Is DeepSeek Building Its Own AI Chip? Inside the July 2026 Reuters Report

The Reuters scoop on July 7–8, 2026 was consistent across multiple outlets. Key details:

• DeepSeek is developing a chip optimized for AI inference, not training.
• The project started approximately one year ago (circa mid-2025) and remains in early stages.
• The company is in discussions with chip design firms, foundries, and memory suppliers.
• Hiring of chip design engineers has ramped in recent months, but not through public job boards — through private recruitment.
• A successful chip would reduce DeepSeek's dual dependency on both Nvidia and Huawei Ascend — notable because DeepSeek V4 already runs on Ascend hardware.

Liang Wenfeng rarely gives interviews. His most substantive public remarks on chips come from two deep-dives with Waves (暗涌) in May 2023 and July 2024. He never announced a chip program. He did say three things that establish a clear strategic logic:

These quotes establish: (1) export controls are the binding constraint; (2) a 4× compute gap makes hardware efficiency existentially important; (3) software-hardware co-design must be developed domestically. Reuters reported on company-level actions — hiring, supplier talks. The distinction matters: founder's long-term strategic framing ≠ official project announcement.

The "Jack Ma said something similar" framing needs calibration: Alibaba's chip program is an eight-year operational reality, not a recent rumor.

• September 2018, Yunqi Conference: Jack Ma personally named the new entity "T-Head" (平头哥 — the honey badger, symbolizing fearlessness). Chip development was declared a group-level strategic priority, not a business unit initiative.
• 2024, Joe Tsai (Chairman): US chip export restrictions "clearly impact" Alibaba Cloud. China is roughly two years behind the US in AI. Long-term belief: China will develop advanced semiconductor self-sufficiency.
• 2026, Eddie Wu (CEO, earnings call): T-Head AI chips cumulative delivery 560K+ units. Annual run-rate revenue in the billions of RMB. IPO for T-Head not ruled out.

Zhenwu Series Roadmap

Two strategic differentiators stand out: the new chips are designed to be CUDA-compatible, lowering migration cost for engineers (unlike Huawei's Ascend which requires rewriting kernels). Manufacturing has shifted from early TSMC to domestic foundries, insulating supply from US restrictions on advanced node fabrication for mainland China.

Custom silicon is a global industrial trend, not a China story. TrendForce data (2026): cloud provider custom AI chip shipments growing at 44.6% YoY, vs. 16.1% for general-purpose GPUs — the first time custom silicon growth has decisively outpaced GPUs.

The one-sentence answer: AI competition has extended from "who has the best model" to "who has the cheapest, most controllable compute."

Driver 1: Economics — Inference Cost Is AI's "Monthly Rent"

The industry uses a housing analogy: training = down payment (one-time, concentrated); inference = monthly rent (continuous, scales linearly with users). When a ChatGPT-class product has hundreds of millions of daily active users, inference spend exceeds training spend.

• Morgan Stanley estimates: a 24,000-GPU Blackwell cluster costs roughly $852M in hardware; an equivalent Google TPU cluster costs roughly $99M.
• SemiAnalysis, Bernstein: at large-scale, multi-year inference deployments, custom ASICs can achieve a 40–65% TCO advantage vs. general-purpose GPUs; hyperscaler inference cost per token can drop 30–40%.
• Nvidia data center GPU gross margins exceed 70%. Custom silicon converts a permanent "GPU tax" into a one-time R&D investment.

Driver 2: Supply Chain Security and Geopolitics

US export controls on advanced AI chips (H100 / H800 / H20, each restricted in turn) force Chinese companies to find alternatives. Even for US companies, GPU allocation from Nvidia is constrained. Security here means supply chain predictability: not being choked by a single supplier or a single government's policy decisions.

Driver 3: Hardware-Software Co-Design

General-purpose GPUs sacrifice efficiency for flexibility. Custom ASICs sacrifice flexibility for efficiency on known workloads:

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

Driver 4: Competitive Moat and Negotiating Leverage

Even without fully replacing Nvidia, custom chips add procurement leverage, differentiate cloud offerings, and build a "model + cloud + silicon" full-stack narrative — the strategic logic behind DeepSeek's chip investment and Alibaba's "golden triangle" positioning.

Driver 5: Energy and Sustainability

Inference chips are optimized for performance-per-watt. In megawatt-scale and gigawatt-scale data centers, electricity and cooling costs rival hardware procurement costs. ASICs eliminate the large swaths of general-purpose circuitry that GPUs carry, delivering meaningfully lower power consumption.

Bottom line: training remains Nvidia's home territory; inference is the battleground for custom ASICs.

Risks and Uncertainties

Chip development has precedents for failure: Meta rebuilt MTIA from scratch; Nvidia will not concede without a software ecosystem fight; rapid architectural evolution in AI models can render a well-optimized ASIC obsolete. DeepSeek is years from production. Readers deserve this nuance alongside the headline.

For teams that need predictable, controllable compute for inference and AI agent workloads today — without waiting for custom silicon to mature — dedicated bare-metal Apple Silicon cloud nodes offer an independent path. No shared-pool neighbors, auditable tenant boundaries, and configurable from the NUKCLOUD pricing page without exposure to chip supply volatility.

• Is DeepSeek really building its own AI chip?
  According to a July 7, 2026 Reuters report citing three sources, DeepSeek is in early stages of developing a custom AI chip optimized for inference. The company has not officially confirmed the project. DeepSeek is reportedly hiring chip engineers privately and talking to foundries and memory suppliers.
• 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. Reuters reported on company-level actions, not a founder proclamation.
• How is Alibaba involved?
  Alibaba's chip unit T-Head (founded in 2018 under Jack Ma's strategy) is already mass-producing Zhenwu AI chips, with 560,000+ units shipped and billion-yuan annual revenue as of mid-2026.
• Why inference chips first, not training chips?
  Inference workloads are repetitive and predictable — ideal for custom ASICs. Training still relies heavily on Nvidia GPUs and the CUDA software stack. Custom ASICs can reduce total cost of ownership by 30–65% at inference scale.
• Is it about national security or saving money?
  Both. Economics is the primary driver — cutting the Nvidia tax and per-token costs at scale. Export controls and supply chain risk accelerate the shift. Nvidia data center GPU gross margins exceed 70%, making the business case for custom silicon compelling for any large-scale AI deployment.

Last updated: July 10, 2026 | Disclaimer: DeepSeek has not officially confirmed the chip project as of this writing.