Good evening, ladies and gentlemen. Welcome to the investor conference call hosted by ManyCore regarding the company's 2026 interim results.

Before we start the call, we'd like to remind you that this call may include forward-looking statements, which are underlined by a number of risks and uncertainties, and they may not be realized in the future for various reasons. Information about general market conditions comes from a variety of sources outside of ManyCore.

This presentation also contains some unaudited non-IFRS numbers that should be considered in addition to, but not as a substitute for, the company's financials prepared in accordance with IFRS. During today's call, management will use Chinese as the main language, including for management's remarks and the Q&A session.

Third-party interpreters will provide simultaneous interpretation into English. The interpretation is provided solely for the purpose of improving meeting efficiency.

In the event of any discrepancies, the management's original statements in Chinese will prevail. Joining us today are Mr.

Victor Huang, Co-Founder and Chairman of ManyCore; Mr. Shen Bei, Chief Financial Officer of ManyCore.

To start, the management will share recent strategic business financial updates of the company. Following that, we'll move on to the Q&A session.

I will now turn the call over to Mr. Huang.

Hello, investors and analysts. Good evening.

I am Huang Chao Victor. Thank you for joining ManyCore Technology's first earnings call since our IPO.

Three years ago, ManyCore began its transformation into a spatial intelligence company, extending the application layer of... This strategic pivot also marked the beginning of our exciting journey toward AGI.

We believe AGI is the inevitable trend, but it's more than just language models. While recent breakthroughs in LLMs are truly inspiring, LLMs alone are not enough to achieve AGI.

To truly get there, we must equip AI with an understanding and awareness of the physical world. We aim to fill in the missing puzzle pieces that LLMs leave out...

and spatial dynamic capabilities beyond just language, and spatial intelligence is a crucial one. First, spatial intelligence is an intelligent agent's ability to perceive, remember, understand, imagine, and take action within the 3D world.

It's functional intelligence shared by all animals and is far older than language. If we stop and think about what our brains process beyond language, we realize that spatial capabilities are...

First, spatial memory or reconstruction. After you walk through a room, your brain doesn't just store a pile of photos.

It creates a structured memory where you can recall and mentally navigate at will. You can close your eyes, point to the restroom in your house, and walk to the sofa without tripping over the coffee table.

Second, spatial imagination or SpatialGen that we launched, the model we launched. If you see a corner, a cropped photo, or a single piece of furniture, your brain instantly fills the rest of the room.

This is a generative spatial prior, the ability to construct a complete, self-consistent world from incomplete observations. Third, spatial understanding, or the Spatial LLM product we open-sourced last year.

You don't just see point clouds. You understand the sofa and a coffee table belong together, and the door can be walked through.

Objects, relations, functions, and affordance, this is semantic layer for the space. Fourth...

When we see something, when we are about to do something, we can bypass language. It can already guide our behaviors.

What is our core strategic track or segment? When we talk about AGI, the ultimate vision is having robots work on our behalf in the physical world.

We can't reach that end game with language models alone. AI must understand the 3D world.

Over the past year, you've probably heard buzzwords like spatial intelligence, world models, and physical AI. These all address the gaps left by LLMs on the path to AGI, aiming to give AI a deeper cognition and understanding of the physical world.

Regardless of the technical path, I believe the core task is simulating the physical world, which we break it down into three categories. Type 1: simulating visual realism through rendering to give humans a genuine sense of the space.

This is our bread and butter. When we founded the company, our focus was on leveraging GPU...

worlds as possible. Back then, we relied heavily on optical simulations like ray tracing and later introduced smaller AI models to improve realism and speed.

Over the last 2 to 3 years, things have completely changed. Everything from image to video generation models are, in our view, just a new piece of rendering engines.

Whether it's home design, architecture, e-commerce, or film, the goal is always to make things look more real. We are not abandoning this core area.

