Direct Answer
China has not conclusively "beaten America at AI," and its strongest models are not free to operate. What Chinese labs have done is strategically important: they have made high-capability models available at low hosted prices or through open-weight releases, then used that distribution to attract developers, cloud workloads, tooling, and research.
Export controls constrained access to advanced chips and manufacturing equipment. They did not stop model progress. The response combined efficient architectures, domestic semiconductor investment, large infrastructure programs, aggressive pricing, open-weight distribution, and a willingness to optimize for the tasks where "good enough and much cheaper" wins.
Video credit: AI educator and entrepreneur Vaibhav Sisinty. Watch the original video and follow Vaibhav on X.
Evidence, Scope, and Credit
The supplied transcript is the narrative spine for this article. Its economic and geopolitical claims are checked against the U.S. Bureau of Industry and Security, Chinese government planning documents, official model releases, the DeepSeek-V3 technical report, model repositories, company engineering posts, current privacy documentation, and reporting about OpenRouter traffic.
Four labels matter throughout: official fact, company claim, third-party measurement, and inference. Model launch charts are company claims. OpenRouter traffic is a third-party sample. "China's master plan" is an interpretation, not a document proving one coordinated commercial strategy across every Chinese lab.
The Headline Claim Check
| Claim | Verdict | What the evidence supports |
|---|---|---|
| Chinese models handle 60% of AI used by U.S. companies. | Misleading | Roughly 60% of U.S.-company token usage observed on OpenRouter, not all enterprise AI usage. |
| U.S. chip controls completely backfired. | Not established | The controls restricted selected chips, equipment, software, and support. Chinese progress continued, but the counterfactual cannot be observed. |
| DeepSeek trained a frontier model for under $6 million. | Incomplete | About $5.576 million was the report's estimate for the final V3 training run, excluding earlier research, experiments, salaries, facilities, hardware ownership, and data work. |
| Kimi K3, Qwen3.8, GLM, and DeepSeek are all free and open source. | Mixed | GLM-5.2 publishes weights under MIT. Several DeepSeek and Qwen families publish weights. K3 weights are promised for 27 July; Qwen3.8-Max is a hosted preview with open weights promised. Inference and operations still cost money. |
| Chinese models beat GPT and Claude. | Task-specific | They lead or compete on selected benchmarks, prices, and workloads. No benchmark proves a universal winner across reliability, safety, coding, writing, multimodality, and operations. |
| The Mate 60 Pro used Huawei Ascend AI chips. | Incorrect | TechInsights identified a Kirin 9000S smartphone system-on-chip fabricated on SMIC's 7 nm N+2 process. Huawei Ascend is a separate AI-accelerator line. |
| Chinese platforms provide a secret training-data advantage. | Speculative | China has enormous consumer and industrial data ecosystems. The exact training data, permissions, and causal advantage of individual models are generally not public. |
| Self-hosted weights automatically make AI private. | False | Self-hosting removes one external inference provider, but privacy still depends on telemetry, logs, network paths, dependencies, access controls, and operator security. |
The Model Map: What Is Actually Available?
| Family | Status on 24 Jul 2026 | What is genuinely open | Practical note |
|---|---|---|---|
| DeepSeek | Mature open-weight family plus hosted services. | Multiple checkpoints, code, and technical reports are public under their stated licenses. | Strong case study in hardware-aware architecture and low-cost inference, but every checkpoint and provider needs its own review. |
| Qwen | Large open-weight ecosystem; Qwen3.8-Max-Preview is hosted now. | Many Qwen checkpoints and tools are downloadable. Alibaba says Qwen3.8 open weights are coming. | Do not confuse the broad open Qwen ecosystem with the current 3.8-Max preview. |
| GLM-5.2 | Weights published through Z.ai and Hugging Face. | The model card lists a 753B-parameter checkpoint under the MIT License. | A free license still leaves very large inference, serving, and security costs. |
| Kimi K3 | Hosted chat, agent, coding, and API access are live. | Moonshot promises full weights by 27 July 2026. | Until the files and license are inspectable, evaluate K3 as a hosted model and describe its open status carefully. |
"Open source" is often used loosely in model marketing. For procurement, split it into five questions: are the weights downloadable, is the license commercially usable, is training code available, are training data disclosed, and can the model be independently reproduced? Most releases satisfy only part of that stack.
