China's AI Century
Can Beijing Use Intelligence to Rewire Global Power by 2030?

Table of Contents

• Executive Summary
• Key Questions Answered
  • Can Chinese AI materially shift global economic gravity away from the West?
  • Can China build frontier AI without the West?
  • Are sanctions slowing China or accelerating independence?
  • Is "good enough AI" economically sufficient?
• Core Findings
  • 1. Semiconductor Independence: Huawei's Ascend Program
  • 2. Export Controls: A Double-Edged Sword
  • 3. Foundation Models and the DeepSeek Disruption
  • 4. Software Ecosystem: The Hidden Moat
  • 5. Manufacturing, Supply Chain, and Production Scale
  • 6. The Broader Domestic Chip Ecosystem
  • 7. System-Level Scaling and the "Good Enough AI" Thesis
  • 8. Energy Efficiency and Compute Economics
  • 9. Global South Influence and Market Strategy
• Contradictions & Debates
  • 1. Ascend 910C Specifications: A Fundamentally Uncertain Product
  • 2. Can Huawei Ever Catch NVIDIA?
  • 3. Production Scale: 100,000 to 1.4 Million
  • 4. Are Export Controls Working?
  • 5. Hardware Specs vs. Software Innovation as the Decisive Factor
• Deep Analysis
  • The Training vs. Inference Asymmetry
  • The "Sanctions Made China Stronger" Thesis
  • Is Cheap Inference More Powerful Than Smarter Models?
  • China's Path to AI Dominance Without Frontier Silicon
• Implications
  • For US Export Control Policy
  • For Global AI Competition
  • For NVIDIA and Western Semiconductor Firms
  • For the Global South
  • For Global Labor Markets and Deflation
• Future Outlook
  • Scenario A — "Sanctioned but Self-Sufficient"
  • Scenario B — "Fragmented AI World" (Base Case)
  • Scenario C — "Chinese Cost Shock"
  • Scenario D — "Taiwan Crisis / Compute Supply Disruption"
  • Scenario E — "Open-Source Dominance"
• Unknowns & Open Questions
• Evidence Map

Executive Summary

China's pursuit of AI supremacy represents one of the most consequential technological and geopolitical contests of the 21st century. This report synthesizes evidence from 16 sources spanning semiconductor analysis, export control policy, foundation model development, and market dynamics to assess whether Beijing can use artificial intelligence to materially rewire global power by 2030.

The central finding is paradoxical: US export controls have simultaneously constrained China's access to frontier AI hardware and accelerated the creation of a domestic ecosystem that is less capable but increasingly viable. Huawei's Ascend chip program, the centerpiece of China's semiconductor independence effort, has made rapid iterative progress—from the 910A (2019) through 910B, 910C, and announced 910D—but remains structurally constrained by manufacturing bottlenecks (SMIC yields below 50% on 7nm DUV process [8], [16]), a software ecosystem 12 years behind NVIDIA's CUDA [8], and critical supply chain dependencies on foreign high-bandwidth memory (HBM) [14], [16].

The most analytically important finding concerns a training vs. inference asymmetry created by the architecture of export controls: US restrictions have successfully targeted arithmetic performance (critical for training), creating a roughly 4-year hardware lead and approximately 3× cost advantage in training price-performance [13]. However, controls have left memory bandwidth (critical for inference) essentially unrestricted, giving the US effectively zero lead in inference hardware [13]. Since inference costs dominate real-world AI deployment, China's AI deployment ecosystem can remain highly competitive even with a meaningful training hardware deficit.

Huawei's Ascend 910C—the current flagship—delivers approximately 50–80% of NVIDIA H100 performance depending on workload and metric, with the most commonly cited figure being ~60% of H100 inference throughput based on DeepSeek testing [3], [4], [6], [8], [11]. Production targets for 2025 range from 400,000 to 1,000,000 Ascend units across variants [7], [8], [16], though source reliability on volume figures is low. Meanwhile, DeepSeek's R1 model—reportedly built for $5.6–6 million using pre-sanctions NVIDIA A100 chips [10], [12]—has challenged the assumption that frontier AI requires billions in capital expenditure, triggering a market rout that erased $600 billion from NVIDIA's market value in a single day [12].

The key unresolved question is whether "good enough AI" at 50–60% of frontier performance, combined with open-source model diffusion, cheaper inference economics, and massive domestic scale, is sufficient to shift global economic and strategic gravity. The evidence suggests this is plausible for commercial deployment and inference-heavy applications, but unlikely for frontier model training through at least 2027. The implications extend far beyond technology: Chinese AI could export deflationary pressure to Western industries, provide the default digital infrastructure for the Global South, and alter military calculations in Asia.

Overall confidence: MEDIUM. The directional narrative—gap narrowing but persistent—is well-supported across sources. However, critical quantitative claims (performance benchmarks, yield rates, production volumes, cost figures) suffer from a pervasive lack of independent verification. No source provides standardized MLPerf benchmark data for any Huawei chip. Most performance figures originate from Huawei-controlled disclosures, promotional publications, or unverified testing.

Key Questions Answered

Can Chinese AI materially shift global economic gravity away from the West?

