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

~20 min read · MACCOME · Last updated: July 9, 2026

Who should read this? Developers asking whether DeepSeek is really building silicon, enterprise leads evaluating domestic compute alternatives, and investors tracking inference cost and supply-chain risk. Global context first: the July 2026 wave—OpenAI's Jalapeño inference chip with Broadcom, Anthropic reportedly talking to Samsung, Zhipu evaluating custom silicon—shows AI competition shifting from model quality to unit economics and controllable compute. Then DeepSeek: a July 7, 2026 Reuters exclusive says DeepSeek is developing a custom AI chip for inference, while DeepSeek V4 already runs on Huawei Ascend and Alibaba T-Head Zhenwu has shipped 560,000+ units. This article delivers the evidence chain, quotes from Liang Wenfeng, DeepSeek CEO, Jack Ma's 2018 T-Head bet through Eddie Wu's 2026 production numbers, global comparison tables, five drivers, inference-vs-training math, risks, and a six-step runbook. Structure: TL;DR → six pain points → rumor breakdown → executive timeline → progress tables → global trend → economics → FAQ.

bolt

TL;DR — 30-second verdict

  • DeepSeek chip rumor: Per three Reuters sources, likely real but early-stage; target is an inference ASIC, not training; no official announcement yet.
  • Liang Wenfeng: Never publicly launched a chip program; in 2024 he called export bans the top challenge and described compute hunger—strategic motive, not a product launch.
  • Alibaba T-Head: Not a rumor—eight years in production; Zhenwu 810E is shipping, 560K+ units, roughly ten-billion-yuan annual revenue run-rate.
  • Global trend: OpenAI Jalapeño, Anthropic × Samsung talks, Zhipu evaluation—custom silicon shipment growth at 44.6% vs 16.1% for general GPUs (TrendForce 2026).
  • Core driver: Economics first—inference is AI's recurring rent; custom ASIC TCO can run 30–65% below GPU clusters at scale; supply-chain predictability is second.

Six Pain Points: Why "AI Companies Building Chips" Dominated July 2026

In one week, Reuters reported DeepSeek's custom inference silicon, The Information said Zhipu is weighing bespoke chips, and Anthropic was linked to Samsung 2nm talks. That is not three isolated headlines—it is a structural shift from who has the best model to who has the cheapest, most predictable compute. If you already read our OpenAI × Broadcom Jalapeño inference chip article, this piece uses the DeepSeek rumor plus Alibaba's eight-year T-Head track record to add a China lens and a global scorecard.

  1. Inference cost is AI's rent: Training is a one-time capex spike; inference scales with daily active users. At ChatGPT-class scale, serving spend has overtaken training. Nvidia datacenter GPUs carry gross margins above 70%—cloud buyers call the margin leakage the Nvidia tax.
  2. Export controls tighten the vise: U.S. restrictions on H100/H800/H20 sales to China rotated for years. Liang Wenfeng, DeepSeek CEO, said in 2024 that advanced-chip export bans are the biggest challenge. Even U.S. hyperscalers face Nvidia allocation risk—supply predictability is now a hard constraint everywhere.
  3. Partner and build-in-house are parallel, not either/or: DeepSeek V4 is tuned for Huawei Ascend 950 with partial Ascend training—but Reuters still reported a custom inference ASIC effort in July. The accurate frame: partnerships ship today; in-house silicon is a long bet.
  4. Hardware–software co-design window: DeepSeek's UE8M0 FP8 format and MLA architecture read as signals for domestic co-design; OpenAI Jalapeño is shaped around ChatGPT serving (KV cache, batching, latency). GPUs trade efficiency for flexibility; ASICs trade flexibility for efficiency on known loads.
  5. Rumor vs production asymmetry: DeepSeek silicon is unconfirmed, but Alibaba T-Head is a volume business (560K+ units, ten-billion-yuan revenue scale). Conflating "Jack Ma talked chips recently" with "Reuters DeepSeek secret project" misleads readers—the former is a 2018 strategy call, the latter is 2026 early R&D reporting.
  6. Downstream impact on builders: Inference price wars raise API volatility and routing complexity. When compute infrastructure wobbles, multi-model routing and a 24/7 Agent control plane matter more than betting on a single chip vendor.

