AI Infrastructure Boom: Optical Modules and Compute Hardware Surge; DAMO Academy Unveils GPU Solver for Billion-Variable PDEs

Explosive Growth in the AI Infrastructure Chain: Optical Communications and Compute Hardware Lead the Rally; Software–Hardware Synergy Enters a New Era of “Hundreds of Millions of Variables”

Recent A-share market performance has displayed pronounced structural characteristics: the AI infrastructure supply chain—represented by optical modules, AI servers, and high-speed connectors—has surged explosively. The ChiNext Index rose nearly 2% in a single day, with optical module leaders such as InnoLight and Accelink repeatedly hitting new all-time highs. This is not short-term thematic speculation, but rather an inevitable reflection of the accelerating global AI compute arms race. Crucially, Alibaba’s DAMO Academy recently unveiled “Mindie”, a GPU-accelerated solver—the first of its kind to achieve real-time, high-precision solutions for partial differential equation (PDE) systems involving over 100 million variables. This milestone marks a major breakthrough for China in the foundational toolchain of AI for Science. The dual resonance of hardware advancement and software breakthrough is driving AI compute demand beyond isolated chip-performance competitions toward a systemic, deep-level co-evolution spanning optical interconnects, advanced packaging, High Bandwidth Memory (HBM), and domestic EDA/CAD tools—the full technology stack.

Optical Communications Lead the Way: 800G/1.6T Modules Reach the Threshold of Mass Delivery

Optical modules serve as the “major arteries” for intra- and inter-data-center data transmission in AI infrastructure. As next-generation AI super-servers—such as NVIDIA’s GB200 NVL72—enter accelerated deployment, compute density per rack now exceeds 10 PFLOPS, rendering traditional 100G/200G optical interconnects a critical bottleneck. Currently, 800G modules have progressed from pilot trials at leading cloud providers to large-scale procurement, while 1.6T modules entered sampling and validation in Q2 2024. According to LightCounting’s latest report, global 800G optical module shipments are projected to grow over 300% year-on-year in 2024—and by 2025, they are expected to account for more than 45% of total data center optical module revenue.

Companies like InnoLight and Accelink, leveraging first-mover advantages in silicon photonics integration, high-speed COB packaging, and thermal management, have secured primary orders from global tech giants including Microsoft, Meta, and AWS. Their record-high stock prices reflect a fundamental market revaluation: Chinese optical communication suppliers are transitioning from catch-up players to global technical standard-setters. Notably, this optical communications boom is not limited to component vendors—it also activates upstream segments (e.g., high-speed PCBs such as Shennan Circuits, laser diode chips such as Yuanjie Technology) and downstream enablers (e.g., liquid cooling solutions such as Galan Co.), generating strong vertical integration effects.

Compute Hardware Evolution: From GPU-Centric Stacking to System-Level Optimization—HBM and Advanced Packaging Emerge as New Battlegrounds

Simply scaling up GPU counts no longer suffices to meet large-model training efficiency demands. Today’s compute bottlenecks are shifting from computation units to memory bandwidth and interconnect latency. HBM3 memory—delivering up to 1.2 TB/s bandwidth (2.5× higher than HBM2e)—has become standard on flagship AI chips such as the GB200 and MI300X. Domestic firms JCET and TFME have achieved small-batch production of HBM3 advanced packaging (e.g., TSV through-silicon vias, hybrid bonding), while Rambus (Lanchip) and VeriSilicon (Chipanalog) accelerate development of CXL-based memory pooling protocol chips to dismantle the “memory wall.”

Meanwhile, the chiplet-based heterogeneous integration paradigm is reshaping the hardware ecosystem. AMD’s MI300 series adopts a 3D packaging architecture combining 5nm CPU cores, 6nm GPU cores, and a 4nm I/O die—enabling efficient collaboration across different process nodes. This trend directly fuels surging demand for domestic EDA tools—including BloomTech’s NanoSpice and Empyrean’s ALPS—in high-end applications such as multi-physics simulation and signal integrity analysis. Previously overlooked “design-enablement” capabilities are now rising to strategic importance, driven by skyrocketing system-level complexity.

Mindie Solver: Breaking the 100-Million-Variable Barrier—A Watershed Moment for Deep Software–Hardware Convergence

DAMO Academy’s “Mindie” GPU solver appears, on the surface, to be a mathematical software breakthrough—but it reveals a deeper shift: the value center of AI infrastructure is moving from “computing fast” to “computing accurately, efficiently, and broadly.” Traditional scientific computing solvers (e.g., PETSc, Trilinos) often hit memory capacity and communication overhead limits when solving fluid dynamics or electromagnetic field simulations on billion-cell grids—requiring hours or even days of iteration on national supercomputing clusters. “Mindie” achieves a qualitative leap via three proprietary innovations:

• Dynamic Sparse Tensor Compression Algorithm: Achieves 92% compression of discretized PDE matrices without compromising numerical accuracy;
• Zero-Copy GPU Memory Cross-Core Scheduling Engine: Eliminates CPU–GPU data transfer latency, achieving 98% GPU memory utilization;
• Adaptive Mixed-Precision Computing Framework: Accelerates non-critical regions using FP16, while preserving FP64 precision in sensitive areas.

Benchmark testing shows that “Mindie” completes a 3D turbulent flow simulation involving 120 million variables in just 12 minutes on a single server equipped with eight NVIDIA H100 GPUs—representing a two-order-of-magnitude speedup over conventional approaches. This means tasks previously dependent on national supercomputing centers—such as weather forecasting, molecular dynamics for drug discovery, and electromagnetic compatibility simulation for chip design—can now run routinely on enterprise-grade AI clusters. Its significance extends far beyond raw metrics: it compels hardware vendors to deliver system-level solutions featuring higher bandwidth (HBM3), lower latency (CXL 3.0), and stronger heterogeneous scheduling capabilities (e.g., NVIDIA GPUDirect Storage), thereby transmitting infrastructure upgrade pressure upstream across the entire value chain.

Valuation Re-Rating Logic: From Thematic Investing to Deterministic Growth

Current market enthusiasm for AI infrastructure has moved beyond conceptual hype into the phase of earnings verification. Optical module makers report robust Q1 order books; server vendors (Inspur Information, Sugon) saw Q2 tender volumes rise over 40% quarter-on-quarter; and HBM packaging and test lines are operating at full capacity. More importantly, breakthroughs like “Mindie” are dismantling the outdated valuation paradigm of “hardware-heavy, software-light”: when domestic EDA tools can support 3nm chip design, and when indigenous solvers can replace commercial ANSYS suites, valuations for related enterprises will pivot from P/E ratios to P/S (price-to-sales) or even customer LTV (lifetime value) frameworks.

That said, geopolitical risks remain a source of volatility. Although China’s Ministry of Commerce export controls targeting Japan focus narrowly on strategic resources like rare earths, they objectively accelerate domestic substitution in semiconductor materials, equipment, and EDA tools. Short-term fluctuations do not alter the long-term trajectory: AI compute power—today’s digital economy equivalent of electricity, water, and coal—exhibits strong policy rigidity, sustained capital investment inertia, and pronounced technological generational lock-in. As fiber-optic cables pulse within server rooms, GPUs hum quietly beneath liquid cooling, and hundred-million-variable equations converge in code—a silent yet monumental revolution in computing power has already moved from blueprints to reality.