DEV Community

Cover image for China's AI Dual Flywheel: Why Mainland Hardware Wins First, Hong Kong Internet Later
Allen Elzayn
Allen Elzayn

Posted on

China's AI Dual Flywheel: Why Mainland Hardware Wins First, Hong Kong Internet Later

#ai #cloud

Originally published at allenarch.dev

On November 9, 2025, UBS dropped Q3 earnings data for China A-shares: +12% YoY growth overall, but the real story is in the splits.

AI-related sectors led the charge:

  • Media: +57%
  • Electronics: +41%
  • Computers: +34%

Meanwhile, Hong Kong tech stocks the ones everyone watches for AI cloud exposure delivered mixed signals. Tencent reports this Thursday. Alibaba hasn't announced its date yet.

The mainland hardware boom is real. The internet AI monetization story? Still loading.

Key data points:

  • UBS Securities: A-shares Q3 earnings +12% YoY, driven by AI-related tech sectors (CNBC, Nov 9)
  • HSBC: Mainland winners = AI infrastructure hardware; Hong Kong winners = internet with AI cloud/models (CNBC, Nov 9)
  • Alibaba: AI spending in e-commerce already break-even, +12% ROAS in testing (CNBC, Oct 16)
  • Baidu: "Domestically developed chips and homegrown software" shield AI push from US export controls (Reuters, May 21)
  • Global capex context: Alphabet $91–93B (2025), Meta ≥$70B (2025, "notably larger" in 2026), Microsoft $34.9B in Q3 alone (+74% YoY) (TrendForce, Oct 30)

Two cycles, different clocks

China's AI story isn't monolithic. It's two cycles running on different timelines, and understanding the gap between them is the whole game.

Mainland hardware suppliers are cashing in now. The capex surge from global hyperscalers, plus China's domestic substitution push, is hitting their order books immediately. Server components, optical transceivers, power modules, liquid cooling all translating to actual earnings.

"Given a demand upswing from AI and self-reliance, the greater tech sector's rapid earnings growth drove overall ex-financials' earnings."

— Lei Meng, UBS Securities China equity strategist

The numbers back it up: electronics +41%, computers +34%, media +57% in Q3. Not narrative, profit.

Hong Kong internet platforms Alibaba, Tencent, Baidu are in a different phase. They're building AI cloud infrastructure and training models, spending heavily (Alibaba alone committed RMB 380 billion over three years), but the revenue from AI products is still ramping.

"In mainland China, hardware manufacturers related to AI infrastructure have benefited the most from the rally. In Hong Kong, internet names with AI-related cloud services and models have benefited the most."

— Herald van der Linde, HSBC head of Asia Pacific equity strategy

Translation: the datacenter gets built first, monetization comes later.

The gap between these cycles matters. Hardware suppliers get paid when the datacenter ships. Internet platforms get paid when enterprises adopt AI tools and scale usage over quarters.

That's a 12–18 month lag.

The ROI split: e-commerce vs cloud

Not all AI spending is created equal. Some verticals are already cash-flow positive. Others are still in the build phase.

Alibaba's e-commerce AI is already profitable. On October 16, Alibaba VP Kaifu Zhang told reporters the spend on Taobao and Tmall is breaking even, with preliminary tests showing +12% return on ad spend.

"It's very rare to see double-digit changes in such metrics."

— Kaifu Zhang, Alibaba VP

That's not hype, that's measurable ROI.

The playbook is straightforward: AI personalizes search, improves virtual try-ons, optimizes ad targeting. Conversion rates lift, take rates improve, and the math closes within quarters.

Source: CNBC, Oct 16

Cloud AI is a different story. Alibaba's cloud division, along with Tencent and Baidu, is burning cash to build datacenters, deploy GPUs, and scale inference capacity. CFO Toby Xu was clear on the August call: AI and consumption are "two major historic opportunities" requiring investments of "historic scale." Near-term margin gets deprioritized.

That means capex/sales stays elevated while ROI accrues slowly:

  1. Datacenter utilization has to climb (still ramping)
  2. Enterprise adoption has to scale (early innings)
  3. Net dollar retention on AI workloads has to improve (2026 story)

E-commerce delivers ROI in quarters. Cloud delivers in years.

The hardware cycle is already paying off

Global hyperscalers are jacking up capex, and that money flows straight to physical components. China's mainland suppliers are cashing in now.