Partially, through inference economics and open-source diffusion rather than frontier training supremacy. China is unlikely to match Western frontier training capability by 2030, but this may matter less than commonly assumed. The evidence shows: (a) Chinese chips already achieve near-parity in inference cost-performance [13]; (b) Chinese open-source models like DeepSeek R1 are achieving global adoption at a fraction of Western development costs [10], [12]; (c) the combination of cheap Chinese hardware + cheap/free models creates a vertically integrated alternative stack [10], [12]; and (d) the Global South represents a massive addressable market where Chinese AI's cost advantages are decisive [5], [10].

Can China build frontier AI without the West?

Not yet at frontier scale, but increasingly "good enough" for commercial and strategic purposes. Huawei's best chip (Ascend 910C) achieves roughly 50–60% of H100 real-world training performance and 60–70% inference performance [3], [4], [6], [8], [11], [14]. Over 90% of 130 notable Chinese language models (2017–2024) were trained on Western hardware [8]. The first model trained entirely on Chinese hardware was iFLYTEK's Spark 3.5, released only in January 2024 [8]. However, software innovations and architectural efficiency are enabling competitive outputs despite inferior hardware [10], [12], [14].

Are sanctions slowing China or accelerating independence?

Both simultaneously—but the acceleration effect may be exceeding the constraining effect. All sources agree that US export controls have constrained China's access to EUV lithography, leading-edge foundry nodes, and advanced HBM [5], [6], [7], [8], [9], [11], [13], [14]. However, the controls have also: (a) forced Huawei to redesign chips for domestic SMIC fabrication [14]; (b) driven Chinese tech giants toward domestic alternatives [1], [2], [3], [4], [7], [8]; (c) catalyzed software efficiency innovations like DeepSeek's architecture [10], [12], [14]; and (d) created a government-backed procurement pipeline that would not have existed otherwise [15], [16]. NVIDIA's share of the Chinese market dropped from 90% pre-export controls to approximately 50% as of May 2025 [8].

Is "good enough AI" economically sufficient?

For inference—increasingly yes. For frontier training—not yet. The Ascend 910C at $12,000–$28,000 versus the H100 at $25,000–$44,000 offers competitive price-performance for inference workloads [8], [14], [16]. DeepSeek's reported $5.6–6 million model development cost [10], [12] suggests software innovation can dramatically reduce hardware requirements. However, training reliability remains a "critical weakness" of Chinese hardware [11], and no source provides data on Chinese chips being used for frontier-scale training (1T+ parameters).

Core Findings

1. Semiconductor Independence: Huawei's Ascend Program

Huawei's Ascend chip family is the centerpiece of China's effort to build an independent AI hardware stack. The program has shown rapid iterative progress but remains structurally constrained.

Hardware Evolution

The Ascend 910C uses a dual-die (dual-chiplet) packaging design built on SMIC's 7nm DUV process with approximately 53 billion transistors [11], [14], [15], [16]. This represents approximately a 3× improvement in FP16 throughput over the 910A within the same effective process generation, achieved entirely through architectural optimization since SMIC's DUV process cannot match the transistor density of TSMC's leading nodes [14], [16].

The Performance Gap

The most consistently cited performance figure—derived from DeepSeek's independent testing—is that the Ascend 910C delivers approximately 60% of NVIDIA H100 inference throughput on large language models [3], [4], [6], [8], [11]. Source 14 provides the most methodologically rigorous estimates, breaking down performance by workload:

• Training performance: 910C ≈ 50–60% of H100 [14]
• Inference performance: 910C ≈ 60–70% of H100 [14]
• Memory bandwidth: 910C ≈ 54% of H100 (~1,800 GB/s vs. ~3,350 GB/s) [14]
• Interconnect bandwidth: HCCS ≈ 44% of NVLink (~400 GB/s vs. 900 GB/s) [14]

The interconnect gap deserves special emphasis. Huawei's HCCS delivers only approximately 56 GB/s per link versus NVIDIA's NVLink at 400–900 GB/s per GPU pair, enabled by NVSwitch chips that Huawei lacks an equivalent for [9]. At the aggregate level (8 GPUs), the gap is approximately 392 GB/s versus 3,200–7,200 GB/s—an approximately 8–18× disadvantage [9]. This bottleneck is particularly consequential for distributed training of large models requiring communication across many GPUs.

The FLOP/s gap between China's leading GPU (Ascend 910C, ~752–800 TFLOP/s FP16) and NVIDIA's B200 (2,250 TFLOP/s FP16) stands at approximately 3× as of 2024 [8], down from an order of magnitude in 2018 [8]. Chinese memory bandwidth has improved at approximately 24% per year from 2017 to 2025, versus approximately 13% per year for non-Chinese chips [8], suggesting faster relative improvement on this critical metric.

Critical caveat: No independent, third-party benchmark of any Ascend chip exists in the available sources. All performance claims derive from Huawei-controlled disclosures, leaked specifications, or limited third-party testing (DeepSeek). Until organizations like MLPerf publish verified Ascend results, all performance figures should be treated with significant uncertainty [1], [2], [3], [4], [5], [6], [7].

Source Disagreements on 910C Specifications

The available sources report substantially different specifications for the same chip:

These discrepancies may reflect different 910C sub-variants or steppings, different information sources, or estimation error. The 910B3 variant introduced HBM3e with bandwidth increased to 1.2 TB/s [16], suggesting iterative memory upgrades are occurring within product lines. The power consumption discrepancy (310W vs. 600W) is particularly significant for data center economics and efficiency calculations, and cannot be resolved from available evidence.

