Chapter 247: External Sales
California, Tesler Headquarters, 12th Floor Briefing Room.
Three slides were cycling through on the conference room screen.
[Autonomous Driving Agenda]
- Lease of Luna Cloud Learning Slots
- Data Risk and Internal Countermeasures
- Cost-Benefit Comparison and External Strategy
Arms crossed, Merkus stared at the monitor and spoke quietly.
“It might look like we’re chasing after Dojin. Is their system worth the risk?”
Autonomous driving division head Eliason pointed to the first slide with a laser pointer.
“It’s a major concern internally, but let’s look at the numbers first.”
Eliason flipped the screen.
A coverage achievement forecast appeared for urban, rainy, and nighttime conditions.
“If we pursue improvements with only a rule-based system or develop our own learning system, full coverage in the Bay Area and greater LA region is expected in at least 24 months. Leasing just 12% of the Luna cluster, however, reduces that timeline to 5–6 months.”
“Definitely… a rule-based system alone can’t keep up with a cloud learning-based one. What about the risks?”
In response to Merkus’s question, Eliason flipped to the next slide.
[Data Leakage Risk Assessment]
- Backend Exposure Possibility: Anonymization + Isolated Sessions → Medium Risk
- Operational Stack Penetration Possibility: Sandbox-Based → Low
- Side-Channel Reverse Tracing: Pattern Inference Possible → Buffer Strategy Insertion Needed
“When leasing the system, we can only learn within a sandbox, and the API, kernel, and stack are all physically isolated.”
“So Dojin’s wary of system penetration?”
Eliason nodded.
“If you know the core structure of the autonomous learning platform, it’s not hard to replicate.”
“What’s the likelihood of that?”
For a moment, Merkus’s eyes gleamed with anticipation.
If they could just get their hands on Luna’s structure, they could significantly cut down on their own infrastructure development time.
But Eliason’s answer wasn’t very encouraging.
“Dojin has implemented very thorough countermeasures—virtual session isolation, scheduler randomization, dummy load insertion, stack access restrictions—so reverse-tracing the system is virtually impossible.”
Merkus rubbed his chin, looking disappointed.
“Well… unless they’re idiots, they’d have countermeasures in place.”
“Right.”
“So conversely, what’s the likelihood of our data leaking to them?”
Eliason’s pupils flickered momentarily.
“Frankly… no matter how isolated the sandbox is, using Dojin’s system means there’s always some level of data leakage risk.”
Merkus’s eyebrows twitched slightly.
Noticing his reaction, Eliason quickly added,
“But we’re only talking about fine-grained pattern inference at the log level. The likelihood of core algorithms or structures leaking is very low.”
“Hmm… really?”
Merkus’s tone still sounded uneasy.
He quietly looked at the Dojin logo on the side of the screen.
“Once we decide to lease Dojin’s system, some data leakage is inevitable. So until our own infrastructure is ready, we’ll use Dojin’s system to secure as much coverage as possible.”
Eliason cautiously added to Merkus’s statement,
“So far, we’ve secured about 19 petabytes of data from real-world driving. Of this, 7.4 PB can be converted for initial learning, and using the Luna cluster slots, we can process more than 60% of that with inference-based methods. Within six months.”
Merkus sighed softly.
“I’m not happy about giving up some of our data in exchange for buying time.”
“But if we don’t do this, Dojin could take the lead in autonomous driving.”
The briefing room fell silent for a moment.
After a pause, Eliason continued.
“Once Dojin’s autonomous driving platform becomes the standard, regulators, insurers, and even consumer perceptions will all be based on Dojin.”
“Microsoft Windows was like that. Same with the Lunar smartphone OS.”
Merkus looked at the screen again.
It was a shame to lose the data,
but they had no choice but to use Dojin’s system.
If development was delayed, they could completely lose control of the autonomous driving market.
Dojin Tech Conference Room.
“I’ll report on Dojin Tech’s last quarter results.”
At Oh Seong-hak’s words, all eyes turned to the screen.
“I’ll start with the mobile division.”
“Please keep it as brief as possible.”
The material looked extensive, and there wasn’t enough time to go over everything in detail.
“I’ll focus on the summary.”
I nodded.
Executive Vice President Oh flipped the slide to the summary page.
