← Back to All Research
Commission a Report
← Back to All Research
Commission a Report
DeepSeek, a research division of the Chinese quantitative trading firm High-Flyer Capital Management, released two models in rapid succession in late 2024 and January 2025 that together constitute the single most disruptive event in AI since GPT-4's launch. Understanding the investment implications requires separating what was technically achieved from the narrative that surrounded the release.
DeepSeek-V3 (released December 26, 2024) is a Mixture-of-Experts (MoE) language model with 671 billion total parameters, of which 37 billion are activated per token. It was trained on 14.8 trillion tokens using a cluster of 2,048 Nvidia H800 GPUs — a chip that Nvidia designed specifically after the Biden administration tightened export controls to exclude H100s from China. The training run consumed approximately 2.788 million H800 GPU-hours. At H800 spot pricing of roughly $2/hour at the time, this yields a total compute cost of approximately $5.576 million — the source of the "$6M" headline. DeepSeek released the weights under a permissive open-source licence.
DeepSeek-R1 (released January 20, 2025) is a reasoning-focused model trained using reinforcement learning, directly competing with OpenAI o1. R1 achieved benchmark parity or superiority with o1 on several mathematical and coding benchmarks and was released with open weights at multiple capability tiers (R1-Zero, R1, and distilled variants). The distilled models — ranging from 1.5B to 70B parameters — run locally on consumer hardware.
The $6M figure is real but partial. It counts only direct GPU compute costs for the final training run. It excludes:
A more honest all-in cost estimate for DeepSeek-V3's development, including amortised hardware, runs, data, and labour, is likely in the $50–150M range. This is still dramatically lower than the $100M+ compute alone cost of GPT-4 training (estimated $63–100M), and the $500M–1B+ estimates for frontier models from 2024 onward. The cost efficiency story is real. The $6M figure is a marketing number, not an accounting one.
DeepSeek did not simply get lucky with a cheap training run. The efficiency reflects genuine architectural innovations:
Multi-head Latent Attention (MLA): DeepSeek's custom attention mechanism compresses the KV cache substantially relative to standard multi-head attention. This reduces memory bandwidth requirements during both training and inference — directly relevant to the hardware cost equation.
DeepSeekMoE with auxiliary-loss-free load balancing: Standard MoE models require auxiliary losses to prevent expert collapse (where a small subset of experts receives most tokens). DeepSeek developed a load balancing strategy that achieves expert utilisation without this auxiliary loss, improving training efficiency and final model quality.
FP8 mixed-precision training: V3 was trained using 8-bit floating point precision for most operations — a level of aggressive quantisation that most large-scale training pipelines had not yet adopted. This roughly doubles training throughput on compatible hardware versus FP16, and the H800 chips support it natively.
Multi-Token Prediction (MTP): An auxiliary training objective that predicts multiple future tokens simultaneously. This is not new in principle, but DeepSeek's implementation improved sample efficiency measurably during the training run.
Taken together, these are not marginal optimisations. They represent a coherent systems-engineering approach to training efficiency that the US frontier labs, with more abundant compute, had less incentive to pursue as aggressively. The export controls, paradoxically, may have forced DeepSeek to innovate in directions that are now being studied and adopted globally.
| Benchmark | DeepSeek-V3 | DeepSeek-R1 | GPT-4o | Claude Sonnet 3.7 | Gemini 2.0 Pro |
|---|---|---|---|---|---|
| MMLU (5-shot) | 88.5% | 90.8% | 87.2% | 88.3% | 87.8% |
| MATH-500 | 90.2% | 97.3% | 76.6% | 78.3% | 82.1% |
| HumanEval (coding) | 82.6% | 85.3% | 90.2% | 93.7% | 84.5% |
| SWE-Bench Verified | 42.0% | 49.2% | 38.8% | 70.3% | 51.2% |
| GPQA Diamond (science) | 59.1% | 71.5% | 53.6% | 65.0% | 62.1% |
| LiveCodeBench | 40.5% | 65.9% | 33.4% | 43.2% | 37.6% |
| AIME 2024 (maths olympiad) | 39.2% | 79.8% | 13.4% | 16.0% | 42.0% |
Note: Benchmarks are as-reported at model launch. Direct comparisons are imperfect due to differing evaluation harnesses and prompt formatting. R1 results reflect the full R1 model, not distillations. Claude, GPT, and Gemini figures reflect the best available versions concurrent with DeepSeek's January 2025 release. The AI frontier has advanced substantially since; consult current leaderboards for live standings.
