Is DeepSeek Building a Custom AI Chip?
Inside the July 2026 Reuters Report

On June 24, 2026, OpenAI and Broadcom unveiled Jalapeño — a custom inference chip built in just nine months. On July 2, Anthropic was reported in talks with Samsung for a 2nm custom chip. Then on July 7, Reuters cited three sources saying DeepSeek is developing its own AI inference chip. This is not a China story. It is an industry-wide shift away from general-purpose GPUs toward custom silicon — and DeepSeek is the latest entrant. This article breaks down what is confirmed, what remains rumor, and why every major AI lab is now building its own chip.

01 The Global Custom Silicon Wave (Context First)

Understanding DeepSeek's chip rumor requires a global baseline. TrendForce data for 2026 shows: hyperscaler custom AI chip shipment growth at 44.6%, far outpacing general-purpose GPU growth at 16.1%. Custom silicon is, for the first time, significantly outgrowing GPUs in shipment velocity.

The inference cost problem is the same for every AI company: inference is the rent payment — continuous, user-scale, and growing. Training is the down payment. Custom ASICs target rent reduction.

02 What Reuters Actually Reported — and What DeepSeek Has Not Confirmed

The Reuters report published July 7, 2026 contains five verifiable claims. Here is a fact-by-fact breakdown:

• Target use case: The chip is designed for inference (inference), not training — consistent with every major custom chip project launched in 2025–2026.
• Project start: Approximately one year prior to the report (mid-2025), described as still in early stages.
• Supply chain contacts: DeepSeek is in talks with chip design firms, foundries, and memory suppliers.
• Hiring approach: Engineers are being recruited privately, not through public job boards — consistent with stealth hardware programs at other companies before announcement.
• Dual-dependency reduction: Success would reduce reliance on both Nvidia and Huawei Ascend — noteworthy since DeepSeek V4 already runs on Ascend.

What has not been confirmed: DeepSeek has issued no press release, blog post, or social media statement confirming the chip program. The project may be restructured, delayed, or cancelled before any public announcement.

Indirect evidence strengthens credibility: DeepSeek's June 2026 Series A raised approximately $7.4 billion (RMB 51 billion), with disclosed uses including custom AI chip development and domestic compute center expansion. The company's UE8M0 FP8 data format has been interpreted by engineers as hardware-software co-design signaling for domestic chips.

Primary sources for this article:

Reuters — DeepSeek developing own AI chip (July 7, 2026)

OpenAI Official Blog — OpenAI & Broadcom Jalapeño Inference Chip Announcement

Wall Street Journal — Alibaba AI chip to fill Nvidia void

03 What CEO Liang Wenfeng Has Said About Chips

Liang Wenfeng (梁文锋), DeepSeek's founder and CEO, has given very few public interviews. The most substantive are two deep-dives with Chinese publication Waves (Angyong) in May 2023 and July 2024. Key chip-relevant quotes:

• On the real constraint: "Our real challenge has never been funding — it is the export ban on advanced chips." (July 2024, Waves interview)
• On the efficiency gap: Domestic AI training efficiency lags international peers by roughly one generation; data efficiency by another. Combined, Chinese labs need approximately four times the compute to achieve equivalent results.
• On the ecosystem problem: "Many domestic chips fail not because of hardware, but because they lack a developer community — only second-hand information. China must have people at the technological frontier."
• On compute hunger: "For researchers, the hunger for compute is endless. We will deliberately deploy as much compute as possible."

Critical distinction: Liang Wenfeng's quotes establish strategic motivation — export constraints, efficiency gaps, compute hunger. The Reuters report covers corporate action — hiring, supplier talks. Do not conflate the two when citing.

04 Alibaba T-Head Is Already Shipping: Jack Ma's 2018 Bet Pays Off

Unlike DeepSeek's early-stage rumor, Alibaba's chip program is a measured, multi-year execution now in commercial production. In September 2018, Jack Ma personally named the new chip unit T-Head Semiconductor (平头哥, "honey badger" — fearless). The chip unit was elevated to group strategic priority from day one.

Key commercial metrics (H1 2026):

• 560,000+ units shipped cumulatively
• Billion-yuan (RMB) annual revenue run rate
• 400+ enterprise customers using Zhenwu clusters; clients include Alibaba Cloud and China Unicom
• T-Head registered capital increased to RMB 1 billion (June 2026)
• Alibaba committed RMB 380 billion (~$52B) over three years to cloud and AI infrastructure including chips and liquid cooling
• Unlike Huawei Ascend (closed ecosystem), Zhenwu is designed to be CUDA-compatible, lowering migration cost for engineers

For deeper technical analysis of the Zhenwu 810E, see:

Caixin Global — Alibaba's New Processor Shows Applications Are Key to AI Chip Success

05 Why Tech Giants Build Custom AI Chips: Cost, Control, and the Nvidia Tax

The economics of inference at scale make custom silicon economically rational for any AI company above a threshold size. Five drivers, ranked by importance:

