Physix FrontierDeepSeek Enters Chip Manufacturing at $59B Valuation: AI Becoming a National Asset
This article is reprinted from Meng Siwei (WeChat Official Account "Fourth Dimension's Dream"), all rights reserved. Click to view the original link.
On the same day, Reuters published two exclusives regarding Chinese AI. One reported that DeepSeek is developing its own AI chips. The other stated that regulators are discussing how to open up access to the most advanced models with top companies. One digs down into infrastructure, the other pulls inward on regulation. While the directions seem opposite, they send the same signal: AI is now being managed as a national-level asset.
I. DeepSeek Plans to Manufacture Its Own Chips
According to three sources familiar with the matter, DeepSeek's self-developed chip targets inference scenarios. The project started about a year ago and is currently in the early stages, engaging with chip design firms, wafer foundries, and memory manufacturers. Hiring for chip engineers is also proceeding quietly. Previously, DeepSeek's models relied mainly on NVIDIA and Huawei chips, with base models trained on H800s. The V4 model released in April adapted to Ascend, and Huawei participated in some training for V4-Flash.
(Original image placeholder)
Actually, viewed in a global context, this is not an isolated case. OpenAI just launched its first inference chip, Jalapeno, co-developed with Broadcom last month. Anthropic has also been reported to be considering self-developed chips. Model companies are collectively extending downwards, betting on the same direction. The computational focus is shifting from training to inference, and inference chips promise lower costs and power consumption compared to general GPUs.
Does this remind you of the judgment I've repeatedly made earlier: In the era of inference, with the LLM paradigm settled, ASICs inevitably rise? This is determined by first principles, regardless of any external factors.
Another timeline is equally noteworthy. DeepSeek is preparing for its first external financing since inception, aiming for around $7 billion with an estimated valuation of $52-59 billion. From rejecting external capital to gradually opening up, from focusing on model R&D to entering the chip arena, this company is changing two long-standing strategic directions simultaneously. The capital needed for chip manufacturing is exactly what this financing aims to solve.
Looking at the domestic market, the competitive landscape is also reshaping. Currently, Huawei holds about half of the ~$50 billion Chinese AI chip market, while Alibaba and Baidu's self-developed chips are expanding their shares. Now that DeepSeek is joining the competition, domestic computing power is moving from "one player carrying the load" to "multiple players racing." Competition isn't bad; true industrial chain maturity often forms through sustained competition and internal iteration.
II. How to Open Up the Most Advanced Models?
According to Reuters, over the past month, regulators have held discussions with companies like Alibaba, ByteDance, and Zhipu. Topics included restrictions on overseas access to the most advanced AI models (including unreleased ones), incorporating AI technology theft into national security legal penalties, and managing funding sources for domestic AI startups.
However, it must be emphasized that these are still at the discussion stage. The scope of restrictions hasn't been finalized and may only apply to future new models. There is no clear timetable for implementation or timing. This point was repeatedly stressed in the original report.
Clues about the direction come from a summary of a May roundtable on open-source AI regulation. The discussion proposed a tiered management approach: basic open-source tools require simple filing, advanced technologies undergo security reviews, and the most sensitive frontier models are restricted from public release or allowed for domestic use only.
Looking at the US, the logic is converging. In June, the US required that foreign citizens could not access Anthropic's most advanced Fable and Mythos models. Among them, Fable reopened after deploying new security measures, while Mythos remains accessible only to certain "trusted" US institutions.
While specific measures differ, the underlying logic is highly consistent. Frontier AI models are being incorporated into strategic asset management frameworks by both the US and China.
At the end of last week's article Token Cost Only 1/10: What Does It Mean That Even Microsoft Is Considering Chinese Models?, I warned about "the boundaries of openness": While the open-source strategy is good, access restrictions in high-sensitivity areas are a long-term ceiling. Now, this boundary has entered formal discussion. The second half of the open-source strategy might involve "internal-external layering": Volume globalization continues to run, with seven Chinese models in OpenRouter's top ten unaffected, and the discussed plans mainly target future frontier models, while frontier localization proceeds in parallel.
While the open-source strategy is good, we must see the boundaries. Access restrictions in high-sensitivity areas are a long-term ceiling, and the commercial path for open weights (free models, paid ecosystem) itself hasn't been fully validated. For secondary markets, the tradable mapping of this narrative remains inference compute and traffic infrastructure, not the valuation stories of model companies.
Meng Siwei, WeChat Official Account: Fourth Dimension's Dream Token Cost Only 1/10: What Does It Mean That Even Microsoft Is Considering Chinese Models?
(Original image placeholder)
OpenRouter's latest weekly usage leaderboard: DeepSeek V4 Flash leads with 5.34 trillion tokens, followed by Xiaomi MiMo, MiniMax M3, Hy3 preview, and GLM 5.2. Seven of the top ten spots belong to Chinese models, with Anthropic's three Claude models taking the remaining three.
For the overseas expansion narrative, this further clarifies the path. If frontier weights tend to stay domestic, then the main drivers for going global are services and compute. Shantou's token exports and Ascend's entry into Korea happen to fit the "selling water, not wells" model. Looking at the two legs of the overseas expansion narrative now, these two legs align better with regulatory winds than "global distribution of open weights."
But we must note that both reports rely on anonymous sources—one is early-stage, the other under discussion—don't trade on them as fait accompli. DeepSeek's chip manufacturing still faces physical constraints in accessing advanced foundries and HBM, which loops back to domestic limitations. The progress of ZXGJ, CXCC, etc., ultimately determines the ceiling for DeepSeek and peers.
Logic and technology—we are still on the road!
Recommended Reading (Original Official Account)
- Where Do Memories for 10 Billion Agents Go? HBM Might Just Be Chapter One
- Meta Partners with Samsung, Anthropic Follows: Is the Foundry Cycle Back?
- NVIDIA Kyber Delayed by a Year, NVL72x2 Cancelled: Who Benefits When NVIDIA Stumbles?
Disclaimer: Materials in this article are compiled from company announcements, news, public institutional research reports, and social media networks. It does not constitute investment advice for any professionals or individual stocks mentioned. Markets carry risks; invest cautiously. Please contact via backend message if images/text infringe rights or violate disclosure rules for deletion.
