DeepSeek's latest moves look contradictory at first glance. One one hand, the company has been quietly developing its own AI inference chip for about a year, according to Reuters. It has been in talks with chip designers, foundries, and memory suppliers. The project remains in early stages, and recruiting has been conducted privately — no public job postings.
On the other hand, DeepSeek — along with Alibaba and other Chinese AI companies — is set to purchase a limited number of Nvidia H200 GPUs. The total is reportedly less than 200,000 units, fewer than half of what these companies applied for earlier this year. The usage restrictions are clear: training only, no inference.
Together, these two moves tell a coherent story: buying buys time; building buys independence.
Training is a one-time investment — intense, expensive, but finite. Inference is a permanent utility bill — every user query, every API call, every token generated adds to the cost. Over the full lifecycle of an AI system, inference can account for 80% to 90% of total compute costs, according to some industry estimates.
Building its own inference chips is DeepSeek's ultimate cost-control play. But custom silicon takes years to develop and billions of dollars to bring to production. In the meantime, the model still needs to be trained. The H200 purchase isn't a contradiction — it's a bridge. DeepSeek isn't making a choice between buying and building. It's doing the math on both.
The Policy Window: What 200,000 H200 Units Actually Mean
The H200 shipments aren't just a supply chain story — they're a political barometer.
The U.S. government approved exports of H200 to China back in December 2025. Nvidia and AMD agreed to remit 15% of revenue from China-bound chip sales to the U.S. government. But approval from Beijing took longer to materialize, held up by concerns over cybersecurity risks and the impact on domestic semiconductor ambitions.
The resulting compromise: limited shipments, less than half of what Chinese firms requested, with strict usage conditions. Training only. Inference must run on non-U.S. chips.
This isn't an opening — it's a pressure valve. The door opens just wide enough to prevent the training pipeline from seizing up entirely. The volume is sufficient to keep the lights on, not to liberate China's AI labs from foreign suppliers.
AI competition always comes back to supply chains. And supply chains always come back to politics — even when the product is just a piece of silicon.
Why Model Makers Are Becoming Chip Makers
DeepSeek isn't alone in this move. The trend is global.
OpenAI unveiled its first custom inference chip, codenamed Jalapeño, last month — designed with Broadcom and already in lab testing. Anthropic reportedly explored chip development earlier this year. Google has TPU. Amazon has Trainium. Meta has MTIA.
Every major AI company eventually faces the same question: buying chips off the shelf leaves too much value on the table.
But DeepSeek is different from OpenAI or Google in one key way. It doesn't operate a cloud business. It doesn't sell chips to third parties. Its sole reason for building silicon — if the project goes all the way — is to make its own models cheaper and faster to run on its own hardware.
A custom inference chip designed specifically for DeepSeek's architecture can strip away unnecessary functional units. Every square millimeter of silicon can be optimized for exactly the operations DeepSeek's models perform. The company's V3.1 model, which introduced a new FP8 data format (UE8M0), has been interpreted by industry observers as a sign that the algorithmic team is already designing with next-generation Chinese chips in mind — designing the model around the hardware before the hardware even exists.
When model companies build chips, they're turning algorithm advantage into hardware moat. That's the deepest competitive trench you can dig.
The Incumbent's Dilemma: Huawei Loses a Customer — and Gains a Potential Rival
DeepSeek has been a showcase customer for Huawei's Ascend chips. In April, DeepSeek announced that its V4 model had been fully adapted to run on Ascend processors, with Huawei confirming its chips participated in part of V4-Flash training. The two teams likely worked together during the development phase.
Now DeepSeek is becoming a competitor.
One source told Reuters that the chip project is intended to reduce dependence on both Nvidia and Huawei. DeepSeek is moving from "choose one supplier" to "become the supplier."
For Nvidia, the impact is limited — it already has minimal share in China. For Huawei, the stakes are higher. DeepSeek was a major Ascend customer. If its own inference chips succeed, the demand for Ascend in inference workloads could shrink significantly.
Meanwhile, Alibaba, Baidu, and ByteDance are also developing their own AI chips. The domestic Chinese AI chip market is becoming crowded — and every new entrant is a potential customer lost for the incumbents.
DeepSeek isn't abandoning Huawei — it's telling every chip vendor that it can build its own future. That's a negotiating position more powerful than any volume purchase.
The Real Barriers: Why Building a Chip Is Harder Than Building a Model
Quiet momentum doesn't mean easy path. DeepSeek faces real engineering hurdles.
Manufacturing. U.S. export controls restrict Chinese chip designers from using the most advanced foreign foundries. DeepSeek's chip would likely rely on domestic fabs with older process nodes — a meaningful performance gap compared to TSMC or Samsung's bleeding-edge processes.
Memory. AI inference chips are memory-hungry. High-bandwidth memory (HBM) is dominated by three suppliers — SK Hynix, Samsung, and Micron — all subject to export restrictions.
Software. A piece of silicon is useless without a full stack: compilers, drivers, operator libraries, framework integration, and continuous optimization. Nvidia spent two decades building CUDA. Even if DeepSeek gets the hardware right, it still needs to make its models run stably and efficiently on its own silicon.
On the bright side, funding isn't the problem. DeepSeek completed its first external funding round in June, raising approximately 51 billion yuan at a valuation around 400 billion yuan. But even with ample cash, a production-ready AI inference chip typically takes years and multiple billions of dollars — and multiple tape-out attempts often fail before one succeeds. The chip is the easy part. The ecosystem is the hard part.
Why Inference: The Structural Shift Everyone Is Missing
DeepSeek chose inference over training as its chip entry point — and that choice is telling.
Training is a one-time sprint. Inference is a perpetual marathon.
Every token generated, every API request served, every conversation held adds to the inference bill. As prices for AI APIs continue to fall — DeepSeek's V4 Flash charges 2 yuan per million output tokens, less than one-twentieth of GPT-5.5's pricing — the cost of inference hardware increasingly determines the profitability of AI service providers.
Goldman Sachs' 2026 compute industry report estimates that inference costs already set the break-even point for large model service providers. Relying solely on external GPUs makes it increasingly difficult to sustain low API prices.
Last month, DeepSeek introduced "peak-valley pricing" — the first such mechanism in China's AI industry — charging higher rates during the day and lower rates at night. That's a clear signal: inference demand is tight enough that the company is using price signals to shift batch workloads to off-peak hours.
Training is a capital expenditure. Inference is an operating expense. And when the OPEX line gets big enough, it starts looking like a strategic problem — not a line item.
The Endgame: Where DeepSeek's Compute Chessboard Is Headed
If you map out DeepSeek's compute strategy as a chessboard, the pieces are already in position.
In the short term: H200 shipments land, easing training pressure. Inference still runs on existing capacity and domestic chips.
In the medium term: Custom inference chips move through tape-out, validation, and production — gradually replacing purchased compute in DeepSeek's own API and agent services.
In the long term: A hybrid model — self-built inference chips paired with externally sourced training compute — reduces single-supplier dependence while avoiding the massive upfront cost of a complete hardware replacement.
DeepSeek's chip project, if it reaches production, would mark a meaningful shift from "application-layer innovation" toward "infrastructure-layer capability" in the Chinese AI industry. But for now, it remains more of a strategic reserve than an immediate supply chain disruption.
DeepSeek isn't building a chip. It's building leverage — and that leverage will pay dividends whether the chip ever ships or not.
P.S. If you're running an AI company and watching DeepSeek's moves, the question isn't "Should I build chips too?" It's "When does my inference bill cross the break-even point where building starts making sense?" For a company serving billions of tokens a day, that day is closer than you think.
