В июне 2026 OpenAI и Broadcom представили Jalapeño — custom inference ASIC с nine-month tape-out. Через две недели, 7 июля, Reuters сообщил, что DeepSeek — китайская lab за R1 и V4 — тихо разрабатывает собственный inference-only AI chip, по словам трёх informed sources. Парадокс: DeepSeek уже работает на Huawei Ascend, но параллельно идёт in-house silicon. Co-design partnerships и internal R&D идут рядом; custom chips на ранней стадии, partnerships активны.
Data-driven материал для dev'ов, infra-инвесторов и CTO, которые следят за AI unit economics и supply-chain risk. Синтез Reuters, WSJ, постов OpenAI, интервью Waves и earnings Alibaba: глобальная chip wave июля 2026, evidence и timeline DeepSeek, что CEO Liang Wenfeng говорил (и не говорил), mass production T-Head, пять драйверов custom silicon, inference vs training, риски, пять FAQ и six-step framework.
Обновлено: 9 июля 2026. DeepSeek официально chip project на момент публикации не подтвердил.
01 Это не только Китай: Jalapeño OpenAI и глобальная волна custom chips
Pain points для infra-команд:
- Nvidia tax: gross margin datacenter GPU > 70 %. На hyperscaler scale inference — recurring rent, не one-time capex.
- Allocation risk: даже US cloud giants стоят в GPU-очередях. Custom ASIC — negotiating leverage, не только nationalism.
- Workload mismatch: general GPU — швейцарский нож; LLM inference — один предсказуемый workload; ASIC выигрывает по unit economics.
- Скорость тренда: TrendForce (2026): рост custom AI chip shipments 44,6 % vs 16,1 % у general GPU — custom silicon впервые обгоняет GPU по growth.
Плотный headline cluster — июль 2026:
- 2026-06-24: OpenAI + Broadcom Jalapeño inference ASIC (9-month tape-out)
- 2026-07-02: Anthropic reportedly в talks с Samsung по 2nm custom silicon
- 2026-07-07: Reuters: DeepSeek разрабатывает inference chip
- 2026-07-07: The Information: Zhipu AI оценивает custom chips
| Компания | Проект | Стадия | Workload | Ключевой сигнал |
|---|---|---|---|---|
| OpenAI | Jalapeño (Broadcom) | Tape-out done | Inference | Deploy конец 2026; ~50% cost savings claimed |
| TPU v6/v7 | At scale | Train + infer | Gemini end-to-end на TPU | |
| Amazon | Trainium3 / Inferentia | Commercial | Both | Anthropic на Trainium at scale |
| Microsoft | Maia 100 | Deploying | Inference | Azure / OpenAI workloads |
| Meta | MTIA | Internal | Inference | Recommendations; уже один redesign |
| Anthropic | Samsung talks | Exploratory | TBD | The Information, июль 2026 |
| DeepSeek | Unnamed inference ASIC | Early R&D | Inference | $7,4B round; no official confirm |
| Alibaba (T-Head) | Zhenwu 810E / M890 | Mass production | Train + infer | 560K+ units shipped; billion-yuan revenue |
| Huawei | Ascend 950 | Mass production | Both | DeepSeek V4 adapted; orders surging |
| Zhipu AI | Custom chip eval | Early | Inference | The Information, июль 2026 |
02 Что Reuters реально сообщил (и что DeepSeek не подтвердил)
7–8 июля 2026 медиа повторили Reuters с consistent facts:
- DeepSeek строит custom AI chip для inference, не training.
- Программа стартовала mid-2025 (~год назад) и остаётся на early stage.
- DeepSeek ведёт переговоры с chip designers, foundries и memory suppliers.
- Chip engineers нанимают privately — не через public job boards.
