DeepSeek pivots to build its own AI chip, aiming to cut reliance on Nvidia

Chinese AI startup DeepSeek is designing its own inference chip, people familiar with the matter told Reuters on July 7, 2026, a move the company hopes will reduce dependence on foreign suppliers such as Nvidia and Huawei and mark a major strategic shift for the firm. The effort is at an early stage and focuses on inference — running trained models for users — rather than training new models from scratch, the sources said.

The shift matters because DeepSeek has grown rapidly as a low-cost alternatives provider and, this year, opened to outside capital: Reuters reported in mid‑June that DeepSeek was set to raise roughly $7 billion in a maiden funding round at a valuation Reuters put between $52 billion and $59 billion, though earlier reporting cited slightly different totals. Designing in‑house inference silicon would help control token costs and hardware margins as usage scales, a strategic aim that also dovetails with growing U.S. scrutiny of the company’s access to advanced chips.

Chip designed for inference, not training

According to Reuters, DeepSeek’s chip programme is targeted at inference workloads — the stage “in which a trained model generates responses for users” — rather than the far more power‑hungry task of training large models from scratch. Focusing on inference narrows the engineering and capital challenge: inference chips can be smaller, cheaper to prototype, and optimized for throughput and power efficiency for deployed models. But the project is “at an early stage,” Reuters cautioned, and DeepSeek has been speaking with multiple chip designers, foundries and memory suppliers as it explores options.

That technical choice also aligns with DeepSeek’s identity. The company first drew attention in January 2025 for an inexpensive but capable model that captured market share quickly, helping Chinese LLM market share rise from 3% to 13% in the two months after DeepSeek’s R1 launch, Reuters reported. An inference chip could lock in that cost advantage by lowering operational spending per token.

Why DeepSeek wants to cut Nvidia and Huawei out

Reuters frames the chip push as a strategic bid to reduce reliance on Nvidia and Huawei hardware, both of which currently underpin much of the global AI stack. The company’s move follows a spate of events this year: Reuters reported on June 16 that DeepSeek was closing more than $7 billion in outside funding, and on June 17 that U.S. officials had considered but held off on blacklisting the firm over national‑security concerns.

Costs and export risk are obvious drivers. Building bespoke inference silicon can lower per‑token costs and insulate performance from export controls or supplier decisions. But the shift also arrives amid an intensified capital cycle: DeepSeek’s apparent embrace of external funds after years of staying private gives it both the cash and investor pressure to control margins and IP flows.

Skeptics note limits and risks. U.S. officials voiced national‑security concerns this summer, and Reuters reported allegations that DeepSeek attempted to use shell companies in Southeast Asia to obtain advanced American chips, a claim the startup has disputed in past coverage. Frontier labs such as Anthropic and OpenAI have publicly warned lawmakers about aggressive data‑extraction attempts and other behaviour they attribute to some Chinese rivals, underscoring geopolitical friction around compute supply chains.

Competitive context matters too. Nvidia remains the dominant supplier for both training and inference in high‑performance settings; Huawei offers an alternative inside China. Western peers have poured billions into chip and stack development: Anthropic and OpenAI have emphasized scale and cloud partnerships rather than bespoke system‑on‑chip strategies. DeepSeek’s differentiator so far has been price; delivering a working inference chip at lower cost would be a concrete way to sustain that edge.

The picture of DeepSeek’s finances is mixed in reporting: Reuters’ June coverage cited “more than 50 billion yuan ($7.40 billion)” raised at a valuation above $50 billion, while later pieces described a slated $7 billion round valuing the company at $52–$59 billion — differences likely reflecting reporting windows and rounding.

Looking ahead, the critical near‑term milestones are technical partnerships and foundry commitments that Reuters says DeepSeek is pursuing, plus any regulatory moves from Washington. Investors and rivals will be watching whether the startup can translate cash and model popularity into chip expertise — and whether that will blunt U.S. leverage over compute supplies. If DeepSeek pulls it off, it changes supplier dynamics; if the effort stalls, the firm risks spending a high fraction of its fresh capital on an uncertain hardware bet.