2. Export Controls: A Double-Edged Sword

US export controls on AI chips to China have evolved through multiple rounds, each closing loopholes opened by the previous round while creating new adaptation incentives:

October 2022 controls: Set dual thresholds at aggregate bidirectional transfer rate <600 GB/s AND aggregated processing <4,800 bit TOPS (equivalent to <300 FP16 TFLOPS) [9]. This forced NVIDIA to create the A800 and H800—deliberately throttled versions of the A100 and H100 tailored for the Chinese market [9]. The H800 matched H100 arithmetic performance exactly but reduced NVLink bandwidth to 400 GB/s (vs. H100's 900 GB/s), legally bypassing the 2022 rules [13].

October 2023 controls: Closed the A800 loophole by removing network bandwidth as a control variable entirely and halving the arithmetic performance ceiling to 50% of A100 levels [13]. This eliminated the H800 and forced NVIDIA to design the H20—retaining only approximately 15% of the H200's arithmetic power while preserving similar memory and network bandwidth to the H200 [13].

December 2024 update: Set a 2 GB/s/mm² bandwidth density limit on standalone HBM exports but explicitly exempted chips with co-packaged HBM, allowing H20 exports to continue unimpeded [13]. The two BIS documents for this update totaled 210 pages [13].

The fundamental asymmetry this creates is the most analytically important finding of this report: Export controls have been highly effective at restricting arithmetic performance (critical for training) but essentially ineffective at restricting memory bandwidth (critical for inference) [13]. This means the US has a roughly 4-year lead in training hardware but essentially no lead in inference hardware [13]. Since inference costs dominate real-world AI deployment—training is a one-time cost, inference is ongoing—the economic impact of controls is less severe than headline performance gaps suggest.

NVIDIA's China revenue impact has been substantial: the company generated approximately $17 billion (13%) of total revenue from China in FY2025 [7], and JPMorgan estimates NVIDIA could lose up to $16 billion in China revenue in 2025 under worst-case scenarios [7]. NVIDIA already forecast a $5.5 billion revenue hit from the H20 export ban [3], [7]. On April 28, 2025, NVIDIA shares fell over 2% following a Wall Street Journal report that Huawei is entering final testing for the Ascend 910D [7].

3. Foundation Models and the DeepSeek Disruption

DeepSeek represents the most dramatic evidence for China's AI competitiveness outside of hardware. Founded in late 2023 by Liang Wenfeng, a hedge fund manager and Zhejiang University alumnus [10], [12], the firm has produced findings that challenge fundamental assumptions about AI development economics:

• R1 model reportedly cost $5.6–6 million to develop [10], [12]—orders of magnitude less than US frontier models, though these figures are unverified and likely exclude prior R&D, salaries, data acquisition, and infrastructure costs
• Built using NVIDIA A100 chips purchased before US export restrictions [12]
• Surpassed ChatGPT in app store downloads shortly after launch [10]
• Open-source release allows others to build on the technology [10], [12]
• Uses inference-time computing (selectively activating model portions per query) to reduce costs [12]
• Achieved #3 ranking on the Chatbot Arena leaderboard [1]
• Marc Andreessen called it "AI's Sputnik moment" [10]

Market impact was severe: NVIDIA fell 17% ($600 billion loss, described as the largest single-day loss in stock market history), ASML fell 6%, Broadcom fell 17%, GE Vernova fell 21%, Vistra fell 28%, and the Nasdaq fell 3% [12].

Critical limitations on DeepSeek evidence: Both primary sources for DeepSeek claims (Sources 10 and 12) are Bitrue cryptocurrency exchange blog posts—not technology research outlets. Neither provides benchmark data (MMLU, HumanEval, GSM8K, etc.) to substantiate performance claims against GPT-4, Gemini, or LLaMA [10], [12]. The $5.6–6 million figure's methodology is unspecified. The cost comparison to US companies spending "billions" [10] is potentially misleading if it compares a single training run to total company R&D budgets.

More credible corroboration comes from Tom's Hardware, which reports that DeepSeek provides native support for Huawei's CUNN kernels and CUDA-to-CUNN conversion [11], suggesting genuine technical depth. DeepSeek's testing of the Ascend 910C also positions the firm as a serious hardware-software integration player [11].

Open-source as strategic weapon: DeepSeek's open-source release of R1 [10], [12] mirrors a broader pattern of Chinese AI firms using open-source to drive adoption. The sources note this "democratizes access to advanced AI for smaller companies and developing nations" [10]. Combined with cheap Chinese hardware, open-source Chinese models create a vertically integrated alternative to the Western AI stack.

4. Software Ecosystem: The Hidden Moat

The software ecosystem may be the most durable and strategically significant constraint on China's AI ambitions. NVIDIA's CUDA platform was introduced in 2007—twelve years before Huawei's CANN framework in 2019 [8]. The resulting gap encompasses:

• Developer tools and libraries: CUDA offers mature, extensively documented tools including TensorRT, cuDNN, and deep integration with PyTorch and TensorFlow [5], [9]
• Cloud provider partnerships: NVIDIA maintains strategic alliances with AWS, Azure, and Google Cloud [5]
• Aggressive co-optimization cycles: 18–24 month hardware-software refresh cadence [5]
• Developer ecosystem: Millions of developers with CUDA experience versus a nascent CANN/MindSpore community

Developers report CANN as bug-prone, unstable, and poorly documented [8]. Working with Ascend chips requires debugging without significant community support, and model optimization depends heavily on Huawei and progresses slowly [6]. Each Chinese AI chip maker bundles proprietary software stacks (Huawei's MindSpore/CANN, Baidu's PaddlePaddle, etc.), fragmenting the ecosystem and forcing developers to spend significant time adapting models across platforms [15].

