The DeepSeek Chip Gambit: Why China's AI Champion Can't Touch Nvidia's Moat

The "open source with impunity" thesis sounds clever. The hardware reality says something else entirely.

The argument is elegant. On Friday, @teortaxesTex laid out the case: DeepSeek will codesign its own AI models and the chips they run on. With that kind of vertical integration, the company can open source its frontier models freely because anyone who deploys them on generic hardware pays a higher compute cost. It is the Google TPU playbook applied to a Chinese AI lab. "Full vertical integration with proprietary hardware is when you can open source with impunity."

The logic works. The question is whether DeepSeek can actually build the hardware. Three weeks of evidence say the gap between the theory and the silicon is measured in years, not quarters.

DeepSeek really is building a chip — but it's inference only

On July 7, Reuters confirmed via three sources that DeepSeek is developing its own AI chip. It is designed for inference — the stage where a trained model generates responses for users — not for training new models. The effort began about a year ago and remains "at an early stage," with DeepSeek reaching out to foundry, design, and memory partners. Hiring has been done quietly, with no public job postings.

This is a real project, funded by a ¥50 billion ($7.4 billion) first external funding round that closed in June. DeepSeek explicitly earmarked funds for chip development, alongside chip purchasing. Liang Wenfeng himself contributed ¥20 billion of his own money. Tencent, CATL, NetEase, JD.com, and China's National AI Industry Investment Fund all joined the round.

But here is what "early stage" means in practice. Competitive AI chips take two to four years from design to production. DeepSeek's reported manufacturing partner is SMIC, whose best process is roughly 7nm class. TSMC is at 3nm and heading to 2nm. Then there is high-bandwidth memory: U.S. export curbs have cut China's access to HBM, a component critical to any AI inference chip. And rival labs are further along. OpenAI has already taped out Jalapeño, its custom inference chip built with Broadcom. Anthropic is building its own. DeepSeek is playing catch-up on someone else's field.

The CSIS recently noted that SMIC will "most likely remain stuck at 7nm or perhaps a flawed 5nm technology node for many years." ASML's CEO put it bluntly: "By banning the export of EUV, China will lag 10 to 15 years behind the West."

The most honest Nvidia bull works at DeepSeek

If you read only one document on this, read the leaked transcript of Liang Wenfeng's investor conference. It is the most candid assessment of the U.S.-China chip gap ever attributed to a Chinese AI CEO.

Wenfeng said it plainly: "Four Huawei cards equal one Nvidia card, while being two years behind." DeepSeek's allocation of 16,000 Huawei Ascend 950 chips is equivalent to about 4,000 Nvidia B-series chips. That is "far from sufficient." His explicit goal: "as long as the price is reasonable, buy as many as available." If he could convert the entire $7.4 billion funding round into Nvidia GPUs within six months, "that would be pure bliss."

The chip development, in other words, is a hedge — not a competitive strategy. It is what you do when you cannot legally buy the silicon you actually want.

Wenfeng also made a claim some have interpreted as a bear case for Nvidia: "Nvidia's CUDA moat is rapidly being dismantled." His evidence is TileLang, DeepSeek's internal high-level language used to train V3. But "dismantling" means DeepSeek can port their software stack to non-Nvidia hardware. It does not mean Huawei Ascend competes on silicon. It means they can survive without CUDA. That is a China AI survival story, not an Nvidia disruption story.

Jensen Huang sees it as health food for the whole industry

On the same day teortaxesTex posted his thesis, a new Jensen Huang interview with Axios dropped. When asked about DeepSeek and Kimi, he did not flinch:

"These Chinese models are excellent. The market misunderstood the impact of DeepSeek the first time, and it has misunderstood the impact of another Chinese model, Kimi, again this time. First of all, great open AI models are good for the whole industry. When more high-quality AI becomes available, even if it is open-source and regardless of where it comes from, there will be more usage. Whenever there is more usage, Nvidia will sell many more computers."

This is not spin. The tokenomics model backs it: global token consumption is running at roughly 288 trillion tokens per day and growing. Supply cover is projected at 1.41 through 2027Q4 — demand still outrunning capacity. The fleet's AI power draw goes from ~31 GW at end of 2026 to ~63 GW at end of 2027. Training share remains at 61%, meaning the fleet is not yet flipping to inference-majority the way many assume. When it flips in 2027, that shift drives another wave of hardware demand, not a reduction.

China consumes roughly 22% of global AI tokens, and its fleet allocates about 59% to training. A DeepSeek inference chip that captures a slice of Chinese inference compute is a rounding error on the global demand curve.

The real Nvidia thesis is what happened today

While the Twitter thread went viral, Nvidia was busy locking in $500 billion in AI infrastructure deals with SK Group, securing HBM memory supply through 2030. It announced a $1.5 billion partnership with Amkor for chip packaging and testing in Arizona. It invested $1 billion in Naver for a Korean AI data center. SK Telecom will build a 2-gigawatt Nvidia Vera Rubin DSX AI factory with HBM4 memory from SK Hynix.

These are not things a company whose moat is "dismantling" gets to do.

The fundamentals back the picture. Nvidia closed Friday at $206.80. The latest quarter (FQ1 ending April 2026) delivered $81.6 billion in revenue, up 85% year over year, with 75% gross margins and $48.6 billion in free cash flow. The company sits on $68.2 billion in net cash against $12.3 billion in total debt. Forward P/E is 16.1 with a PEG of 0.575. Fifty-eight analysts carry a consensus strong buy with a mean price target of $303, implying 46% upside.

What to watch

DeepSeek's inference chip is real, and part of a broader trend: OpenAI, Anthropic, Alibaba, and Baidu are all building custom inference silicon. Over five to ten years, this could redirect a slice of the global inference TAM away from Nvidia GPUs. But China's inference TAM was never legally Nvidia's to capture anyway thanks to export controls. What matters for Nvidia is the rest of the world.

The teortaxesTex thesis is a brilliant description of what vertical integration should do — but it applies fully to exactly one company today: Google, whose TPU v7 serves tokens at roughly 7.5× cheaper than a B200 rental because Google owns the silicon, the stack, and the cloud. DeepSeek is not Google. It has SMIC instead of TSMC. It has restricted memory instead of HBM4. It has 16,000 Huawei chips instead of a 2-gigawatt AI factory.

The day DeepSeek announces a tape-out on a competitive node with unrestricted HBM access, the debate changes. Until then, the chip gambit is a survival strategy, not a threat.