AI Market Watch Asia · Fortnight 20: DeepSeek reportedly passes a $1B annualized revenue run rate while seeking about $7.45B at a roughly $75B valuation and preparing a Shanghai STAR Market listing

DeepSeek reportedly passes a $1B annualized revenue run rate while seeking about $7.45B at a roughly $75B valuation and preparing a Shanghai STAR Market listing

It counts as a record and a strategic shift: the round would be one of the largest private raises by an Asian AI lab, and the planned STAR Market listing would take a Chinese frontier lab public on a disclosed revenue run rate.

DeepSeek reportedly passed a $1B annualized revenue run rate while pursuing roughly $7.45B in new financing at a reported 500 billion yuan (~$75B) valuation and preparing a Shanghai STAR Market listing, per the L1 report. The revenue figure is the first disclosed run-rate milestone at this scale for a Chinese frontier lab, and the round-plus-listing combination would make DeepSeek the first independent general-purpose large-model company to pursue an A-share listing on a disclosed run rate.

The $1B run-rate figure roughly doubles the ~$500M ARR reported in late August, per L3 coverage. The company has raised API prices up to 12x on V4 Pro and V4 Flash since mid-August, citing compute constraints, per L3 coverage. Separately, OpenRouter data showed DeepSeek-V4-Flash leading global weekly token usage at 8.83 trillion tokens, up 570% week-over-week, per L3 coverage. Founder Liang Wenfeng reportedly said more than 70% of compute remains allocated to training rather than inference, per the L1 report.

The reported round would follow a June 2026 external raise of approximately RMB 50–51 billion (US$7.4–7.5B), which closed the shift from High-Flyer Quant parent funding, per L3 coverage. Prior investors mostly held indirect stakes via a limited partnership controlled by Liang Wenfeng; the new round opens direct ownership to a wider investor base, per L3 coverage. CATL, Tencent, NetEase, JD.com, IDG Capital, and a national AI industry fund participated in the June round, per L3 coverage. Over the first seven months of 2026 the company reported roughly RMB 475 million in revenue and RMB 715 million in net losses, per L3 coverage.

The STAR Market path has a named sponsor: CITIC Securities, with IPO materials targeted for filing in 2026 and listing in 2027, per L3 coverage. If completed, it would follow Zhipu AI and MiniMax, which listed in Hong Kong, per L3 coverage. DeepSeek raised infrastructure spend from about RMB 1.2 billion in 2025 to roughly RMB 11 billion in the first seven months of 2026 and is building a gigawatt-scale compute center in Ulanqab, per L3 coverage.

The revenue mix is the diligence question the numbers leave open: price-led expansion versus recurring inference demand. DeepSeek's own record offers a caution — independent testing of V4 Flash found it failed 46.2% of real-world agent tasks despite topping leaderboards, per L3 coverage. On compute supply, DeepSeek released native Huawei Ascend support for TileLang and five core libraries, covering kernel development and communication functions used in large-model workloads, per L3 coverage. The company is recruiting around 150 senior backend engineers as existing infrastructure nears its limits, per L3 coverage.

Alibaba unveils the Zhenwu V900 AI chip, claimed at 3x the compute of the M890, and its chip unit T-Head open-sources its SAIL migration stack, citing 650+ customers

It counts as a quantified deployment claim and a strategic shift: the chip comes with specific compute and cluster-scale claims (3x, 500,000 cards), named production customers (Ant Group, Xiaohongshu, XPeng), and open-sourced tools meant to cut the cost of moving off CUDA.

The single most consequential development of this fortnight is Alibaba's unveiling of the Zhenwu V900 AI accelerator at the 2026 Yunqi conference on September 22, which the company claims delivers three times the compute of its predecessor M890 and scales to a single cluster of 500,000 accelerator cards, alongside T-Head's expanded open-sourcing of the SAIL software stack. It ranks above routine product and funding updates because it pairs a quantified hardware claim with named production customers — Ant Group, Xiaohongshu, and XPeng Motors — and a 650-plus customer count, giving analysts a countable deployment base rather than a spec sheet alone.

At the same conference, Alibaba Group CEO Eddie Wu said the self-developed M890 AI superpod already supports inference for models with roughly 2 trillion parameters and has been scaled up in Alibaba Cloud data centers this quarter. The V900 announcement followed weeks after Alibaba priced a US$10.2 billion Hong Kong share placement earmarked entirely for AI infrastructure.

A day after the chip launch, chip unit T-Head expanded open-source access to T-Head SAIL, its CUDA-like stack linking PyTorch to Zhenwu hardware, adding the sailify migration tool, Triton-for-sail, and accelerators DeepGEMM-for-sail and FlashAttention-for-sail. T-Head says Zhenwu serves 650-plus customers across 20-plus industries, and Xiaohongshu built its own migration agent on the open code. T-Head lists 39 quantized Qwen, DeepSeek, and Kimi models on ModelScope with 348,000-plus downloads to date.