The practical impact is measurable: Over 90% of 130 notable Chinese language models released between 2017 and 2024 were trained on Western hardware [8]. The first model trained entirely on Chinese hardware was iFLYTEK's Spark 3.5 LLM, released in January 2024 [8]. DeepSeek's next AI model has reportedly been delayed primarily due to the effort required to run training or inference on Huawei's chips instead of NVIDIA's [15]—a tangible, immediate cost of the software ecosystem gap.

Emerging bright spots: DeepSeek's decision to optimize its V4 frontier model for Ascend hardware is described as "the most significant validation of the 910C ecosystem, proving Chinese AI can be built on Chinese silicon" [14]. The CANN software ecosystem, while far smaller than CUDA, is described as now "production-grade for Transformer workloads" driven by this optimization [14]. Some experts predict that as AI models converge on Transformer architectures, NVIDIA's software ecosystem advantage may decline [11], though this prediction carries only medium confidence (0.5 per the source [11]).

5. Manufacturing, Supply Chain, and Production Scale

Manufacturing is identified as the single largest constraint on China's AI hardware independence.

Yield Crisis

SMIC's reported yields at 7nm fall below 50%, compared to TSMC's approximately 90% at the same node [8]. Source 16 reports that SMIC's N+2 process yield has improved from approximately 20% to 40–50% [16], though these figures carry very low confidence (0.4) and lack independent verification. Source 3 reports an even lower yield of approximately 30% [3], sourced from a single crypto-exchange publication. The true figure is unknown but almost certainly well below TSMC's levels, creating a compounding cost penalty that likely requires state subsidies to make Chinese chips price-competitive [8].

Without access to ASML's EUV lithography equipment—restricted by export controls—SMIC is confined to older DUV techniques that inherently limit transistor density and yield [3], [5], [7], [8].

Production Volumes

Production figures vary widely across sources and carry low confidence:

Source 16 explicitly acknowledges that 910C specifications and shipment figures are "largely mysterious" and based on "unreliable information" [16]. The most consistently reported figure across sources is the 70,000-unit initial order valued at approximately $2 billion [1], [2], [4].

Supply Chain Dependencies

Despite progress toward localization, critical dependencies remain:

• HBM memory in the 910C is still sourced from overseas manufacturers (primarily SK Hynix and Samsung), not domestic Changxin Memory [14], [16]. Both suppliers are subject to US pressure on technology exports to China [14].
• Shadow supply chain: TSMC is under investigation for potentially violating export controls; an estimated over 2 million Ascend 910B logic dies may have been manufactured by TSMC for Huawei shell companies after controls were imposed [8]. This suggests a shadow supply chain that partially mitigates official restrictions but is legally precarious.
• Samsung disruption risk: Samsung has reportedly paused production of Baidu's 4nm Kunlun chip designs [15], illustrating how US pressure on allied nations can disrupt Chinese supply chains.
• Localization rate: The chip's overall localization rate is reported as exceeding 90% [16], though the source does not define what counts as "localized" and the claim may be promotional.

6. The Broader Domestic Chip Ecosystem

China's AI chip effort extends well beyond Huawei, though Huawei remains the clear leader:

Baidu (Kunlun P800): Delivers 345 TFLOPS FP16 [15]. Baidu unveiled a 30,000-chip cluster powered by third-generation Kunlun P800 processors in 2025. Qianfan-VL models (3B, 8B, 70B parameters) were all trained on Kunlun chips [15]. Kunlun secured orders worth over 1 billion yuan (~$139 million) from China Mobile for AI projects [15]. Baidu's stock increased 64% over the year, partly attributed to the Kunlun reveal [15].

Alibaba (T-Head PPU): Features 96GB HBM memory and PCIe 5.0, pitched as a rival to NVIDIA's H20 [15]. A China Unicom data center runs over 16,000 PPU chips out of 22,000 total chips—a significant deployment [15].

Cambricon (MLU 590): Delivers 345 TFLOPS FP16, built on 7nm with FP8 support introduced in 2023 [15]. Returned to profitability by late 2024; share price jumped nearly 500% over 12 months [15]. This financial turnaround demonstrates that China's domestic chip path can yield commercially viable products [15].

Key observation: Most domestic offerings are "barely comparable to Nvidia's A100 from 2020" [15], with a multi-year gap to the H100. The 910C at 50–60% of H100 training performance [14] represents the best-case domestic ceiling.

7. System-Level Scaling and the "Good Enough AI" Thesis

Chinese chip makers are pursuing system-level strategies to compensate for single-chip performance gaps:

• Huawei's Atlas 950 SuperPoD: Planned for 2026 H2, targeting 8,192 Ascend chips producing 8 EFLOPS FP8, 1,152 TB total memory, and 16.3 PB/s interconnect bandwidth [6], [15]. The Atlas 960 plans to scale to 15,488 chips [15].
• CloudMatrix 384: Huawei's super node (384 Ascend 910C chips) is claimed to be stronger than NVIDIA's GB200 NVL72 [16], though this claim carries very low confidence (0.35) and lacks independent verification.
• Baidu's cluster approach: A 30,000-chip Kunlun P800 cluster [15], demonstrating scale-as-compensation at the thousand-chip level.
• MoE architecture advantage: Mixture-of-Experts architectures (like DeepSeek V3) can achieve better scaling on constrained interconnects because they require less all-reduce communication per training step [14].

