HSAI (Trans): SGI guide raised to 300 mn; breakeven in 2027

Compiled by Dolphin Research: $Hesai(HSAI.US) FY26Q2 earnings call transcript.

I. Key takeaways

1. Q3 guidance

a. Total revenue of RMB 1.1bn–1.15bn, with lidar shipments of approx. 0.8mn–0.85mn units.b. SGI revenue in the high eight-digit RMB range (near RMB 100mn); Kosmo contributes from Q3 and is booked under SGI.c. Mix shift: Q3 revenue outside ADAS lidar is expected to approach or exceed half of total, marking a significant structural change.

2. Full-year guidance update

a. SGI revenue guidance raised materially to RMB 200mn–300mn from RMB 100mn, driven by faster-than-modeled commercialization and validation of actuator modules.b. SGI revenue target for 2027 is approx. $100mn (circa RMB 700mn), with breakeven expected in the same year.c. Lidar shipments unchanged at 3.0mn–3.5mn units; robot lidar to exceed 0.5mn units in 2026 (vs. ~0.24mn in 2025).d. H1 shipments were approx. 1.1mn units, nearly doubling YoY; historically about two-thirds of annual volume fell in H2 last year.

3. This quarter’s key financials

a. Total: revenue of RMB 861mn ($127mn), +22% YoY, the ninth straight YoY increase; GPM at 40%.b. Segment: lidar revenue RMB 816mn ($120mn), OP of RMB 66mn ($10mn); SGI first revenue contribution of RMB 45mn ($7mn), OP loss of RMB 64mn.c. Shipments: total lidar shipments >0.628mn units, nearly +80% YoY; ADAS >0.485mn units, approx. +60% YoY; robot lidar >0.142mn units, approx. +193% YoY.d. Opex: S&M RMB 15mn, G&A RMB 67mn, R&D RMB 231mn, with R&D growth mainly from focused SGI investment.e. Profit: GAAP net profit RMB 71mn ($10mn), +60% YoY, fifth consecutive GAAP-profitable quarter; Non-GAAP net profit RMB 101mn ($15mn).

4. GPM and per-vehicle lidar value

a. 2026 full-year GPM outlook remains near 40%; management expects a healthy margin structure to persist.b. Per-vehicle lidar value is moving up: L2 shifting from 1 long-range unit to '1 long-range + 1 rear blind-spot'; entry L3 may be 1 ATX + 2 FTX blind-spot, and higher configs (ETX + multiple FTX) can reach ~$500–$1,000 per vehicle.c. SGI long-term GPM target is ~40%; Kosmo carries structurally higher margins than pure hardware due to cloud services, gradually boosting group margins.

5. Related-party transaction cap

Plan to raise the annual cap under the product supply framework agreement with Sharpa from RMB 100mn to RMB 300mn, subject to approval at an EGM; this year’s transactions under the framework are expected to be mainly actuator modules.RMB 300mn is the cap under the agreement, not the expected actual amount.

II. Call details

2.1 Management highlights

1. Platform strategy: See–Understand–Act

a. The company is evolving from a lidar vendor into a full-stack infrastructure platform for robotics and physical AI, with three layers: See (lidar to perceive the physical world), Understand (Kosmo turns real environments into reusable AI-ready 3D spatial assets), and Act (actuator modules convert intelligence into precise physical actions).b. All three layers have commercial traction: lidar scales profitably and generates cash; Kosmo prototypes shipped in Jul 2026 with initial orders, contributing revenue from Q3; actuator modules began revenue in Q2 and are ramping quickly.

2. ADAS lidar: awards and multi-lidar trend

a. Won a mass-production award from Great Wall Motor using high-end ultra-long-range ETX, with SOP expected by end-2026; ETX offers more than double ATX’s detection range, while ATX has become the cost-performance benchmark.b. Blind-spot product FTX is gaining momentum; Li Auto L8 and L9, each with four Hesai lidars, are in mass production, and the newly launched Li Auto L6 (~RMB 250k MSRP) offers a four-lidar option, bringing high-end configs into mainstream price bands.c. Intl and JV: after Mercedes-Benz, secured a key award from Volkswagen, covering multiple models of its China JVs, and expanded collaboration with GAC Toyota; with Chinese OEMs going abroad, overseas volumes are expected to be meaningful by 2027.

d. Share: per Gasgoo, Hesai held 44% share of China’s long-range ADAS lidar market in Jun 2026, ranking No.1 domestically for 17 consecutive months.e. Regulation: China’s first mandatory national safety standard for L3/L4 will take effect in 2027; in the U.S., New Jersey proposed requiring commercial AVs to equip two independent sensing technologies beyond cameras, typically lidar and radar.