The pattern is clear: DeepSeek-R1 is genuinely frontier-competitive on mathematical and scientific reasoning, surpassing or matching o1 in the areas o1 was designed to excel at. DeepSeek-V3 matches GPT-4o on general language tasks. The lagging areas are enterprise-grade coding agents (SWE-Bench, where Anthropic's Claude leads significantly) and instruction-following reliability — areas that matter most for commercial deployment.
As of late March 2026, DeepSeek-R2 has not been formally released, but multiple indicators suggest it is in late-stage development. Community reports and API endpoint testing by developers suggest DeepSeek is testing a model with substantially improved reasoning and reduced hallucination rates. Credible estimates place R2 training at 3–4x the compute of V3, using an expanded cluster that likely includes Huawei Ascend 910C chips as well as H800s. If R2 ships at the efficiency ratio of V3 relative to its contemporaries, it would represent a material step-change in Chinese frontier capability. DeepSeek has also filed patents suggesting work on a next-generation MLA variant with further KV cache compression. The release timeline based on development cadence points to mid-2026.
Framing China AI as "DeepSeek" is an analytical error of the same type as framing US AI as "OpenAI." The Chinese ecosystem includes multiple well-funded labs with frontier-competitive models across different capability profiles.
Alibaba's Qwen 2.5 series, released in September 2024, represented the first major indication that China's frontier model development had achieved parity on general-purpose language tasks. Qwen 2.5-Max (the MoE flagship) posted MMLU scores of 87.7% and outperformed GPT-4o on several Chinese-language and bilingual benchmarks. Qwen 2.5-Coder 32B achieved results competitive with GPT-4 Turbo on HumanEval, a significant milestone for a domestically-produced coding model.
The Qwen 3 series, released Q1 2026, further closes the gap. Qwen 3-235B (MoE, 22B active parameters) benchmarks comparably to Claude Sonnet 4.6 on coding tasks and surpasses it on Chinese-language instruction following. Critically, Alibaba releases Qwen models under Apache 2.0 — the most permissive commercial licence available — making Qwen 2.5 and 3 the preferred base for Chinese enterprise AI deployment and a significant open-source contributor globally.
Baidu's ERNIE series has historically been the most commercially deployed Chinese AI system, backed by Baidu's dominant Chinese search market position. ERNIE 4.0 (2024) closes much of the quality gap with international models on Chinese-language tasks but continues to lag on scientific and coding benchmarks. The more relevant story is ERNIE Speed and Turbo — Baidu's inference-optimised variants — which power real-time Chinese search AI at scale. Baidu's AI search integration (Ernie Bot embedded into Baidu Search) has over 300M monthly active users, making it the world's largest deployed consumer AI assistant by user count outside ChatGPT.
Baidu's strategic position differs from DeepSeek and Alibaba: it is primarily an inference-at-scale operator rather than a frontier training lab. The competitive risk to Baidu is disruption from more capable open-source models (Qwen, DeepSeek) deployed by competitors, rather than from US export controls.
ByteDance's AI research division, often operating under the "Seed" brand, has been one of the most prolific publishers of AI research globally. Their Doubao product (consumer AI assistant) has rapidly accumulated users, and ByteDance's unique advantage is its unparalleled data moat: TikTok, Douyin, and Toutiao generate behavioural and engagement signal at a scale no other Chinese lab can match. ByteDance released Doubao Pro 32K in 2024, which benchmarks competitively with Qwen 2.5-72B on general tasks. The more significant development is ByteDance's inference infrastructure: its BytePlus cloud ARM offers Doubao inference to third-party developers at aggressive pricing, directly competing with Alibaba's Tongyi cloud.
Moonshot AI, founded in 2023 by former Google Brain researchers and valued at approximately $3B as of mid-2024, has pursued a distinctive strategy focused on ultra-long context. Kimi's context window — reportedly 200K tokens at launch and extended to 1M+ in 2025 — predated the long-context race at Western labs by several months. The Kimi k1.5 reasoning model, released February 2025, demonstrated competitive reasoning performance with DeepSeek-R1 on mathematical benchmarks while being trained on fewer parameters, suggesting strong architectural efficiency. Kimi is the primary AI assistant for long-document analysis and research in Chinese enterprise settings.
Zhipu AI, spun out of Tsinghua University's KEG Lab, operates the GLM (General Language Model) series and produces the ChatGLM commercial deployment. GLM-4 (2024) is competitive with GPT-3.5 Turbo on Chinese benchmarks. Zhipu's significance is less about frontier performance and more about academic-to-commercial pipeline: its deep connections with Chinese universities mean it is the default academic AI infrastructure for Chinese researchers, with implications for talent pipeline and research publication flow.