We continue to invest heavily here, and we're going to roll out a similar AI rendering product. Also, the world model, which is a popular product in the market, is something that we're focusing on.

Type 2: simulating how the human brain thinks about space, which we started 2 to 3 years ago. In the past, for the past 30 to 40 years in computer graphics, the spatial understanding was largely manual.

In our early days, we were building a lot of tools. We hired thousand, tens of thousands gig workers just to manually label floor plans across the...

Today, this is almost entirely automated by AI. AI will definitely replicate human memory, reconstruction, and spatial understanding.

This is where we have been deploying resources heavily. Type 3: simulating the laws of physics.

This is where simulators come in. We co-open-sourced the Spire simulator with partnerships like NVIDIA and Adobe, also collaborated with Cosmos.

As the barriers to software development continue to drop, we anticipate more... is to remain hardware and simulator agnostic.

We will generate the data needed by the simulator, which is part of our strategy. Today, we just launched our brand-new 3D generative model, Lux3D.

Users can simply upload an image or type a few prompt to generate a multi-format 3D assets. These assets highly accurately reproduce both the visual features and the PBR material...

of the real-world objects. This is a crucial module in our generative model lineup.

Additionally, we recently started internal testing of our world model product. We are taking an explicit 3D approach.

Building on static 3D scene generation, we're targeting core features like dynamic interaction, spatial consistency, and editability. This will power practical applications across AI video, gaming, embodied AI training, and digital twins.

Number three: what about the synergy between our core and new business? Within our spatial intelligence track, our legacy platforms like Kujiale and Coohom fundamentally serve as tools that allow humans to express their mental designs of space.

This has generated massive amounts of data over the past decade, but it's still far from enough. The future of spatial intelligence is no longer about humans manually operating these tools.

Instead, our goal is for AI systems to possess the same spatial reasoning capabilities as a human operating a computer. This is where we've been working hard toward.

Future spatial data will come from mainly two sources: first is human creativity expressed through tools like Kujiale and Coohom; the second is direct data capture from the physical world. Both will coexist for a long time.

We need this data to train our foundational models. Moving forward, other software and robots will require spatial intelligence to navigate and understand our world.

This spatial intelligence could be delivered as an API, embedded system, or local application. Anything is possible.

Therefore... be complementary and mutually reinforcing.

For instance, SpatialVerse, our synthetic data platform for model and robot training that we launched a few years ago, is essentially our proprietary model augmented with human-in-the-loop services. We are continuously optimizing our models to reduce the need for manual intervention and increase automation, with the goal of achieving full automation.

Number four: why we choose this track over others? In the vast blue ocean of AI and the...

a company could take, but for ManyCore, our choice is absolute. We are fully committed to spatial intelligence and going all-in on developing our own proprietary world models.

First, the ceiling is incredibly high. On the road to AGI, the era of physical AI is just the beginning.

It will profoundly reshape how we live and produce goods in the real world. The potential here is massive.

We aim to drive industry transformation while securing long-term, strong returns. Second, we're building a business with an economic moat.

Conventional business model will ultimately converge into underlying foundation models, and competition at the model layer is brutal. Only the...

and high gross margins. The second tier competes purely on price, and the rest won't survive.

Therefore, rapidly pushing the upper limits of our model capabilities at controllable cost is our absolute strategic priority. Third, it aligns with our DNA and heritage.

Since Day 1, our business model has been built around monetizing GPU computing power. In the long run, the combination of algorithms, model, and computing power is guaranteed to be a high-margin, high-retention business, which suits us perfectly.

These three elements create a compounding flywheel. Early investments snowball over time, creating an insurmountable lead.

The true value of this business is that it accelerates over time. Early investments compound.

Number five: commercialization strategy. Thanks to our technical foundation and productization capabilities, we already have a solid commercial foothold.

Our proven revenue streams include: number one, rendering services. We charge on per-image, per-video, or real-time rendering basis.