The "good enough AI" thesis holds that China does not need to match NVIDIA chip-for-chip but merely needs hardware sufficient to run competitive AI models at scale. Supporting evidence includes: DeepSeek R1's frontier-class results despite constrained hardware [1]; the Ascend 910B roughly matching the A100 in raw FP16 TFLOPS (320 vs. 312) [9]; 60% inference performance potentially acceptable for many commercial deployment scenarios [3], [4], [11]; and lower power consumption potentially reducing total cost of ownership [4], [5].

Against this thesis: training performance (the bottleneck for frontier model development) is where the gap is largest, and no source provides data on Chinese chips being used successfully for frontier-scale training [1], [2], [3], [4], [5], [6], [7], [8]. The CUDA ecosystem advantage may impose a significant "software tax" on workloads ported to CANN/MindSpore [2], [6], [8].

8. Energy Efficiency and Compute Economics

Power consumption data across sources is contradictory:

If the 310W figure for the 910C is accurate, the chip's performance-per-watt would actually be superior to the H100 (800 TFLOPS at 310W ≈ 2.58 TFLOPS/W vs. ~990 TFLOPS at 700W ≈ 1.41 TFLOPS/W). This seems inconsistent with the claim that it only achieves 60% of H100 inference performance, creating an internal contradiction that none of the sources address [3], [4], [14], [15].

The Ascend 910D's claimed 1.2 PFLOPS at 350W [5] would represent dramatically superior energy efficiency per FLOP if verified. However, the Ascend 910B—the chip with the clearest data—operates at approximately 400W delivering approximately 320 TFLOPS (~0.8 TFLOPS/W) versus approximately 989 TFLOPS at 700W for the H100 (~1.4 TFLOPS/W), suggesting NVIDIA is actually more energy-efficient per FLOP at the 910B generation [6].

Confidence in energy efficiency claims: LOW. The data is contradictory and lacks independent verification. Power efficiency data is insufficient across sources to make definitive claims about China's compute sovereignty advantages from cheaper electricity or grid infrastructure.

9. Global South Influence and Market Strategy

Huawei's international strategy explicitly targets non-Western markets: China, Middle East, Russia, and countries less aligned with US trade policies [5]. The Ascend 910D is reportedly 30–40% cheaper than comparable NVIDIA solutions [5], making it potentially attractive for price-sensitive markets.

Adoption barriers remain significant outside China due to political concerns and lack of trust in Chinese technology [7]. Huawei's market cap declined from approximately $500 billion in 2020 to approximately $160 billion in 2025 due to sanctions [5], while NVIDIA's reached approximately $2.2–2.6 trillion [5].

No sources provide data on actual Global South adoption of Chinese AI hardware or models. The combination of cheap Chinese models (open-source) + Chinese hardware creates a vertically integrated alternative to the Western AI stack, but whether this is being adopted in practice remains unknown [5], [7], [10].

Chinese government procurement provides structural advantages: China Mobile ordered over 1 billion yuan (~$139 million) of Kunlun chips [15], and a China Unicom data center runs 16,000 Alibaba T-Head PPUs out of 22,000 total chips [15]. However, telecom operators reportedly prefer to mix chips from multiple suppliers rather than fully commit to Huawei [15], suggesting residual hedging behavior.

Contradictions & Debates

1. Ascend 910C Specifications: A Fundamentally Uncertain Product

The most significant unresolved contradictions concern the basic specifications of Huawei's current flagship chip:

These discrepancies may reflect different sub-variants, different information quality, or promotional vs. conservative estimates. Both sources 14 and 15 are rated "High" confidence by the chunk reports, making this a genuine disagreement requiring resolution through independent benchmarking.

2. Can Huawei Ever Catch NVIDIA?

Synthesis: The middle-ground evidence—that the gap has narrowed substantially but remains significant—is most consistently supported. Source 2's claim that Huawei will "never catch NVIDIA" is an editorial conclusion not supported by comparative technical analysis. Source 1's claim that the 910C "matches" the H100 contradicts the preponderance of evidence from more credible sources. The most important analytical nuance is that the gap is far narrower for inference than for training, and the question of whether matching frontier training is necessary for strategic AI objectives remains open.

3. Production Scale: 100,000 to 1.4 Million

Sources report production targets ranging from 100,000+ to 1.4 million Ascend chips for 2025 [1], [4], [7], [8], [16]. The most grounded estimates cluster around 400,000–1,000,000 combined 910B+910C units [8], [16], while the 1.4 million figure comes from the least reliable source (anonymous Reddit post) [1]. Even the most conservative confirmed figure—the 70,000-unit initial order [1], [2], [4]—represents meaningful commercial deployment if sustained.

4. Are Export Controls Working?

This is the most consequential debate in the report:

For controls working: Huawei was forced from TSMC N7+ to SMIC N+2 [11]. The 910C delivers only 50–60% of H100 training performance [14]. Over 90% of Chinese models were trained on Western hardware [8]. SMIC yields remain well below industry norms [8]. NVIDIA's H20—designed to comply with rules—caused dissatisfaction among Chinese buyers due to reduced performance [3].