01.AI, founded by Kai-Fu Lee (former Google China president) and backed by Sinovation Ventures, released the Yi model family as a bilingual (English/Chinese) open-source series. Yi-34B achieved competitive results with LLaMA 2 70B in early benchmarking and Yi-1.5 extended the lineage. The Yi models are widely used as foundation models for fine-tuning across the Chinese enterprise AI market. 01.AI's strategic positioning — open-source, bilingual, commercially accessible — mirrors Meta's approach with LLaMA and serves a similar function in the Chinese ecosystem.
Minimax (formerly named Hailuo) has carved out a distinct niche in multimodal AI, particularly video generation. Its Hailuo video model became a viral sensation in late 2024, generating photorealistic video clips from text prompts at quality levels competitive with Sora's early demos. On language benchmarks, Minimax's MoE model is competitive but not frontier-leading. The investment angle on Minimax is its video and audio generation capabilities — an area where Chinese labs have demonstrated systematic strength and where Western labs have been slower to commercialise.
| Lab / Model | MMLU | Coding (HumanEval) | Reasoning (MATH) | Long Context | Primary Strength |
|---|---|---|---|---|---|
| DeepSeek-V3 | 88.5% | 82.6% | 90.2% | 128K | General + Math |
| DeepSeek-R1 | 90.8% | 85.3% | 97.3% | 128K | Scientific reasoning |
| Qwen 2.5-Max | 87.7% | 79.1% | 85.3% | 128K | Bilingual, enterprise |
| Qwen 3-235B | ~89% | ~84% | ~91% | 128K | Coding, CN enterprise |
| Kimi k1.5 | ~86% | ~78% | ~92% | 1M+ | Long context, reasoning |
| ERNIE 4.0 Turbo | ~84% | ~72% | ~78% | 128K | CN search at scale |
| GLM-4 | ~80% | ~68% | ~70% | 128K | Academic CN ecosystem |
Qwen 3 and Kimi k1.5 results are approximate, based on available published benchmarks and developer reports as of Q1 2026. Tilde (~) indicates interpolated estimates from partial public data.
The aggregate picture is of a Chinese AI ecosystem that, across its top three or four labs, is operating at or near GPT-4-class capability on benchmark evaluations, with DeepSeek-R1 exceeding that tier on specific mathematical and scientific reasoning tasks. The gap to the current US frontier (Claude Opus 4.6, GPT-5.4, Gemini 3 Pro) remains meaningful, particularly on complex software engineering, nuanced instruction following, and agentic multi-step tasks. But that gap is closing faster than the export control regime was designed to prevent.
The US semiconductor export control regime targeting China's AI development has undergone three major tightening cycles:
October 2022 (Biden Round 1): BIS restricted export of Nvidia A100 and H100 GPUs to China, based on chip-to-chip interconnect bandwidth thresholds. The controls applied performance limits defined by aggregate performance (TOPS) and inter-chip bandwidth — the first time hardware performance specifications had been weaponised as export control criteria.
October 2023 (Biden Round 2): After Nvidia successfully sold the H800 and A800 (chips designed to comply with the 2022 thresholds by reducing inter-chip bandwidth), BIS revised the rules to close the loophole. Additional restrictions were placed on a broader class of chips, data centre equipment, and software tools. Restrictions were extended to cover 21 additional countries as potential transshipment routes.
January 2025 (Biden "AI Diffusion Rule"): Days before the Trump inauguration, the Biden administration published a tiered framework for AI chip export controls, establishing three tiers of countries with differentiated access to AI hardware and model weights. The rule was substantially revised and scaled back by the Trump administration in subsequent months, with some tier-2 country restrictions relaxed under bilateral agreements.
The honest answer is: partially, with diminishing effectiveness over time.
What worked: Chinese frontier labs do not have access to H100 or H200 GPUs in the volumes required to match the largest US training runs. The most capable AI hardware produced by Nvidia — Blackwell B200, GB200, and the emerging Rubin architecture — remains inaccessible to Chinese buyers. This creates a compute ceiling that is real and meaningful at the absolute frontier. Anthropic training Claude Opus 4.6 on clusters of Rubin chips that China cannot purchase represents a genuine capability advantage that benchmarks may not yet fully reflect, but that will manifest in future model generations.
What failed:
SMIC's N+1 process node — which entered volume production in 2022 — is approximately 7nm equivalent in density and power efficiency, though it uses deep ultraviolet (DUV) lithography rather than extreme ultraviolet (EUV), which imposes yield penalties and limits the most aggressive design rules. The Kirin 9000S chip in Huawei's Mate 60 Pro (revealed September 2023 in a significant embarrassment for the US export control regime) was manufactured on this node.
