Goldman Sachs Tech Tour 2026: Bullish on China AI Inference Demand Surge and Accelerating Commercialization of Quantum Computing and Robotaxis

Overview

Based on Goldman Sachs' China Private Tech Tour 2026 held across multiple locations in April, this report features in-depth interviews with founders and executives covering AI foundation models, software, hardware, quantum computing, satellite communications, brain-computer interfaces, and robotaxis. The core conclusion is that China's technology industry is undergoing a comprehensive upgrade from underlying infrastructure to upper-layer applications. Specifically, the implementation of Generative AI (GenAI) has significantly boosted demand for AI servers, optical modules, storage, and specialized chips. Meanwhile, frontier technologies such as quantum computing and robotaxis are crossing the 'Valley of Death' and entering a phase of accelerated commercialization. Goldman Sachs maintains a positive outlook on China's technology supply chain and related innovative enterprises.

Core views

Regarding AI infrastructure, the report notes significant structural differences between Chinese market demand and overseas markets. As domestic foundation models largely rely on open-source ecosystems, training compute demand remains relatively moderate, while inference demand has surged due to an explosion in application-layer adoption. This trend directly drives Superpod deployments, scaling from 64 GPUs to 256 or even 1,024 GPUs, thereby fueling strong demand for 800G and above high-speed optical modules. Although upstream laser components previously experienced supply tightness, the supply chain has gradually optimized through long-term agreements and technological iteration. Notably, in optical chip materials, Thin Film Lithium Niobate (TFLN) is gradually replacing Silicon Photonics (SiPh) as the preferred solution for high-end interconnects due to its performance advantages in 3.2T+ transmission rates and lower manufacturing costs. At the chip and computing architecture level, domestic suppliers are breaking monopolies through differentiated competition. Beyond traditional GPGPUs, new AI chips featuring Reconfigurable Parallel Processing (RPP) architectures are seen as superior solutions for inference scenarios due to higher energy efficiency, flexibility, and greater memory bandwidth. These chips are not only used in cloud servers but are also expanding into AI PCs and autonomous driving. Concurrently, the importance of storage has become prominent; AI SSDs and memory controller ICs designed specifically for AI inference have become critical breakthrough points for improving overall system efficiency. In frontier technology fields, quantum computing demonstrates diverse technological pathways (superconducting, NMR, cold atom, photonic). China exhibits high self-sufficiency in core components and software, moving from the R&D phase toward early commercialization for solving complex computational tasks (e.g., portfolio management, power dispatch). For Robotaxis, driven by mature map-less NOA algorithms and declining hardware costs, China's Robotaxi fleet size is projected to surge from 5,000 units in 2025 to 705,000 units by 2030, with leading companies actively expanding into overseas markets such as the Middle East and Europe. Additionally, non-invasive Brain-Computer Interfaces (BCI) have achieved substantive progress in rehabilitation and cognitive behavioral therapy applications.

Analysis framework

Goldman Sachs employed a typical 'Supply Chain Field Research + Supply-Demand Framework Analysis' methodology. First, through frequent face-to-face founder visits, they obtained first-hand information on technology roadmaps and commercial implementation progress to validate macro judgments. Second, they applied a supply-demand framework to dissect specific sub-sectors: in optical modules and servers, the focus was on capacity bottlenecks (e.g., laser shortages) and technology iteration cycles (e.g., TFLN replacing SiPh); in chips, they compared the cost-performance ratios of different architectures (GPGPU vs. RPP) in specific scenarios (inference vs. training). Finally, combining policy guidance (such as GenAI penetration targets in China's 15th Five-Year Plan) with market penetration S-curves, they predicted scaling timelines for emerging industries like Quantum and Robotaxi. This logic of 'micro-validation + macro-mapping' ensures robust conclusions.

Methodology notes

• Industry/Sector Analysis FrameworkSupply-demand framework

  Assessing industry prosperity by analyzing supply-side capacity and technical barriers alongside demand-side growth rates and structural changes

  The report frequently mentions 'supply tightness' and 'strong demand' for optical modules, as well as the structural shift in AI server demand from 'training' to 'inference.' This is a classic supply-demand analysis method used to judge price elasticity and investment priority.

• Industry/Sector Analysis FrameworkPenetration S-curve

  Using an S-shaped curve to describe market acceptance of new technologies from introduction and growth to maturity

  The report's forecast of Robotaxi fleet size growing from 5,000 to 705,000 units, and its assessment of quantum computing moving from R&D to commercialization, implicitly identify these technologies as being in the rapid ascent phase of the S-curve, used to predict the timing of market scale explosions.

