24 июня 2026 OpenAI и Broadcom представили Jalapeño — custom inference ASIC, собранный за девять месяцев. Через неделю, 7 июля, Reuters со ссылкой на трёх источников написал, что DeepSeek разрабатывает собственный inference-чип — хотя уже крутится на Huawei Ascend. Параллельно T-Head Alibaba отгрузил 560 000+ Zhenwu chips с billion-yuan annual revenue. Это не национализм — это unit economics. Для AI-разработчиков, infra engineers и инвесторов: что Reuters реально написал, прошлые ремарки Liang Wenfeng, восьмилетняя дуга T-Head, глобальная волна custom silicon, пять драйверов, inference vs training, риски, FAQ. Обновлено: 10 июля 2026.
01

Не только Китай: Jalapeño OpenAI и глобальная волна custom chips

Custom AI silicon — уже глобальный феномен. TrendForce (2026): shipments hyperscaler custom chips растут на 44,6%, general-purpose GPUs — на 16,1%. Впервые custom silicon обгоняет по growth rate.

КомпанияChip projectСтадияWorkloadKey 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: 24 июня OpenAI Jalapeño; 2 июля Anthropic–Samsung 2nm talks; 7 июля Reuters on DeepSeek; 7 июля The Information on Zhipu. См. также deep dive по Jalapeño.

02

Что Reuters реально написал (и что DeepSeek не подтвердил)

Bottom line: можно писать «Reuters и другие сообщают, что DeepSeek запустил custom inference chip program». Нельзя писать «Liang Wenfeng officially announced chip development». Тегайте: sources familiar / early stage / unconfirmed.

30-second summary: скорее всего real, но early. Нет CEO announcement. T-Head уже в mass production. Economics drives shift; geopolitics accelerates.

01

Inference-only ASIC: заточен под serving, не под training clusters.

02

Старт ~mid-2025: описано как «about a year ago»; всё ещё early stage.

03

Supply chain talks: переговоры с chip designers, foundries, memory vendors.

04

Quiet hiring: chip engineers recruited privately, не на public job boards.

05

Dual-dependency play: снижение reliance на Nvidia и Huawei Ascend — DeepSeek уже runs on Ascend.

Credibility factorAssessment
Source tierHigh. Reuters standard «three people familiar with the matter»
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 — hardware-software co-design signal
Contradictory takesPartnership и in-house R&D run in parallel — Ascend live, custom silicon early
03

Что CEO DeepSeek Liang Wenfeng говорил про chips и compute

Liang Wenfeng дал мало публичных интервью. Самый ценный source — два deep dive с Waves (暗涌) в мае 2023 и июле 2024. Он никогда не анонсировал chip program, но обозначил 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 ещё ~1× — ~4× total compute для parity.

02

Missing tech community: domestic chips lack first-hand developer community; кто-то должен стоять на frontier.

03

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

04

Founder remarks ≠ product launch: Reuters описывает company actions (hiring, vendor talks), не CEO announcement.

05

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

04

T-Head Alibaba уже шипит — ставка Jack Ma 2018 окупается в 2026

Не пишите «Jack Ma недавно сказал, что Alibaba будет делать chips». Точная дуга: Jack Ma задал T-Head strategy в 2018, Joe Tsai объяснил export-control pressure в 2024, CEO Wu Yongming раскрыл mass-production numbers в 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: новые Alibaba chips support Nvidia CUDA ecosystem, easing engineer migration (unlike Huawei 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

Почему tech giants строят custom AI chips: cost, control и «Nvidia tax»

One-liner: AI competition сдвинулась от «who has the best model» к «who has the cheapest, most controllable compute».

01

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

02

Supply chain resilience: US export controls, allocation shortages, single-vendor risk — не только «national security», но и predictable supply.

03

Hardware-software co-design: general GPUs trade efficiency for flexibility; ASICs — наоборот для known workloads. Jalapeño targets real ChatGPT serving (KV cache, batching, latency).

04

Bargaining power: даже partial self-supply strengthens Nvidia negotiations и enables full-stack «model + cloud + chip» story.

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 — custom ASIC battleground.

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

Six-step decision guide:

01

Separate rumor from announcement: пишите «reportedly» until DeepSeek confirms.

02

Split training vs inference planning: frontier training still needs Nvidia; inference — 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 и 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 и iOS CI still need stable macOS hosts 24/7.

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

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

Последнее обновление: 10 июля 2026 · Sources: Reuters, OpenAI official, WSJ, Caixin Global, Waves interviews, Alibaba/T-Head public disclosures