3. Robot lidar

a. Working with 50+ embodied intelligence companies globally; recent orders from Unitree, Robbyant, Galbot, Galaxea, Dexmal, among others.b. JT128 has become the preferred lidar for humanoid and quadruped robots; in some high-performance warehouse logistics setups, a single unit can mount up to 15 JT128s, and JT128 ASP is higher than JT16 used in lawn mowers.c. Ramp: Q2 robot lidar shipments were nearly 3x YoY; annual growth was +47% in 2023, +67% in 2024, +426% in 2025, with 2026 full-year expected to grow another 2–3x.d. Morgan Stanley estimates robot lidar demand by 2050 could be ~6x that of autos.

4. Picasso chip

a. Picasso is a full-color ultra-sensitive SPAD SoC that fuses depth and color on a single chip; time-aligned multimodal output at the same timestamp is an ideal input for world models.b. Picasso SoC is SOP-ready; since Apr 2026 launch, Full Color ETX with Picasso has secured initial awards including Cargo Bot and is under customer validation, with RFI/RFQ ongoing with leading robotaxi operators and global OEMs.

5. Kosmo (spatial intelligence platform)

a. An integrated system combining AI spatial camera, AI algorithms, 3D spatial assets, and cloud services to capture, reconstruct, and understand the physical world; it targets the sim-to-real data bottleneck where robots need geometrically precise, physically reliable digital environments.b. Performance: reconstructing a 200 m² restaurant is ~5x more efficient than leading alternatives, with fidelity high enough to keep 4 mm text on menus legible.c. Commercial progress: prototypes shipped in Jul, with initial orders in seven days from top humanoid robot companies including Galbot; since the Apr preview, 200+ potential partners have engaged, spanning culture/tourism, film/TV, gaming, and advertising.

d. Biz model: the AI spatial camera is the entry point, with algorithms, cloud, and an accumulating 3D asset library at the back end; each deployment adds assets, a thicker library unlocks more applications, more applications attract more users and generate more assets, creating a flywheel with recurring revenue, network effects, and operating leverage.

6. Actuator modules (Act layer)

a. Performance: torque and power density are ~3x current leading products, volume is 37% smaller, transmission efficiency exceeds 95%, and validated over 2 million duty cycles.b. Entry path: start with dexterous hands, the hardest part of humanoids, prove the architecture in the most demanding scenarios, then expand to other body parts; full-body modules are targeting SOP in H2 2026, extending to shoulder and wrist joints.c. Capacity and orders: dedicated lines are fully running; cumulative shipments exceeded 10k modules by end-Q2, ramping near 10k per month soon, with 2027 shipments expected to reach six figures.d. Customers and validation: supplying Sharpa, an AI robotics company; in Aug 2026 at a DQ store in Shanghai, Sharpa humanoids are expected to complete what the company deems the world’s first ‘zero-retrofit’ commercial deployment, autonomously producing Blizzard ice cream for a full shift.e. TAM: actuators can account for over half of a robot’s BOM, with a single humanoid possibly requiring 100+ modules, implying long-term demand potentially approaching trillions of units per internal estimate; JP Morgan views 2026 as a milestone year for humanoids.

2.2 Q&A

Q: After raising SGI guidance, can you detail Sharpa’s collaborations with NVIDIA and Google Gemini Robotics, and other potential partners? What is the long-term revenue potential of actuators, and how does Hesai achieve strategic synergy with Sharpa?

A: Hesai is building a full-stack infrastructure platform for robotics and physical AI, enabling robots to see, understand, and act; this requires years of R&D, engineering, and real-world validation.Our mission since inception has been consistent: empower life with robots. Lidar was the first commercial product to prove we can build a leading technology platform and scale globally, but it is not the endpoint.

The CEO is also a co-founder of Sharpa, a humanoid robotics company that has gained substantive global recognition, recently adopted by Gemini Robotics and NVIDIA; Sharpa will soon begin its first real commercial deployment at a DQ store in Shanghai.It is not a lab demo or a pop-up showcase, but robots autonomously performing tasks in a normal operating store, marking a step from demo to real operations.