What "Free" Actually Costs
A model license can cost zero while the system around it remains expensive. The right comparison is total cost per accepted result.
| Cost layer | Hosted API | Self-hosted weights |
|---|---|---|
| Model access | Input, output, cache, tool, or subscription charges. | Possibly zero license fee, subject to the license. |
| Compute | Included in the provider's rate. | GPU rental or purchase, power, cooling, storage, networking, and spare capacity. |
| Operations | Provider handles the serving fleet. | Deployment, quantization, batching, scaling, upgrades, monitoring, and incidents. |
| Governance | Provider contracts, retention, residency, subprocessors, and audit controls. | Internal access, supply-chain review, provenance, logs, and policy enforcement. |
| Quality | Evaluation, retries, human review, and fallback models. | The same, plus runtime tuning and regression testing. |
accepted-result cost =
inference or infrastructure
+ engineering and operations
+ retries and fallbacks
+ human review and repair
+ security and complianceOpen weights become attractive when utilization is high, workloads are stable, the team can operate the stack, and domain tuning creates a measurable advantage. For intermittent workloads, a hosted API can remain cheaper even when its token price looks higher.
What Export Controls Changed
The U.S. controls introduced in October 2022 were broader and more precise than "China was banned from Nvidia." The Bureau of Industry and Security restricted selected advanced-computing items, semiconductor-manufacturing equipment, supercomputer end uses, and certain support by U.S. persons. The rules were updated later and licensing policy continues to evolve.
- The controls created real constraints. Access to frontier accelerators, manufacturing tools, software, and service became harder, more expensive, or legally restricted.
- The constraints did not freeze Chinese AI development. Labs optimized architectures, acquired allowed hardware, invested in domestic chips, expanded data-center capacity, and shipped competitive models.
Saying the controls "did not even slow China down" goes beyond the evidence. Progress under constraint is visible; the unobserved no-control timeline is not.
DeepSeek's Real Efficiency Lesson
DeepSeek-V3 matters because it combined several engineering choices into a strong cost-performance result. Its technical report describes a mixture-of-experts architecture, Multi-head Latent Attention, load balancing without auxiliary loss, multi-token prediction, FP8 training, and hardware-aware co-design.
The model has 671 billion total parameters and activates 37 billion for each token. Its sparse design means the full network does not perform equal work on every token. MLA reduces the memory burden associated with attention and long contexts. These are meaningful gains, not magic that removes compute.
Infrastructure Plus Distribution Is the Strategy
China's 2017 New Generation Artificial Intelligence Development Plan treated AI as a national and industrial priority. The 14th Five-Year Plan also called for national data and computing hubs. That does not mean every company follows one centrally scripted product plan, but it creates long-term alignment around chips, data centers, research, talent, cloud services, and adoption.
- Release capable weights or price hosted access aggressively.
- Developers integrate the model into tools, agents, and internal systems.
- Compatibility, fine-tunes, tutorials, and community knowledge accumulate.
- High-volume users buy cloud inference, managed deployment, hardware, support, or adjacent services.
- Usage evidence and ecosystem feedback improve the next release.
Alibaba is the clearest example: Qwen sits inside a much larger cloud, developer-tool, commerce, and enterprise platform. The model can be a distribution layer for the ecosystem even when a checkpoint is free to download.
What the 60% Adoption Number Really Tells Us
Reporting published on 24 July says Chinese-developed models accounted for roughly 60% of token usage attributed to U.S. companies on OpenRouter. That is striking. It is also narrower than the viral version.
OpenRouter is a multi-model routing marketplace used heavily by developers and AI-native teams. Its data does not include every direct OpenAI, Anthropic, Google, AWS, Azure, Alibaba, or self-hosted deployment. The share therefore shows a sharp preference inside one important routing market, especially where price and switching matter. It does not establish a 60% share of all U.S. corporate AI.
Company evidence supports the direction without proving the universal number. Pinterest says internally hosted, fine-tuned open models can deliver similar performance at less than 10% of the cost of leading proprietary models for selected workloads. It also says Pinterest uses a mix of internal, open, and third-party proprietary models. That mixed stack is the useful pattern.
Privacy Depends on the Deployment Path
| Deployment | Who may process the data? | Main risk | Best use |
|---|---|---|---|
| Consumer chat product | The product provider under consumer terms and settings. | Training settings, human review, retention, memory, and connected apps. | Public or low-sensitivity experiments. |
| Commercial API | The API provider and documented subprocessors. | Retention, region, logs, account compromise, and contract gaps. | Production workloads with a DPA and restricted data. |
| Third-party host for open weights | The inference host, its cloud, and subprocessors. | The model may be open while the hosting path is opaque. | Teams that need easier deployment and can validate the host. |
| Self-hosted open weights | Your organization and infrastructure providers. | Misconfiguration, telemetry, vulnerable dependencies, provenance, and weak patching. | Stable, high-volume, sensitive work with capable operations. |
The transcript's comparison with U.S. consumer assistants needs nuance. OpenAI says individual ChatGPT and Codex content may be used for training unless the user opts out; business and API data are not used for training by default. Google ties Gemini improvement and human review to product and activity settings, with separate temporary-chat and business controls. Anthropic lets consumer users choose whether chats improve Claude; five-year de-identified retention applies when that option is enabled, while commercial terms differ.
