Source: raw/US_AI_Dominance_Is_Over_-_Here_s_Why.md
Creator: Nate B Jones | Platform: YouTube | Feed: AI News & Strategy Daily | Nate B Jones | URL: https://www.youtube.com/watch?v=JBzz53HqMEs
Against the “China has caught up” narrative around Kimi K3 and Qwen 3.8, Jones argues that “Chinese models” is a broken shorthand — the labs share a country of origin, not a price, deployment path, hardware burden, or best use. His answer up front: yes, serious AI users should test Chinese models, but you must make three separate decisions — the task, the model, and the deployment path — and measure cost per accepted result rather than token price. Bias note: the video plugs his own Ringer multi-agent product and his Substack deep-dive, so the swarm recommendations double as marketing.
Key Takeaways
- The models are wildly different products. DeepSeek pursues extreme inference economics (15/M output) and called it open weight while the weights were not yet downloadable at recording. Z.ai’s GLM 5.2 is MIT-licensed and very strong at long-horizon coding, but its released BF16 checkpoint is ~1.5TB. MiniMax M3 is 427B total / 23B active per token, spanning coding, computer use, and image/video understanding, under a custom license requiring attribution and notice, prior written authorization above $20M annual revenue, and prohibiting military use.
- The US government’s own read (CAISI, May): DeepSeek V4 Pro is “the most capable Chinese model” it tested, “about eight months behind” the leading US frontier — and it ranged from “53% cheaper to 41% more expensive per correctly solved task across seven benchmarks.” Token price and finished-work cost can point in opposite directions; DeepSeek has changed prices since, so treat the numbers as an indicator of where to test, not a verdict.
- Cost per accepted result is the gold standard: include input, output, reasoning, tool calls, retries, latency, and infrastructure. A cheap model gets expensive via long reasoning traces, unnecessary tool calls, late failures, and human cleanup; a premium model like Fable 5 can be cheaper if it succeeds in one pass — and Opus 5 launched with the same token-efficiency focus, which is how frontier models now compete.
- The distillation fight matters: Anthropic says DeepSeek, Moonshot, and MiniMax used about 24,000 fraudulent accounts to generate more than 16 million Claude exchanges while evading access restrictions; a later White House memo described “industrial-scale extraction by China-based entities” without naming labs or validating Anthropic’s numbers — then Kimi K3 was named by the White House, and K3 was separately publicly accused of distilling Fable 5. These are allegations, not court findings, and the public record doesn’t show how much released capability came from the alleged campaigns.
- Self-host only when four conditions all hold (see below) — otherwise use a first-party API or managed third-party host.
- Verdict: use Chinese models selectively — as specialists, challengers, and in some workflows the default — but “country of origin is just the very beginning of having a conversation about due diligence.”
The Three Decisions
1. Name the task (and your failure tolerance)
- High-volume, price-sensitive API work — extraction, classification, first-pass research, test generation, reviewable work: DeepSeek absolutely belongs in the test set; a 15–30x price cut lets the machine “check and recheck and check again.” Be cautious wherever one ambiguous answer drives an unrecoverable action.
- Local/personal-hardware work — smaller Qwen models and the R1-distilled models are the real local-AI story (not the same artifact as full R1 or V4 Pro): private notes, offline work, predictable availability, trading capability for control and privacy. The right comparison is not small-Qwen vs Kimi K3 — it’s whether the small model is good enough for the exact job you want to keep on your machine.
- Long-horizon coding/research/agent work — test GLM, Kimi, Qwen, and MiniMax against the best American frontier models with the same tools and the same acceptance standard. Don’t change your bar: GLM 5.2 is particularly strong on coding, but a model can write an excellent patch and still lose a repository-level task by picking the wrong tool, missing a test, or failing after 20 minutes.
- Ambiguous/frontier-judgment work — the strongest American frontier systems remain the baseline to beat, with more mature tool integrations, enterprise controls, and operational evidence; Chinese offerings counter on price, open weights, and deployment choice.
- Sensitive work — decide where the data may go before choosing the model.
2. Name the artifact and deployment
- “Open” is not one property. Ask separately: Can I call it through an API? Can I download the weights? Does the license permit my use? Can my hardware run it at the context length and speed I need?
