Chapter 246: Backend Data

“Tell me more.”

It wasn’t Park Cheol-jin who answered my question, but Chilbong.

As the leader of the entire semiconductor development and TF team, he understood the situation more clearly than anyone else.

“Recently, the performance of Luna revealed during the live broadcast of the Polar Orbiter mission and the autonomous driving test seems to have definitely stimulated the market.”

Chilbong manipulated the computer in the conference room to show the search volume data for the past two weeks.

[Global search trends related to Luna (LUNA): +310% increase]

[Changes in mentions of Black Robin, Dojin AI chipset, and LAD autonomous driving]

-KTalk Feed: +482%

-Twister: +760%

-Weibo: +370%

“Looks like the mood is set.”

Chilbong nodded.

“It seems that many companies have truly felt the economic value of AI through this autonomous driving live broadcast.”

As Chilbong flipped through the materials,

the test scene of LAD based on Luna began to play as a real-time clip.

“There are currently three types of Black Robin chipsets being produced, right?”

-Black Robin 1: For cloud AI learning

-Black Robin E: For real-time inference in vehicles (installed in EV-Nova)

-Black Robin L: Combat-use embedded chipset (AGE, SWAN drone, etc.)

Chilbong displayed the status chart on the screen and added further explanation.

“The cumulative production to date is 5,500 units of Black Robin 1 for cloud use, 60,000 units of Black Robin E for vehicles, and 230,000 units of Black Robin L for combat use.”

“The quantity for cloud use seems too low…… Is it because of the yield?”

Chilbong quietly nodded.

“Yes, there are many constraints in production due to yield and process stability issues.”

Chilbong pressed the remote control to display the yield trend of the Black Robin chipset.

[Yield changes for Black Robin 1]

-First 3 months: Average 15.3%

-Recent 1 month: Average 23.7% (some lots reaching over 30%)

“Most of the Black Robin 1 produced so far have been made using the LUNA-EVX prototype equipment. Ah, by the way, we recently finalized the equipment name from ‘LUNA-EVX prototype’ to ‘LUNA-EVX1.’”

Since I already knew about the equipment name, I lightly nodded.

“The production cost must be quite high.”

“Yes. With a 2.7nm process, only 32 to 35 chipsets can be produced from a single wafer. Moreover, 75% of them are defective……”

“So, all the Black Robin 1 produced are being applied to the Luna cluster?”

“Correct. Currently, the Luna cluster consists of 860 nodes, with a total of 5,440 chipsets installed. The actual operating rate is about 94%.”

I asked while looking at the cluster configuration diagram on the screen.

“This seems too weak to reveal Luna to the outside world.”

“I agree.”

Chilbong nodded in agreement with my words.

“Actually, the learning response shown in the LAD video is limited to simple traffic scenarios. To reveal the full version of Luna, at least a learning cluster with over 50,000 Black Robin 1 chipsets needs to be completed.”

“Hmm…… Considering stability, parallel learning of multiple scenarios, real-time streaming, and certification tests in various countries…… That means we need at least 75,000, right?”

“In fact, that’s why the disclosure of the Luna platform keeps getting delayed.”

As Chilbong said, it took over a year just to secure 5,500 Black Robin 1 chipsets.

Even if we increase the production capacity, it would take at least another 2 to 3 years to reach the target quantity.

But……

“Won’t the yield and production issues be somewhat resolved once the LUNA-EVX2 equipment development is completed?”

The currently operating EVX1 equipment is limited to a 2.7nm process.

However, the LUNA-EVX2 being developed by Team Leader Park Sang-bae is a next-generation equipment capable of implementing a 2.0nm-level process.

“Definitely, once EVX2 is completed, the yield and production issues will be significantly improved.”

“And the Black Robin 2 chipset will also become producible.”

While the computational performance of the current Black Robin 1 is around 500 TOPS, Black Robin 2 is expected to have much higher performance.

With a more precise process, there will also be significant improvements in power efficiency and heat control.

Chilbong pressed the remote again to display the comparison chart between Black Robin 1 and 2.

[Black Robin 1 vs. 2 Comparison]

-Process: 2.7nm → 2.0nm

-Maximum computational performance (TOPS): 500 → 2,000

-Power consumption: 320W → 180W

-Heat index: 1.0x → 0.6x

“According to simulation results, Black Robin 2 has four times the performance and half the power consumption compared to Black Robin 1. This means learning efficiency is improved by nearly five times, just from simple calculations.”