DeepSeek Chip Rumor: What Reuters Reported—and What Remains Unconfirmed

On July 7–8, 2026, outlets followed a Reuters exclusive. The consistent claims:

  • DeepSeek is developing a custom AI chip aimed at inference, not training
  • The program reportedly started around mid-2025 ("about a year ago") and remains early-stage
  • DeepSeek is talking to chip design houses, foundries, and memory suppliers
  • It has stepped up chip-design hiring in recent months—often off public job boards, via direct outreach
  • Success would reduce reliance on both Nvidia and Huawei Ascend, even though DeepSeek already deep-integrated Ascend

Editorial guardrail: Write "Reuters and follow-on outlets report DeepSeek has started a custom inference chip program." Do not write "Liang Wenfeng officially announced a chip." Tag copy with sources say / early stage / not confirmed by the company.

Credibility assessment

Dimension Assessment
Source tier High. Reuters' standard "three people familiar with the matter" phrasing; cross-checked by major business press
Official confirmation None. As of July 9, 2026, DeepSeek has issued no press release or social post confirming the project
Circumstantial evidence Strong. June 2026 external funding round of roughly $7.4 billion (~51 billion yuan) with stated uses including custom AI chips and domestic compute expansion; IDC planning hires; UE8M0 FP8 interpreted as hardware–software co-design
Contradictory takes Some analysts argue DeepSeek will lean on Ascend near-term and downplay custom silicon. Partnership plus in-house R&D is the balanced read

What Has Liang Wenfeng Said—and How It Relates to the Rumor

Liang Wenfeng, DeepSeek CEO, rarely gives on-the-record interviews. The most cited sources are two long-form Waves profiles from May 2023 and July 2024. He never announced "DeepSeek will build chips" in those sessions—Reuters describes company behavior (hiring, supplier talks), not a founder product launch.

Key quotes (compute and silicon context)

format_quote

"Our real challenge was never capital—it is the export ban on advanced chips." — July 2024, Waves interview

format_quote

Compared with the best overseas labs, domestic training efficiency is roughly half, and data efficiency another half—you need about 4× the compute for the same outcome. — Liang Wenfeng, Waves

format_quote

"Many domestic chips fail because they lack a real technical community—only second-hand information. China needs people at the frontier." — Liang Wenfeng, Waves

format_quote

"For researchers, hunger for compute is endless… we deliberately try to deploy as much compute as we can."

Those lines establish strategic motive: compute limits, export control, and the need for hardware–software alignment. In copy, separate long-term founder statements from official project announcements.

Alibaba / T-Head: Eight Years of Execution, Not a Fresh Headline

Readers often ask whether "Jack Ma said the same thing." Clarify: Alibaba's chip program is a multi-year strategy already in production—not a July 2026 rumor.

Jack Ma era (2018): strategic origin

  • At the September 2018 Cloud Summit, Alibaba merged C-SKY (Zhongtian Micro) with Damo Academy silicon teams into T-Head Semiconductor Co., Ltd.
  • Jack Ma personally approved the brand name—"T-Head" (honey badger) means "fear nothing"
  • Zhang Jianfeng (Xing Dian) framed chips as a group-level strategic priority
  • Early scope: AI accelerators (Hanguang line), embedded IP, cloud-edge integration; later expanded to server CPUs (Yitian) and RISC-V IP (Xuantie)

Do not write "Jack Ma recently called for chips." Accurate framing: Jack Ma set T-Head strategy in 2018; Joe Tsai explained export-control pressure in 2024; Eddie Wu disclosed production metrics in 2026.