This is a highly mature cash cow business for us. Type 2: spatial data collection and generation.

As a spin-off of our large models, this business has already established a solid operational baseline. Type 3: large model services.

We monetize through monthly, annual subscriptions as well as API access. I think the market opportunity...

In the short term, we aim to empower roles like 3D designers. This goes beyond our traditional stronghold in interior spatial design, extends to film, television, gaming, AR/VR, surveying, mapping, mechanical...

market, tens of millions of professionals across these sectors. In the long run, we want our foundational models and tech stack to serve as the spatial intelligence module embedded with brains of all computer systems and robotics.

This market is tremendously big. In closing, I want to emphasize that our top priority during this transitional phase is to build a team and infrastructure capable of pushing our large models to SOTA performance.

We have the... the upper limits of our model capabilities, even up to SOTA level.

Spatial intelligence still is in its infancy. Only those who can achieve SOTA capabilities will earn a seat at the table in the future.

For us right now, the most critical goal is staying at the table until AGI becomes reality. We firmly believe in the first principle of entrepreneurship: create genuine value first.

At this stage, our relentless focus is on raising the ceiling of our model capabilities. Knowing that the long-term commercial upside is immense, we remain committed to long-term vision, an open ecosystem, working alongside industry partners to bring spatial intelligence into the real world.

Thank you all. I will now hand it over to Shen Bei.

Thank you, Victor. As Victor outlined, ManyCore is firmly committed to spatial intelligence track, with a focus on advancing, iterating our model capabilities.

I'll talk about two topics: first, I'll review our financial and operational performance; second, I'll share our strategic and business outlook moving forward. For the first half, total revenue reached...

products are now widely adopted through thousands of industries, from spatial design, 3D content creation, intelligent agent training, cultural tourism. Our revenue from our new AI applications products came in approximately 31 million, up 117% year-over-year.

Our gross margin stood at 84%, an expansion of roughly 1 percentage points year-over-year. Driven by the top-line growth, gross margin expansion, enhanced cost, operational efficiency, our adjusted net profit for the first half reached...

5.2 million, matching the full-year level, representing a year-over-year of 200% growth. We have successfully built a full-stack tech architecture spanning computing power, models, and applications.

Last year, we launched the industry's first largest spatial intelligence model dedicated to 3D scene recognition and generation. It features two capabilities: structured understanding and visual generation.

In May, this model officially completed the generative AI service filing. Tracking our model's token consumption, ending July, for that month, our average daily token consumption was around 2.4 billion, and we will continue to track the token consumption by this model.

Moving on to applications, our Aholo spatial intelligence platform integrates the core 3D capabilities of our models, data, tools. It opens up features like spatial reconstruction, understanding, generation, and editing to thousands of industries.

We are also using AI to power our existing products. We built a new cross-industry product and applications powered by the spatial intelligence models.

We also launched our AI intelligent design platform, launched on Kujiale, allows users to generate customized spatial plans in just a few minutes. This helps sales acquire customers more easily.

In the first half, platform generated over 20 million RMB, and also the MAU in June grew by 50% compared with December last year. In May, our LuxReal AI video generation was launched.

Powered by 3D generative AI technologies, LuxReal rapidly generates... high spatiotemporal consistency, allowing for extreme, controllable editing throughout the entire workflow.

It's tailor-made for professional creation of AI shorts. The monthly-over-month increase in registered users increased by over 240%, and we will actively expand in overseas countries.

Our SpatialVerse service generates high-fidelity, physically accurate, synthetic virtual datasets that mimic the real-world physical properties and spatial relationships. These diverse datasets allow developers to train AGIC models in virtual environments, enhancing the cognitive abilities of intelligent robots and AR/VR systems.

In the first half, SpatialVerse secured approximately 6.8 million RMB in total order value. We have been building our own dedicated GPU clusters since 2012.