SMIC's path to 5nm and below is constrained by the absence of ASML EUV systems — the Netherlands, under US pressure, has denied export licences for EUV equipment to China. Without EUV, SMIC can continue to improve its DUV-based processes through multi-patterning techniques (printing the same layer multiple times with precise alignment), but this is yield-constrained and economically challenging at scale. Independent analysts estimate SMIC is approximately 3–5 years behind TSMC on equivalent node performance with the current tooling. That gap is not closing at the rate optimists projected in 2023.
Huawei's Ascend 910B is the primary domestically-produced AI training accelerator in deployment across Chinese hyperscalers including Baidu Cloud, Alibaba Cloud (for domestic government-facing workloads), and Huawei Cloud. Independent benchmarking — conducted primarily through inference latency and throughput tests rather than formal training benchmarks — suggests:
The strategic implication: China's AI training infrastructure is running at approximately 60–80% efficiency relative to what an equivalent dollar spend on US hardware would achieve, and improving. The 20–40% gap remains meaningful for the absolute frontier, but it is not the 10:1 disadvantage that export control advocates originally projected.
Nvidia's stock declined approximately 17% in the session following DeepSeek's R1 release announcement on January 20, 2025 — an intraday loss of roughly $600B in market capitalisation, at the time the largest single-day market cap loss in stock market history. The selloff was driven by a coherent but ultimately flawed chain of reasoning: if models can be trained at a fraction of the compute cost, the demand for Nvidia GPUs will collapse.
The flaw in this reasoning is the Jevons Paradox applied to AI compute. When the cost of AI capability falls, the applications it enables expand. Lower training costs mean more experiments, more fine-tuned models, more inference deployments, and ultimately more total GPU-hours consumed — not fewer. The short-term shock to inference economics is real: if Chinese open-source models can be self-hosted at near-zero marginal cost, the inference API businesses of mid-tier Western labs face structural compression. But this does not reduce demand for the chips used in training frontier models or deploying inference at hyperscale.
The more legitimate concern — which the market has not fully priced — is whether the efficiency improvements demonstrated by DeepSeek will slow the upgrade cycle for inference hardware. If operators can serve twice the queries per GPU-hour, they need fewer chips to serve the same demand, which could delay refresh cycles. Nvidia's DGX Cloud and inference-oriented SKUs are more exposed to this risk than its flagship training products.
OpenAI's public response to DeepSeek was measured but the internal disruption was significant. Sam Altman acknowledged DeepSeek-R1 as "impressive" while noting OpenAI would "obviously deliver much more advanced intelligence." More revealing was the acceleration of model releases in the subsequent months: o3 (January 2025), o3-mini, and the rapid iteration cadence through GPT-5 variants — all of which suggests DeepSeek's release acted as a competitive accelerant.
OpenAI's structural advantage over DeepSeek remains: first-mover enterprise contract lock-in, the Microsoft Azure distribution moat (M365 Copilot users, Azure OpenAI Service), and the fact that GPT-5.4's benchmark performance on enterprise-relevant tasks (SWE-Bench, browser automation, instruction following) exceeds DeepSeek-V3 meaningfully. However, the release of DeepSeek under open weights permanently changed the competitive calculus for the inference-as-a-service market, where OpenAI earned most of its 2024 API revenue.
Google's response was strategically interesting. DeepSeek's open-source release amplified pressure on Google to open its own models — a debate that had been ongoing internally since Meta's LLaMA releases. Google accelerated the Gemma family (its open-weight series) and deepened developer commitments to the Gemini API. Google's actual competitive position relative to DeepSeek-V3 is strong: Gemini 3 Pro significantly outperforms V3 on multimodal tasks and enterprise software engineering. The DeepSeek shock was more of a narrative problem for Google than a capability threat.
For Meta, DeepSeek's release was validating. Meta has pursued an open-source-first model release strategy since LLaMA 1 — a strategy that was internally controversial and externally criticised as strategically irrational. DeepSeek demonstrated that open-source weights at frontier capability are achievable and create significant downstream ecosystem value. LLaMA 4, released in April 2025, benefited substantially from community optimisations enabled by the open-weights movement that DeepSeek amplified. Meta's AI strategy — using LLaMA as a platform that others build on, while deploying its own models across 3B+ users — is now widely viewed as defensible.
Microsoft's response was to integrate DeepSeek models directly into Azure AI Foundry — effectively monetising the Chinese model's popularity without endorsing it strategically. Azure customers can now deploy DeepSeek-R1 on Microsoft-managed infrastructure, which creates an interesting dynamic: the efficiency gains DeepSeek brings reduce per-query inference costs for Azure customers, which increases query volumes, which generates Azure compute revenue. Microsoft benefits from DeepSeek's efficiency even as it competes with the model.