Jack Ma vs Joe Tsai vs Eddie Wu

Executive Role Public silicon-related statements
Jack Ma 2018 strategic sponsor Named T-Head; elevated chips to group strategy; reduced public appearances after stepping down as chairman in 2019
Joe Tsai Chairman 2024 podcast: U.S. chip export limits "clearly affect" Alibaba Cloud; China AI ~two years behind the U.S.; long-term belief China will develop advanced semiconductors; export rules cited among reasons Alibaba Cloud IPO was paused
Eddie Wu CEO FY2026 earnings call: T-Head AI chips delivered 470,000+ cumulative units; ten-billion-yuan annualized revenue scale; T-Head IPO not ruled out

Zhenwu product line

SKU Timing Highlights
Hanguang 800 2019 Early AI inference accelerator
Zhenwu 810E January 2026 launch Train + infer unified; 96 GB HBM2e; performance between Nvidia A800 and H20; in mass production
Zhenwu M890 2026 144 GB memory; 800 GB/s die-to-die link; ~3× 810E throughput
Zhenwu V900 Planned Q3 2027 216 GB memory; 1200 GB/s interconnect
Zhenwu J900 Planned Q3 2028 Next-gen parallel compute architecture iteration

Commercial metrics (2026): Cumulative shipments 560,000+; annualized revenue on the order of ten billion yuan; customers include Alibaba Cloud internally, China Unicom, and reportedly 400+ enterprises on Zhenwu clusters; T-Head registered capital raised to 1 billion yuan (June 2026); Alibaba pledged 380 billion yuan over three years to cloud and AI infrastructure.

Relationship with Nvidia: WSJ reporting says newer Alibaba chips aim for CUDA compatibility to lower engineer migration cost—a different software path than Huawei. Manufacturing has shifted from early TSMC flows toward domestic foundry (industry consensus points to SMIC 7nm-class mature nodes).

July 2026 Global Progress Scorecard

Custom silicon is a global phenomenon, not China-only. English readers weight unit economics and the Nvidia tax; China-focused readers care about domestic alternatives—a complete brief covers both.

Company Chip program Stage Workload Key numbers / events
DeepSeek Unnamed custom inference ASIC Early R&D Inference ~$7.4B funding; quiet hiring; no official confirmation
Alibaba (T-Head) Zhenwu 810E / M890 Mass production Train + infer 560K+ units shipped; ten-billion-yuan revenue run-rate
Huawei Ascend 950 family Production Train + infer DeepSeek V4 tuned; Reuters noted order surge
OpenAI Jalapeño (with Broadcom) Tape-out done; deploy pending Inference Nine-month design-to-tape-out; Azure deploy by end 2026 (see our Jalapeño deep dive)
Google TPU v6/v7 Large-scale production Train + infer Gemini end-to-end on TPU
Amazon Trainium3 / Inferentia Commercial Train + infer Anthropic runs large Trainium fleets
Microsoft Maia 100 Rolling out Inference Azure / OpenAI serving workloads
Meta MTIA Internal deploy Inference Recommendation-heavy; one redesign cycle already
Anthropic Samsung custom talks Exploratory TBD The Information, July 2026
Zhipu AI Evaluating custom silicon Early Inference The Information, July 2026

TrendForce (2026): hyperscaler custom AI chip shipment growth at 44.6% versus 16.1% for general-purpose GPUs—custom silicon is outpacing GPU growth for the first time on that metric.

Five Drivers Behind the Custom Chip Wave (Ranked)

One-line answer: Labs are not chasing silicon for its own sake—AI competition moved down-stack into compute economics and supply control.

  1. Economics: inference is the rent bill — Morgan Stanley once sized a 24,000-GPU Blackwell cluster near $852M in hardware versus roughly $99M for a comparable Google TPU footprint (hardware-only). SemiAnalysis and Bernstein peg custom ASIC TCO advantages of 40–65% vs GPUs; hyperscaler serving can cut per-token cost 30–40%.
  2. Supply-chain security and geopolitics — Export rules, domestic substitution incentives, Nvidia allocation. "Security" here means predictable supply: not hostage to one vendor or one nation's policy.
  3. Hardware–software co-design — DeepSeek UE8M0 FP8, OpenAI Jalapeño serving optimizations, Google TPU bound to JAX. GPUs pay for flexibility; custom chips buy efficiency on known graphs.
  4. Negotiating leverage and moats — Even partial in-house silicon strengthens procurement talks and supports "model + cloud + chip" full-stack narratives (Alibaba's triangle, OpenAI infrastructure story).
  5. Energy and sustainability — Inference ASICs optimize performance-per-watt. At megawatt datacenter scale, power and cooling rival chip purchase price; ASICs drop unused GPU circuitry.