- Mixture-of-experts is why enormous models can still be economical: the router activates a fraction of parameters per token (MiniMax M3: 23B active of 427B), but total parameters still drive storage, memory, networking, and serving burden — a model can be cheap through an optimized API and absurd on your laptop. Don’t convert active-parameter counts into a hardware estimate; ask about checkpoint size, precision, context requirements, and serving topology.
- The 4-condition self-host test: (1) weights available and license permits your use; (2) the data/sovereignty requirement is real enough to need the model on-prem; (3) the workload justifies the AI infrastructure economics; (4) you have a team that can operate the stack safely. Self-hosting also needs a named owner for authentication, model-server patches, capacity, and observability — otherwise you’re converting vendor risk into an operating problem you discover when something breaks.
- Three deployment paths: first-party API (provider controls service, price, logging, availability), third-party host (run open weights in your chosen region and contract — an American host running Qwen weights sits under American jurisdiction), self-host (maximum control, total responsibility).
3. Trace the data and the exit path
- China’s interim rules govern generative-AI services offered to the public inside mainland China — they don’t automatically govern raw weights privately operated outside China. DeepSeek’s first-party privacy policy says personal data is processed and stored in China and inputs may be used to improve its technology; Alibaba says Model Studio customer data is not used for training and offers deployment scopes that can exclude mainland China.
- Self-hosting removes provider-side runtime filters but not refusal behavior or political bias learned in the weights — you may need to fine-tune, and you must check inference calls, tools, telemetry, and logs. Moving the weights changes who can see the data; it doesn’t make the rest of the software supply chain trustworthy by magic.
- Own your evals and preserve an exit path: identify who receives the prompt, where it’s stored/processed, governing law and contract, retention and training terms — and whether you can lift prompts, tests, and tools to another model if the provider changes its price.
Distillation and Policy Context
- Normal distillation is ordinary and disclosed: a teacher generates examples/answers/reasoning traces, a student trains on them. OpenAI sells distillation as a product; frontier labs use it internally (a Mythos teacher model teaches the released versions of Mythos and Fable); DeepSeek openly fine-tuned Qwen-derived students on 800,000 samples curated from R1.
- A distilled student is not the teacher in a smaller file: it inherits reasoning patterns on examples it saw but loses breadth, reliability, and safeguards — hence “big model smell.” “DeepSeek runs locally” conflates the local student with the frontier teacher.
- The controversy begins at authorization and conduct: training on API outputs against contractual restrictions, fake accounts, access-control evasion at scale. Chinese entities are typically not authorized to distill American models.
- Why it matters even if you reserve judgment: chips are physical, model outputs are information. Restricting a teacher’s GPUs doesn’t control what it can teach through an API (his Napster analogy) — capability moves through synthetic data, fine-tuning, RL, and distillation, imperfectly but fast and cheaply enough. That means capability gaps can close faster than hardware policy suggests, and frontier providers tighten access, monitoring, and account controls in response.
- Export controls didn’t invent Chinese efficiency — MoE, low-precision training, sparse attention, quantization, and hardware utilization reflect a broader decades-long tendency; controls raised the strategic incentive. “The Americans don’t have to worry about chips, they have to worry about power.” China’s AI-plus policy supports open-source sharing of models, tools, and datasets; open weights also reduce dependence on American APIs, though the policy doesn’t state that motive explicitly.
Incoming — Qwen3.8-Max and Ling-3.0-flash (2026-08-05 addition)
Two data points landed in the same week, both relevant to this framework’s licence-and-capability axes.
Qwen3.8-Max (Alibaba) — announced 2026-08-03, per raw/newsletter-theneurondaily-com-e5b35850e2.md (The Neuron Daily, secondary): a 2.4-trillion-parameter model whose coding agent reportedly works unsupervised for 10+ days straight, “from empty folder to finished product.” Open weights were said to be releasing the following week, alongside a smaller sibling. If the open-weights release lands as described, this is the largest model yet to reach open weights and a direct test of this article’s thesis about capability-per-licence-risk.
The 10-day autonomous-run claim is the sort of long-horizon assertion this wiki has repeatedly found to be the least reproducible part of a model announcement — treat it as unverified marketing until someone outside Alibaba runs it. A separate, independently-surfaced report (from an ungrounded search result, flagged as such at ingest) said Qwen 3.8-Max was already available in Hermes Agent at a 20% discount; that is not confirmed here.