- Успех снизит зависимость от Nvidia и Huawei Ascend.
| Измерение | Оценка |
|---|---|
| Source tier | High — Reuters standard «three people familiar with the matter» |
| Official confirmation | None на дату research |
| Circumstantial evidence | Strong — ~$7,4B external round (июнь 2026) с custom chips; IDC hiring; UE8M0 FP8 как hardware-software co-design |
| Contradictory takes | Часть analysts видит near-term рост reliance на Ascend. Точная формулировка: partnership и in-house R&D параллельно |
Пишите «Reuters reports DeepSeek launched inference chip program.» Не пишите «Liang Wenfeng officially announced chip production.» Tag: sources / early stage / unconfirmed.
| Дата | Milestone |
|---|---|
| 2023–2024 | Интервью Waves Liang Wenfeng: export bans, compute hunger |
| 2025-01 | DeepSeek R1 на Nvidia H800 (export-banned с late 2023) |
| Mid-2025 | Chip program reportedly starts |
| 2026-04 | V4 adapted to Ascend; V4-Flash partial Ascend training |
| 2026-06 | ~$7,4B funding round; custom chips в disclosed use of funds |
| 2026-07-07 | Reuters exclusive на inference ASIC |
03 Что говорил CEO Liang Wenfeng — и Alibaba T-Head уже shipping
Liang Wenfeng редко выступает публично. Лучшие sources — два Waves interview (май 2023, июль 2024). Он никогда не объявлял DeepSeek chip program. Reuters описывает company behavior — hiring, supplier talks — не founder launch event.
Четыре цитаты Liang про chips и compute:
- Export bans, не money: «Our real challenge has never been funding—it is export controls on advanced chips.» (июль 2024)
- ~4× compute gap: domestic training и data efficiency каждый отстают ~1×; вместе нужно roughly 4× compute для того же результата.
- Tech frontier: domestic chips не имеют developer community; China нужны teams на frontier, не second-hand information.
- Compute appetite: hunger researchers за compute бесконечен — мы deploy as much capacity as we can.
Alibaba T-Head: восемь лет execution, не fresh rumor. Jack Ma назвал Pingtouge (T-Head) на Apsara conference сентябрь 2018, merge Damo Academy и C-SKY. Chairman Joe Tsai (podcast 2024) связал export controls с cloud strategy; CEO Wu Yongming (FY2026 earnings) disclosed 470K+ AI chips delivered и billion-yuan annualized chip revenue.
| SKU | Timing | Highlights |
|---|---|---|
| Hanguang 800 | 2019 | Early inference ASIC |
| Zhenwu 810E | Jan 2026 | Train+infer; 96GB HBM2e; between A800 and H20; in production |
| Zhenwu M890 | 2026 | 144GB; 800GB/s die-to-die; ~3× 810E |
| Zhenwu V900 | Planned 2027 Q3 | 216GB; 1200GB/s interconnect |
| Zhenwu J900 | Planned 2028 Q3 | Next-gen parallel compute architecture |
Commercial metrics 2026: 560K+ cumulative shipments; billion-yuan annualized revenue; 400+ enterprise customers на Zhenwu clusters; registered capital raised to 1B yuan (июнь 2026); Alibaba pledged 380B yuan over three years для cloud и AI infra. WSJ: new chips CUDA-compatible для easier migration (unlike Huawei stack). Manufacturing shifting от TSMC к domestic foundries (industry указывает SMIC 7nm-class flows).
04 Почему tech giants строят custom AI chips: cost, control и Nvidia tax
AI competition shifted от «best model» к «cheapest, most controllable compute».
Пять драйверов (ranked):
- Economics — inference is rent: training — down payment; inference — monthly rent. На ChatGPT-scale DAU inference spend exceeds training. Morgan Stanley cited ~$852M для 24K Blackwell cluster vs ~$99M для equivalent TPU cluster (hardware only). SemiAnalysis/Bernstein: custom ASICs могут дать 40–65% TCO advantage over GPUs at scale; 30–40% lower per-token cost для hyperscalers. Nvidia datacenter GPU gross margin >70% — in-house silicon converts permanent GPU tax в one-time R&D.
- Supply chain resilience: US export controls на H100/H800/H20; allocation queues даже для US buyers. Security здесь — predictable supply, не только cyber risk.
- Hardware-software co-design: DeepSeek UE8M0 FP8 и MLA; OpenAI Jalapeño tuned для KV cache, batching, latency; Google TPU bound to JAX/TensorFlow. GPUs trade efficiency за flexibility; ASICs наоборот для known workloads.
