DeepSeek’s Custom AI Chip Plan Joins a Wider Silicon Migration

Chinese AI startup DeepSeek is designing its own chip to handle user queries, a custom inference processor meant to reduce its reliance on Nvidia and Huawei hardware. The July 7 report on DeepSeek’s chip effort, citing three people familiar with the matter, describes the project as about a year old and in early-stage talks with unnamed chip designers, foundries, and memory suppliers. DeepSeek declined to comment when reached.

The news lands inside a migration that already has names. OpenAI unveiled Jalapeño, its first in-house accelerator built with Broadcom, on June 24, 2026. Anthropic is now in early talks with Samsung over a custom chip, per The Information as reported by TechCrunch on July 2, 2026. DeepSeek’s move is the clearest signal yet that controlling silicon has stopped being optional for leading AI labs; it is now the cost of staying in the game.

Why Inference, Not Training

The DeepSeek chip is being built for inference, the layer of AI computing where an already-trained model answers a user query. Training is the headline cost front, the place where the largest GPU bills live. Inference is the daily operational cost, multiplied across every ChatGPT-style message, code completion, or translation the company’s models serve. As user adoption of AI products compounds, running those models efficiently has become the fastest-growing and most expensive line item in AI infrastructure.

DeepSeek arrives at inference silicon from a base built around efficiency. The lab’s V3 research paper, cited in a CSIS congressional testimony, said the model was trained on 2,788,000 H800 GPU-hours, an estimated $5.576 million compute bill for the final pretraining run. Its successor, V4, released in April 2026, was adapted specifically for Huawei’s Ascend architecture, a Mixture-of-Experts design that activates roughly 37 billion of its total ~1 trillion parameters per inference pass.

The R1 foundation model that preceded V4 was trained on Nvidia H800 chips, a part Washington banned for export to China in late 2023, per the Reuters wire. DeepSeek has used both Nvidia and Huawei silicon across different model generations.

A chip tuned to that workload shape would lean on low-precision arithmetic, optimized attention kernels, fast on-chip SRAM, and high-bandwidth external memory. SemiAnalysis describes what any custom chip program still owes its buyer beyond the silicon itself: manufacturing access, compiler tools, runtime libraries, memory supply, packaging capacity, and developer adoption.

What the inference focus actually buys:

• Inference economics compound. A cheaper per-query cost turns into margin and into who can serve frontier models at scale.
• Efficiency is in DeepSeek’s brand DNA. V3’s reported training bill was the efficiency story that moved markets in early 2025; an inference chip carries that thesis into silicon.
• Western competitors target the same layer. OpenAI’s Jalapeño is built for LLM inference, evidence that the inference market is the obvious first move across the field.

The Wider Silicon Migration

DeepSeek entering the design ring formalizes a pattern OpenAI, Google, and Meta have owned for years. The closest US analogue to DeepSeek on training-cost efficiency is Anthropic, which told TechCrunch its compute strategy still centers on a diversified hardware stack of chips from Google, Amazon, and Nvidia, and that it had nothing further to add on the Samsung talks. Anthropic has hired Clive Chan, an early member of OpenAI’s silicon team, to lead its own program, per The Information as relayed by Technology.org.

OpenAI set the recent tempo. Its in-house accelerator, Jalapeño, went from initial design to manufacturing tape-out in nine months, what OpenAI calls the fastest ASIC development cycle it has ever completed, with Jalapeño rollout and partner details published on OpenAI’s announcement page. The chip was developed with Broadcom handling silicon implementation, board integration, high-performance networking, and scalable production systems; Celestica contributed board, rack, and system expertise.

Deployment of Jalapeño begins by the end of 2026, at gigawatt scale with data center partners including Microsoft, expanding across multiple generations. The architecture is positioned as a blank-slate inference platform, not a general-purpose accelerator adapted from earlier AI workloads. Engineering samples are running ML workloads at production target frequency and power, including GPT-5.3-Codex-Spark.

China’s Chip Stack Splits Open

DeepSeek’s chip push arrives as China’s domestic AI chip market has shifted decisively toward Huawei. Reuters notes that US export curbs on Nvidia’s cutting-edge parts have helped Huawei gain around half of China’s $50 billion domestic AI chip market. The shift is sharp inside 18 months.

Huawei is projected to book roughly $12 billion in AI chip revenue in 2026, up from $7.5 billion in 2025, per the Huawei 2026 AI chip revenue and Nvidia’s zero share report. Major China orders include Alibaba, ByteDance, and Tencent. Nvidia CEO Jensen Huang, in an interview with the Special Competitive Studies Project’s ‘Memos to the President’ podcast, put Nvidia’s China share bluntly: ‘In China, we have now dropped to zero.’

DeepSeek’s V4, released in April 2026, was optimized specifically for Huawei’s Ascend architecture and CANN software framework rather than Nvidia’s CUDA, per Tom’s Hardware. Huawei’s Ascend 950PR is the primary procurement target for China’s largest tech firms, and Huawei engineers reportedly collaborated with DeepSeek ahead of V4’s launch. That single product line lifted Huawei’s profile inside China’s AI stack to a dominant share.