Years of sustained investment in GPU infrastructure, coupled with experience in hardware/software co-optimization, enable us to deliver highly efficient services competitively. Our first-half CapEx was 27.6 million, up 120% year-over-year.

Looking ahead, ManyCore's mission is to make every space a computable world. Physical AI will be the next global wave in artificial intelligence, and spatial intelligence will serve as the bridge connecting the digital and physical worlds.

On this journey, we'll continue to accelerate our model iterations and capability upgrades. We're dedicated to leveraging our 3D digital assets and model capabilities to evaluate how intelligent systems perceive, understand, imagine, and interact with the 3D physical world.

Specifically, our strategic and business priorities for the second half include: first, continuously enriching our digital assets using... more integrated hardware and software applications for high-definition 3D reconstruction.

This will expand our ultra-high-fidelity, multi-scenario 3D digital assets in both 3DGS and PBR formats. These form assets will underpin our model development and capability enhancements.

Second, accelerating the iteration of our world models, maximizing our technical edge in spatial understanding and generation. We're transitioning from static spatial generation to interactive spatial generation, rapidly approaching the L4 era of...

such as LuxReal and our AI intelligent design platform, powered by our proprietary spatial intelligence models. For our global users, we continue to serve as the picks and shovels provider for the spatial intelligence, aim to propel the real-world application of this stack across spatial design, media, entertainment, agent training, cultural tourism, industrial digital twins, and gaming.

Thank you all for listening. We'll now open the floor for Q&A.

Hi, management. Thank you, Victor and Mr.

Shen, for your prepared remarks. I got two questions.

You talked about the Lux3D video generation model, the LuxReal, and also the Aholo. In terms of application, how different are they?

For the 31 million revenue contribution from AI-related products, does it include AI Design Factory from Kujiale or SpatialVerse? Is it included in the 31 million revenue?

Hi, Selina. Now let me take your first question.

Aholo as our spatial intelligence platform... 3D is one of the large language models under this umbrella.

The Lux3D is under the umbrella of Aholo. Aholo is a brand, our spatial intelligence brand, and there are several products underneath that, that includes video generation, understanding model, and a few other modules.

Going forward, we will launch a variety of spatial intelligence models. They're all going to be placed under the brand of Aholo.

You talked about a token consumption. Is that consumed by Lux3D?

Correct. The token consumption covers the total consumption of various models deployed across different platforms, and Lux3D is one of the models that contributes to the token consumption.

Revenue from AI-related products, the AI design platform contributed 11 million. I think it was up from 6.8 million last year.

Like I said, it's intelligent design revenue is within the 31 million revenue from AI-related products.

Thank you for your answer.

Good evening, management. Can you hear me?

Hi, Victor. Hi, Bei.

Thank you for the opportunity. I got two questions.

One is on the roadmap or the evolution of the spatial intelligence sector. You talked about that we have achieved L1 or L2 of a spatial intelligence with spatial reconstruction.

My question is, in terms of the implementation and also customer's willingness to pay, moving toward L4 of spatial intelligence, where are the bottlenecks, in terms of the tech path? Also, how do we monetize, or when could we see monetization for the L4 of spatial intelligence, and do you expect to have a timeline that it will contribute to total revenue?

Also, you talked about some of the progress transitioning towards large language models, and we have already shown there have been some initial results. Looking ahead, what about our business model?

Right now, we are providing data for robot companies to train with, so that's our business model. Are there any opportunities where we can power these customers with our model capabilities to empower them to have a better inference, so that will ultimately translates into a more encouraging top-line revenue growth?

Let me address this question by saying that in reconstruction and generation, spatial reconstruction and generation, over the past a few months, we have a couple of millions of revenue, signed into the contract. That these capabilities can be implemented, and if you follow us on WeChat public account, there are some demos, and customers are willing to adopt these solutions, particularly in the US market, these products are quickly adopted.

Spatially built upon our large language models coupled with human intervention, it involves some inference plus human intervention. We're seeing significant growth, particularly in the US and European markets.

