Why Did the Company America Sanctioned Only Get Stronger? — The Sanctions Paradox GLM-5 Just Proved
Summary
America's Entity List tried to strangle Chinese AI firm Zhipu, but two years later the company surged 80% on the Hong Kong Stock Exchange and dropped a frontier open-source model. GLM-5, trained exclusively on Huawei chips with 744 billion parameters, slashed Silicon Valley's price tag by six-fold and slapped an MIT license on top. Did sanctions actually breed a monster?
Key Points
The Sanctions Paradox — Blockade Accelerated Self-Reliance
The U.S. Entity List cut off Zhipu's NVIDIA chip imports, but this triggered a pivot to Huawei's Ascend chip ecosystem. China's semiconductor-software-AI vertical integration accelerated, and following DeepSeek V3.2, GLM-5 achieved frontier-level performance, raising questions about the long-term efficacy of sanctions. Historically, the Soviet space program and Iran's military industry followed identical patterns under technology blockades.
The Open-Proprietary Gap Has Vanished
The MMLU benchmark gap between open-source and proprietary models collapsed from 17.5 points to just 0.3. GLM-5 scored 77.8% on SWE-bench, surpassing GPT-5.3 Codex (77.3%) and approaching Claude Opus 4.6 (79.4%). Combined with MIT licensing and 6x lower pricing, the business case for proprietary models is weakening significantly.
AI Three Kingdoms — The End of One Best AI
Within 6 days in February 2026, Claude Opus 4.6, GPT-5.3 Codex, and GLM-5 launched consecutively, splitting benchmarks three ways. Claude leads SWE-bench, GPT dominates Terminal-Bench, and GLM wins on cost efficiency. The concept of a single best model has evaporated, and enterprises are moving toward multi-model strategies.
6x Price Disruption Shaking AI Business Models
GLM-5's input token price of $0.80 per million is over 6x cheaper than Claude Opus 4.6's $5.00. DeepSeek V3.2 also offers 85% savings versus GPT-5.1. This pricing pressure forced Anthropic to launch Claude Sonnet 4.6 at a 40% discount, and frontier API prices may halve by late 2026.
The Dawn of AI Sovereignty
Training a frontier model solely on Huawei chips is a geopolitical statement beyond a technical achievement. 93% of executives say AI sovereignty must factor into business strategy in 2026. For countries between the U.S. and China — India, Middle East, Southeast Asia — an alternative AI infrastructure has been confirmed. The shift toward a multipolar AI chip ecosystem has begun.
Positive & Negative Analysis
Positive Aspects
- Democratization of AI Access
The emergence of a frontier-grade MIT-licensed model means startups in Africa, universities in Southeast Asia, and SMEs in Europe can freely leverage cutting-edge AI. A 6x drop in API costs is a game-changer for budget-constrained organizations, reopening AI projects previously abandoned due to cost.
- Healthy Price Competition
Pricing pressure from GLM-5 and DeepSeek is driving healthy market repricing. Anthropic launching Claude Sonnet 4.6 at 40% lower cost is a prime example. Both consumers and enterprise users benefit from this competitive structure.
- Multipolarization of AI Infrastructure
NVIDIA GPU monopoly was AI's hidden vulnerability. Huawei Ascend's success in training frontier models proves infrastructure diversification is possible. Interest and investment in AMD, Intel, and national chip projects are rising.
- Accelerated Innovation Through Competition
Three frontier models launching within 6 days created a healthy competitive landscape. Competition drives faster progress than monopoly, with tangible quality improvements like hallucination reduction from 90% to 34%.
Concerns
- AI Safety Governance Fragmentation
Frontier-grade models under MIT license can be used without safety guardrails. While OpenAI and Anthropic invest in safety research, open-source models allow anyone to strip safety measures, creating fundamental tension between democratization and safety.
- Benchmark Overestimation Risk
Similar SWE-bench and MMLU scores don't guarantee equivalent real-world performance. GLM-5 is only 2 weeks old with limited enterprise deployment data. Multiple 2025 models with strong benchmarks underperformed in production.
- Deepening Tech Bloc Formation
U.S. sanctions have accelerated tech decoupling, potentially forming an AI Curtain splitting U.S.-centric and China-centric ecosystems. Different chips, frameworks, and standards would cause compatibility issues, research collaboration breakdowns, and slower innovation.
- Commercial Sustainability of Open-Source AI
MIT-licensed free distribution is strategic but long-term revenue models remain unclear. Frontier model training costs hundreds of millions per generation. Government subsidy dependence could lead to politically-driven rather than market-driven innovation.
Outlook
Short-term (6 months to 1 year): Price war between open-source and proprietary AI intensifies, with frontier API prices potentially dropping to half current levels by late 2026. Medium-term (1-3 years): The concept of one best AI model disappears entirely as enterprises adopt multi-model strategies. NVIDIA's dominance weakens as AI infrastructure multipolarizes. Long-term (3-5+ years): AI's Linux moment may materialize — open-source becomes the frontier default, proprietary models reposition as premium enterprise options. Worst case is complete U.S.-China AI ecosystem severing, but reality will likely settle somewhere in the middle.
Sources / References
- GLM-5 vs Claude Opus 4.6 vs GPT-5.3 Codex: Three-Way Comparison — AI Free API
- GLM-5 Achieves Record Low Hallucination Rate — VentureBeat
- World's Strongest Open-Source LLM Trained Solely on Huawei Chips — Trending Topics EU
- Zhipu Releases GLM-5 Under MIT License — The Decoder
- Opus 4.6, Codex 5.3, and the Post-Benchmark Era — Interconnects
- The February Reset: Three Labs, Four Models, and the End of One Best AI — Medium / Data Science Collective
- 2026 Is the Year Open-Source AI Reaches Frontier — Swfte AI