DeepSeek's release crystallised a structural shift in the AI market that had been building since LLaMA 1: the commoditisation of model intelligence below the absolute frontier. The practical consequence is a bifurcated market:
Anthropic occupies a distinctive position in the post-DeepSeek competitive landscape. The standard objection to Anthropic as an investment thesis has been that it faces commoditisation pressure from open-source models on the low end and OpenAI/Google competition on the high end, with a $61.5B valuation (post the $2.75B Amazon funding round — part of a total $8B Amazon commitment) that appears to require sustained frontier model superiority to justify. DeepSeek's arrival accelerated the low-end commoditisation concern. But it also clarified why Anthropic's specific positioning is structurally more defensible than any Chinese model, regardless of benchmark parity.
The defensibility argument rests on four factors that cannot be replicated through architectural innovation or open-weight releases:
Anthropic developed Constitutional AI (CAI) — a training methodology in which the model's safety properties are baked into the training process through a constitutional document that governs its reasoning, rather than being applied as post-hoc output filters. This is not merely a safety feature; it is an architectural choice that produces qualitatively different model behaviour in adversarial or ambiguous contexts. No Chinese lab has published a comparable alignment methodology, and the safety community's consensus is that CAI-style approaches produce meaningfully more reliable behaviour on open-ended deployment scenarios.
For enterprise and government customers — who represent the highest-value, most defensible market segment — this matters not as a marketing claim but as an operational requirement. Regulated industries (finance, healthcare, government) cannot deploy models whose alignment properties are opaque or unverified. Anthropic's interpretability research programme, which has made the most significant published progress on understanding internal model representations of any lab, is the technical foundation for this assurance.
Anthropic's mechanistic interpretability team — led by researchers including Chris Olah, who pioneered the neural network visualisation field — has published what is broadly considered the most significant empirical work on understanding transformer model internals. The "dictionary learning" / sparse autoencoder approach to extracting interpretable features from Claude's residual stream has revealed what appear to be coherent internal representations of concepts, emotions, and reasoning states. This research has practical implications for enterprise trust: Anthropic can increasingly explain why Claude behaves as it does in specific contexts, not just describe the output.
No Chinese lab — and no other US lab — has published comparable interpretability work. This is a medium-term moat that becomes more valuable as regulatory pressure for AI explainability increases, which it will.
Anthropic holds FedRAMP High authorisation for Claude deployment in US government contexts — a certification process that takes years and requires extensive third-party audit of security controls, data handling, and model behaviour. Chinese models are effectively permanently excluded from this market regardless of their benchmark performance. The market for AI deployment in US federal agencies, defence contractors, intelligence community primes, and regulated financial institutions is worth tens of billions of dollars annually and is structurally inaccessible to any Chinese-origin model.
Beyond formal certifications, Anthropic's enterprise trust proposition rests on data residency guarantees, audit trails, and the credibility of its safety commitments — all of which Chinese labs cannot credibly offer to Western enterprise buyers regardless of technical capability.
| Model | Context Window | Input ($/MTok) | Output ($/MTok) | Primary Use Case | SWE-Bench |
|---|---|---|---|---|---|
| Claude Haiku 4.5 | 200K tokens | $1.00 | $5.00 | High-volume, cost-sensitive | ~45% |
| Claude Sonnet 4.6 | 1M tokens | $3.00 | $15.00 | Enterprise default, agent tasks | 70.3% |
| Claude Opus 4.6 | 1M tokens | $5.00 | $25.00 | Frontier reasoning, research | 80.8% |
| Opus Fast Mode | 1M tokens | $30.00 | $150.00 | Latency-critical frontier | 80.8% |
On SWE-Bench Verified — the most commercially relevant benchmark for enterprise software use cases — Claude Opus 4.6 at 80.8% leads the field. DeepSeek-R1 at 49.2% and V3 at 42.0% are materially behind. This is not a gap that open-weight efficiency gains alone can close; it requires the kind of sustained RLHF and agentic scaffold investment that Anthropic has been systematically building.
Anthropic's $2.75B funding round from Amazon (part of the larger $8B Amazon commitment structured across multiple tranches), combined with Google's total investment of approximately $2B, values the company at $61.5B as of the most recent round. At a reported annual revenue run rate approaching $2–3B (with steep growth trajectory driven by enterprise contracts and the M365 Copilot integration reaching 400M+ Microsoft seats), the multiple is demanding but not irrational for a company with genuine frontier capability and a structural government/enterprise trust moat.
Community speculation around upcoming Anthropic model codenames — the names "Mythos" and "Capybara" have circulated in AI developer communities as potential internal project names for next-generation Claude models — should be treated cautiously. Anthropic does not confirm model names in advance, and the source of these codenames is unverified. What is documented is Anthropic's research publication cadence suggesting ongoing work on substantially extended reasoning capabilities, improved agentic reliability, and further interpretability tooling. A plausible 2026 frontier model from Anthropic should be expected to widen the SWE-Bench and complex reasoning gap relative to DeepSeek further, not narrow it.