• 1. Unit economics — inference is the rent payment. Training is a one-time capital outlay. Inference scales with users, 24/7. Morgan Stanley estimated a 24,000-GPU Blackwell cluster costs approximately $852 million in hardware; an equivalent Google TPU cluster costs approximately $99 million (Reuters Breakingviews). Nvidia's data center GPU gross margin exceeds 70%. Custom ASICs convert a permanent per-unit "Nvidia tax" into a one-time R&D investment. At scale, SemiAnalysis and Bernstein estimate custom ASICs deliver 40–65% TCO reduction versus general-purpose GPUs on mature inference workloads.
• 2. Supply chain resilience. US export controls on H100, H800, and H20 GPUs have sequentially cut Chinese AI labs' access to frontier compute. Even US companies face Nvidia allocation queues. Supply chain security means predictability — not being at the mercy of a single vendor or a single government's policy.
• 3. Hardware-software co-design. DeepSeek's UE8M0 FP8 and MLA architecture are optimized for specific hardware characteristics. OpenAI's Jalapeño is designed around real ChatGPT serving patterns — KV cache sizes, request batching, latency targets. General-purpose GPUs sacrifice efficiency for flexibility; custom ASICs trade flexibility for efficiency on a known workload.
• 4. Competitive leverage and differentiation. Even a partial Nvidia alternative strengthens negotiating position on pricing, allocation, and roadmap. Building a chip also enables a full-stack narrative: model + cloud + silicon (Alibaba's "golden triangle"; OpenAI full-stack infrastructure).
• 5. Energy and sustainability. Inference chips prioritize performance per watt. In gigawatt-scale data centers, power and cooling costs equal or exceed hardware costs. ASICs eliminate large portions of general-purpose GPU circuitry that AI workloads never use.

06 Inference Chips vs Training GPUs: Why the Industry Is Splitting

Almost every custom chip program in 2025–2026 targets inference, not training. This is not accidental:

Conclusion: Training remains Nvidia's domain. Inference is the primary battleground for custom ASICs. DeepSeek targeting inference specifically is the economically rational choice.

Risks worth noting for early-stage programs:

• High early-stage failure rate: From tape-out to stable mass production typically takes 2–4 years; early programs can fail due to yield, architecture drift, or foundry constraints.
• Meta's full rebuild: Meta's MTIA chip program was completely redesigned from scratch before reaching production — a cautionary data point.
• Architecture change risk: If the next model generation shifts architectural assumptions, an ASIC optimized for today's patterns may become obsolete quickly.
• Partnership parallel track: DeepSeek continues Huawei Ascend collaboration while pursuing custom silicon — a rational dual-track hedge, not contradiction.

07 Six-Step Decision Framework for Tech Leaders

As custom silicon reshapes AI infrastructure, here is a practical framework for technology decision-makers evaluating the impact on their organizations:

1. Quantify your current inference cost (calculate your Nvidia tax): Review the last 12 months of GPU compute spend, separating training from inference. If inference exceeds 50%, custom silicon alternatives warrant serious evaluation.
2. Profile your workload characteristics: Analyze request pattern stability, batch size predictability, and latency requirements. Stable, predictable inference is the best-fit target for ASIC migration.
3. Assess CUDA ecosystem dependency depth: Map your inference codebase's reliance on cuDNN, NCCL, and other CUDA-proprietary APIs. Higher dependency means higher migration cost and longer timelines.
4. Research vendor roadmaps by compliance requirement: Alibaba Zhenwu (CUDA-compatible, domestic foundry), Huawei Ascend (closed ecosystem, regulatory-friendly for China), international ASICs (Google TPU, AWS Trainium) — filter by your business jurisdiction and compliance needs.
5. Design a hybrid compute architecture: Do not replace everything at once. Maintain Nvidia GPUs for flexible training and experimentation; add custom ASICs for stable high-volume inference; include bare-metal physical machines for low-latency AI Agent workloads.
6. Build supply chain diversification: Maintain at least two independent compute supply chains. Review export control developments quarterly and update contingency plans accordingly.

08 FAQ

Q: Is DeepSeek really building its own AI chip?
A: According to a July 7, 2026 Reuters report citing three sources, DeepSeek is in the early stages of developing a custom AI chip optimized for inference. DeepSeek has not officially confirmed the project. DeepSeek is reportedly hiring chip engineers privately and talking to foundries and memory suppliers.

Q: Did Liang Wenfeng announce a chip program?
A: No. In 2024 interviews, he described export controls on advanced chips as DeepSeek's biggest challenge — establishing motivation, not announcing a program. The Reuters report covers corporate action, not a founder declaration.

Q: How is Alibaba different from DeepSeek on chips?
A: Completely different stage. Alibaba's T-Head has been building chips since 2018 and is in mass production with 560,000+ Zhenwu chips shipped. DeepSeek is reportedly in early-stage R&D with nothing publicly confirmed.

Q: Why inference chips, not training chips?
A: Inference workloads are repetitive and predictable — ideal for ASIC optimization. Training requires CUDA flexibility and peak throughput that Nvidia still dominates. Economically, inference scales with users 24/7; it is the rent payment, not the down payment.

Q: Is this about national security or saving money?
A: Both. Unit economics is the primary driver — cutting the Nvidia tax and per-token costs at scale. Export controls and supply chain risk accelerate an already economically-motivated shift.

Last updated: July 10, 2026 | Disclaimer: DeepSeek has not officially confirmed the chip project as of this writing. All references are based on public reporting from Reuters, WSJ, and Caixin. Verify against latest news before publishing.

The DeepSeek and Alibaba stories reveal a structural shift: inference cost has become the defining variable in AI commercialization. Whether custom ASIC or bare-metal Apple Silicon, every path is pursuing the same goal — running more AI inference at lower cost with greater control. For AI teams running production workloads today, GPU cloud volatility — unpredictable performance in shared resource pools, restricted environments, no persistent state — creates real operational risk. Generic GPU cloud has three well-known limitations: shared resources with unpredictable performance at peak; restricted environments that block private dependencies; high per-token cost at scale. ZUKCLOUD bare-metal Mac mini cloud nodes address a specific and growing need: dedicated Apple Silicon physical machines, zero hypervisor overhead, 7×24 AI Agent uptime, and flexible day/week/month pricing — a stable, controllable inference environment while the custom ASIC market matures.