Moore Threads, China's leading domestic GPU maker, posts H1 2026 revenue of 1.736 billion yuan, up 147% year-over-year, with robust H2 outlook.

# Moore Threads H1 2026: Domestic GPU Commercialization Reaches Critical Inflection

If 2025 was the "infrastructure foundation year" for China's AI compute, then 2026 is the "realization year" for domestic GPUs.

On August 9th, Moore Threads (688795.SH) delivered a mid-year report that caught the market's attention. The numbers speak for themselves: H1 2026 revenue of 17.36 billion yuan, surging 147.42% year-over-year—already exceeding full-year 2025 revenue. Gross profit totaled 9.89 billion yuan, up 103.78% year-over-year. Most notably, parent company net loss collapsed by 95.73%, with a remaining loss of just 11.56 million yuan.

In today's market environment, many tech companies rely on "grand vision" narratives to drive stock prices. Moore Threads, however, is delivering tangible "real cash" and "repeatable, verifiable business loops." Let's unpack the logic beneath this mid-year report—what engine is really driving this locomotive, and how wide is the road ahead?

Many observers seeing 147.42% revenue growth immediately suspect "low prior-year comparables." But when we extend the timeline and pair this with the 95.73% loss collapse, it becomes clear this is not simple base-effect math, but concentrated release of scale economics.

The GPU chip business is fundamentally high fixed cost, low marginal cost. R&D spend and tape-out costs are sunk; once production scales, per-chip costs plummet. Moore Threads' mid-year revenue surge directly pushed gross margins to new highs—the dramatic loss narrowing is proof.

What's more telling: this profit improvement is not "squeezed margins." The report shows H1 R&D investment of 769 million yuan, up 38.16% year-over-year, with cumulative R&D since 2022 approaching 5.9 billion yuan. This reveals something crucial: Moore Threads narrowed losses while *intensifying* R&D—achieving profitability improvement through scale economics (the inherent high fixed cost, low marginal cost structure of GPU chips), not cost-cutting. This profit improvement driven by surging revenue has extraordinary staying power, not temporary contraction. As H2 shipments scale further, unit costs will continue to decline; quarterly breakeven is highly likely before year-end 2026.

Stock price upside often comes from market mispricings of "ceilings." Moore Threads' upside depends on one question: how many S5000 compute clusters can it sell?

The financial report provides a clear signal—the MTT S5000 has achieved volume sales and completed deliveries across Beijing, Wuxi, Hangzhou and other cities. This is not laboratory PPT; it is tangible engineering execution.

The S5000's key specs: 1000 TFLOPS AI compute per card, 80GB memory, 784GB/s inter-card bandwidth, full precision support from FP4 to FP64. Critically, it achieves 95% linear scaling efficiency in clusters, with training precision on par with mainstream international products. For large model customers, raw card performance matters less than cluster stability.

The "train domestic models on domestic chips" milestone further shatters market bias. On the Kuayao cluster, Moore Threads completed full-stack training of a MoE-236B foundation model on 250 trillion tokens, supporting "Peking University's EvoPhys-World" which topped the Stanford leaderboard for 37 days. Previously, the market assumed domestic GPUs only handle inference, not high-end training. Now the S5000 proves through action: the "ultimate exam" for large model training—domestic compute is ready.

Equally important: customer composition is shifting. The report notes accelerated penetration of internet giants and carriers—critical infrastructure customers. This signal is vital. Government and financial orders are "baseline grain," but internet megacorps and operators are the "growth engine." These customers have long certification cycles and demanding requirements, but once in the supply chain, stickiness is extremely high and order continuity is rock-solid. This penetration expansion is the true foundation for Moore Threads' sustained high-growth revenue.

In domestic-replacement narratives, the nightmare is "looks good, doesn't work"—customers buy chips but lack ecosystem and software, turning them into bricks. Moore Threads' insight: solve the customer's "afraid to use" pain point with ecosystem.

100% core math library compatibility, 3000+ PyTorch operators, mainstream large models (DeepSeek, Kimi, GLM) achieve Day-0 adaptation. Customers migrating from CUDA to MUSA face near-zero switching cost. This "zero-cost migration" strategy is classic internet playbook—capture the beachhead with free, frictionless experience.

More powerful: developer ecosystem flywheel. As of H1, MUSA developers exceeded 800,000. From tens of thousands to 800,000 may seem modest in capital markets, but in software ecosystems, this is the self-growth inflection point. More developers mean more optimized models, richer applications, attracting more hardware purchases. Once this virtuous cycle locks in, competitors pay enormous cost to break it.

Additionally, the company's "compute-rendering-simulation" integrated strategy pre-positions it for embodied AI. From cloud large model training to simulation to edge deployment, Moore Threads has completed the full loop—not just a compute supplier, but infrastructure architect for the coming physical AI era.

One financial report detail deserves close attention: in the in-development projects table, "Huashan chip" and "Lushan chip" investment notably increased.

Per founder Zhang Jianzhong's prior comments, next-generation chips based on the "Huagang" architecture may debut by end-2026. If Huashan launches first, it directly targets the tens-of-billions-yuan enterprise AI training market, benchmarking top-tier international compute cards, penetrating core compute center supply chains. If Lushan debuts, it's a direct challenge to premium consumer graphics—notably, no domestic vendor has arm-wrestled international giants in high-end consumer cards. Moore Threads could create a "Black Monkey King" breakout moment.

Near-term: S5000 penetration acceleration among internet and carrier customers drives certain revenue growth. Mid-term: volume production pares costs; positive net profit is a matter of time. Long-term: Huagang architecture new products open a second growth curve, pivoting from "domestic replacement" to "global competition."

Many institutional investors favor Moore Threads not solely for sustained revenue growth, but because it has built the deepest moat in domestic GPUs: the MUSA ecosystem. MUSA developers now exceed 800,000, achieving 100% core math library compatibility with CUDA and 3000+ PyTorch operator compatibility, with mainstream large models achieving Day-0 adaptation. This "zero-cost migration" convenience drops switching barriers to minimum.

Ecosystem flywheel dynamics are the major theme: more developers → richer application optimizations → better hardware sales → thriving ecosystem. This "software-defined hardware" competitive moat cannot be broken by simple chip spec stacking. The commercial value of an 800,000-developer community—like Android or Harmony ecosystems of past eras—will underpin long-term customer repurchase.

Investment is betting on certainty. While the market still debates concepts, Moore Threads' 147.42% growth already announces: the domestic GPU commercialization curtain has only just risen.