The safe rule is not "Western closed model bad, Chinese open model private" or the reverse. Read the exact plan, provider, region, and settings. Keep confidential data out until the contract, architecture, retention, access, and deletion path are understood.
A Practical Buyer Checklist
- Name the exact artifact: model ID, checkpoint hash, quantization, runtime, provider, and region.
- Read the license: commercial use, modification, redistribution, acceptable-use terms, and attribution.
- Map the data path: prompts, files, tool outputs, caches, logs, backups, telemetry, and support access.
- Check jurisdiction: residency, transfers, subprocessors, DPA, deletion, and breach terms.
- Run behavioral evals: quality, refusals, political sensitivity, factuality, citations, prompt injection, and multilingual behavior.
- Measure accepted-result economics: cache hits, tokens, latency, retries, reviewer minutes, repair work, and uptime.
- Keep a second route: fallback provider, exportable prompts, portable tools, and no undocumented vendor-only dependency.
- Limit autonomy: sandbox tools and require approval for spending, sending, publishing, deleting, credentials, and production changes.
What Other Countries Should Learn
The policy lesson is not that every country must train the world's largest chatbot. It is that durable AI capability is a stack:
- Reliable energy, data centers, networks, and access to compute.
- Research institutions, engineering talent, and open technical education.
- Local-language data governed by clear rights and privacy rules.
- Public procurement that rewards deployed outcomes, not demo days.
- Cloud and inference markets that let startups test without owning a data center.
- Open standards, portable applications, and enough model diversity to avoid one-vendor dependence.
For India, Europe, and other large markets, the practical target is selective sovereignty: control high-value data, retain deployment options, support local-language and domain models, and buy global frontier intelligence where it is genuinely better. Full self-sufficiency is not required to avoid full dependence.
Video Chapters
| Time | Topic | Editorial note |
|---|---|---|
| 00:00 | China just beat America at AI | A provocative thesis, not a settled measurement. |
| 02:25 | Why everyone thought China would lose | The supply chain is more complex than three companies controlled by one country. |
| 04:11 | The U.S. chip controls | The rules target selected chips, equipment, end uses, and support. |
| 05:14 | Domestic chips | The Mate 60 Pro used the Kirin 9000S, not an Ascend accelerator. |
| 06:22 | DeepSeek efficiency | MoE and MLA matter; the $6M figure is final-run compute. |
| 08:41 | The memory trick | MLA compresses attention state; it does not summarize a document in the ordinary sense. |
| 10:23 | Data-center strategy | National computing hubs are a real policy direction, still under construction. |
| 12:16 | Data advantage | Plausible at ecosystem level; exact training data remain undisclosed. |
| 13:55 | Why models are free | Separate zero-price weights from paid inference and operations. |
| 16:04 | Distribution | The strongest section: low-cost models pull workloads into clouds and tools. |
| 19:01 | National AI mission | Policy coordination exists, but firms retain distinct incentives. |
| 20:26 | Distillation and smuggling claims | Separate documented cases, allegations, and normal technical learning. |
| 21:20 | Data privacy | Compare deployment routes and contracts, not nationality alone. |
| 21:46 | Sovereign AI | Prioritize control over data, infrastructure options, skills, and portability. |
| 23:29 | The ecosystem lesson | The most durable takeaway from the video. |
Bottom Line
Chinese AI labs are forcing the market to separate frontier prestige from production value. A model does not need to win every benchmark to win millions of routed tasks. It needs to be capable enough, cheap enough, available enough, and easy enough to integrate.
Treat "free" as a distribution strategy, not a total-cost claim. Treat OpenRouter's 60% as a market signal, not national market share. Treat export controls as a constraint with adaptation, not a simple victory or failure. And treat every model - Chinese, American, open, or closed - as one component in a governed stack that can be tested, routed, replaced, and audited.
Sources
- Vaibhav Sisinty: The Real Reason China's Best AI Is Free
- U.S. Bureau of Industry and Security: October 2022 controls
- DeepSeek-V3 Technical Report
- Moonshot AI: Kimi K3 technical release
- Alibaba Group: Qwen3.8-Max-Preview announcement
- Z.ai: GLM-5.2 model card and weights
- eWeek: Chinese-model share on OpenRouter
- Pinterest Engineering: On the reprioritization of open-source AI
- TechInsights: SMIC 7 nm N+2 in Huawei Mate 60 Pro
- State Council of China: 2017 AI development plan
- NDRC: 14th Five-Year Plan and long-range objectives
- OpenAI: How data is used to improve model performance
- Google: Gemini Apps Privacy Hub
- Anthropic: Consumer model-improvement and retention controls