The central investment debate triggered by DeepSeek's efficiency demonstration is whether the $660–690B annual hyperscaler capex cycle — committed by Microsoft, Google, Amazon, and Meta through 2026–2028 — can be justified if training costs are falling this rapidly.
PRZC's assessment: the capex supercycle is not broken by DeepSeek's efficiency gains, but it is restructured by them. The Jevons dynamic is operating: lower per-capability costs expand the total market for AI deployment, which increases aggregate compute demand. However, the composition of that demand is changing:
Mid-tier closed-source inference providers — any company whose value proposition is serving API calls at GPT-3.5/4 capability levels without frontier differentiation. DeepSeek-V3 and Qwen 2.5-Max are open-source alternatives at this capability tier. The inference margin is structurally zero in this segment.
AI wrapper SaaS companies — businesses that have built on top of foundation model APIs without proprietary data assets, defensible workflow integrations, or network effects. These companies face a double squeeze: API costs may not fall as fast as their competitors' self-hosted alternatives, and their "AI-powered" differentiation is replicable.
Nvidia inference-oriented SKUs — not Nvidia overall, but specifically the inference-optimised products (L40S, H100 NVL, future Blackwell inference SKUs) face slower refresh cycles as efficiency improvements reduce required chip density for a given query volume.
Frontier training labs with safety moats (Anthropic, Google DeepMind) — the combination of frontier capability, interpretability research, government trust certifications, and enterprise-grade safety architecture cannot be commoditised by open-source releases. Chinese competition makes these moats more valuable, not less, by clarifying what the enterprise trust premium is actually paying for.
Nvidia at the training frontier — Blackwell and Rubin architectures continue to advance faster than SMIC/Huawei alternatives can close the gap. The export control regime, even in its imperfect form, maintains a training compute advantage for US-accessible labs. Nvidia's pricing power at the frontier training tier is structurally intact.
TSMC — the single most irreplaceable node in the global AI hardware supply chain. Every AI training workload — US, Chinese, or otherwise — runs on TSMC-manufactured silicon at the leading edge. Chinese chip development (Huawei, SMIC) is accelerating but remains dependent on TSMC indirectly (through the equipment supply chains that TSMC's processes define). TSMC's moat is geopolitically fragile but commercially unassailable in the 5-year investment horizon.
Enterprise AI platforms with proprietary workflow integrations — Microsoft Copilot's advantage is not the underlying model; it is the integration with Excel, Teams, Outlook, SharePoint, and Active Directory identity infrastructure. Chinese models cannot replicate this through open weights. The enterprise AI platform layer benefits from commoditised underlying models (lower API costs) while maintaining defensible integration moats above the model layer.
AI cybersecurity — DeepSeek's open-weight release has already been used by threat actors to build uncensored jailbroken variants. The expansion of the open-source model ecosystem increases the attack surface for AI-enabled threat vectors. Defence-oriented AI security companies — detecting AI-generated content, defending against AI-assisted phishing and social engineering, and managing AI agent security — have a structural demand tailwind regardless of which lab's model is used.
Not all AI companies face equal commoditisation risk. The relevant dimension is distance from the foundation model layer:
| Layer | Examples | China Competition Risk | PRZC Stance |
|---|---|---|---|
| Training frontier models | Anthropic, OpenAI, Google DeepMind | Low-Medium (benchmark parity possible; trust/safety moat protects enterprise) | OVERWEIGHT |
| AI hardware (frontier training) | Nvidia (H200/Blackwell/Rubin) | Low (export controls + TSMC dependency protect near-term) | NEUTRAL-OVERWEIGHT |
| AI hardware (inference) | Nvidia inference SKUs, AMD Instinct | Medium (efficiency gains compress refresh cycle) | NEUTRAL |
| Foundry / Advanced packaging | TSMC, ASE, Amkor | Low (SMIC 3–5 years behind; packaging bottleneck) | OVERWEIGHT |
| Enterprise AI platforms | Microsoft Copilot, Salesforce Einstein | Low (workflow integration moat, enterprise trust) | OVERWEIGHT |
| Mid-tier inference APIs | Undifferentiated API providers | High (DeepSeek-V3 open weights compete directly) | UNDERWEIGHT |
| AI wrapper SaaS | Various consumer/SMB AI tools | High (no proprietary data or integration moat) | UNDERWEIGHT |
Every global AI thesis — US, Chinese, European — runs through the Taiwan Semiconductor Manufacturing Company. TSMC manufactures the logic chips that power every frontier AI training cluster globally. Nvidia's Hopper, Blackwell, and Rubin architectures are fabricated at TSMC's 4nm and 3nm nodes. Apple's M-series chips. AMD's MI300X. Google's TPUv5. Qualcomm's Oryon. As of mid-2024, virtually all of this production ran through TSMC's fabs in Hsinchu Science Park and Tainan — both in Taiwan.
