Is DeepSeek Building Its Own AI Chip?
Inside the July 2026 Reuters Report — And Why Every AI Lab Is Doing the Same

Before diving into DeepSeek, the framing matters. "AI companies building custom chips" is a global trend, not a story about Chinese tech nationalism. TrendForce data from 2026 shows cloud-vendor custom AI chip shipments growing at 44.6% year-over-year — far outpacing general-purpose GPU growth of 16.1%. Custom silicon has, for the first time, meaningfully outrun GPUs in growth rate.

Within two weeks in July 2026, four independent chip announcements hit the wire:

• June 24: OpenAI + Broadcom announce Jalapeño (inference ASIC, nine-month design cycle)
• July 2: Anthropic reportedly in talks with Samsung for a 2nm custom chip (The Information)
• July 7: Reuters: DeepSeek developing a custom inference chip
• July 7: The Information: Zhipu AI (China) also evaluating custom chip development

If your article headline only says "China chip," you are leaving half the story — and half the search traffic — on the table. The relevant query for global readers is: "why is every AI lab building custom silicon now?"

The July 7, 2026 Reuters exclusive, citing three people familiar with the matter, reported the following:

• The chip targets inference, not training. DeepSeek is not trying to build an H100 competitor for pre-training workloads.
• The project started approximately one year ago (mid-2025) and remains in early stages. No chip name, architecture details, or tape-out timeline has been disclosed.
• DeepSeek is in talks with chip design firms, foundries, and memory suppliers — the three necessary upstream partners for building an ASIC from scratch.
• Hiring is happening privately, not through public job boards. Chip design engineers have been quietly recruited in recent months.
• If successful, it would reduce dependence on both Nvidia and Huawei Ascend, which is notable because DeepSeek V4 already runs on Ascend and V4-Flash used Ascend for partial training.

Correct framing: "According to Reuters and multiple media citing sources, DeepSeek has initiated a custom inference chip project." Do not write "Liang Wenfeng officially announced chip development." The disclaimer "DeepSeek has not officially confirmed the chip project as of this writing" should appear in any responsible article.

Liang Wenfeng rarely speaks publicly. The most valuable source is two long-form interviews with Chinese tech publication Waves (Anyin) in May 2023 and July 2024. He has never publicly announced a chip program, but three quotes establish the strategic motivation clearly:

• "Our real challenge has never been funding — it's the export ban on advanced chips." (July 2024, Waves interview) — This establishes the constraint, but it is a motivation statement, not a project announcement.
• "Compared to overseas labs, domestic training efficiency lags by roughly 1x and data efficiency by another 1x — meaning we need approximately 4x the compute to achieve the same results." — This number quantifies the competitive cost penalty and creates internal pressure for hardware-software co-design.
• "Many domestic chips fail to gain traction because of the absence of a surrounding technical community. China inevitably needs someone to stand at the frontier." — This frames participation in the chip ecosystem as a mission, not just a business calculation.

Key distinction: "Founder's strategic framing" is not the same as "official project announcement." Reuters reported on company actions (hiring, supplier talks). Liang's quotes explain why those actions are strategically rational. Both have news value; they are not the same type of claim.

If DeepSeek is "rumor + early stage," Alibaba T-Head is "eight years of execution + mass production." This contrast is the most useful framing device in any article on this topic.

2018: Jack Ma's Strategic Decision

At the September 2018 Yunqi Conference, Alibaba merged its chip research teams into T-Head Semiconductor (平头哥, named personally by Jack Ma after the honey badger — "fearless"). The mandate: AI chips (Hanguang series), embedded processors, server CPUs (Yitian), and RISC-V IP cores (Xuantie). This was a declared group-level strategic priority, not a skunkworks project.

Three Voices, Three Eras

Zhenwu Series Roadmap

Commercial Data Points (H1 2026)

• Cumulative shipments: 560,000+ units (H1 2026)
• Annualized revenue: billions of yuan
• Customers: Alibaba Cloud internally, China Unicom, reportedly 400+ enterprise customers on Zhenwu clusters
• Capital: T-Head registered capital raised to RMB 1 billion (June 2026)
• Group investment: Alibaba committed RMB 380 billion over three years to cloud and AI infrastructure (chips, compute, liquid cooling)
• Manufacturing: Shifted from TSMC toward domestic foundries (industry consensus points to SMIC 7nm and comparable nodes) to address US restrictions on TSMC advanced AI chip work for mainland customers

Accurate framing: Jack Ma laid the foundation in 2018, Joe Tsai explained the strategic urgency in 2024, Wu Yongming reported the production results in 2026. This is eight years of consistent execution, not a recent announcement.

The short answer: AI competition has expanded from "who has the best model" to "who has the cheapest, most controllable compute." Five structural forces are driving this shift.

1. Economics: Inference is the "Monthly Rent" of AI

The industry analogy: training = home down payment (one-time, capital-intensive); inference = monthly rent (ongoing, scales linearly with users). When products like ChatGPT reach hundreds of millions of daily active users, inference spending exceeds training spend.