“At this rate…… We need to put more pressure on Team Leader Park Sang-bae.”

As I said this laughing, Chilbong also smiled.

“Even without you pressuring him, he’s already working day and night on equipment development. The development team and budget are also being heavily invested.”

I glanced at Executive Vice President Oh Seong-hak.

His expression wasn’t good, which meant a lot of money was being spent.

“Ahem…… Well, let’s talk more about equipment development later. So, there are quite a few places wanting to buy our chipsets, right?”

“Yes, especially big tech companies like Tesler, Gogle, and Metacube are making ‘immediate purchase’ inquiries. Additionally, there have been test sample requests from two North American cloud companies, three European OEMs, and three Chinese mobility companies.”

Chilbong moved to the next slide.

The negotiation status with each company was organized item by item.

[External Inquiries and Negotiation Status]

1. Tesler: Bulk supply of Black Robin E + cloud learning slot lease proposal

2. Gogle: Sharing of Black Robin 1 roadmap and request for negotiation on Luna API sandbox access

3. Metacube: Multimodal inference farm consulting + joint chipset benchmark verification proposal

4. 3 EU OEMs: Black Robin E, L mixed package PoC proposal

5. Chinese mobility companies……

As soon as I saw the slide, the corners of my mouth naturally turned up.

The other proposals were attractive enough, but what stood out the most was Tesler’s cloud learning slot lease proposal.

“Looks like Tesler has finally admitted that rule-based models have their limits.”

In response to my words, Chilbong nodded.

“Yes, they’ve decided internally to switch to a cloud-based learning structure. It seems your autonomous driving video was the decisive factor.”

“In that situation, there must be many shortcomings in building their own AI learning platform.”

“Especially securing chipsets would be the biggest issue.”

I nodded at this.

Chilbong added further explanation.

“Currently, most of the GPUs Tesler owns are from Ensidia’s N100 series. Some are being shifted to AMMD, but in terms of quantity, performance, and heat issues, they all fall short of the Black Robin level.”

Worldwide, we are the only ones to achieve 500 TOPS with a single chipset.

Although Ensidia’s N100 also has a maximum performance of around 350 TOPS, since it’s manufactured with a 4nm process, the production quantity is low and the cost is high.

‘It’ll take another two years for the mega-hit NH100 to come out……’

A monster chipset with a performance of 1,200 TOPS.

But by then, we’ll already have the mass production system for Black Robin 2 in place.

So, there’s no need to worry.

“Anyway, for Tesler, building their own cloud is realistically not easy. So, they’re willing to take the risk of outsourcing their internal backend data to us and making this proposal.”

For Merks, it must have been a decision that hurt their pride.

Right now, securing competitiveness by leasing even a part of our cluster would be more urgent than maintaining pride.

“So, what are the lease terms?”

Chilbong moved to the next slide.

The specific terms proposed by Tesler were displayed on the screen.

[Summary of Tesler Cloud Lease Terms]

1. Lease scale: 12% of Luna cluster resources

2. Lease period: Up to 1 year (with early termination clause)

3. Data security: All learning data anonymized and processed only in isolated virtual sessions within the Luna cluster

4. Ownership of results: Inference model results solely owned by Tesler

5. Technical access restrictions: No access rights to the Luna operating stack

6. Option clause: Early withdrawal possible if own cluster is secured within 1 year

I carefully reviewed the materials from start to finish.

It was a typical ‘temporary solution’ proposal, nothing special.

“A one-year contract means they don’t plan to use it long-term from the start.”

“Yes, they intend to withdraw as soon as their own infrastructure is ready.”

“What’s the possibility of them secretly replicating our technology?”

“Luna’s cluster structure and learning protocols are 100% proprietary, making internal access fundamentally impossible. And the leased slots will only operate in a virtualized sandbox environment.”

I tapped my fingers on the table.

“If we can get a look at even some of Tesler’s real-world learning data, it wouldn’t be a loss.”

Unlike us, who have only sold 30,000 electric vehicles,

Tesler has already put over a million electric vehicles on the road.

Considering the scale and diversity of the data collected from them,

actually…… leasing the cluster would be more than possible.

Chilbong smiled and nodded, seeming to agree with my thoughts.

“Real-time access to Tesler’s data would be difficult, but backend feedback data like error patterns during learning, sensor conflicts, and route deviation factors can be collected in log form. Reflecting that would significantly accelerate the commercialization of fully autonomous driving.”

As Chilbong said, the data Tesler accumulates through cluster leasing will ultimately become the fuel to perfect our technology.