For Hesai, Sharpa is a valuable real-world testbed. We already supply lidar to a diverse robot customer base and expect over 0.5mn units shipped this year; through Sharpa we gain first-hand insight into what execution systems must deliver in practice.This creates a feedback loop: we put technology on robots, and robots tell us what the market truly needs—from performance metrics and form factor to reliability, edge cases, duty cycles, and total ownership cost.

Two examples: for humanoids, lidar requires resolution, FOV, compactness, and robustness over simply maximizing range; for actuators, the bar goes beyond paper specs—reliability, consistency, durability, and performance under real loads are critical.Such needs are clear only when you build and operate yourself, which differentiates a simple parts supplier from a supplier with first-hand knowledge of real operations; supplying Sharpa and offering manufacturing services provides a high SNR test and learning platform before broader market rollout.

As disclosed in the continuing connected transaction announcement and circular with Sharpa, we plan to raise the annual cap under the product supply framework agreement from RMB 100mn to RMB 300mn, subject to shareholder approval at an EGM; transactions this year are expected to be mainly actuator modules.Long-term opportunities are much broader than Sharpa itself—our products are designed for third-party customers at scale; if physical AI is entering rapid adoption, our role is to provide underlying infrastructure, akin to selling shovels in a gold rush.

Q: Can you elaborate on the positioning of humanoid actuator modules, their fit within the broader robotics strategy, and the long-term financial model? Have shipments started to customers beyond Sharpa?

A: That is really three questions: do we want to do it, is the market ready, and can we do it well.First, do we want to: yes. Actuators can be over half of a robot’s BOM, and a humanoid may need 100+ modules, implying long-term demand near trillions of units by our internal estimate, with current margin levels around 40%—exactly the market we want.

Second, is the market ready: yes. Humanoids are entering scale-up and real deployments this year; every robotics company is seeking good joints, yet they are hard to find, and many suppliers do not truly understand what leading robotics players need or why those metrics matter.Joints are both critical and hard, requiring power, precision, compactness, and durability.

(The third sub-question on ‘can we do it well’ was interrupted by a technical line issue and was not completed on the call.)

Q: What drove the SGI guidance uplift? How should we view the revenue mix between actuator modules and Kosmo within SGI?

A: 2026 SGI guidance was raised from RMB 100mn to RMB 200mn–300mn; the 2027 target is approx. $100mn (circa RMB 700mn), with SGI expected to reach breakeven in 2027.The uplift is due to commercial demand and technical validation exceeding initial model assumptions, underscoring our evolution from a lidar company to an infrastructure platform for robotics and physical AI.

This year SGI is mainly driven by robot actuators, built on a decade-plus of engineering experience, maturing quickly, and starting revenue in Q2; Sharpa is a major source of actuator demand, likely the first to deploy humanoids into restaurants with a scalable path, with demand exceeding expectations.This is why we propose raising the annual cap under the continuing connected transaction framework with Sharpa from RMB 100mn to RMB 300mn; to clarify, RMB 300mn is the cap and still requires shareholder approval at the upcoming EGM.

Kosmo is the other SGI pillar and is advancing fast: prototypes shipped in Jul, initial orders came within seven days, and Q3 revenue contribution is on track.We are not disclosing mix granularity now; directionally, actuators are expected to be the bulk of 2026 SGI revenue, while Kosmo contributes a smaller eight-digit RMB amount this year, as commercialization begins later and higher-value cloud revenue accumulates with deployments.

By 2027 the mix should look different: as Kosmo extends from AI spatial cameras to cloud processing, subscriptions, and licensing of high-quality 3D spatial assets, its contribution will increase, with a rising share of recurring cloud revenue and structurally higher margins.Thus, this guidance raise mainly reflects faster-than-expected actuator commercialization in 2026–2027, with Kosmo poised to become another key contributor as the platform scales.

Q: What is Kosmo’s business model, who are current and potential customers, and what is the delivery cadence?

A: Since the Apr preview, Kosmo has engaged 200+ potential partners across robotics, film/TV, gaming, culture/tourism, and luxury, and the scope is still expanding; this breadth indicates it is not a niche product serving a single vertical.Think of 2D cameras as the infrastructure for how the world has been recorded and consumed for decades—Kosmo is the same idea, but in 3D and AI-driven; wherever a normal camera can be used today, 3D digitization of the physical world presents a larger opportunity.

We cannot yet provide specific unit pricing or order sizes, as the ultimate market size is uncertain and many customers are in prototyping and pilots; more relevant is who is testing seriously, who has ordered, and how fast they move.Prototypes began shipping in Jul, and within seven days of the first batch we received strong feedback and locked initial orders.

