In the spring of 2026, the chief executive of a Shenzhen AI chip startup moved into a budget hotel next to SMIC's Southern Semiconductor fab. His inference processor had passed validation, its software stack fully optimized. His tape-out window slipped from Q1 to Q2, then went on indefinite hold. His entire business rested on allocation from a single production line, and he had no answers for his investors.
That is the physics of the market DeepSeek just decided to enter.
Within a few weeks, three of the most-watched AI labs in the world made the same move. OpenAI shipped Jalapeño, its first custom inference chip co-designed with Broadcom. Anthropic reportedly opened early-stage talks with Samsung to build one on the 2nm SF2 process. And Reuters surfaced a project inside DeepSeek that reportedly began about a year ago — a company that hired chip designers through private outreach, never posting a single job to a public board.
Read together, it looks like a global consensus: frontier labs must own their inference silicon.
It isn't. For DeepSeek, this is the same move for opposite reasons, and the constraint that decides whether it works appears on no lab's balance sheet.
Start with what the Western move actually is. It is a war on Nvidia's margin.
When Midjourney switched inference from Nvidia GPUs to Google TPUs, its monthly bill fell from $2.1 million to under $700,000 — a 67% cut for identical work. AWS Trainium and Google TPUs already deliver 40-60% savings over H100s on inference. Renting general-purpose GPUs means paying Nvidia's margin on top of your actual compute.
That margin is a real, addressable cost. Custom silicon lets a lab optimize for its own model architecture and take that margin back. This is a financial engineering problem, and money solves it.
DeepSeek is fighting something the free market does not govern.
For DeepSeek, Nvidia is not the enemy because Nvidia is not even in the room. As Radio Free Mobile's Richard Windsor puts it: "Nvidia is at zero in China and staying there." Export controls closed the door; the H800 that trained R1 was later banned outright.
So DeepSeek has already made the pivot the West is only starting. It moved to Huawei's Ascend, releasing its V4 model optimized for Ascend chips, with Huawei confirming its processors handled part of V4-Flash training. Huawei now holds roughly half of China's $50 billion AI chip market.
And that is precisely the problem DeepSeek is now trying to escape. Sources say it doesn't want to become too reliant on Huawei. The vertical turn is not about escaping Nvidia — it already escaped Nvidia. It's about refusing to trade Nvidia dependence for Huawei dependence.
The vertical turn is not about escaping Nvidia — it already escaped Nvidia. It's about refusing to trade Nvidia dependence for Huawei dependence.
But here is where the strategy hits a wall that no amount of the company's fresh $7.4 billion can move.
China's entire advanced-AI-chip capacity flows through one bottleneck: SMIC's N+2 line, the country's only operational node capable of mass-producing advanced AI processors. By the end of 2026 it targets about 70,000 wafers a month — roughly 2.6 million AI chips a year.
Domestic demand for 2026 is around 4.2 million units. That is a 1.6-million-chip deficit, structural and unbudging.
Now the punchline. Huawei has locked an estimated 43% of that N+2 capacity through a five-year master supply agreement. Its HiSilicon engineers are reportedly embedded in the fab, having driven over 60% of the node's yield improvements. The company DeepSeek wants to depend on less is the company that owns the line DeepSeek would need to build its escape.
Everyone else fights over scraps. Cambricon holds 9-11%. Alibaba's T-Head gets 5-8% of residual capacity. GPU startups like Biren and Moore Threads sit at the absolute margins, unable to secure consistent wafer starts despite competitive designs. Allocation here is not a commercial transaction — it follows national strategic alignment and state equity, not price.
And expanding the line is not a near-term option. Without EUV, SMIC leans on DUV multi-patterning that doubles or triples production time; per-wafer costs run 40-50% above old international norms; yields only recently crossed the 50% break-even line. Building an equivalent fab takes three to five years. The deficit is locked in.
This is why the optimistic case matters, and why it isn't enough.
DeepSeek has been laying algorithmic groundwork for domestic silicon for a while. Its V3.1 introduced the UE8M0 FP8 data format, understood to be designed specifically for next-gen domestic chip hardware — as the reporting frames it, the algorithm team was thinking about the chip while writing the model.
And the payoff, when the co-design works, is real. Look at MoWorld, the domestic-NPU world model from Moxin Tech, built with strategic shareholder Huawei: a 14-billion-parameter MoE model running at over 50 FPS on domestic NPUs, at 30% of the cost of a comparable GPU deployment.
That is the whole thesis of the domestic stack in one number. Tune the model to the silicon and you can undercut GPU economics by 70%. This is the same bet, in hardware form, that Chinese open-weight models are making on price — that "good enough at radically lower cost" beats frontier supremacy.
But notice what that 30%-of-GPU-cost figure quietly assumes: that you have the NPUs. That someone got the wafers. MoWorld's economics are downstream of foundry allocation and HBM access — and separate export controls have cut China's supply of high-bandwidth memory, a component critical to inference chips.
The cost story is genuine. It is also entirely hostage to a constraint upstream of the model, upstream of the design, upstream of the $7.4 billion.
Which lands us on the reframe. A model company can raise $7.4 billion and still not become a semiconductor company. Design is slow — years, and billions. Manufacturing is the real gate, and DeepSeek doesn't own it.
Windsor is blunt about the ceiling: DeepSeek "has almost no chance of selling silicon outside of China unless it gets access to leading-edge manufacturing." Bottled inside a supply-constrained domestic market, this chip is a cost center, not a product line.
So why do it at all? Because the deeper signal is that DeepSeek does not expect the squeeze to end.
A company famous for doing more with less compute — for owning no hardware — is now planning gigawatt-scale data centers in Inner Mongolia and designing its own accelerator. That reversal only makes sense if scarcity is permanent. Vertical integration is a hedge against a supply constraint DeepSeek has concluded will never lift.
And the loser is not Nvidia, already at zero in China. The loser is Huawei.
Huawei's dominance was always a gift of US export controls — Nvidia's absence handed it half the market. But that dominance depends on its customers having nowhere else to go. When DeepSeek, Alibaba, and Baidu all design their own chips, Huawei's grip on its captive market loosens from the inside.
The irony is exact. Huawei controls the line those chips would have to be built on. The company best positioned to strangle its rivals' escape is the one they are all trying to escape.
Which brings us back to a chief executive in a budget hotel, watching his tape-out window vanish. DeepSeek has the money he lacked. It does not have the one thing that mattered to him — a guaranteed place on the only line that counts. Until it does, the chip is a bet on scarcity, not a way out of it.