• Valuation Method

  Total Addressable Market (TAM) Update and Unit Economics (UE) Analysis

  The report demonstrates commercial viability by updating the TAM for Robotaxis and emphasizing improvements in Unit Economics (UE). This is a qualitative valuation method based on market sizing and return-on-investment logic.

Asset mapping & comparison

Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).

• xFusion

  Beneficiary Logic: As a leading AI server provider, directly benefits from the explosion in China's inference demand and Superpod deployments

  Strengths

  Strong supply chain integration capabilities, better positioned to handle component shortages

  Weaknesses

  Recent memory cost increases may suppress short-term margins

  Comparison

  More focused on China's rapidly growing domestic inference market compared to international giants

  Risks

  Supply chain disruption risks due to geopolitical tensions

• Liobate

  Beneficiary Logic: Leader in Thin Film Lithium Niobate (TFLN) technology, benefiting from specification upgrades in high-speed optical modules

  Strengths

  TFLN offers superior transmission performance over Silicon Photonics at 3.2T+ speeds with lower manufacturing costs

  Weaknesses

  Market acceptance of the new technology route still requires time to validate

  Comparison

  Possesses a generational technological advantage in the high-performance optical chip sector

  Risks

  Risk of significant cost reductions in Silicon Photonics due to continuous technological advancements

• Deeproute.AI

  Beneficiary Logic: Leader in map-less NOA technology with accelerating Robotaxi fleet expansion

  Strengths

  VLA model supports rapid entry into new cities without high-definition maps

  Weaknesses

  High R&D investment and regulatory uncertainty

  Comparison

  Faster expansion speed compared to competitors relying on HD maps

  Risks

  Autonomous driving safety incidents and tightening regulatory policies

• SpinQ / CAS Cold Atom

  Beneficiary Logic: Pioneers in quantum computing commercialization with autonomous and controllable technology

  Strengths

  Own core IP in superconducting and cold atom technology routes with high self-sufficiency

  Weaknesses

  Quantum computing remains in early stages; large-scale general-purpose computing is still some way off

  Comparison

  QPUs offer irreplaceability in specific complex computing tasks compared to GPUs

  Risks

  Technology iteration failure or being overtaken by other routes (e.g., photonic quantum)

Key data

• Expected Growth Rate of 800G Optical Modules in China MarketTriple-digit growthStarting from a low base, primarily driven by domestic cloud service providers
• China GenAI Penetration Target90%Government '15th Five-Year Plan' target, by 2030
• DePIN Computing Cost Reduction MagnitudeReduced to 10% of traditional public cloud costsAchieved through Decentralized Physical Infrastructure Networks
• China Robotaxi Fleet Size Forecast (2025E)5,000 unitsCurrent base
• China Robotaxi Fleet Size Forecast (2030E)705,000 unitsHigh-growth expectation
• TFLN Chip Process Node180nmMore mature than SiPh with lower manufacturing costs

Impact & implications

For investors, this implies that the investment thesis for Chinese tech stocks is shifting from pure 'thematic speculation' to 'earnings realization.' The AI infrastructure chain (especially optical modules, AI servers, and domestic inference chips) will benefit from confirmed increases in capital expenditure, while the quantum computing and Robotaxi sectors possess significant long-term growth option value. Furthermore, with the development of the HarmonyOS ecosystem and DePIN, monetization capabilities in the software and services layers are strengthening, helping to alleviate margin pressure caused by intense hardware competition.

Risks

• Geopolitical tensions leading to restrictions on semiconductor and equipment imports
• AI application implementation falling short of expectations, resulting in low Return on Investment (ROI) for computing power
• Severe shortages in the supply chain of key technologies (e.g., laser components, high-end GPUs)
• New technologies like Robotaxis facing strict regulatory approvals and safety incident risks
• Macroeconomic volatility affecting corporate IT spending and consumer willingness to pay

What to watch

• Actual capital expenditure (Capex) execution by Chinese Cloud Service Providers (CSPs) on AI servers and optical modules
• Rate of market share gain for domestic AI chips (especially RPP architecture) in the inference market
• Progress in mass production of TFLN optical modules for 3.2T and above speed products
• Scale of commercial operations and improvement in Unit Economics (UE) for Robotaxis in major Chinese cities
• Concrete commercial order wins for quantum computing in vertical sectors such as finance and energy