Ling-3.0-flash (inclusionAI / Ant Group) — genuinely shipped, MIT-licensed, verified against the HuggingFace API. See the full article. It is the cleaner test case for this framework precisely because the licence is a plain MIT rather than a custom community licence with a distillation clause — the variable that dominates the decision here.
Token Share vs Spend Share — the Vercel Chart (2026-09-29 addition)
Source: Matthew Berman, “The Most Important Chart In AI Right Now” (youtube.com/watch?v=2w7ZdceZT-g, raw/The_Most_Important_Chart_In_AI_Right_Now.md, fetched 2026-08-27). Berman shows charts and posts from Vercel, Gavin Baker, Bindu Reddy and Christian Catalini; all figures are as he reads them.
-
Open weights now win on volume but not on money. A Vercel chart covering June to August shows open-weight models taking a growing share of the tokens on its platform.
- DeepSeek’s token share (25.2%) now edges Anthropic’s (24.5%).
- Its share of spend is only 2.8%, against Anthropic’s 64.6%: “Anthropic spend share is 23x DeepSeek’s”.
-
Gavin Baker’s prediction: frontier tokens will be “60 to 90% of all economic value, but only 10 to 25% of tokens.”
-
Cost per completed task confirms this article’s central test. Kimi K3 is half GPT-5.6 Sol’s per-token price, but a completed task costs 84¢ vs 96¢, “because Kimmy K3 needs many more tokens to come to the same conclusion”. The same effect shows up in September’s open-weight releases: GLM 5.3 Flash uses about 47K tokens per task against GPT-5.6 Luna’s 20K. See the September model landscape.
-
The third-party-host and fine-tune paths are already in production at US companies. Berman’s list:
- Thomson Reuters and Airbnb on Qwen;
- Harvey on Kimi K3, fine-tuned into a legal model, with Harvey’s own chart showing Legal Agent Bench 19.7;
- Cursor on Kimi K2.5;
- Perplexity on DeepSeek.
The reasons he gives are cost, privacy and the ability to fine-tune on your own data. He also counts over 151,000 Qwen derivative models, which means someone may already have built the fine-tune you need.
-
A new risk for the “exit path” decision (Berman’s opinion): as Chinese models become co-designed with Chinese chips, US firms building on them may end up dependent on Chinese hardware. Around the same time, SemiAnalysis reported GLM 5.3 Flash serving 100 trillion tokens a day “on purely Chinese chips” (Berman,
raw/Cancel_your_subscriptions_Ox-Alpha_is_here_GLM_5.3_Flash.md).
What this changes in the framework: nothing in the three decisions, but it adds a benchmark for decision 1. If a Chinese model’s completed-task cost is within about 15% of a US frontier model’s, as with K3 against Sol, choose on data path and exit path rather than on price.
October 2026 Update — DHS ban, export controls, and US open models as an alternative
Sources: The Neuron (raw/newsletter-theneurondaily-com-4f8dd7da56.md, Oct 19; raw/newsletter-theneurondaily-com-df67e8a9fe.md, Oct 20; raw/newsletter-theneurondaily-com-fd43280340.md, Oct 18) and The AI Daily Brief (raw/America_s_Open_Model_Comeback_Is_Starting.md, raw/What_the_Best_Business_AI_Users_Are_Doing_Different.md, raw/AI_s_First_Full_Field-Level_Disruption.md, raw/Why_AI_Companies_Are_Betting_Personality_Beats_Intelligence.md).
Decision 3 (data and exit path) got harder for anyone selling to the US government.
- DHS banned DeepSeek from federal civilian systems. A July memo, made public Monday Oct 19, orders agencies to remove DeepSeek models and stop buying DeepSeek products, including third-party services built on them. NLW notes the order applies “even if it’s hosted by a hyperscaler on US soil”, and that no technical report shows risk from US-hosted weights. The practical effect: a startup built on DeepSeek may be ineligible for federal work.
- Distillation became a joint advisory. On Sept 8 the NSA, FBI and CISA said DeepSeek and Alibaba used distillation “at industrial scale”; Beijing hinted at retaliation (The Neuron).
- China’s own brakes. On Friday Oct 16 China proposed export controls on 23 AI items, including weights for models with 2 trillion or more parameters or a top-10 world ranking. Bloomberg and the AP say Kimi K3 and Qwen3.8-Max exceed the size line. API and app access would continue, and already-public technology is exempt (The Neuron). Future Chinese flagships may arrive as APIs only, which removes the self-host path for them.
