DeepSeek Didn't Break the Chip Ban — It Walked Around It Five Ways

October 20, 20266 min read

On January 27, 2025 I opened my phone like everyone else and saw the same headline: a Chinese AI app had hit number one on the US App Store. By the end of that day, Nvidia had reportedly lost about $589 billion in market value. That is not a typo in the story. It is the market's way of admitting that almost three years of US export controls — built to stop exactly this — had not worked the way Washington hoped.

I make a living around enterprise cloud and procurement. When a policy is sold as a hard wall, I still ask where the doors are. This story is about following the chips.

▶ Watch the video: Why America's Chip Ban Couldn't Stop DeepSeek

The chokepoint

The US did not ban "AI" in the abstract. It went after the hardware that makes frontier models practical. Reuters reported in September 2022 that US officials ordered Nvidia to halt sales of its top AI chips to China, with Commerce Department rules following in October 2022. The idea was simple: if you control the best accelerators, you slow the other side's training runs.

Nvidia's response was also simple. Reuters noted an A800 tailored for China in late 2022. Then, in October 2023, CNBC reported a wider ban that caught more parts, including the H800. Each tightening pushed Chinese buyers and US suppliers into a cat-and-mouse game over what still counted as legal.

Five ways around the wall

Way one — chips built to fit the rules. Compliance is not the same as capability. Downgraded or specially labelled parts can still move in volume. Nvidia later took a $5.5 billion charge tied to its H20 line, CNBC reported in April 2025 — a reminder that "China editions" are a business, not a sideshow. The Financial Times reported in August 2025 that Nvidia and AMD would pay 15% of certain China chip revenues to the US government; CNBC in December 2025 described H200 sales to China with a 25% cut. The wall kept moving. So did the price of walking through it.

Way two — the back door. Rules that focus on direct sales to China do not automatically close every route. CNBC and Reuters covered a Singapore server fraud case in early 2025, with a ministerial quote on the record. In May 2026, outlets reported Washington closing an overseas-subsidiary loophole. That sequence matters: loopholes get named only after they have been used.

Way three — the accusation. In February 2026, Reuters carried a US official's claim that DeepSeek trained on Blackwell — an allegation, with Saif Khan quoted. Nvidia has called smuggling reports "far-fetched," per PhoneArena in late 2025. DeepSeek has reportedly denied the Blackwell claim, with reporting in 2026 pointing to H800 and Ascend 910C instead. I treat this block as contested facts, not proof. Markets still traded on it.

Way four — need fewer chips. The DeepSeek-V3 technical report on arXiv is the boring hero of the story. If you can train and serve with less silicon per unit of intelligence, the ban's math changes. You are not trying to match America's biggest cluster; you are trying to get more from whatever you can source. Meta's public Llama materials are part of the same open-model ecosystem that makes efficiency gains legible to outsiders.

Way five — build your own. This is the door Washington cannot fully close without slowing everyone, including allies. Reuters wrote about a Huawei Mate 60 Pro teardown in September 2023; Tom's Hardware reported DeepSeek V3.2 running on China-native chips in late 2025; MIT Technology Review covered V4 in April 2026. Liang Wenfeng's interviews, circulated via outlets like Transformer and the Straits Times, frame a firm that treats chip scarcity as a design constraint. Jensen Huang reportedly described China-related revenue falling from most of a slice to near zero and back toward negotiation — Tom's Hardware cited a 95% to zero arc in October 2025. Reuters in May 2026 said H200 approvals for roughly ten Chinese firms had not yet meant deliveries, with China once about 13% of Nvidia's revenue; by August 2026, reports described first small H200 shipments — on the order of 10,000 units each to ByteDance and Tencent. The ban shapes who pays whom. It does not erase demand.

The honest caveat

Export controls raised the cost and the creativity of Chinese AI. They did not freeze time. Allegations about smuggling and denials about Blackwell belong in that bucket: serious if proven, noisy until then. The video labels allegations as such on screen, and I am doing the same here.

What it means if you own chip stocks

The investment question is not "did the ban fail?" It is which moat still holds. Nvidia's CUDA ecosystem and scale still matter. So does the cycle of new generations every year or so. A single app-store moment can wipe hundreds of billions off one name because traders are repricing monopoly and geography at once. If you own semis, you are not just betting on demand for AI. You are betting on how much of that demand must flow through US-designed silicon, and how fast alternatives compound.

What I'm watching next

Whether reported H200 shipments scale beyond the first small tranches Reuters and Tom's Hardware described in mid-2026, and what share of revenue China regains.
The next rule patch after the May 2026 subsidiary loophole closure — and what "compliant" SKUs cost after charges like the H20 write-down.
Efficiency releases (DeepSeek V4 and peers) on native Chinese hardware — proof that way four and way five reinforce each other.

The chip ban was never a single switch. It was a stack of rules, priced workarounds, contested supply chains, and models that need less metal. DeepSeek's rise did not prove controls are useless. It proved that when you wall off the best tools, the other side optimises, reroutes, and sometimes builds substitutes. The market's one-day $589 billion verdict was fear that the wall had more doors than investors assumed.

Sources: CNBC and Reuters (DeepSeek App Store, Jan 27 2025; Nvidia market-cap loss); Reuters and US Commerce/BIS (export controls, 2022); Reuters and CNBC (A800, H800, H20 charge, H200 terms); Financial Times via CNBC (China revenue share to US government); Reuters/CNA and CNBC (Singapore server fraud); CNBC/Reuters (subsidiary loophole, May 2026); Reuters (Blackwell allegation, Feb 2026; H200 approvals and shipments, 2026); PhoneArena (Nvidia on smuggling); China Academy/Implicator (DeepSeek denials); DeepSeek-V3 technical report (arXiv 2412.19437); Meta Llama 3.1 model card; Liang Wenfeng interviews; Reuters/FT (Mate 60 teardown); Tom's Hardware (V3.2, Jensen Huang quote, H200 shipments); MIT Technology Review (DeepSeek V4).

For information only. Not financial advice.

← Back to Insights