- Bargaining power и differentiation: даже partial self-supply strengthens Nvidia negotiations и cloud marketing («model + cloud + silicon» full stack).
- Energy: inference ASICs optimize performance per watt — power и cooling rival chip purchase cost на gigawatt-scale datacenters.
| Dimension | Training | Inference |
|---|---|---|
| Workload | Dynamic, experimental | Static, predictable requests |
| Software moat | CUDA (cuDNN, NCCL, Nsight) | Hand-tuned kernels per model |
| Chip goal | Peak FLOPS + programmability | Throughput, latency, $/token |
| Spend pattern | One-time cluster capex | 24/7 opex at larger scale |
| Leaders | Nvidia H100/B200 | TPU, Trainium, Maia, Jalapeño, rumored DeepSeek ASIC |
Bottom line: training остаётся home turf Nvidia; inference — custom ASIC battleground.
05 Hard numbers, risks и что может пойти не так
- DeepSeek funding: ~$7,4B external round (июнь 2026); disclosed uses include custom AI chips и domestic compute expansion
- T-Head shipments: 560K+ units; billion-yuan annualized revenue (H1 2026)
- ASIC TCO band: 40–65% advantage vs GPUs at multi-year inference scale; 30–40% per-token savings для hyperscalers
- Custom silicon growth: 44,6% vs GPU 16,1% (TrendForce 2026)
- Zhenwu 810E: 96GB HBM2e; performance between Nvidia A800 и H20
- Alibaba infra pledge: 380B yuan over three years (chips, compute, liquid cooling)
Risks:
- Не пишите DeepSeek chips «confirmed» до official release.
- Meta MTIA уже saw full redesign — architecture shifts могут void ASIC bets.
- Foundry geopolitics: SMIC-class flows lag TSMC 3nm на leading-edge AI silicon.
- CUDA compatibility ≠ mature optimization stacks на day one.
06 FAQ, six-step framework и wrap-up
- Q1: DeepSeek правда строит свой AI chip? A: Reuters (7 июля 2026, three sources) reports early-stage inference chip. No official confirmation.
- Q2: Liang Wenfeng объявил chip program? A: Нет. Он cited export controls как main challenge в 2024 — не product launch.
- Q3: Какова роль Alibaba? A: T-Head (Jack Ma strategy 2018) mass-produces Zhenwu — 560K+ shipped, billion-yuan revenue 2026.
- Q4: Почему inference first? A: Predictable workloads suit ASICs; training still needs CUDA depth.
- Q5: Security или savings? A: Both — economics leads; export controls accelerate existing trend.
Шесть шагов для engineering teams:
- Chip news radar: track Reuters, OpenAI blog, Alibaba earnings для DeepSeek confirmation.
- Split train vs infer budgets: rumored DeepSeek silicon inference-only — model TCO separately.
- Watch stack signals: UE8M0 FP8, Ascend ports, T-Head CUDA compatibility — migration cost follows software, not press releases.
- API fallback chains: LiteLLM или OpenRouter чтобы supplier shifts не требовали app rewrites.
- ASIC lock-in risk: non-Transformer futures могут obsolete fixed-function silicon — keep model flexibility.
- Stable local compute: run Cursor agents, OpenClaw и iOS CI на dedicated bare-metal nodes — не shared VPS, которые stall когда supply chains move.
Primary sources (re-check links before citing in production):
OpenAI: Jalapeño inference chip with Broadcom
Caixin Global: Alibaba Zhenwu 810E analysis
SCMP: Joe Tsai on export restrictions
Shared cloud IDE и multi-tenant VPS setups struggle с context bleed, sleep-interrupted long inference jobs и missing 24/7 launchd supervision когда GPU allocation shifts. Для production Cursor agents, OpenClaw gateways и iOS CI/CD CALMVPS bare-metal Mac Mini rental offers dedicated Apple Silicon, 120-second provisioning и monthly elasticity — swap env vars на isolated node когда chip supply или API pricing moves, без rebuilding entire stack.