- Chips: IDC says Nvidia’s share of China’s AI-chip market fell to 34% in early 2026 while Huawei rose to 28% (The Neuron).
DeepSeek’s October releases (decision 1, cheap bulk work). V4.1-Flash-1017 (Oct 17, MIT licence) reports DeepSWE v1.1 77.0 at unchanged prices (0.60 off-peak). DeepSeek also opened agent Search and Fetch tools at $0.75 per 1,000 successful calls, free through DeepSeek Harness, Claude Code or Codex until Nov 1. Both run on DeepSeek’s servers; The Neuron’s advice is public-web research only, and check IT policy first. Details: October landscape.
Token share vs spend share, again.
- Vercel CEO Guillermo Rauch reported a record day on Vercel’s AI Gateway: “Open models used 78.4% of tokens, while closed used 21.6%” (via NLW).
- Ramp economist Eric Kerazian: open-source models “remain less than 5% of business spend” (via NLW).
- The Information’s reader survey: 72% use open-weight models with some regularity, 46% use them for their work, but only 21% use Chinese open-weight models (via NLW).
- Together these confirm the pattern above: open models win on volume, not on money, and most of the open volume in businesses is not Chinese.
The new argument for open weights is fine-tuning, and US options now exist.
- “You cannot fine-tune Luna” is the reply NLW quotes to “it doesn’t actually seem cheaper than using Luna”.
- Satya Nadella’s “reverse information paradox”: “You essentially pay for intelligence twice, once with money and again with… the proprietary knowledge you must reveal”.
- Microsoft says MAI tuned for McKinsey’s tasks beat GPT-5.5 on quality “while being 10x lower on cost” (vendor claim).
- October added US-made open models to test alongside the Chinese set: Meta’s Muse Spark Open (816B / 42B active, AA 39), Reflection Beam (501B) and NVIDIA Nemotron 3.5 Super VL. None tops Kimi K3 or GLM 5.3 on the charts relayed so far. See the October landscape.
Try It
- Run the individual-scale version of his protocol: take 20 examples of real work, include the ugly edge cases, pick one Chinese model, run the test, and measure cost per accepted result across the whole chain.
- His own result doing this: building his Ringer multi-agent swarm (Fable orchestrates, cheaper models execute), Qwen proved a genuinely useful swarm participant for many use cases. (Disclosure: Ringer is his product.)
- Before any self-host conversation, score the 4 conditions honestly — if any fails, price a first-party API or a third-party host in your jurisdiction instead.
Open Questions
- Whether Kimi K3’s weights actually shipped “very soon” after recording, as Moonshot promised. Partial answer (2026-09-29): Berman lists K3 as an open-weights model and says Harvey fine-tuned it into its legal model, which suggests the weights are available. ^[inferred] That is still secondary; confirm on Moonshot’s release page.
- Which specific seven benchmarks CAISI used, and current DeepSeek pricing vs the May figures — he flags the numbers as already-moving. Later data points (secondary):
- In August Berman cites DeepSeek V4 Flash at 18¢ per million output tokens.
- DeepSeek V4.1 Flash (Sept 9) lists input at 0.30 peak and output at 1.20, per Berman reading DeepSeek’s pricing page (
raw/Deepseek_did_it_again....md).
- The evidentiary status of Anthropic’s fraudulent-accounts numbers — described in the video as serious allegations the White House memo did not validate.
Related
- Kimi K3 (Moonshot) — the release that triggered this analysis
- GLM 5 Series (Z.ai) — the MIT-licensed coding-strong family discussed here
- GPT-5.6 Series — the US frontier baseline these models get tested against (as of this analysis, July 2026; OpenAI replaced it with GPT-6 Astra, Sol and Luna in September 2026)
- Cheap-Executor Delegation — the frontier-orchestrator-plus-cheap-workers pattern his Ringer swarm implements
- Cost and Intelligence Levers — the wider cost-vs-capability trade space around cost-per-accepted-result
- Task-Relative Redaction (Nate B Jones) — same creator on controlling what data reaches which model
- Frontier Model Landscape, September 2026 — GLM 5.3 Flash, DeepSeek V4.1 Flash and the other September releases, with per-task costs