Against controls working: DeepSeek built a competitive model for under $6 million using pre-sanctions hardware [10], [12]. Chinese firms are developing software innovations that reduce hardware dependency [10], [12], [14]. NVIDIA's China market share fell from 90% to ~50% [8], meaning domestic alternatives are gaining ground. The $600 billion NVIDIA market cap loss [12] suggests investors believe Chinese AI can compete despite sanctions. Shadow supply chains (2M+ TSMC-manufactured dies for Huawei shell companies [8]) partially circumvent controls.

Assessment: Controls are imposing real costs on Chinese hardware capabilities but are failing to prevent competitive AI development, particularly in inference and consumer-facing applications. The most important unintended consequence is that controls have forced the creation of a domestic ecosystem that would not have emerged organically.

5. Hardware Specs vs. Software Innovation as the Decisive Factor

Source 9 emphasizes hardware specs (TFLOPS, bandwidth, process node) as the primary determinants of AI capability. Sources 10 and 12 implicitly argue that software innovation (inference-time computing, efficient architectures, open-source) can compensate for hardware inferiority. Source 14 notes that architectural innovations like MoE can partially offset interconnect constraints. This debate is unresolved and represents one of the most consequential analytical questions for China's AI trajectory.

Deep Analysis

The Training vs. Inference Asymmetry

The most analytically important finding from the source set is the training/inference asymmetry created by export controls [13]. This has profound implications that are underappreciated in most Western analysis:

Training requires massive arithmetic throughput for matrix multiplications during backpropagation. BIS controls have successfully restricted arithmetic performance, giving the US a roughly 3× cost advantage (approximately 4-year temporal lead) in training hardware [13].

Inference—serving trained models to end users—depends primarily on memory bandwidth for reading model weights and KV caches. BIS controls have left memory bandwidth essentially unrestricted [13]. The H20, despite having only approximately 15% of the H200's arithmetic power, retains similar memory bandwidth to the H200 [13].

The practical implication: China can serve (infer from) frontier models at near-US-parity costs. China's disadvantage is concentrated in the ability to train new frontier models from scratch. Since inference costs dominate total AI spend in production (training is a one-time cost, inference is ongoing), the economic impact of controls is far less severe than headline performance gaps suggest.

This asymmetry suggests a bifurcated strategy is already emerging: Chinese developers use NVIDIA GPUs for training (when available) and Huawei Ascend chips for inference [8]. ByteDance and Tencent are named as companies using this hybrid approach [8]. As inference demand grows relative to training demand—a natural consequence of AI deployment scaling—this could make Chinese hardware progressively more competitive in the workload that matters most for economic deployment, without ever matching NVIDIA for training.

This is the most important contrarian insight of this report: China may not need to match Western frontier training capability if it can dominate the economics of inference at scale.

The "Sanctions Made China Stronger" Thesis

The evidence strongly supports this contrarian angle. Before sanctions, Chinese firms purchased NVIDIA chips at scale with no incentive to develop alternatives. Export controls have:

1. Forced Huawei to redesign the Ascend line for SMIC's DUV process rather than TSMC's leading nodes [14]
2. Created a domestic software ecosystem (MindSpore/CANN) that would not have existed otherwise [14,15]
3. Driven Chinese tech giants toward domestic alternatives—ByteDance ordered 100,000 Ascend 910B units [3]; Baidu, ByteDance, and Tencent are collaborating with Huawei on Ascend integration [4]; ByteDance and Tencent use hybrid approaches [8]
4. Created a massive government-backed procurement pipeline with regulatory mandates requiring domestic chips in sensitive industries [7,15,16]
5. Catalyzed software efficiency innovations like DeepSeek's architecture, potentially more durable than hardware advantages [10,12,14]
6. Forced NVIDIA to accept permanent market share loss—from 90% to ~50% of the Chinese market [8]

However, the thesis has important limits. The severity of yield constraints (below 50% at SMIC [8]) and the absence of EUV access suggest that sanctions are still imposing meaningful costs on China's AI hardware ambitions. The shadow supply chain (2M+ TSMC-manufactured dies for Huawei shell companies [8]) complicates the narrative, suggesting that China's progress may partly depend on circumventing the very controls meant to constrain it.

Is Cheap Inference More Powerful Than Smarter Models?

The DeepSeek disruption suggests that cost efficiency may matter more than raw capability for many commercial applications. DeepSeek R1 achieved competitive results for $5.6–6 million [10], [12]—though these figures are unverified—while surpassing ChatGPT in app store downloads [10]. The model's open-source release enables deployment on any compatible hardware [10], [12], creating adoption pathways that bypass hardware restrictions entirely.

The inference-time computing approach (selectively activating model portions per query [12]) reduces per-query costs, making inferior hardware viable for competitive model serving. If this efficiency trajectory continues, Chinese AI could deliver 80–90% of frontier model capabilities at 10–20% of the cost—a "good enough" proposition that many commercial users would accept.

China's Path to AI Dominance Without Frontier Silicon

Combining the evidence from all 16 sources, China's most viable path to AI influence by 2030 does not require matching Western frontier hardware. Instead, it involves:

1. Dominating inference economics through cheaper chips (Ascend at 50–80% of NVIDIA pricing [14,16]) and lower power requirements
2. Open-source model diffusion that achieves global adoption regardless of hardware constraints [10,12]
3. Massive domestic scale (potentially 1 million+ Ascend chips installed by 2026 [8,16]) that creates network effects and drives software ecosystem maturation
4. Algorithmic innovation (MoE architectures, inference-time computing, model distillation) that reduces hardware requirements [10,12,14]
5. Global South infrastructure built on Chinese hardware + Chinese models at price points NVIDIA cannot match [5,10]

This path would not produce technological supremacy in the traditional sense, but it could produce something potentially more consequential: the default AI infrastructure for most of the world's population.