Early demand is concentrated in two areas. First is robotics, the most immediate pull: humanoids are entering early mass production and urgently need high-quality input data to scale.The ceiling is real—robots can be flawless in simulation but fail in factories or homes; training on low-quality spatial data is like reading a blurry textbook hundreds of times and still not learning correctly.

Kosmo converts real environments into high-fidelity, physically reliable 3D assets, providing richer training sets, better generalization, and reducing collapse risk from poor data.Second is media and digital content—gaming, film/TV, luxury, culture/tourism, and AI short-form ads—content has remained 2D for decades and needs a dimensional upgrade.

The common goal is to make the physical world immersive and reusable: a store, set, concert, or historical site becomes a reusable 3D spatial asset rather than a one-off shoot; film production is the clearest example, where Kosmo brings higher quality, consistency, and lower cost than green screens and on-location shoots.We are not yet able to disclose those in deep talks, but they are household names globally.

On biz model, Kosmo is a system of ‘AI spatial camera + algorithms + 3D assets + cloud’, with devices as the entry and compounding value at the back end; more devices mean more input, a thicker asset library unlocks more applications, draws more customers, and generates recurring revenue via cloud usage and asset licensing.Early commercialization pull is real—orders and prototypes are just the first turn of the flywheel, and recent feedback validates both the tech and the business blueprint; we are optimistic about the next batch of orders.

Q: Can you update Q3 and full-year 2026 guidance? Any change to the 3.0mn–3.5mn lidar shipment outlook?

A: Q3 total revenue is expected at RMB 1.1bn–1.15bn, with lidar shipments of approx. 0.8mn–0.85mn units, and SGI revenue in the high eight-digit RMB range; actuators are ramping fast and will soon approach ~10k units per month.Importantly, Q3 is expected to be the first quarter where revenue outside ADAS lidar approaches or exceeds half of total, another clear step toward our role as infrastructure for robotics and physical AI.

We maintain full-year lidar shipments at 3.0mn–3.5mn units; H1 shipments were approx. 1.1mn, nearly doubling YoY.Note ADAS follows auto seasonality, with H2 typically much stronger—about two-thirds of last year’s full-year volume was in H2, so this year’s H2 ramp is a normal cadence.

Three drivers underpin lidar growth. First is penetration: lidar is no longer just about NEV upstarts—traditional OEMs like Geely and Changan are adding more lidar into their EV architectures; penetration was near 20% in 2025 and is expected to reach ~30%–40% this year.Per Gasgoo, we have ranked No.1 in China’s long-range ADAS lidar market for 17 straight months.

Second, do not cap penetration at 100% per vehicle because it is no longer one-per-car; China’s L3/L4 safety standard was released in Aug and takes effect Jul 2027, and sensors must be integrated at the design stage.We are already seeing multi-lidar setups for wider coverage and redundancy, and have secured 3–6 lidar per vehicle awards at Li Auto, Xiaomi, and Changan; multi-lidar is a key trend for 2026, including in high-end L2.

Third is robotics: humanoids, lawn mowers, and robotaxis are all ramping, with embodied intelligence especially noteworthy, and we work with 50+ leading players including Unitree.Morgan Stanley estimates robot lidar TAM could be 6x autos, and we think it may be larger—any robot moving in the real world needs to know where it is, what is around it, and how it is changing; lidar is becoming a core sense for robots.

Q: What are the upside or downside risks to the 2026 robot lidar shipment guide? Beyond lawn mowers, delivery vehicles, and two-wheelers, which downstream segments are accelerating? Does the U.S. FCC’s recent stance on China-made robot equipment impact the robotics business?

A: We still expect full-year robot lidar shipments to exceed 0.5mn units (vs. ~0.24mn in 2025), with deliveries on track and demand broadly distributed rather than concentrated in a single niche.Regarding FCC, we do not see material direct impact at present; most of our robot lidar shipments this year are for lawn mowers, with customers primarily Chinese manufacturers serving overseas markets centered on Europe, while U.S. lawn mower robot penetration remains relatively low.

We will monitor rules closely and stay fully compliant, and we are not changing shipment outlook.More importantly, lawn mowers are just one entry into a much larger opportunity.

The core logic is straightforward: any robot moving in the physical world must do what humans do—see, understand, and act in real time; lidar is becoming robots’ foundational sense, and robots without reliable 3D perception are effectively walking blind into an unfamiliar room.Per GGII, YO, and Frost & Sullivan, we lead across humanoids and quadrupeds, robotaxis, delivery vehicles, and lawn mowers.