TrendForce's October numbers tell the story:

  • Alphabet: $91–93 billion in 2025 (third upward revision, was $52.5B in 2024)
  • Meta: at least $70 billion in 2025, "notably larger" in 2026
  • Microsoft: $34.9 billion in Q3 alone, +74% YoY, with 2026 growing "even faster"
  • Amazon: on track for ~$118 billion (raised from $100B forecast)

Source: TrendForce, Oct 30

That capital gets spent on GPUs/ASICs (Nvidia Blackwell, Google TPU, AMD MI300X, Huawei Ascend 910C), HBM and memory, optical transceivers (100G/800G for datacenter interconnects), server chassis and power supplies, liquid cooling systems, and networking switches.

China's mainland suppliers play in everything except top-tier GPUs (export controls). But the rest of the bill of materials? They're competing and winning.

Take optical transceivers, the components that convert electrical signals to light for datacenter networking. Innolight (中际旭创) ranked #1 globally in 2024 with $3.3 billion revenue, up 122% YoY. New Yi Sheng (新易盛) jumped from 7th to 3rd place globally with $1.2 billion revenue, up 179%. Seven of the top 10 global optical module vendors are now Chinese companies.

This isn't just market share, it's capturing the upgrade cycle. Cignal AI reports 800G optical transceiver shipments will grow 60% in 2025, with 1.6T entering volume production in 2H25. When hyperscalers build out AI datacenters, these are the components going into every rack.

Domestic substitution accelerates this.

"Domestically developed chips and increasingly efficient homegrown software will form a strong foundation for long-term innovation in China's AI ecosystem."

— Shen Dou, Baidu VP

Reuters reported Huawei's Ascend 910C is prepping for mass shipments. It's not Blackwell-class (roughly 60% of H100 performance for inference), but for Chinese AI labs, it's the only option.

When Huawei ships Ascend at scale, mainland component suppliers win twice: they supply global hyperscaler builds and domestic Chinese datacenters (China's total computing power reached 246 EFLOPS in mid-2024, ranked #2 globally).

The earnings data confirms it. Hardware cycle is live.

The optical transceiver numbers are particularly telling. These components sit at the intersection of global hyperscaler spending and China's domestic buildout. When Innolight reports 122% revenue growth and New Yi Sheng jumps 179%, that's not inventory build it's actual datacenter deployment translating to component orders. The 60% shipment growth forecast for 800G modules in 2025, followed by 1.6T volume ramp in 2H25, gives us a leading indicator: as long as those shipment numbers hold, the hardware capex cycle is intact. When those numbers roll over, that's the signal to watch for cycle 1 peaking.

The internet monetization lag

Alibaba Cloud, Tencent Cloud, Baidu AI Cloud are in build mode. They're burning capital to scale infrastructure and wait for enterprise customers to actually adopt and pay.

HSBC flagged Hong Kong's AI winners as "internet names with AI-related cloud services and models," which is accurate, but timing is everything. Cloud AI revenue scales when enterprises move from pilot to production, datacenters fill up, and net dollar retention starts expanding. None of that happens overnight.

Chinese enterprises are still early in cloud AI adoption especially compared to U.S. companies. It's a multi-quarter ramp, and the platforms are building capacity ahead of demand.

Alibaba's RMB 380 billion commitment, Tencent's and Baidu's similar investments they're all betting enterprise adoption accelerates. The risk: if it lags, datacenter utilization stays low and ROI gets pushed out. The reward: if it scales, they own the stack and capture margin.

Right now, the expense is visible. Revenue is still loading.

The tell will be 2026 earnings. Watch for higher net dollar retention on AI workloads (existing customers spending more), growing backlog (contracted revenue not yet recognized), and improving gross margin on cloud AI (utilization kicking in).

If those metrics show up, the second flywheel is spinning. Until then, it's a capex story.

The sequence to watch

If the dual-flywheel thesis holds, here's what should show up over the next 12–18 months:

1H26: Hardware margin expansion outpaces internet. Mainland suppliers (optical, power, liquid cooling) see gross margin improve from volume leverage and better product mix. Inventory turns accelerate as orders pull forward. Operating leverage kicks in.

Hong Kong platforms still show elevated capex/sales ratios (building datacenter capacity), flattish or down cloud gross margin (low utilization early in ramp), and high R&D spending (model development, toolchains).

2H26: Internet monetization inflects. Hong Kong platforms start reporting accelerating cloud AI revenue growth, net dollar retention trending up, and gross margin expanding as utilization climbs.

If that happens, the second flywheel spins, and HK internet multiples re-rate.

