What China Catches Up On First

The DeepSeek team shipped V4 on Friday. 1.6 trillion parameters, 49 billion active. Day-zero support on Huawei Ascend silicon, which is the detail I keep coming back to. The pricing card is more honest than the benchmark deck: $0.55 per million input tokens. Roughly a tenth of frontier US lab pricing for a model that, on the public benchmarks the team chose, is within a few points of Claude Opus 4.6 and GPT-5.5.

I won't argue the benchmarks. I want to walk through the five-layer stack we have been using to underwrite the AI race, because V4 is a clean chance to mark the score on each layer. The layers, bottom up, are:

China is winning two of them today, contesting two, and walled off from one. That is a more useful map than either "China caught up" or "China is a decade behind."

Layers one and two: China already won, and the win compounds.

Rare earths first. China produces around 70% of global mining supply and refines closer to 85%. The export-control regime announced last October was paused in November under the Trump-Xi truce, running through November 10, 2026. The suspension is the wrong number to anchor on. The infrastructure for licensing, end-use checks, and 0.1%-by-mass tracing on permanent-magnet materials is now built. Every server-grade Nvidia and Huawei accelerator on Earth ships with magnets that, on paper, fall under that scheme when the pause lapses. The control is a switch, and the documentation chain to legally import a US AI server already runs through MOFCOM.

Frankly, this layer does not move on a five-year horizon. You cannot speed-run a mine. Project Vault, the $12B US strategic minerals reserve announced in February, adds real capacity in 2030. What it adds in 2026 is approximately zero.

Energy is layer two, and the numbers are worse than most US investors I talk to have priced. Chinese data centers pay roughly half what US data centers pay for power. Permitting runs in months, not years. Domestic transformer lead times in China are 48 weeks. The US average is 143. Morgan Stanley projects a 44-gigawatt US grid shortfall over three years. China's reserve margin has not dipped below 80% nationwide in two decades. If you came up in engineering, you recognize this immediately. The bottleneck on training the next model is not algorithm research. It is the upstream resource graph, and on the US side that graph runs out at the substation well before the fab.

US utility capex plans jumped to $1.4T this year. That is a real response. It also lands online in 2029 and 2030, and the model the next round of GPUs is going to train on does not wait that long.

Some people use Claude to write emails. Others use it to basically run their entire business while they play Wordle.

This isn't just ChatGPT's cooler cousin. It's the AI that's quietly revolutionizing how smart people work – writing entire business plans, planning marketing campaigns, and basically becoming the intern you never have to pay.

The Hustle's new guide shows you exactly how the AI-literate are leaving everyone else behind. Subscribe for instant access.

Get our Claude playbook

Layer three: chips, the chokepoint that keeps the score honest.

Here is where the catch-up story breaks down. SMIC is stuck at 7nm at scale, with 5nm pilot-run yields hovering around 20%. A 20% yield is not a manufacturing process, it is an experiment with a manufacturing label on it. The lithography chokepoint is still ASML, EUV is still embargoed, and SMIC's multi-patterning DUV workaround is expensive enough that the Huawei Ascend roadmap is constrained by yields, not demand.

Huawei's stated target is 1.6 million high-end logic dies for AI accelerators in 2026, with a planned doubling of 7nm capacity. If they hit it, the gap closes. If yields stay where they are, the gap holds. This is the layer the export-control regime was actually designed to keep open, and on the published numbers, it is doing its job.

Here's the thing. The chip layer is the gating layer for everything above it. You can have all the rare earths and all the gigawatts you want. If you cannot stamp out the dies, you cannot run the training cluster. China has the input side. It does not yet have the fab side. That is the piece I want every US-based AI investor to sit with this week, because the rest of the stack is where the optimism lives, and the rest of the stack is downstream of this one bottleneck.

Layer four: the model layer is where the gap is closing fastest.

DeepSeek V4 is the cleanest evidence I have seen that the model layer is closing. Near-frontier benchmarks at a fraction of the price, running on domestic chips on day zero. Qwen 3.5 from Alibaba in February, Kimi K2.6 and GLM-5.1 in April. Most US AI valuations are pricing the model layer as a moat. It is also the layer with the shortest catch-up timeline.

The mechanism is distillation. A student fits a stronger teacher's outputs at a fraction of the compute, the trick Hinton wrote up in 2015, and every serious lab does some version of it. xAI's CEO admitted to it under oath last week in federal court. The US has alleged the same about DeepSeek for two years. The technique that makes catch-up cheap is now in the public record on both sides of the Pacific.

A frontier-lab moat measured in months is a different asset than one measured in years. Anthropic raising at a $1T mark this week is pricing a model layer the rest of the world is closing on at roughly six-month intervals. The math gets uncomfortable if you write it out.

Layer five: applications, where China is walled off.

Distribution is the last layer, and US labs still own it in the West. ChatGPT, Claude, Cursor, M365 Copilot, the enterprise diligence tools every fund I know is now licensing. None of those run on Chinese models in any market that matters to a US-based fund. WeChat, Doubao, and Alipay's AI agents are massive inside China and invisible outside. The wall is real. It is also a regulatory artifact, and regulatory artifacts move faster than infrastructure ever does.

If I had to bet on the one layer where catch-up gets shocking, it is this one. Not because the technology collapses the wall, but because a single executive order, in either direction, can rewrite who ships in which market.

Honest score today: China wins one and two, the US wins three, four is a coin flip on a six-month horizon, and five is a wall that holds until it does not. The order of catch-up that worries me most is models first, then chips, then applications. The order that worries me least is rare earths and energy, because those are already lost.

I do not have visibility into the layers that will actually decide this. I cannot see SMIC's true yield curve. I have no read on Huawei's stockpile. I cannot guess what either administration will do with export controls in the back half of the year. I am underwriting on lagging data, and the lag is exactly long enough to make me wrong.

The number that mattered this month was not Anthropic at $1T. It was DeepSeek at $0.55 per million tokens, on Huawei silicon, with an April 24 ship date. I have not seen that number in any deck I have read this quarter, and I think we will look back on that as an oversight.

— SWEdonym