Inference Chip vs Training GPU: Why the Market Is Splitting

Bottom line: training stays Nvidia territory; inference is the custom ASIC battleground.

Dimension Training Inference
Workload Dynamic, experimental, architecture churn Static model, predictable request patterns
Software stack Deep CUDA moat (cuDNN, NCCL, Nsight) Hand-tuned kernels for fixed models
Silicon needs Peak FLOPS plus programmability Throughput, latency, cost per token
Economic scale Large one-time cluster spend 24/7 continuous, often larger opex
Examples Nvidia H100/B200 leadership TPU, Trainium, Maia, Jalapeño, rumored DeepSeek ASIC

Risks and Uncertainty: Early Programs Can Fail

  • Meta MTIA already rebooted once—custom silicon is not a straight line
  • Architecture drift: training architectures change fast; inference is steadier, but a post-Transformer shift raises ASIC sunk cost
  • DeepSeek unconfirmed: as of this writing, zero official statement—do not label the project "proven"
  • Manufacturing bottlenecks: leading-edge foundry capacity, HBM supply, tool export limits under sanctions
fact_check

Pre-publish checklist

  • Never write "Liang Wenfeng officially announced chips"—use "reportedly / sources say"
  • Keep training vs inference separate—the table above is the anchor
  • Jack Ma timing: 2018 strategy, not "Jack Ma just said"
  • Dual-currency note: ~$7.4 billion ≈ 51 billion yuan
  • Expect news flow every 2–4 weeks—watch Reuters and company earnings calls

Custom Silicon Timeline (Selected Milestones)

timeline
2023–2024  Liang Wenfeng (Waves): export bans top challenge; compute hunger
2025-01    DeepSeek R1 ships; trained on Nvidia H800 (export-blocked late 2023)
mid-2025   Reported start of custom chip program
2026-04    DeepSeek V4 on Huawei Ascend; V4-Flash partial Ascend training
2026-06    DeepSeek external round ~$7.4B; uses include custom AI chips
           OpenAI + Broadcom announce Jalapeño inference ASIC (9-month tape-out)
2026-07-07 Reuters exclusive: DeepSeek developing custom inference chip
           The Information: Zhipu evaluating custom silicon
2018-09    Alibaba forms T-Head (Jack Ma names brand)
2026-01    Alibaba Zhenwu 810E enters mass production

Six-Step Runbook: What Builders Can Do While Silicon Strategies Play Out

Chip strategy is a hyperscaler chess match—but application teams can act today by reducing single-vendor compute dependence and keeping Agent infrastructure stable. This pairs with our Huawei openPangu Ascend full-stack article and ds4 high-memory Mac inference decision guide.

  1. Build a compute cost dashboard: Split training vs inference spend; track cost per million tokens, GPU utilization, API bill trends—when inference exceeds ~60% of AI opex, prioritize routing and caching before capex bets.
  2. Multi-vendor routing architecture: Do not bind production Agents to one model or one chip ecosystem; configure primary, fallback, and degrade paths (see the OpenRouter routing matrix).
  3. Separate rumor from purchasable: DeepSeek custom silicon is early; T-Head Zhenwu, Huawei Ascend, and Nvidia each have clearer procurement paths—buy on production status + software maturity, not headlines.
  4. Model co-design ROI: Fixed serving graphs (stable model, batch inference) favor ASIC TCO; experimental stacks (frequent architecture changes) need GPU flexibility.
  5. Keep the control plane always-on: When compute layers fluctuate, Gateway, CI runners, signing hosts, and Agent schedulers need stable dedicated environments—not shared inference clusters.
  6. Quarterly supply-chain review: Refresh every 2–4 weeks from Reuters, WSJ, and earnings-call silicon commentary; update an internal risk register.

