitals, already in operation for many years.

Training Large Model is like hosting a major sporting event in this city.

The influx of hundreds of thousands of tourists puts a huge strain on the city's capacity. But because it's a megacity, it can manage the influx with careful planning.

But what about a different ecosystem? Imagine you're in a small town next to a big city, and you're asking, "Can I host the World Cup too?"

Of course we can try, but it's not as simple as just inviting two football teams to play. The entire town, from its underground pipe network to its commercial ecosystem, needs to be rebuilt. Once it becomes a megacity, then it's ready.

What's worse, people all over the world have gotten used to the rules of this old city. When they write code, they assume that CUDA can be called by default, that Nvidia graphics cards are used by default, and that the test cases mentioned in the papers can only run on Nvidia.

Not using Nvidia? Then try it yourself.

The code that was running fine suddenly started generating a thousand errors after I removed the CUDA dependencies. After fixing the first error, another thousand appeared.

A model that can run doesn't mean it can run fast; a small model that can run doesn't mean Large Model can run; a single GPU that can run doesn't mean multiple GPUs can run; and multiple GPUs that can run doesn't mean Thousand-Card Cluster can be stably trained for dozens of days.

The most terrifying thing about Large Model is that they are not only successful if they can "run", but they must also be stable, efficient, and robust.

It is essential to maximize the utilization rate of each card while maintaining extremely high Computing Power costs.

Otherwise, if you spend the same amount of money, others train for one month while you train for three months. Others burn ten million while you burn thirty million.

The final model results may not even be as good.

Zhao Wenyuan devoted himself to the study for two days, and the more he studied, the uglier his expression became. The biggest difficulty was not that it was technically impossible.

On the contrary, many things can theoretically be done.

The biggest problem is that the workload is too large.

This is not a problem that a genius programmer could solve by secluding himself for three months.

This is an ecosystem built up by the industry over more than a decade, by countless companies, developers, papers, and open-source projects.

It's like the Foolish Old Man Who Moved Mountains. Let alone Zhao Wenyuan alone, or the small group of people in the Source Code Technology model team, even if Han Luyi gave him a development team of several hundred people, it would probably take several years to create a usable version of the most core part.

Is this realistic?

That's unrealistic.

Zhao Wenyuan scratched his hair hard, then looked at the hair he had pulled out and touched it with heartache.

He opened WeChat and found a Chinese colleague from his time at Google.

That colleague did adaptation work for TPU back then. Google's approach of developing its own Chip and frameworks faces the same dilemma; they just managed to overcome it by investing heavily back then.

Zhao Wenyuan wrote a long message, briefly explaining his predicament.

How can we adapt to new Hardware at the training framework level? What are the most difficult shortcomings to overcome in the early stages of a self-developed Chip ecosystem? If a startup team wants to run Large Model on domestically produced AI Chip, are there any realistic approaches?

This may involve some confidential information. Zhao Wenyuan hadn't been in contact with this former colleague for two years, and he didn't expect to receive a reply anyway.

To my surprise, they replied half an hour later.

Upon seeing the message, Zhao Wenyuan laughed in anger.

The reply was brief, "Don't even think about it, it's hopeless."

It turns out that people who are extremely popular really do smile.

After a while, the other party sent an even longer message.

"The TPU path is not something startups can replicate. If you just want to save money, just buy Nvidia cards. If you want to support domestic Chip, then let the Chip manufacturers build the ecosystem; you can only do application-level adaptation at most. You can't spend tens of millions of dollars to support the ecosystem of Hardware manufacturers, can you?"

Tens of millions? US dollars? My friend, you're being conservative.

Zhao Wenyuan locked his phone, turned on his computer, and looked at the document he had written on the plane.

<Roadmap for Domestic Computing Power Adaptation of Tangyuan Model>

It's somewhat ironic.

He deleted the title and retyped it.

<Tangyuan Model Domestic Graphics Card Compatibility Assessment: Not Feasible at the Present>

Zhao Wenyuan was the only one left in the entire office building. He looked out into the darkness through the glass, his own face reflected in the glass, looking somewhat unfamiliar.

Just then, the phone vibrated, and the screen lit up again.

He looked down and saw it was a message from Han Luyi.

"Wenyuan, I just landed in Haicheng. Let's check the Model Training progress on Monday morning."

Zhao Wenyuan hesitated for a moment, unsure how to reply to Han Luyi.

End of Chapter
Prev
NovelSummaries
Next

Explore the wiki

Han Luyi
Zhao Wenyuan
Gu Siyue

Comments

No comments yet.
AboutStatsAPIChangelogStatusContactPrivacyTermsDMCA
© 2026 RandomTranslator·v3.26·RSS