On June 24, 2026, OpenAI and Broadcom unveiled Jalapeño—a custom inference ASIC built in nine months. One week later, on July 7, Reuters cited three sources saying DeepSeek is developing its own inference chip, even while already running on Huawei Ascend. Meanwhile, Alibaba's T-Head has shipped 560,000+ Zhenwu chips with billion-yuan annual revenue. This is not nationalism—it is unit economics. For AI developers, infra engineers, and investors, this guide covers what Reuters actually reported, Liang Wenfeng's past remarks, T-Head's eight-year arc, the global custom-silicon wave, five drivers, inference vs training, risks, and FAQ. Last updated: July 10, 2026.
01

This Isn't Just China: OpenAI's Jalapeño and the Global Custom Chip Wave

Custom AI silicon is now a global phenomenon. TrendForce (2026) shows hyperscaler custom chip shipments growing at 44.6%, far outpacing general-purpose GPUs at 16.1%—custom silicon is winning on growth for the first time.

CompanyChip ProjectStageWorkloadKey Signal
DeepSeekUnnamed inference ASICEarly R&DInference$7.4B funding; private hiring; unconfirmed
Alibaba (T-Head)Zhenwu 810E / M890Mass productionTrain + infer560K+ shipped; ~$1.4B annual revenue
HuaweiAscend 950 seriesMass productionTrain + inferDeepSeek V4 adapted; orders surging
OpenAIJalapeño (Broadcom)Tape-out doneInference9-month design cycle; deploy late 2026
GoogleTPU v6/v7At scaleTrain + inferGemini runs end-to-end on TPU
AmazonTrainium3 / InferentiaCommercialBothAnthropic uses Trainium at scale
MicrosoftMaia 100DeployingInferencePowers Azure / OpenAI workloads
MetaMTIAInternalInferenceRecommendations; once scrapped and rebuilt
AnthropicSamsung custom chip talksExploringTBDThe Information, July 2026
Zhipu AIEvaluating custom chipEarlyInferenceThe Information, July 2026

Key dates: Jun 24 OpenAI Jalapeño; Jul 2 Anthropic–Samsung 2nm talks; Jul 7 Reuters on DeepSeek; Jul 7 The Information on Zhipu. See also our OpenAI Jalapeño deep dive.

02

What Reuters Actually Reported (And What DeepSeek Hasn't Confirmed)

Bottom line: You can write "Reuters and others report DeepSeek has started a custom inference chip program." You cannot write "Liang Wenfeng officially announced chip development." Tag it: sources familiar / early stage / unconfirmed.

30-second summary: Likely real but early. No CEO announcement. T-Head is already at mass production. Economics drives the shift; geopolitics accelerates it.

01

Inference-only ASIC: Optimized for serving, not training clusters.

02

Started ~mid-2025: Described as "about a year ago"; still early stage.

03

Supply chain talks: Engaging chip designers, foundries, and memory vendors.

04

Quiet hiring: Chip engineers recruited privately, not on public job boards.

05

Dual-dependency play: Would reduce reliance on both Nvidia and Huawei Ascend—DeepSeek already runs on Ascend.

Credibility FactorAssessment
Source tierHigh. Reuters "three people familiar with the matter" standard
Official confirmationNone as of this writing
Circumstantial evidenceStrong. ~$7.4B (~51B RMB) June 2026 round earmarked for chips and domestic compute; IDC hiring; UE8M0 FP8 format seen as hardware-software co-design signal
Contradictory takesPartnership and in-house R&D run in parallel—Ascend is live, custom silicon is early
03

What DeepSeek CEO Liang Wenfeng Has Said About Chips and Compute

Liang Wenfeng (DeepSeek CEO) has given few public interviews. The most valuable source is two deep dives with Waves (暗涌) in May 2023 and July 2024. He never announced a chip program, but framed the strategic motive.

"Our real challenge has never been capital—it is export controls on advanced chips." — Liang Wenfeng, Waves interview, July 2024

01

4× compute overhead: Domestic training efficiency lags ~1×, data efficiency another ~1×—~4× total compute needed for parity.

02

Missing tech community: Domestic chips lack a first-hand developer community; someone must stand at the frontier.

03

Endless compute hunger: Researchers always want more capacity; DeepSeek deploys as much compute as it can.

04

Founder remarks ≠ product launch: Reuters describes company actions (hiring, vendor talks), not a CEO announcement.

05

Co-design signals: UE8M0 FP8 and MLA architecture optimizations point toward hardware-specific tuning.