A common misconception is that TSMC's Chinese fabs represent meaningful AI supply chain exposure. They do not. TSMC operates two China facilities: Fab 16 in Nanjing (16nm/28nm, ~60,000 wafers/month) and Fab 10 in Shanghai (180nm legacy, 8-inch). Neither node is capable of producing competitive AI accelerators, and both are legally prohibited from supplying Chinese AI customers with advanced chips under successive BIS export control rounds (October 2022, October 2023, and the November 2024 directive halting all sub-7nm AI chip shipments to Chinese entities). The Nanjing fab's VEU authorisation was revoked effective December 31, 2025 and replaced by an annual licence regime. China's share of TSMC revenue fell from ~22% at peak to 11% in 2024 ($9.9B of $90.1B total) — a direct consequence of these controls. The Sophgo incident (2024), in which ~3 million 7nm dies ordered by a Chinese fabless firm were diverted to Huawei's Ascend 910B programme, involved TSMC's Taiwan fabs, not its China facilities.
Geographic concentration in Taiwan remains the primary structural risk, but it is not static. TSMC's Arizona Fab 1 began N4 high-volume production in Q4 2024 — the first leading-edge AI-capable production outside Taiwan. TSMC Kumamoto (Japan) commenced 12/16nm production in December 2024 at 55,000 wafers/month, with a second fab targeting 6/7nm by end-2027. TSMC Dresden (Germany, 28/16nm, ~40,000 wafers/month) is on track for ~2027. TSMC has announced $165B of Arizona investment covering six fabs through the early 2030s. Taiwan will remain the dominant site for the most advanced nodes (N2 and below) through at least 2028, but the concentration risk narrative is measurably less acute than it was in 2022–2023.
China's stated position on Taiwan has not changed. The PLA has conducted increasingly realistic military exercises simulating a blockade. A blockade scenario, even without a full invasion, would disrupt global AI hardware supply for a period measured in years, not months — but the disruption would fall disproportionately on N3 and below nodes, which remain Taiwan-only. Arizona and Kumamoto provide partial buffer for automotive, consumer, and some HPC-adjacent applications at 4nm and above.
The market continues to price Taiwan risk as a tail scenario with a probability in the 5–10% range on a 10-year horizon, based on analyst surveys. PRZC's assessment is that this probability is directionally correct but that the impact at realisation is systematically underweighted: a Taiwan disruption does not merely slow AI development — it potentially resets the global hardware stack for a multi-year period. There is no hedge against this scenario for AI infrastructure long positions; there is only the option to reduce exposure to companies with higher Taiwan dependency and accept lower expected returns in the base case.
US-China technology decoupling in AI has two distinct and partially contradictory dynamics:
Hardware decoupling (accelerating): The semiconductor supply chain is bifurcating. US fab capacity is expanding (CHIPS Act subsidies to Intel, TSMC Arizona, Samsung Texas). Chinese fab capacity is expanding independently (SMIC expansion, Huawei foundry investments). The two chains are increasingly distinct, though they share equipment suppliers at several critical nodes (ASML DUV, Applied Materials CVD tools). Full decoupling of the silicon supply chain will take 10–15 years and will not be completed before the current AI investment cycle resolves.
Software and model decoupling (failing): Open-source AI model weights cross borders in seconds. DeepSeek-R1 was downloaded from Hugging Face hundreds of thousands of times in the first 48 hours after release. No export control regime can prevent the transfer of a mathematical artifact encoded as floating-point numbers. The US government's attempts to control the diffusion of AI model weights — including the Biden-era "AI Diffusion Rule" and proposed weight security frameworks — have no effective enforcement mechanism at the software layer. Chinese model capabilities will continue to be available to US developers, and US open-source models (LLaMA, Mistral, Gemma) will continue to be available to Chinese developers. Software decoupling is aspirational, not operational.
DeepSeek-R1's release altered the national security community's framing of AI risk in a significant way. Prior to January 2025, the dominant framing was: "China is falling behind in AI due to compute restrictions; the US maintains a decisive lead." Post-DeepSeek, the framing has shifted to: "China has achieved frontier-competitive AI capability with fewer resources than expected; the compute restriction approach alone is insufficient."