[Dojin Tech Q1 2019 Results]
- Mobile Division
(1) Revenue: 68.7 trillion won (2) Operating Profit: 11.2 trillion won (3) Details
- Smartphone Sales: 93 million units • Quantum 4-S, A and Quantum 3 Series • Quantum Pad S2, A2
- Camera Modules: 170 million units
- Mobile Lenses: 220 million units
- Actuators: 190 million units
- Battery Division
(1) Revenue: 9.4 trillion won (2) Operating Profit: 1.05 trillion won (3) Details
- Mobile Battery Shipments: 156 million cells
- Automotive Battery Pack Shipments: 63,000 packs
- Major Clients: Quantum Series, EV-Nova, Prime (PRIME), Tesler—supplying to 3 companies
EVP Oh added an explanation.
“Automotive batteries are primarily supplied to the EV-Nova series, and we’re also expanding export contracts with global companies like BYL and Zike. However, due to CAPA issues, we’re taking a conservative approach to additional orders.”
I nodded and instructed him to move to the next slide.
I had already gotten a sense of the earlier content through interim reports.
What I was really curious about was the next item.
- Semiconductor Division
(1) Total Investment: 27.3 trillion won (2) Accumulated Losses: 11.8 trillion won (3) Details
- Production Items: Mobile AP (Black Robin A), AI Computing Processors (Black Robin 1, E, L Series), HBM Memory Line
- Technology Development Status: LUNA-EVX2, Black Robin 2, Neuromorphic Chipset, etc.
EVP Oh’s expression, which had been bright while reporting on the mobile and battery divisions, suddenly darkened when it came to semiconductors.
“Ahem… the semiconductor division is still in the red this quarter. Especially with the expansion of LUNA-EVX facilities and technology development, and shipping the internal Black Robin series at near cost, losses exceed 10 trillion won.”
In other words, profits from smartphones, camera modules, lenses, and mobile batteries were being funneled into semiconductors.
But now, with some external sales starting to happen, there was a glimmer of recovery.
I turned to Lee Hye-ryeong, who was acting as division head in place of the absent Vice President Hong Chung-ki, and asked,
“What’s the current level of external semiconductor sales?”
Hye-ryeong briefly checked the documents in her hand and spoke quietly.
“External sales are still minimal. So far, only about 15,000 Black Robin E chipsets have been shipped, and the rest are for PoC or pilot testing.”
“So the only real sales are for the E series?”
Hye-ryeong nodded.
“As you know, Mr. President, Black Robin 1 is already struggling to meet internal demand. External shipments have been halted. All units are being used internally for Luna cluster upgrades and LAD commercialization.”
I already knew this.
Listening, EVP Oh asked me,
“Mr. President, do you still plan to not sell the Black Robin L externally?”
Black Robin L was a combat-ready embedded chipset, a core component used in strategic weapons like the AGE and SWAN drones.
So I answered firmly,
“There are no plans to sell the L series externally.”
The conference room fell silent for a moment.
EVP Oh was quiet for a moment, then nodded.
“…Understood.”
He looked like he had more to say,
but as someone who knew the strategic value of Black Robin L better than anyone, he didn’t press further.
“So, the only thing the semiconductor division can monetize right now is the automotive ‘Black Robin E’?”
In response to EVP Oh’s question, Hye-ryeong looked up and answered immediately.
“That’s not the case.”
“Then… there are other methods?”
“Yes.”
Hye-ryeong brought up information about the LUNA-EVX expansion on the screen.
She moved the pointer and continued her explanation.
“As of last week, with seven additional LUNA-EVX setups completed, daily wafer processing capacity increased from 12 to 60 wafers.”
“Sixty wafers? So the number of chipsets we can produce daily has gone from 105 to over 500?”
“Actually, we’re looking at up to 700.”
“Seven hundred?”
The number was higher than expected, and my eyes widened.
AI semiconductors have large dies, so on average, only about 35 chipsets can be made per wafer.
Even then, with a yield rate of only 25%, only about nine are usable.
“Has the yield improved?”
At my question, Hye-ryeong smiled and nodded.
“Yes. With Black Robin 1, we used to have frequent pattern damage due to power routing and heat dissipation issues. But by preemptively applying the hotspot simulation engine from EVX2 to the EVX1 setup, we’ve optimized circuit wiring and incorporated a lithography correction algorithm.”
“So what’s the yield now?”
“The effective yield has increased from about 25% to 35%.”
Seven hundred a day means 21,000 a month.
The target of 75,000 units could be achieved in just four months.
In other words, in four months, the Luna platform could be publicly released,
and external sales of the highly profitable Black Robin 1 chipset could begin.