Three Hard Numbers (Citable)

  • ~$7.4 billion / 51 billion yuan: DeepSeek June 2026 external funding round; disclosed uses include custom AI chips and domestic compute expansion (Reuters / public financing coverage).
  • 560,000+ units / ten-billion-yuan revenue scale: Alibaba T-Head Zhenwu cumulative shipments and annualized revenue magnitude in H1 2026 (Eddie Wu earnings call / press).
  • 44.6% vs 16.1%: TrendForce 2026 hyperscaler custom AI chip shipment growth vs general GPU growth—custom silicon growth outpacing GPUs on that measure.

Security vs Cost Savings: How to Frame Both Fairly

Angle Reader How to write it
Geopolitics / decoupling U.S.–China tech watchers Stress export controls, domestic substitution, supply autonomy
Business / investing AI economics audience Stress TCO, gross margin, per-token cost, capex payback
Engineering Builders Stress co-design, ASIC vs GPU, inference architecture
Enterprise security Procurement Stress data sovereignty, supply resilience, third-party dependence

Waiting solely for domestic chips to mature has downsides: early programs fail (Meta MTIA), software migration costs are understated, and Agent control planes cannot idle—a Gateway outage costs more than a 5% inference price swing. Renting Nvidia APIs forever invites price spikes, quotas, and geopolitical shocks. The pragmatic path: multi-vendor compute plus a stable, dedicated control-plane environment.

Teams running OpenClaw Gateway, coding Agents, CI runners, or local model experiments on owned Macs hit procurement lead times, rack constraints, and peak-scale ceilings. VMs often sacrifice Metal and graphics stack fidelity. MACCOME Mac cloud hosts provide dedicated Apple Silicon bare metal, flexible lease terms, and six regional nodes—a steadier production base for AI Agent automation while hyperscaler silicon headlines change weekly. Your control plane should not jitter with every Reuters alert.

FAQ

Is the DeepSeek custom chip report credible?

Reuters on July 7, 2026 cited three people familiar with the matter—a high-credibility bar—but DeepSeek has not officially confirmed. The project is early-stage and targets inference, not training. As of July 9, 2026, label it "reportedly," not "confirmed."

Has Liang Wenfeng publicly announced a chip program?

No. In a 2024 Waves interview he said export bans on advanced chips are the biggest challenge and discussed deploying compute and a 4× efficiency gap—but he did not announce custom silicon. Reuters describes hiring and supplier talks, not a founder launch.

Who at Alibaba has spoken about chips—Jack Ma, Joe Tsai, or Eddie Wu?

Jack Ma set 2018 strategy and named T-Head. Joe Tsai has stressed export-control impact on Alibaba Cloud. Eddie Wu disclosed mass-production metrics on 2026 earnings calls. Alibaba chip work is a mature business—not a fresh rumor. Avoid "Jack Ma recently called for chips."

Why inference chips first instead of training GPUs?

Inference workloads are stable, large, and run 24/7—ideal for ASIC tuning. Training still needs CUDA depth and flexibility where Nvidia leads. The rumored DeepSeek part, OpenAI Jalapeño, and Alibaba Zhenwu all prioritize inference or train-infer unified serving economics.

Are hyperscalers building chips for national security or to save money?

Both—but economics rank first: cutting inference cost (the Nvidia tax) and supply risk is urgent; export rules accelerate existing motives. Custom ASICs can lower TCO 30–65% at scale vs GPUs. For stable Agent infrastructure, see MACCOME Mac cloud rental rates.

Disclaimer: DeepSeek has not officially confirmed a custom chip program. Information is current through July 9, 2026, drawn from Reuters, WSJ, OpenAI announcements, Waves interviews with Liang Wenfeng, Alibaba earnings materials, and public industry analysis. Re-check headlines before republishing.

Sources: Reuters (July 7, 2026 DeepSeek chip report), OpenAI Jalapeño announcement, WSJ (Alibaba AI chip), Caixin Global (Zhenwu 810E), Waves (Liang Wenfeng interviews), TrendForce (custom silicon growth rates).