DeepSeek Pursues Its Own AI Chip as Funding Reports Point to a Larger Infrastructure Push

An inference chip aimed at controlling computing costs

DeepSeek is developing its own artificial intelligence chip for inference, the work of generating answers from trained models, according to reports citing three people familiar with the project. The effort has reportedly been underway for about a year, with the Chinese AI developer recruiting semiconductor engineers and discussing production requirements with chip design firms, foundries and memory suppliers.

The immediate objective is to reduce dependence on Nvidia and Huawei, the two suppliers whose processors DeepSeek has used to build and operate its systems. The project remains at an early stage. DeepSeek has not publicly confirmed it, and no manufacturing partner, production schedule or chip specifications have been announced.

The hardware effort is emerging alongside reports of a much larger infrastructure expansion. One account describes a planned data center containing at least 160,000 Huawei AI chips, while funding reports range from an initial proposed $7 billion raise to a subsequent round that could approach $15 billion. Those accounts describe different financing stages and cannot be treated as a single confirmed transaction.

A proprietary processor would extend DeepSeek's focus on making AI less expensive to operate. Its DeepSeek-V3 and R1 models brought attention to the company's ability to deliver capable systems with an emphasis on computing efficiency. Designing hardware introduces a different challenge: DeepSeek would need to turn that efficiency strategy into a processor that can be manufactured, supported by software and operated reliably at scale.

For now, this is a development project rather than a replacement for existing equipment. Its significance lies in the direction of DeepSeek's spending and engineering effort, not in any demonstrated chip performance or immediate loss of business for its suppliers.

Why DeepSeek is targeting inference

The proposed chip is intended primarily for inference rather than training. Training creates or updates a model by processing large amounts of data. Inference happens afterward, each time a user submits a question and the trained model generates a response.

That distinction matters financially. Training can involve concentrated periods of intensive computing, while a popular service must keep answering requests throughout the day. The continuing expense depends on how much work each processor can handle, how much electricity it consumes and how consistently the system stays available.

Computing can account for more than half of expenses at many AI companies, although DeepSeek has not disclosed a comparable breakdown of its own costs. Rising demand for its models has made infrastructure one of its largest operating expenses. A processor designed around its actual serving workloads could potentially reduce the cost of each answer.

There is no public evidence yet that the proposed chip will deliver those savings. A useful comparison would have to include manufacturing costs, memory, electricity, software development and the number of requests the system can process at an acceptable speed. A cheaper processor does not necessarily mean a cheaper service if it requires more equipment or runs the model less efficiently.

The project also should not be read as a plan to replace every processor DeepSeek uses. A chip designed for answering requests may leave the company dependent on other hardware for training new models and for tasks its own design does not support.

Huawei remains central to the expansion

DeepSeek has relied on both Nvidia and Huawei hardware, with reports describing a recent shift toward Huawei. A version of its V4 model released earlier in 2026 was built around Huawei's Ascend processors. Another account places the preview releases of V4-Pro and V4-Flash in April.

The reported plan for a data center with at least 160,000 Huawei AI chips suggests that purchasing existing processors and developing proprietary hardware are proceeding together. That is consistent with the chip project's early stage: DeepSeek still needs equipment to serve users while its engineers work on a design that has not reached confirmed production.

Huawei's position makes the effort more than a story about Nvidia. One market estimate puts China's domestic AI chip market at about $50 billion and Huawei's share at roughly half. On those figures, Huawei's portion would be around $25 billion. This is a broad market estimate, not a measure of DeepSeek's purchases or a forecast of revenue its chip could displace.

A successful design could give DeepSeek more control over its own computing capacity while reducing purchases from either supplier. It would not necessarily become a product sold to other companies. No commercial sales strategy has been disclosed, and a processor built for DeepSeek's internal workloads could remain an internal tool.

The effort follows the wider Chinese AI chip push described in October 2025, when Huawei, Alibaba and other domestic developers were expanding alternatives to Nvidia. DeepSeek's reported move adds a model developer to that competition, with hardware requirements shaped by the service it operates.

Design is only the first part of the challenge

DeepSeek has reportedly increased hiring in chip architecture, verification and software enablement. Those roles cover distinct parts of the process. Architects decide how a processor performs its work. Verification engineers check whether the design behaves correctly. Software engineers make it possible for models and applications to use the hardware.

The company has also reportedly held discussions with chip design businesses, foundries and memory suppliers. A foundry manufactures processors from another company's design. Memory suppliers provide components that store model data and move it to the processor quickly enough to keep the system working efficiently.

These discussions do not establish that DeepSeek has secured production capacity. No foundry or manufacturing partner has been confirmed. There is also no announced date for tapeout, the point at which a completed chip design is submitted for fabrication.