For us, we position this as a inference business. The Lux3D we just launched today, as well as many other categorized under the inference business, they are not labor-intensive business or like a specific project.

That's that. When are we able to implement that, I think it take time, and it's going to go stage by stage.

In short term, it's going to empower designers across industries for the initial stage, and moving forward, we hope these could be integrated into robots and equipment. By that time, the volume would be significantly larger.

If we are able to equip all of these capabilities into these robots or equipment, the market is going to be tremendous. For SpatialVerse, we mainly in the start, we sell synthetic data for revenue, and going forward, we're going to leverage this to train our models.

As our models iterate, we expect [unclear] or model-as-a-service. This has been repeatedly proven by many AI model companies.

Therefore, in SpaceVerse and also embodied AI sector, I'm sure we are able to deliver a model that lead the market and quickly produce reoccurring revenue very soon.

Thank you. That's very clear.

I do look forward to the more updates from the SpatialVerse or related products, and also new model launches.

Thank you, Cathy.

Thank you, Mr. Huang, and thank you, Mr.

Shen. Around spatial intelligence model, compute, and application, we have been making significant progress, as evidenced in the launch of Lux3D and etc.

I want to know, as we iterate our business, what about our investments in R&D and compute? What scale of investments are we looking at, and what about the timeline and or a planning for that?

On compute, since we started in 2012 after I created the company leaving NVIDIA, we were building in-house software and hardware cluster. Our CapEx in the first half grew by 120% year-over-year to 27.6 million RMB.

It's pretty hard to acquire GPUs or chips on the market, but we'll continue increase compute investment to improve our compute efficiency. With regard to R&D, we will hire top-tier, best-of-the-notch talent, and our total engineering headcount won't grow significantly.

What we are looking right now is the top-tier engineer. Simply put, we're going to be prudent with the expansion of our engineering headcount.

With our compensation packages, we maintain an open attitude. Any add from Mr.

Shen?

We will continue investment for growth. That is a strategic direction for our company.

Compared with the large language model companies, our models do not require significant compute resources, even the size of our parameters. Our in-house compute resources is able to accommodate our training need.

Of course, in the second half, we will acquire more compute resources to support our training and inference workloads.

Good evening, management. I am very pleased to have this opportunity.

Two questions: number one, do we have any metric to disclose on ARR, for example LuxReal? Can management share some color on that?

If we look at our LuxReal and Aholo, they have been making very good progress. What about our customer base?

Can you share some of the profiles of these customers, and also what's the percentage split for the customer How many of them are from overseas countries?

We currently do not disclose the ARR. We are still in [unclear] stage.

Our traditional or legacy software and our newly rollout software are now working in hand or embedded together. I think it would be confusing.

We actually launched a AI-native products which leveraged our model capabilities, and its ARR is growing very quick. We were not trying to exaggerate the ARR growth.

I think our daily revenue reached 10 to 20 thousand every day. I think by the end of the year, we can give more update on the MRR of our AI-native products.

Looking at the growth pace month-over-month, we are seeing [unclear]. As the user profile, LuxReal is for AI short designers and creators, and also 3D designers.

This is a growing sector that's quickly growing. Literally, a few people could form a studio and produce very high-quality AI short videos.

For Aholo, it was just launched. The target customers or consumers are designers who design games or for some exhibition-related designs.

We're seeing wide range of use cases with Aholo. For the reconstruction, it has been used in cultural tourism, mechanical engineering, and also construction protection, building protection.

A wider range of use cases. We recently signed contracts with factories, enabling their [unclear].

For us, from 3D generation, perception, here and there, we believe this is a generalized capability, and we're actively exploring the use cases of these capabilities. These products are relatively new, and there's significant possibilities.

I think in a few months' time, with further progress, I can also update you with use cases where we can monetize these products.

Thank you, management, for your answer. Also, wish you all the best.