Implications

For US Export Control Policy

The evidence suggests export controls need fundamental redesign:

• The training/inference asymmetry [13] means current controls constrain China's ability to train frontier models but not to deploy AI at scale. If the policy goal is to limit China's AI deployment capability, controls need to target memory bandwidth.
• The December 2024 exemption of co-packaged HBM [13] preserves current H20 export viability but leaves a significant vulnerability if future administrations tighten this loophole.
• Shadow supply chains (2M+ TSMC-manufactured dies for Huawei shell companies [8]) represent enforcement failures that undermine control effectiveness.
• The "sanctions as accelerator" effect [5], [7], [14], [16] means controls have a diminishing marginal impact: each additional restriction forces further domestic adaptation rather than capitulation.
• Market share loss is permanent—NVIDIA's 90%→50% Chinese market share decline [8] is unlikely to reverse even if controls are relaxed, because Chinese firms have already invested in porting workloads to CANN/MindSpore.

For Global AI Competition

• The assumption that massive capital expenditure is required for frontier AI is challenged by DeepSeek's cost claims [10], [12], though verification is needed
• Two parallel AI ecosystems are emerging: a US-led stack optimized for frontier training and high-performance inference, and a Chinese stack optimized for cost-efficient inference and open-source deployment
• The "good enough" segment—where most commercial AI applications live—could be captured by Chinese hardware and models, particularly in price-sensitive markets
• NVIDIA faces a structural threat not from Chinese hardware matching its specs, but from Chinese software innovations reducing demand for frontier compute

For NVIDIA and Western Semiconductor Firms

• The Chinese market—previously NVIDIA's second-largest—is being permanently lost to domestic alternatives [7], [8]
• The $16 billion worst-case revenue loss estimate [7] could accelerate lobbying for relaxed export controls
• Jensen Huang's acknowledgment of Huawei as a "serious competitor" [3] may reflect genuine competitive concern or strategic positioning for policy advocacy
• Western chip companies face deflationary pressure on pricing as Chinese alternatives establish cost benchmarks

For the Global South

• Open-source Chinese models offer accessible AI capabilities without Western licensing costs [10], [12]
• Chinese hardware (Ascend 910 series) provides a sanctions-resistant infrastructure option at 30–40% lower cost [5]
• The combination of cheap Chinese models + Chinese hardware creates a vertically integrated alternative stack that developing nations may find preferable to Western options [5], [10]
• No quantitative data exists on actual Global South adoption—this is a critical unknown

For Global Labor Markets and Deflation

The sources do not directly address labor market impacts, but the implications of Chinese AI cost efficiency extend to global wage pressure:

• If Chinese AI delivers 80–90% of frontier capabilities at 10–20% of cost, Western AI companies face deflationary pressure on pricing
• Cheaper AI inference accelerates automation across white-collar and manufacturing sectors
• Chinese manufacturing competitiveness, already formidable, is further enhanced by AI integration (though direct evidence on AI + manufacturing integration is absent from this source set)
• Price compression in AI services could compress wages in AI-adjacent sectors globally

Future Outlook

Scenario A — "Sanctioned but Self-Sufficient"

Probability: LOW-MEDIUM (15–25%)

By 2030, SMIC yields improve above 60%, Huawei closes the per-chip compute gap to within 80% of NVIDIA's latest generation, and annual Ascend production exceeds 2 million units. HBM supply chains are secured through domestic Changxin Memory or continued South Korean access. The CANN/MindSpore ecosystem matures sufficiently to support frontier training. Chinese AI labs train models competitive with Western counterparts on domestic hardware. Chinese open-source models and hardware become the default infrastructure for the Global South.

Key dependencies: Breakthrough in yield improvement, successful HBM localization, software ecosystem reaching critical mass. All three must succeed simultaneously. Requires EUV-equivalent fabrication breakthroughs or effective workarounds.

Scenario B — "Fragmented AI World" (Base Case)

Probability: MEDIUM-HIGH (40–50%)

Two parallel AI ecosystems solidify by 2030: a US-led stack (NVIDIA + OpenAI + cloud hyperscalers) optimized for frontier training and high-performance inference, and a Chinese stack (Huawei Ascend + DeepSeek/Qwen/Ernie + domestic cloud) optimized for cost-efficient inference and open-source deployment. The Global South adopts Chinese models for price-sensitive applications while Western enterprises remain on US infrastructure. Training hardware gaps narrow to approximately 1.5–2× but persist. Chinese firms continue using hybrid approaches: Western hardware for frontier training when accessible, domestic hardware for inference and smaller models [8]. The software ecosystem gap (CUDA vs. CANN/MindSpore) remains the most durable asymmetry.

This is the trajectory most consistent with current evidence trajectories across all 16 sources.