04

Alibaba's T-Head Is Already Shipping—Jack Ma's 2018 Bet Pays Off in 2026

Do not write "Jack Ma recently said Alibaba will make chips." The accurate arc: Jack Ma set T-Head strategy in 2018, Joe Tsai explained export-control pressure in 2024, CEO Wu Yongming disclosed mass-production numbers in 2026.

FigureRolePublic chip-related stance
Jack Ma2018 strategic sponsorNamed T-Head, elevated chips to group strategy
Joe TsaiChairman2024 podcast: US export limits hit Alibaba Cloud; long-term faith in domestic semiconductors
Wu YongmingCEO2026 earnings call: 470K+ T-Head AI chips delivered; billion-yuan annual revenue; IPO possible
ModelTimelineHighlights
Hanguang 8002019Early AI inference chip
Zhenwu 810EJan 2026Train + infer; 96GB HBM2e; between Nvidia A800 and H20; in production
Zhenwu M8902026144GB memory, 800GB/s interconnect, ~3× 810E
Zhenwu V900Planned Q3 2027216GB, 1200GB/s interconnect
Zhenwu J900Planned Q3 2028Next-gen parallel compute architecture

WSJ: new Alibaba chips support the Nvidia CUDA ecosystem, easing engineer migration (unlike Huawei's path). Manufacturing shifted from TSMC toward domestic foundries (industry points to SMIC 7nm-class flows).

A

560,000+ units shipped (H1 2026).

B

~$1.4B annual revenue; 400+ enterprise customers on Zhenwu clusters.

C

T-Head registered capital raised to ~$140M (Jun 2026); Alibaba pledged ~$52B over three years for cloud and AI infrastructure.

05

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

One-line answer: AI competition has moved from "who has the best model" to "who has the cheapest, most controllable compute."

01

Economics—inference is the rent: Training is the down payment; inference is monthly rent. At ChatGPT-scale DAU, inference spend exceeds training. Custom ASICs can cut total cost of ownership (TCO) 30–65% at scale; per-token costs down 30–40%. Nvidia datacenter GPU gross margins exceed 70%—in-house silicon converts the permanent "GPU tax" into one-time R&D.

02

Supply chain resilience: US export controls, allocation shortages, and single-vendor risk—not just "national security" but predictable supply.

03

Hardware-software co-design: General GPUs trade efficiency for flexibility; ASICs do the opposite for known workloads. Jalapeño targets real ChatGPT serving (KV cache, batching, latency).

04

Bargaining power: Even partial self-supply strengthens Nvidia negotiations and enables full-stack "model + cloud + chip" stories.

05

Energy: Inference ASICs optimize performance-per-watt—critical at gigawatt-scale datacenters.

DimensionTrainingInference
WorkloadDynamic, experimental, architecture churnStatic model, predictable request patterns
Software moatCUDA stack (cuDNN, NCCL, Nsight)Hand-tuned kernels for fixed models
Chip priorityPeak FLOPS + programmabilityThroughput, latency, cost per token
EconomicsLarge one-time capex24/7 at scale—bigger ongoing spend
VerdictTraining stays Nvidia territory; inference is the custom ASIC battleground.

Morgan Stanley (via Reuters Breakingviews): a 24,000-GPU Blackwell cluster costs ~$852M in hardware; an equivalent Google TPU cluster ~$99M.

Six-step decision guide:

01

Separate rumor from announcement: Write "reportedly" until DeepSeek confirms.

02

Split training vs inference planning: Frontier training still needs Nvidia; inference is where ASICs win.

03

Track parallel paths: DeepSeek on Ascend is live; custom silicon is early.

04

Model TCO, not sticker price: Focus on per-token cost and multi-year capex ROI.

05

Price early-project risk: Meta scrapped and rebuilt MTIA; architecture shifts can obsolete ASIC designs.

06

Decouple local agents from cloud inference: Chip economics mostly affect API pricing; local Cursor/Codex agents and iOS CI still need stable macOS hosts 24/7.

Disclaimer: DeepSeek has not officially confirmed the chip project as of this writing. Sources: Reuters, WSJ, OpenAI official blog, Waves interviews, Alibaba filings. Not investment advice.

Reality check: API-only local agents mean linear token spend and model availability risk under export controls; personal Macs for mixed training and agents hit unified-memory limits and sleep interruptions; macOS VMs break the EULA and restrict Xcode signing. For iOS CI/CD, local LLM inference, and AI agent automation in production, KVMNODE dedicated Mac Mini M4 cloud rental is usually the better fit: Apple Silicon unified memory for Metal inference, 24/7 uptime, flexible daily/weekly/monthly billing. See pricing.

Last updated: July 10, 2026 · Sources: Reuters, OpenAI official, WSJ, Caixin Global, Waves interviews, Alibaba/T-Head public disclosures