Key indicators:

Metric What it signals Where to find it
800G/1.6T optical transceiver mix Upgrade cycle acceleration; 1.6T volume ramp in 2H25 signals demand pulling forward Innolight, New Yi Sheng quarterly earnings and commentary
Optical module ASP trends Pricing power = tight supply; ASP compression = competition/oversupply Supplier earnings calls (gross margin discussion)
Liquid cooling attach rate GPU density in new builds; higher attach = more advanced AI workloads ODM/server vendor disclosures
Alibaba Cloud revenue growth Enterprise AI adoption pace Quarterly earnings
Tencent Cloud gross margin Utilization and cost efficiency Quarterly earnings
Baidu AI Cloud backlog Forward visibility on revenue Earnings call commentary

When does cycle 2 spin up?

The mainland hardware cycle is validated. Q3 earnings showed media +57%, electronics +41%, computers +34%. Real numbers, not forecasts.

"We think 'growth' may remain a key investment theme. We highlight better risk/reward in the ChiNext board, due to its accelerating earnings with long-term resilience and valuation."

— Lei Meng, UBS Securities China equity strategist

ChiNext's largest members include CATL, Innolight, and Sungrow Power the hardware suppliers riding the capex wave.

The Hong Kong internet cycle is projected. It depends on enterprise AI adoption accelerating, datacenter utilization climbing, and NDR and backlog inflecting upward.

Base case: 2H26. If enterprise adoption follows the typical cloud ramp (18–24 months from pilot to production scale), Hong Kong platforms start showing meaningful AI revenue contribution in late 2026.

Bull case: mid-2026. If Chinese government procurement and SOE mandates accelerate adoption (plausible, given the state AI push), the monetization phase pulls forward 2–3 quarters.

Bear case: 2027. If enterprises stay cautious (macro headwinds, ROI unclear), datacenter utilization stays low, and the second flywheel doesn't spin until 2027 or later.

The timing matters for positioning. Mainland hardware is paying off now. Hong Kong internet is a 2026 call option.

The substitution wildcard

Export controls are forcing China to build its own stack. Near-term, that's a handicap Huawei's Ascend delivers roughly 60% of H100 performance. Long-term? It's a wildcard.

Nvidia's CUDA has 15+ years of tooling, libraries, and community code. Huawei's CANN framework works, but code has to be rewritten, performance optimization starts from scratch, and debugging cycles are longer. For Chinese AI labs racing to keep up with GPT-5-class models, that's a real disadvantage.

Source: Reuters, Apr 21

But compute scarcity breeds innovation. When you can't throw more GPUs at a problem, you optimize model architecture (sparsity, distillation, quantization), improve software efficiency (kernel fusion, memory management), and rethink training recipes.

China's AI labs, locked out of Blackwell-class hardware, might develop techniques that deliver similar output at lower compute cost. If that happens, the "handicap" becomes a competitive edge.

We'll know by 2027 whether forced substitution hurt or helped.

Three things to watch

Tencent earnings (Nov 14). Watch for cloud revenue growth acceleration, commentary on AI workload adoption by enterprise customers, and signals on datacenter utilization rates. If Tencent guides higher on cloud AI, that's the first signal the second flywheel is spinning up.

Alibaba Singles Day results (mid-Nov). Alibaba already said AI e-commerce is break-even and expects "very significant" positive impact on GMV during Singles Day (Nov 11). If GMV growth beats and management attributes it to AI, that validates the fast-ROI playbook.

Chinese government procurement announcements. Beijing's mandate for state-funded datacenters to use domestic chips is just the start. Watch for provincial-level AI infrastructure programs, SOE cloud migration mandates, and university procurement shifts. If government spending accelerates, it pulls forward the enterprise adoption curve.

The framework

Most coverage treats "China AI" as one trend. It's not. The data splits into two cycles with different ROI timelines:

Mainland hardware: triggered by global capex surge and domestic substitution. Beneficiaries are server, optical, power, and cooling suppliers. Timing: earnings impact now (Q3 2025 validated it). Risk: capex slowdown or supply chain shifts.

Hong Kong internet: triggered by enterprise AI adoption and datacenter utilization. Beneficiaries are Alibaba, Tencent, Baidu cloud platforms. Timing: monetization inflects 2H26 (projected). Risk: adoption lags, utilization stays low, ROI gets pushed out.

Knowing which cycle you're analyzing matters more than the headline.

The hardware cycle is real. The internet cycle is loading. The gap between them is the trade.

Resources:

Further Reading:

Connect

Top comments (0)