Scenario C — "Chinese Cost Shock"

Probability: LOW (10–15%)

DeepSeek's efficiency innovations prove generalizable, not a one-time anomaly. Chinese AI delivers 80–90% of frontier capabilities at 10–20% of cost by 2030. Chinese open-source models become the default for cost-sensitive applications globally. Western AI companies face severe deflationary pressure on pricing. NVIDIA's revenue base erodes as price-sensitive markets shift to Chinese alternatives. The $500 billion US "Stargate" initiative [12] and projected $1 trillion total US AI investment [12] prove to be overcommitment if Chinese cost efficiencies are validated.

Key dependencies: DeepSeek's $5.6–6 million cost claims [10], [12] must prove broadly replicable. Requires sustained software innovation compensating for hardware gaps. Depends on the thesis that cheap inference is more powerful than smarter models.

Scenario D — "Taiwan Crisis / Compute Supply Disruption"

Probability: LOW but IMPACT VERY HIGH (5–10%)

A Taiwan crisis disrupts TSMC production, cutting off advanced chips to both China and Western companies simultaneously. In this scenario, China's domestic Ascend production—while inferior—becomes the only available AI hardware at scale. The shadow supply chain (TSMC-manufactured dies for Huawei [8]) is eliminated. China's ~400,000–1,000,000 annual Ascend production [8], [16] becomes strategically decisive. However, China also loses access to any remaining NVIDIA hardware imports, and the HBM supply chain (SK Hynix, Samsung) faces potential disruption.

This scenario is the most extreme test of the "good enough AI" thesis. If Chinese chips are sufficient to maintain AI operations during a supply disruption, the strategic calculus fundamentally changes. If they are not, China faces a severe AI capability gap precisely when it might need it most.

Scenario E — "Open-Source Dominance"

Probability: LOW-MEDIUM (15–20%)

Chinese open-source models (DeepSeek R1, Qwen, future releases) become the global default for most commercial AI applications by 2030. The combination of open-source availability, competitive performance, and dramatically lower development costs creates a self-reinforcing adoption cycle. Western proprietary models (GPT, Claude, Gemini) retain advantages in frontier capabilities but lose market share in the much larger commercial inference market. Chinese firms monetize through cloud services, hardware sales, and ecosystem lock-in rather than model licensing.

Key dependencies: Chinese open-source models must maintain competitive quality. Requires continued government support for open-source strategy. Depends on whether Western firms respond with their own open-source offerings (as Meta has with LLaMA).

Unknowns & Open Questions

The following critical questions cannot be answered from the available source set and represent the most important gaps for future research:

Evidence Map

References

1. ↩ Huawei's Ascend 910C chip matches NVIDIA's H100 - https://reddit.com/r/deeplearning/comments/1ihecl0/huaweis_ascend_910c_chip_matches_nvidias_h100
2. ↩ Huawei vs NVIDIA: Chip Performance Comparison - https://asapdrew.com/p/huawei-vs-nvidia-chip-performance
3. ↩ Huawei vs NVIDIA: Ascend Chip Performance 2025 - https://bitrue.com/blog/huawei-vs-nvidia-ascend-chip-performance-2025
4. ↩ Huawei Ascend 910C vs Nvidia H100: A Big Leap Towards AI Independence | LinkedIn - https://linkedin.com/pulse/huawei-ascend-910c-vs-nvidia-h100-big-leap-towards-ai-qureshi-xi84e
5. ↩ Huawei's Ascend 910D: The Silent Challenger to Nvidia's AI Crown – A Deep Global Perspective (2025) - https://semiconductorsinsight.com/huawei-ascend-910d-vs-nvidia-h100
6. ↩ Comparing Ascend 910B and NVIDIA H100 - https://github.com/lzwjava/jekyll-ai-blog/blob/main/notes/2026-03-28-ascend-910b-vs-h100-en.md
7. ↩ Why Huawei's New AI Chip Isn't a Global Threat to Nvidia Yet - https://tecknexus.com/why-huaweis-new-ai-chip-isnt-a-global-threat-to-nvidia-yet
8. ↩ Why China Isn't About to Leap Ahead - https://epochai.substack.com/p/why-china-isnt-about-to-leap-ahead
9. ↩ GPU Performance Datasheets: NVIDIA & Huawei/HiSilicon - https://arthurchiao.art/blog/gpu-data-sheets
10. ↩ What is DeepSeek AI? - https://bitrue.com/blog/what-is-deepseek-ai
11. ↩ DeepSeek research suggests Huawei's Ascend 910C delivers 60% NVIDIA H100 inference performance - https://tomshardware.com/tech-industry/artificial-intelligence/deepseek-research-suggests-huaweis-ascend-910c-delivers-60-percent-nvidia-h100-inference-performance
12. ↩ DeepSeek AI: Chinese Innovation Cripples BTC, Nvidia Impact - https://bitrue.com/blog/deepseek-ai-chinese-innovation-cripples-btc-nvidia-impact
13. ↩ US export controls on China and their impact on AI - https://epoch.ai/gradient-updates/us-export-controls-china-ai
14. ↩ Huawei Ascend 910C - Awesome Agents AI Hardware Analysis - https://awesomeagents.ai/hardware/huawei-ascend-910c
15. ↩ Who Will Fill Nvidia's AI Chip Void? - https://recodechinaai.substack.com/p/who-will-fill-nvidias-ai-chip-void
16. ↩ A Brief Introduction to Huawei Ascend Cloud - https://medium.com/@huaweiclouddevelper/a-brief-introduction-to-huawei-ascend-cloud-cbef8f25bc34