• Morgan Stanley estimate: a 24,000-GPU Blackwell cluster costs approximately $852M in hardware; equivalent Google TPU cluster approximately $99M (hardware basis, Reuters Breakingviews cited)
• SemiAnalysis / Bernstein estimates: custom ASICs offer a 40–65% total cost of ownership (TCO) advantage in large-scale, multi-year inference deployments
• Per-token cost reduction at hyperscaler scale: 30–40%
• Nvidia data center GPU gross margin: above 70% — custom chips convert a permanent "GPU tax" into a one-time R&D investment

2. Supply Chain Security and Geopolitics

• US export controls on advanced AI chips to China (H100, H800, H20 — successive rounds of restrictions)
• Chinese regulators encouraging domestic compute procurement; security concerns around Nvidia hardware
• Even US companies face Nvidia allocation constraints — custom chips solve a supply reliability problem regardless of geopolitics

3. Hardware-Software Co-design

• DeepSeek UE8M0 FP8 format, MLA architecture: optimized for specific hardware characteristics
• OpenAI Jalapeño: designed around ChatGPT's real serving patterns (KV cache, batching, latency targets)
• Google TPU: deeply integrated with TensorFlow/JAX

General-purpose GPUs trade efficiency for flexibility. Custom chips trade flexibility for efficiency on known workloads — and at inference scale, that trade is overwhelmingly favorable.

4. Competitive Moat and Negotiating Leverage

Even without fully replacing Nvidia, custom chips provide: stronger negotiating position on GPU procurement; differentiated "full-stack" compute story for cloud customers; building blocks for a model + cloud + chip narrative (Alibaba's "golden triangle," OpenAI full-stack infrastructure).

5. Energy and Sustainability

Performance-per-watt matters at megawatt and gigawatt scale. ASICs eliminate the general-purpose circuitry that GPUs carry for flexibility — resulting in substantially lower power consumption for the same inference throughput. At data center scale, power cost rivals hardware procurement cost.

For English-language readers, the "unit economics / Nvidia tax / TCO" frame resonates more than "supply chain independence." Both are real; lead with economics for maximum reach.

Training remains Nvidia's home turf. Inference is where custom ASICs compete and often win. Almost every "AI company building chips" story refers to inference chips — this is not coincidence, it is economics choosing the attack surface.

Six-Step Framework: Evaluating an AI Company's Chip Credibility

1. Check fundraise use-of-proceeds: Is "custom chip" explicitly listed? (DeepSeek's June 2026 raise includes it)
2. Look for recruiting signals: Chip design engineers, EDA toolchain specialists being hired privately
3. Track supplier contacts: Foundry, HBM memory supplier conversations (Reuters' core source signal)
4. Read the software side: Custom data formats, operator optimizations targeting specific hardware (DeepSeek UE8M0 FP8)
5. Verify production data: Shipment volumes, revenue, named customers (Alibaba Zhenwu 810E has clear numbers)
6. Separate management statements: Distinguish "strategic motivation articulation" from "project announcement" — Liang Wenfeng is the former; Wu Yongming's earnings call is the latter

While the AI chip landscape evolves, stable, predictable compute remains the production-grade choice for most teams deploying inference workloads today. General GPU cloud services suffer from over-provisioning noise and Nvidia allocation cycles. JEXCLOUD's bare-metal Apple Silicon nodes sidestep both: dedicated hardware, 7x24 availability, no over-subscription, 120-second provisioning — well-suited to latency-sensitive, predictable inference workloads where performance consistency matters more than theoretical peak throughput. For pricing and node specifications, see the JEXCLOUD pricing page.

Q1: Is DeepSeek really building its own AI chip?
According to a July 7, 2026 Reuters report citing three sources, DeepSeek is in the early stages of developing a custom chip for AI inference. DeepSeek has not officially confirmed the project. The company is reportedly hiring chip engineers privately and in talks with foundries and memory suppliers. Treat as "reported, early-stage, unconfirmed."

Q2: Did DeepSeek CEO Liang Wenfeng announce a chip program?
No public announcement. In 2024 interviews he said export controls on advanced chips were DeepSeek's main challenge, not funding. Those quotes establish strategic motivation, not a project announcement. Reuters reported on company-level actions, not a founder declaration.

Q3: How is Alibaba involved?
Alibaba's chip unit T-Head (founded 2018, named by Jack Ma) is mass-producing Zhenwu AI chips. As of H1 2026: 560,000+ units shipped, annualized revenue in the billions of yuan, 400+ enterprise customers. The Zhenwu 810E is CUDA-compatible, positioned between Nvidia A800 and H20 in performance, and manufactured domestically. This is an eight-year program, not a reaction to recent events.

Q4: Why inference chips first, not training chips?
Inference workloads are repetitive and predictable — ideal for ASICs. Training still relies on Nvidia's GPU compute and the CUDA software stack, where the moat is deep. Inference is also the larger ongoing cost for deployed AI products ("monthly rent"), making the economics of custom silicon far more compelling at scale.

Q5: Is it about national security or saving money?
Both, but economics is the primary driver. Cutting the "Nvidia tax" and reducing per-token cost by 30–65% at scale is the first mover. Export controls and supply chain risk are real accelerants — but they are accelerating an economic trend that would exist regardless. Security and savings reinforce each other here rather than competing.

Last updated: July 10, 2026 | Sources: Reuters (July 7, 2026), OpenAI official blog (Jalapeño announcement), WSJ, Caixin Global, SCMP, Waves (Liang Wenfeng interviews 2023/2024), Alibaba FY2026 earnings call | Disclaimer: DeepSeek has not officially confirmed the chip project as of this writing.