Reuters published an exclusive this morning, citing three people familiar with the matter, that Chinese AI startup DeepSeek is developing its own AI chip designed to reduce dependence on Nvidia and Huawei processors. Nvidia fell 1.5% in premarket on the news. The Philadelphia Semiconductor Index closed down more than 5%. Intel dropped 10%, AMD fell 8%, Applied Materials dropped 10%, Marvell fell 6%. The DeepSeek announcement was one of three simultaneous pressures on the sector Tuesday, alongside Samsung's paradoxical earnings selloff and the Iran escalation, but it is the one with the longest structural tail and the most important implications for how investors think about the US-China AI race.
What DeepSeek Is Actually Doing
Reuters' sources say DeepSeek is building its own AI inference chip, aimed at reducing its reliance on the Nvidia H100s and H800s it has been using, and on Huawei's Ascend 910C, which has been China's primary domestic alternative since US export controls cut off access to the most advanced Nvidia products. DeepSeek has not confirmed the report publicly, and details on timeline, architecture, and manufacturing partner remain sparse.
The inference focus is significant. Training chips and inference chips have very different design requirements. Training requires massive parallelism, high memory bandwidth, and the ability to handle enormous matrix multiplications across billions of parameters. Inference, which is the process of running a trained model to generate outputs, can be optimized for lower power consumption, lower latency, and specific workload profiles. Designing a competitive inference chip is considerably more tractable than designing a competitive training chip, which is why companies from Apple to Amazon have pursued custom silicon for inference while continuing to use Nvidia for training.
DeepSeek's core competitive advantage has always been efficiency. Its R1 and V3 models demonstrated in January 2026 that a Chinese lab could produce frontier-quality reasoning capabilities at a fraction of the compute cost that Western labs assumed was necessary. That efficiency breakthrough, achieved partly through algorithmic innovation and partly through architectural choices, is exactly the kind of expertise that translates well into custom inference silicon design. A chip built specifically to run DeepSeek's model architecture efficiently could, if successful, allow DeepSeek to serve its user base at dramatically lower marginal cost without needing to source chips from either Nvidia or Huawei.
The US-China AI Race: Where It Actually Stands
The DeepSeek announcement is the latest chapter in a competition that has been developing differently than the market's initial assumptions suggested. When the original DeepSeek-R1 shock hit markets in January 2026 and briefly wiped nearly $600 billion from Nvidia's market cap in a single day, the initial interpretation was that Chinese AI had caught up and that Nvidia's compute advantage was irrelevant. That interpretation proved too strong. The more accurate reading was that China had found ways to do more with less, partly because US export controls forced algorithmic innovation as a substitute for raw compute.
The export control regime has created a specific competitive dynamic. China cannot legally access Nvidia's H100, A100, or the newer B100 and B200 chips. Huawei's Ascend 910C provides a domestic alternative that is roughly comparable to the H100 in some benchmarks but significantly behind in others, particularly for the most demanding frontier training runs. The result is a bifurcated world: the US and its allies can train larger, more capable foundation models because they have access to more compute, while China has developed extraordinary expertise in efficiency, compression, and architectural innovation because compute scarcity forced those capabilities.
DeepSeek building its own chip is the logical extension of that dynamic. If you have been forced to extract maximum value from constrained compute resources for two years, the next step is to design the chip that your optimized software actually needs, rather than trying to adapt your software to the chips you can access. TSMC, the primary manufacturer for leading-edge AI chips, is off-limits to Chinese entities under US restrictions, so DeepSeek would need to work with SMIC, China's leading foundry, which operates at 7nm and is working on 5nm but remains approximately two to three process generations behind TSMC's 3nm and 2nm nodes. That is a real constraint on what is achievable, particularly for training-class chips. For inference chips, where power efficiency and latency optimization matter more than raw transistor density, the gap is more manageable.
The strategic picture is one of two parallel AI industrial complexes, each with structural advantages the other cannot easily replicate. The US complex has access to the most advanced semiconductor manufacturing, the most powerful training chips, and the deepest capital markets for AI investment. The China complex has a forced efficiency culture, a massive domestic user base, and the state industrial policy capacity to build vertically integrated AI supply chains across compute, models, and applications. Neither is going away. Neither is definitively winning.
What It Actually Means for the Semiconductor Trade
The market's immediate reaction to the DeepSeek chip news was to sell semiconductor stocks broadly. That reaction reflects a concern that is real but imprecise in its targeting. If DeepSeek successfully develops a competitive inference chip manufactured at SMIC, it removes demand for Huawei's Ascend products within China's domestic market. It does not directly reduce Nvidia's total addressable market in the near term, since Nvidia is already unable to sell its most advanced chips to Chinese entities under existing export controls.
The more structurally significant implication is a long-term one: every major AI lab in the world, whether it is Google, Amazon, Microsoft, Anthropic, Meta, or now DeepSeek, is working toward reduced dependence on merchant silicon from Nvidia. That convergence does not mean Nvidia's business deteriorates in the short term. It means the revenue trajectory that current multiples require Nvidia to sustain through the late 2020s becomes harder to underwrite with high confidence. At 40-plus times forward earnings, Nvidia is not priced for a world in which its largest customers are designing around it. It is priced for a world in which the current dependency structure persists.
For memory stocks, the DeepSeek development is more ambiguous. A successful DeepSeek inference chip would still require DRAM and NAND, just optimized for the workloads DeepSeek's architecture runs. The structural shift from dependence on Nvidia to custom silicon does not reduce total AI compute demand; it changes which company captures the processing value. Memory sits below that layer in the stack and is needed regardless of whose chip is doing the processing.
The Investor Implication: Two Races, One Portfolio
The US-China AI competition is not zero-sum in the way the semiconductor selloff implies. China developing its own AI chips does not make Micron's HBM demand go away. DeepSeek building efficient models does not make Anthropic less valuable. What it does mean is that the AI market is resolving into two distinct ecosystems, each with internal demand for compute, memory, and software infrastructure.
For investors with US-listed portfolios, the most direct implication is that the assumption of permanent US AI supremacy should be held less confidently than it was twelve months ago. DeepSeek's efficiency innovations in early 2026 forced the entire US AI industry to take Chinese AI seriously as a technical competitor. The chip announcement today says that China is now moving to address the hardware dependency that has been its primary structural constraint. That transition will take years and faces real manufacturing limitations, but the direction is clear.
The second implication is about portfolio construction. The US-China AI race is increasingly a reason to own the infrastructure layer, specifically memory, power, and networking, rather than any single AI application or model provider. Memory is needed by both sides. Power infrastructure is needed by both sides. The model layer, where the competitive dynamics are most intense and most uncertain, is where concentration risk is highest. Today's selloff is, in part, a market recognizing that insight more viscerally than it did last month.
The chips were already selling off before DeepSeek's chip news arrived this morning. Samsung's earnings paradox, the Iran re-escalation, and the structural leverage unwind from the Q2 rally provided more than enough pressure. DeepSeek added the one thing those other catalysts lacked: a specific, credible narrative about why the structural AI chip demand story might be more complicated in 2028 than it looks in 2026. That is not a reason to exit the AI trade. It is a reason to think more carefully about where within it you are positioned.
Explore more exclusive insights at nextfin.ai.
