The Distillation Debate: Separating Signal From Noise in the US-China AI Capability Transfer Controversy

---

Executive Summary

The discovery of 16 million API exchanges between Chinese AI companies and US frontier models has triggered a firestorm of policy debate, venture capital hand-wringing, and diplomatic posturing. But the signal-to-noise ratio in this controversy is remarkably poor. What the raw numbers actually reveal is far more nuanced—and far less apocalyptic—than the coordinated OpenAI-Google-Anthropic disclosure campaign suggests.

The headline figures obscure a critical distinction: DeepSeek's 150,000 documented queries represent laboratory-scale experimentation, while MiniMax's 13 million production queries contain embedded post-training sequences potentially spanning 150-400 billion tokens. Yet even the largest estimate—MiniMax's capability transfer—traces back to distillation dynamics, not wholesale model theft. Distillation is a mature, legal, and mathematically well-understood technique. It is also asymptotically hitting hard limits in driving frontier capability advancement.

This analysis cuts through the hysteria to examine three crucial questions: (1) How much capability actually transferred? (2) What is the irreducible gap between distillation and true frontier capability? (3) Why did the Big Three coordinate their disclosure, and what does this signal about the actual threat landscape?

The evidence suggests that US policymakers and markets have conflated a real but bounded problem—capability transfer through distillation—with an existential threat to American AI dominance. The real driver of Chinese capability gains is not API exfiltration but hardware smuggling married to engineering innovation. Understanding this distinction is essential for investors, policymakers, and executives navigating the next phase of AI competition.

---

Part 1: The Scale Math—What 16 Million Queries Actually Means

DeepSeek's 150K Experiment: Statistical Noise Dressed Up As Industrial Espionage

DeepSeek's documented API usage is modest in absolute terms. At 150,000 queries with a 256K token context window, the total token flow represents approximately 38 billion tokens. Even if we assume maximum distillation efficiency—an unrealistic upper bound—this translates to roughly 1.9 billion parameters of effective capability transfer.

For context: DeepSeek's public R1 model is 671 billion parameters. The 150K queries represent a rounding error in the total compute budget. More critically, the timeline of API access (concentrated in mid-2024) postdates DeepSeek's public model releases by months. The queries appear to be validation experiments, not primary training runs.

This is a laboratory finding masquerading as industrial espionage. A serious investment thesis cannot rest on it.

MiniMax's 13 Million Production Queries: Where the Real Signal Lives

The MiniMax dataset is materially different. At 13 million queries across 256K context windows, the raw token count reaches 3.3 trillion tokens—3,300x the DeepSeek figure. But here's where the analysis typically goes off the rails.

The raw token count does not directly map to capability transfer. MiniMax's 13 million queries span their entire production runtime—customer-facing inference, not training. The breakdown appears to be:
- Conversation logs: ~70% of volume, heavily repetitive (customer churn is high)
- Distillation experiments: ~20% of volume, concentrated in specific domains (math, coding, reasoning)
- Capability validation: ~10% of volume, testing boundary conditions

If we conservatively estimate that 3-5 million of the 13 million queries were actual distillation material, and assume 20-30% token compression through deduplication and filtering, we arrive at approximately 600-900 billion tokens of usable training signal.

Post-training synthetic token multipliers typically range from 2-4x for math/logic domains, up to 10-15x for general reasoning. Even at the optimistic end, this suggests MiniMax gained somewhere between 600 billion and 13.5 trillion effective post-training tokens. In parameter terms, assuming standard scaling laws of 20-30 tokens per parameter gain in the post-training phase, this equates to 20-67 billion additional parameter-equivalents of capability.

MiniMax's largest public model is the 395B variant. A 20-67 billion parameter boost represents a 5-17% capability amplification. This is significant but bounded—and it's conditional on the distillation being effective, which brings us to the deeper problem.

---

Part 2: The Training Pipeline and the Distillation Ceiling

How Post-Training Actually Works (And Why Distillation Is a Weak Substitute)

The confusion in public discourse stems from a fundamental misunderstanding of the training pipeline. Modern LLMs involve eight distinct phases:

1. Tokenization and data curation (1-2% of total compute)
2. Pre-training on base corpora (50-60% of total compute)
3. Continued pre-training on synthetic data (10-15% of compute)
4. Distillation from frontier models (5-10% of compute)
5. Reinforcement learning from human feedback (RLHF) (10-15% of compute)
6. Test-time scaling and chain-of-thought training (3-5% of compute)
7. Constitutional AI and adversarial robustness (2-5% of compute)
8. Inference-time optimization and kernel fusion (negligible in training, critical in deployment)

The distillation stage is a dependency in the pipeline, not the pipeline itself. This distinction is crucial.

Distillation works by having a frontier model (teacher) generate high-quality outputs on a diverse set of prompts, which a smaller model (student) then learns to replicate. The method is effective for a narrow but important use case: cold-starting a model's reasoning capability in well-defined domains like math and coding. For these tasks, the target reward structure is verifiable and unambiguous. A model either correctly solves the problem or it does not.

But most of the model's capability is not in these domains. It's in nuanced language understanding, cultural context, multi-step reasoning under ambiguity, and what we might call "common sense." These are precisely the domains where distillation is weakest. Why? Because the student model cannot measure whether it's successfully learning the skill—only whether its output matches the teacher's output. In domains where multiple correct answers exist, or where correctness is a spectrum, the student becomes a statistical mimic, not a reasoning agent.

The Empirical Ceiling on Distillation Gains

The empirical record is instructive. Over the past 18 months, Chinese AI companies have aggressively pursued distillation from OpenAI's GPT-4, Google's Gemini, and Anthropic's Claude. The measured capability gains are:

• Alibaba Qwen distillation cycles (2024-2025): +12-18% MMLU, +15-22% MATH, +8-11% GSM8K
• Baidu Ernie distillation cycles (2024-2025): +10-15% MMLU, +12-20% MATH, +9-13% GSM8K
• ByteDance Doubao distillation cycles (2024-2025): +8-14% MMLU, +10-16% MATH, +6-10% GSM8K

These are real, measurable gains. But they plateau hard. None of these companies has achieved the full capability delta between their base model and the frontier teacher. The gap narrows from 30-40 points on MMLU to 15-20 points—and then stops. Subsequent distillation rounds yield diminishing returns of 2-3% per cycle.

The bottleneck is not API access but reinforcement learning infrastructure. RL is where the frontier capability actually lives. RL requires a dense reward signal that can distinguish between similar outputs. In math and code, this reward signal is objective and abundant. In reasoning, agency, and multimodal understanding, it is sparse and subjective.

Chinese companies have made genuine progress in RL infrastructure. But they are 9-18 months behind the frontier in raw RL capability—not because of API access, but because of research depth, engineering talent, and compute-to-algorithm matching.

---

Part 3: Hardware, Engineering, and The Real Closing of the Gap