The roughly 50% market share figure is contested ground. Bernstein estimated Huawei at about half of China’s domestic AI chip market in 2026, with Nvidia declining to about 8% over the same window, per a Huawei Central summary. Alibaba and Baidu both have in-house silicon programs aimed at the same customers. The remaining share is what DeepSeek, the cloud giants, and a long tail of Chinese designers are now fighting over.

A Geopolitical Squeeze on Every Supplier

The Reuters report frames DeepSeek’s chip work against US export controls that bar Chinese designers from accessing the most advanced overseas foundries and have cut off high-bandwidth memory supply. SMIC, China’s leading foundry, can manufacture on a 7nm-class N+3 process without EUV lithography, but yields trail TSMC’s equivalent nodes by a wide margin, per Tom’s Hardware.

Even if DeepSeek produces a competitive chip design, the company still needs manufacturing access, compiler tools, runtime libraries, memory supply, packaging capacity, and developer adoption, per SemiAnalysis. Reuters flags the same multi-year hurdle. DeepSeek cannot tap TSMC’s most advanced lines through normal channels.

An inference-focused chip partially answers the supply question: fewer chips needed per query, less reliance on any one supplier, and a workload profile tuned for the silicon DeepSeek can actually buy from domestic partners. The design-and-build cycle still runs multi-year by industry standards, and DeepSeek would have to clear that timeline under conditions Western labs do not face.

Money for the Moonshot

DeepSeek’s chip effort runs alongside the lab’s first outside funding round. Reuters put the round at $7 billion with a valuation between $52 billion and $59 billion. A subsequent report on DeepSeek’s $7.4 billion round and Liang’s check said the actual close totaled more than 50 billion yuan, roughly $7.4 billion, making DeepSeek China’s most valuable AI startup.

Founder Liang Wenfeng committed about 20 billion yuan of his own fortune, the round’s biggest check by a single party. Tencent was considering 10 billion yuan; battery giant CATL was looking at about 5 billion yuan, the largest external sums named so far.

The structure flips the usual venture playbook. Investors did not buy equity in DeepSeek directly; their money went into a limited partnership controlled by Liang, with a five-year lock-up and zero voting rights, per Forbes and The Information reporting. China’s National Artificial Intelligence Industry Investment Fund, a state vehicle, took the only direct equity route and walked away with actual voting rights and no lock-up. The arrangement preserves Liang’s grip on the company while giving DeepSeek the war chest the chip program will likely need.

US frontier labs sit well above DeepSeek on the funding curve. OpenAI closed a $122 billion round in March 2026 at an $852 billion valuation. Anthropic raised $65 billion in late May at $965 billion, per the Forbes report. DeepSeek’s deal is the largest single financing in Chinese startup history; the gap between that sum and the US frontier labs explains why DeepSeek is reaching for in-house silicon rather than trying to out-bid them for finite Nvidia inventory.

The Hurdles Between Blueprint and Chip

DeepSeek’s chip is, today, a set of conversations with design firms, foundries, and memory suppliers and stealth hires, with the company declining to comment when Reuters asked. Anthropic’s program is at an earlier stage still; The Information, via TechCrunch, reported that Anthropic had not yet decided what its chip will be used for, how it will fit into the server, or how powerful it will be as of early July 2026. The nine-month cycle OpenAI showed off with Jalapeño is the upper end of what is possible when capital, infrastructure partners, and own-AI design help all line up, per the Samsung talks report and what Anthropic hasn’t decided.

DeepSeek’s path runs through partners that are not yet public, on a domestic foundry and memory supplier base that still trails TSMC’s best, for a chip that will only narrow the question of who controls the underlying silicon rather than end it. Reuters also flags the unresolved training picture: the R1 foundation model was trained on Nvidia H800 silicon that the US banned for export to China in late 2023, and questions of how any subsequent DeepSeek training runs will land remain open.

Frequently Asked Questions

Is DeepSeek actually making its own chip?

Reuters reported on July 7, 2026, that DeepSeek is designing a custom AI inference chip and has been in early-stage talks with unnamed chip designers, foundries, and memory suppliers for about a year. DeepSeek declined to comment when reached.

Why focus on inference and not training?

Inference is the cost paid every time a trained model serves a user query; cheaper-per-query silicon compounds with scale. DeepSeek’s V3 and V4 already ship optimizations tuned for that workload. Training is the rarer, larger upfront GPU bill.

How does this report move Nvidia?

Nvidia shares slipped about 2% in premarket trading after the Reuters wire, per the same report. Nvidia CEO Jensen Huang separately put the chipmaker’s China share at ‘zero.’

Will this let DeepSeek bypass US export controls?

Only partly. A custom inference chip lowers per-query dependence on foreign silicon, but US export rules still bar Chinese designers from the most advanced overseas foundries and from the bulk of high-bandwidth memory supply, so any DeepSeek silicon must run on domestic partners.