Efficient Training: MoE, Distillation, Compute Trends — 2026-09-10
The U.S. government issued a major advisory naming six Chinese AI firms for malicious distillation campaigns, signaling a shift toward legal and economic responses to IP theft in AI training. Meanwhile, DeepSeek announced significant price cuts for its Flash models, intensifying the cost war in inference and potentially altering the economics of model deployment. <!-- /headline --> **US Names Six Firms in AI Distillation Crackdown as DeepSeek Slashes Prices** <!-- /headline -->
Efficient Training: MoE, Distillation, Compute Trends — 2026-09-10
The U.S. government issued a major advisory naming six Chinese AI firms for malicious distillation campaigns, signaling a shift toward legal and economic responses to IP theft in AI training. Meanwhile, DeepSeek announced significant price cuts for its Flash models, intensifying the cost war in inference and potentially altering the economics of model deployment.
<!-- /headline -->US Names Six Firms in AI Distillation Crackdown as DeepSeek Slashes Prices
<!-- /headline -->Top developments
US Agencies Warn of Industrial-Scale AI Distillation by Chinese Firms
On September 9, 2026, the NSA, CISA, and FBI released Advisory AA26-251A, explicitly naming six AI companies (including DeepSeek and Moonshot AI) for engaging in "industrialized extraction" of capabilities from American models. The advisory details how these firms use API abuse and synthetic data generation to distill proprietary capabilities without authorization. This marks a significant escalation from previous warnings, providing concrete guidance for model buyers on provenance checks and API monitoring to mitigate the risk of using compromised models.
DeepSeek Launches Aggressive Price Cuts for Flash Models
On September 10, 2026, DeepSeek officially launched a massive price reduction for its Flash series models, with idle-time calls dropping to as low as 1 yuan per million tokens. This move is widely interpreted as a strategic effort to dominate the low-cost inference market, forcing competitors to reconsider their pricing structures. The reduction highlights the growing divergence between frontier model training costs and the commoditization of efficient, smaller-scale model serving.
AI Spend Per Employee Slumps at Top Firms
Data released on September 9, 2026, indicates that AI spending per employee at top technology firms slumped in August 2026. This trend is attributed to falling token costs, the adoption of cheaper models, and a potential plateau in enterprise AI adoption rates. For training-efficiency researchers, this signals a market shift where cost efficiency and model size optimization are becoming more critical than raw capability gains alone.
Local view
Local Chinese media outlets reported extensively on the U.S. advisory, with sources like Bannedbook and Sohu framing the U.S. actions as an attempt to suppress Chinese AI innovation through "industrialized extraction" accusations. Simultaneously, domestic coverage of DeepSeek’s price cuts emphasized the competitive advantage this provides Chinese developers against international providers, highlighting a narrative of technological self-reliance and cost dominance.
Context & numbers
- Training Cost Trends: Epoch AI data indicates that training compute costs for the largest AI models continue to double approximately every eight months, with spending growing at a rate of 2.4x per year.
- Frontier Model Costs: Frontier pretraining budgets crossed the $500 million mark in 2025, with estimates suggesting costs will reach $1–3 billion per model by 2027.
- Distillation Efficiency: Recent research suggests that post-trained Mixture-of-Experts (MoE) models can skip up to 50% of expert FLOPs via self-distillation techniques with marginal accuracy loss, offering a potential path to reduced inference costs.
On the radar
- Regulatory Response: Watch for potential legislative or executive actions in the U.S. following the AA26-251A advisory, particularly regarding export controls on API access to frontier models.
- Competitor Pricing: Other major labs (OpenAI, Anthropic, Google) may announce price adjustments in response to DeepSeek’s September 10 cuts.
This content was collected, curated, and summarized entirely by AI — including how and what to gather. It may contain inaccuracies. Crew does not guarantee the accuracy of any information presented here. Always verify facts on your own before acting on them. Crew assumes no legal liability for any consequences arising from reliance on this content.
Powered by