Moving from architecture to tapeout and then mass production typically takes well over a year, and producing a competitive AI processor can take years. Even after initial samples arrive, the company would need to test them, address defects and establish dependable production. Software readiness would be another requirement before the chip could carry substantial service traffic.

US export restrictions complicate that process. Reports describe limits on Chinese access to Nvidia's most advanced processors, advanced overseas chip manufacturing and high bandwidth memory. That memory supplies data at the high rates AI processors need. Restrictions affecting either fabrication or memory could constrain the performance of a finished design or delay deployment.

Funding accounts describe different stages

The financing picture requires particular care. The initial account describes DeepSeek seeking its first external funding round, targeting about $7 billion at a valuation of $52 billion to $59 billion. At that stage, outside investment was presented as a departure from the company's previous reluctance to accept it.

A separate account describes that first round as already completed in June, raising approximately $7.4 billion at a valuation above $50 billion. It then describes a new round seeking at least $12 billion, potentially approaching $15 billion, after an original target of about $7.5 billion at a roughly $75 billion valuation.

The accounts therefore differ both in the status of the first round and in the scale of the funding under discussion. They may reflect successive developments, but DeepSeek has not publicly reconciled them. The proposed $7 billion round should not be presented as still pending without acknowledging the later report of a completed transaction.

In the account of the larger subsequent round, battery maker CATL and Tencent are contributing the biggest shares. It also says DeepSeek plans to restructure for an initial public offering in early 2027 after the financing closes. CATL declined to comment, while Tencent and DeepSeek did not respond.

Those reported plans would create room for spending on data centers, processors and the engineering team needed for a proprietary chip. They do not establish how much money would be assigned to the chip project. No budget, investor allocation or confirmed listing timetable has been disclosed.

How the reported developments fit together

The reports place the chip effort alongside model releases and financing plans, but several dates remain approximate. The clearest sequence separates reported events from proposals that have not been confirmed:

• 2024: Founder Liang Wenfeng described chip export controls as a real challenge in an interview.
• About a year before the chip reports: DeepSeek reportedly began the development effort and discussions with potential design, manufacturing and memory partners.
• April 2026: A funding account places the preview releases of V4-Pro and V4-Flash in April. Separate reporting describes a V4 version built around Huawei Ascend processors earlier in 2026.
• June: The later financing account says DeepSeek completed its first external round, raising about $7.4 billion at a valuation above $50 billion.
• Early 2027: DeepSeek is reportedly considering an initial public offering after a further financing round and restructuring.

There is no corresponding date for chip production in that sequence. DeepSeek has not announced a prototype, a tapeout milestone or a launch. The reported early 2027 listing ambition is a financial plan, not a deadline for the processor.

The distinction also prevents model progress from being mistaken for hardware progress. A model running on Huawei processors demonstrates work on that platform; it does not demonstrate that DeepSeek's own silicon is ready.

What this means for Nvidia and the chip market

Nvidia shares reportedly slipped about 0.5% in premarket trading following news of the project. That was a modest market reaction, not evidence of a measurable change in orders or revenue. With no chip ready for deployment, any commercial effect remains prospective.

DeepSeek's effort fits a wider movement among AI developers toward proprietary hardware. Reports also describe an OpenAI inference chip developed with Broadcom and exploratory chip work at Anthropic. The shared incentive is to gain more control over the recurring cost of running models, although these projects differ in their maturity and access to manufacturing.

DeepSeek faces an additional constraint: owning a design would not remove restrictions on how it is manufactured or which memory it can obtain. Reducing dependence on Nvidia could still leave the company reliant on external fabrication, memory and other components.

The next meaningful developments would be confirmation of a manufacturing partner, details of the chip architecture, progress toward tapeout and evidence of performance on DeepSeek's models. Completed financing would clarify the resources available. None of those hardware milestones has an announced date.

The strongest conclusion is therefore narrower than an immediate competitive breakthrough. DeepSeek is reportedly trying to control more of the infrastructure behind its AI services, while continuing to depend on established suppliers. Whether that effort reduces costs will depend on the finished system, not simply on who owns the processor design.

Key Points

• DeepSeek is reportedly developing an inference chip after about a year of work, but has not publicly confirmed the project.
• The effort could reduce reliance on both Nvidia and Huawei. No foundry, production partner or launch date has been announced.
• Export restrictions affecting advanced manufacturing and high bandwidth memory could limit the project's performance or timing.
• Funding accounts describe an initial proposed $7 billion raise, a reported $7.4 billion completed round, and a subsequent round that could approach $15 billion.
• A separate expansion plan includes at least 160,000 Huawei AI chips and a reported ambition for an early 2027 listing.
• No public benchmarks or production milestones yet establish that DeepSeek's chip can lower the cost of serving its models.