Humanoid demand is particularly strong; we collaborate with 50+ embodied intelligence companies globally, and JT128 is rapidly becoming the preferred product for navigation, obstacle avoidance, and safe interaction with people and objects.Warehouse/logistics is also advancing quickly, with high-performance setups mounting up to 15 JT128s per unit for full coverage; JT128 ASP is higher than JT16 used in lawn mowers, with tighter performance requirements and higher lidar value per unit.

Robot lidar is not a subordinate to auto and can be structurally larger long term—more environments, potentially more machines.We often say lidar is the ‘shovel’ in the physical AI gold rush; we do not need to bet which segment scales first, as all—humanoids, warehouse robots, robotaxis, or lawn mowers—require reliable spatial perception, and Hesai is positioned to grow with the ecosystem.

Q: Xiaomi introduced RoboSense in its latest model as a lidar supplier; how do you view Hesai’s share in Xiaomi’s future models? As more suppliers enter, how will competition, pricing pressure, and share evolve?

A: First, Xiaomi’s move toward more lidar per car is positive for the industry—even at L2, they added a rear lidar where none existed, indicating OEMs and consumers recognize lidar’s value; our long-term view remains that every smart car will carry at least one lidar and many will carry several for coverage and redundancy.Competition is not about one-per-car, and rising per-vehicle lidar counts can multiply the opportunity; the pie is bigger than many expected.

On competition, multi-sourcing is standard in autos, especially post-ramp; we do not comment on share allocations for specific future models, which vary with performance, cost, capacity, and platform needs.It is better to let data speak: per Gasgoo, we have ranked No.1 in China’s long-range ADAS lidar market for 17 straight months with ~40%–50% share, achieved while maintaining healthy pricing and margins.

As a leader, the goal is not to take 100% share at any cost; destructive price wars are worst for everyone, squeezing room for tech, quality, and safety.We prefer to lead in technology, deliver more value to customers, and earn reasonable returns commensurate with that value, underpinned by in-house chips, deep systems engineering, large-scale automation, a broad product matrix, automotive-grade reliability, and years of mass-production experience.

These keep us competitive even if we are not the cheapest—customers buy safety and reliability at the moment cameras fail, not a small box.We are also investing in the next cycle: Picasso full-color ultra-sensitive SPAD SoC integrates depth and RGB at the chip level, and to our knowledge we are the only company in China pushing native chip-level full-color lidar toward mass production on a 2H26 timeline.

Longer term, the camera vs. lidar debate will fade, and they will fuse into one integrated system; the only question left is whether families are safe when the car drives itself.Moreover, outsiders fixate on shipment or revenue share and miss profit share—think of Apple’s profit share in smartphones far exceeding shipment share, achieved not by price wars but by technology, product, brand, and differentiated value.

We aim to keep ~40%–50% leading share, but leadership is about taking the economic value, maintaining healthy GPM, and reinvesting profits into innovation.We will keep working toward that.

Q: With industry-wide pricing pressure and competition, how will lidar ASP trend over the next few quarters? What is the margin outlook?

A: We are constructive on margin resilience; 2026 full-year GPM outlook remains near 40%, and we expect margins to stay healthy.Do not focus solely on blended ASP—L3 shifts the lens from ‘lidar price per unit’ to ‘total per-vehicle lidar value’: L2 has moved from typically one long-range unit to adding a rear blind-spot unit, totaling two; entry L3 may be 1 ATX plus 2 FTX, and higher-end configs use ETX plus more FTX, lifting per-vehicle lidar value to ~$500–$1,000.

In other words, the lidar value delivered per car is rising, already reflected in models launched this year.Margins are price minus cost: on price, we lead in share and will not forgo pricing to chase more share—competition is driven by brand, technology, and quality; on cost, in-house chips, system integration, automation, and scale continue to lower costs, targeting stable ADAS margin structure.

Mix helps too: robot lidar and Intl generally carry higher margins than domestic ADAS; Q2 robot lidar grew nearly 3x YoY, and global business will keep expanding, becoming an important support for group margins as it scales.SGI began contributing revenue in Q2 and products are still ramping; while near-term financials are not yet commented on, SGI’s long-term GPM is ~40%, and Kosmo’s cloud services provide structurally higher margins that will increasingly accrete to group GPM.

<End of text>

Risk disclosure and statement:Dolphin Research Disclaimer and General Disclosure