This shift has concrete policy implications. The US intelligence community's 2025 annual threat assessment elevated AI-specific language around Chinese language model capabilities. Congressional pressure for broader export controls — including restrictions on cloud computing access to Chinese entities, and controls on AI training data transfers — has intensified. The Trump administration's approach has been more transactional (using AI export controls as a bargaining chip in broader trade negotiations) than the Biden approach (which treated AI controls as a strategic priority in their own right), creating policy uncertainty for hardware manufacturers with significant China revenue exposure.
For investment portfolios, the national security framing has one clear implication: Anthropic's government and intelligence community contract exposure is a structural advantage, not merely a revenue diversification. As AI capability becomes a national security priority, the labs with established FedRAMP authorisation, cleared personnel, and auditable alignment architectures will capture a disproportionate share of the government spending that the security framing unlocks. This is not priced in Anthropic's current valuation at any reasonable public comparables analysis.
PRZC's analytical discipline requires explicit articulation of what would invalidate the above conclusions. The primary disconfirmation conditions for our OVERWEIGHT themes:
DeepSeek's efficiency shock is real, its implications are consequential, and the market's initial reaction — a panicked sell-off in Nvidia and a brief narrative triumph for "China wins AI" — was predictably wrong in both directions. The correct framing is more nuanced and ultimately more useful for investment positioning:
China's AI ecosystem has compressed the capability gap with US frontier labs to a degree that requires serious strategic reassessment. The compute disadvantage imposed by export controls is real but insufficient to prevent frontier-competitive capability development. The algorithmic innovation demonstrated by DeepSeek, Qwen, and the broader Chinese research ecosystem is substantive, not borrowed, and reflects a genuine engineering culture operating effectively under resource constraints.
At the same time, the factors that matter most at the highest-value segment of the enterprise AI market — alignment architecture, interpretability, government certification, and the trust infrastructure built around them — are not reducible to benchmark scores and cannot be replicated through open-weight releases. Anthropic's position in this segment is structurally more defensible than is reflected in the standard framing of Chinese AI as a symmetric competitive threat.
The investment implications favour concentration in frontier training capability and enterprise trust infrastructure, caution on inference-layer businesses without workflow moats, and heightened awareness of Taiwan as the single point of failure in the entire global AI hardware chain — a risk that no amount of benchmark analysis or competitive positioning can mitigate if it materialises.
| Theme / Asset Class | Stance | Primary Driver | Key Risk |
|---|---|---|---|
| Frontier training labs (Anthropic, Google DeepMind) | OVERWEIGHT | Safety moat, enterprise trust, interpretability | Chinese benchmark parity on SWE-Bench |
| Nvidia (training tier) | NEUTRAL / OVERWEIGHT | Scaling law demand, export control hardware gap | Efficiency gains slow refresh cycles |
| TSMC | OVERWEIGHT | Irreplaceable node; no substitute on horizon | Taiwan geopolitical scenario |
| Enterprise AI platforms (MSFT, CRM) | OVERWEIGHT | Workflow integration moat above model layer | Open-source model self-hosting adoption |
| Mid-tier inference API providers | UNDERWEIGHT | DeepSeek/Qwen open weights eliminate margin | None identified that changes the thesis |
| AI wrapper SaaS (no proprietary data) | UNDERWEIGHT | Commoditised differentiation layer | Vertical-specific network effects could persist |
| AI cybersecurity | OVERWEIGHT | Open-weight expansion increases attack surface | Regulatory pace risk |
This report was prepared by PRZC Research. It is provided for informational and educational purposes only. Nothing in this report constitutes investment advice, a recommendation to buy or sell any security, or an offer or solicitation of any kind. The views expressed reflect the analyst's assessment as of the publication date (March 30, 2026) and are subject to change without notice.
Benchmark data for Chinese AI models (DeepSeek, Qwen, Kimi, GLM) reflects results as reported in published technical papers, official model cards, and credible independent evaluations available at time of writing. Direct benchmark comparisons across models evaluated with different harnesses and prompt formats are inherently imperfect; readers should consult primary sources and up-to-date leaderboards (LMSYS, HuggingFace Open LLM Leaderboard) for current standings. The AI benchmark landscape changes rapidly; figures in this report should not be assumed current beyond the publication date.
PRZC Research is incorporated in the Republic of Seychelles. It does not operate as a regulated investment adviser in the United Kingdom, European Union, United States, or any other jurisdiction requiring registration. This report does not represent research produced by a regulated firm and should not be relied upon as regulated investment research. Past research track records referenced in PRZC publications are provided for transparency; past analytical performance does not guarantee future accuracy.
PRZC Research and its analysts may hold positions in securities discussed in this report. No position disclosures are made with respect to private companies (Anthropic, DeepSeek/High-Flyer, Moonshot AI) as there is no public market for these securities.
PRZC Research | przc.re | March 30, 2026