Engineering notes / Muhammad Kashif
Chip Bans, Open Weights, and What Sovereign AI Means for Builders
US chip export bans were supposed to freeze China’s AI progress. The notes I keep say the pressure did something else: it pushed domestic chips and extremely efficient open models that now rival paid US systems on many t…
Chip Bans, Open Weights, and What Sovereign AI Means for Builders
By Muhammad Kashif · Draft · AI automation & cybersecurity builder, Pakistan
US chip export bans were supposed to freeze China’s AI progress. The notes I keep say the pressure did something else: it pushed domestic chips and extremely efficient open models that now rival paid US systems on many tasks. For a builder in Pakistan, that is not a spectator sport. It changes which models I can run, what I trust with client data, and how I think about national infrastructure.
DeepSeek and the cost curve
DeepSeek trained for a reported roughly six million dollars using a mixture-of-experts approach — the metaphor in the notes is “only call the marketing team” when you need that specialty — plus memory tricks. Other Chinese labs land competitive models in a roughly five-to-ten-million-dollar band. That is not “AI is free.” It is “frontier-adjacent capability is no longer priced like a national lab only.”
When a strong model is cheap to train, it becomes cheap to serve, and open weights spread. That is the strategic surprise the bans accelerated.
Kimi, Qwen, GLM: open first, monetize later
Kimi K3, Qwen 3.8, and GLM show up as free and open, with monetization deferred into cloud and ecosystem plays — an Android-style network-effect bet. Opinion notes also flag a huge token-price gap: US closed models around twenty to twenty-five dollars per million tokens versus Chinese open-weight serving around one to two dollars in the comparisons recorded.
I use that gap carefully. Price is not permission. I still ask where inference runs, what is logged, and whether a client’s regulated data can sit there. Cheap and opaque is not a bargain.
National missions and corporate adoption
Since 2017, Baidu, Alibaba, Tencent, and iFlytek were each assigned national AI platforms in the story the notes tell. That is industrial policy, not a hackathon. Separately, the notes claim about sixty percent of US companies run some Chinese AI, with privacy and propaganda risks argued from both sides. I cannot verify that percentage from my laptop — I treat it as a warning to inventory what our stacks actually call.
A related opinion in the notes: a US ban on Chinese models could backfire if companies relocate offshore and countries are forced to choose sides. Procurement politics is now part of architecture reviews.
Lithium, EnergyX, and the physical layer
AI is not only weights. Lithium supply is infrastructure. China dominates processing in the notes’ framing. EnergyX’s direct-lithium extraction work shows up with a DOE grant around five million dollars and a GM-led fifty-million-dollar round, plus a public share offering narrative around eleven dollars. I am not offering investment advice. I am noting that energy and materials constraints will shape who can train at scale next.
Lessons I apply locally
Every country needs a sovereign AI story — the notes say this about India; the same logic presses on Pakistan. Sovereign does not mean “never use a foreign API.” It means knowing which workloads must stay on controlled infrastructure, investing in local talent who can fine-tune open weights, and refusing to build critical public services that die if a single foreign endpoint blinks.
Ecosystem matters more than the chatbot. Data pipelines, evaluation harnesses, secure hosting, and payment rails decide whether open weights become an industry or a demo day.
My builder checklist after these notes
- Inventory models in production and their jurisdictions.
- Prefer open weights for sensitive offline experiments.
- Keep a paid Western API for workloads that need those vendors’ compliance stories — and document why.
- Watch energy and chip politics as leading indicators, not trivia.
Chip bans did not pause history. They rearranged it. My job is to rearrange my stack on purpose instead of by accident.
On client calls I now ask a blunt question: “Which model families are you willing to ban, and which workloads must stay in-region?” Ambiguous answers become architecture risks. Clear answers become allowlists I can enforce in code.
Open weights are not a personality contest between countries. They are a supply-chain fact. Builders who ignore the supply chain ship on luck. I would rather choose my dependencies while I still can.