For teams evaluating non-Nvidia capacity in China, T-Head's software lead says customers' first question is how much existing code and tuning experience carries over, similar to the lock-in behind Nvidia's CUDA dominance. Separately, Alibaba open-sourced a Qwen 2.7B model and launched a dedicated Qwen mobile AI solution alongside the chip news.

Research Highlights

• MiniMax Adds M3.1-Flash-Preview Coding Model to MiniMax Code — MiniMax launched M3.1-Flash-Preview in MiniMax Code with five reasoning settings, but has not published benchmarks, API access or pricing.
• Unitree open-sources UnifoLM-WLA-1.0, a humanoid foundation model, to counter post-IPO doubts about its AI capability. — Unitree open-sourced a humanoid foundation model claiming top open-source benchmark results, addressing investor doubts about its AI 'brain' after a volatile Shanghai IPO debut.
• StepFun releases Step 5 Preview, claims top-3 global open-source rank at a fraction of Claude Opus 5's cost — StepFun's Step 5 Preview, a 600B-parameter MoE model, ranks top three on Artificial Analysis' open-source leaderboard at about one-eighth Claude Opus 5's per-task cost, weights due October 15.

Notable Developments

• Elice Group — plans an October KOSDAQ listing to raise 156.4–201.1 billion won for regional modular AI data centers, with sites secured for 40 MW and 80 MW. First-half revenue was 34.4 billion won with 10.3 billion won of operating profit, and cloud services were 76% of sales.
• Xiaomi — released the MIT-licensed MiMo-V2.6-Pro (1.02T total / 42B active parameters) and MiMo-V2.6-Flash omni-modal models, with RL training costs disclosed at about $2.62M and $850K. Pro scored 46.32 on the Artificial Analysis Intelligence Index, the highest logged for an open-weight model, but trailed Claude Opus 5 and GPT-5.6 on coding benchmarks.
• D-Robotics — raised a $400M Series C, including a strategic investment from an unnamed global internet company, to expand its Sunrise robotics chips and software stack. Cumulative Sunrise shipments have passed 8 million units, and the S600 has more than 20 embodied-AI customers, including Ubtech and Fourier.
• Anew Labs — closed an oversubscribed $290M first round at a $1.5B post-money valuation, led by HongShan, IDG Capital and Hillhouse Ventures. The ByteDance spinout's raise is one of four mega-rounds that account for about 60% of China's 2026 AI drug-discovery financing.
• Manus — is planning a $500M raise at about a $4B valuation, according to Bloomberg, its first round since Beijing forced Meta to unwind its $2B acquisition. Tencent is now its top shareholder, and the company is weighing a Hong Kong listing.
• GigaAI — is close to a Pre-IPO round at a pre-money valuation of about RMB 20B (~$2.8B) after converting to a joint-stock company. It plans to file under Hong Kong's Chapter 18C, which would make it the first world-model startup to go public.
• Huawei — detailed the Ascend 960 superpod, which scales to 4,096 NPUs and 8 EFLOPS FP8 and uses near-package optics in place of about 48,000 800G optical modules. It confirmed one chip generation per year through the Ascend 980 in 2029, and said its CANN software stack has more than 5,200 monthly active developers.
• SiliconFlow — brought its 2026 fundraising close to RMB 2.9B with state-backed investors ahead of its Hong Kong IPO. Its filing shows 2025 revenue up 653% to RMB 55.33M, while gross margin fell to -24% and the public-cloud token service ran at a -119% gross margin.
• Transwarp — completed a Hong Kong main-board listing (06727.HK) alongside its existing STAR Market shares and entered Stock Connect on its first trading day. 2025 revenue rose 20.6% to RMB 447M, and the company plans a cloud version of its GPU-native database for overseas markets in the second half of 2026.
• Inferact — benchmarked Moonshot AI's Kimi K3 at 709 tokens/sec on 16 Google TPU v7 chips versus 452 on 16 Nvidia GB200s running the same vLLM engine. It attributes the gap to a hand-written, open-sourced Pallas megakernel, with accuracy unchanged on both platforms.

By Country

🇰🇷 Korea

• Q&Smart — Q&Smart launched DaVinciQ, an AI discussion platform for middle and high school students that produces six-dimension thinking reports teachers can reference when drafting school-record comments.
• DeepX — DeepX outlined its physical-AI edge chip strategy at the CES 2027 Media Event in San Francisco, citing KTL tests showing up to 26.25% lower power draw for its certified inference chips.
• Aim Intelligence — Aim Intelligence reached the Asia-Pacific Top 10 of NVIDIA's Inception Grand Challenge 2026 from roughly 10,000 entrants, pitching a control loop linking its Stinger red-teaming and Starfort runtime monitoring.
• Liner — Liner signed an MOU with KAIST di-Lab on September 21 to pilot its Liner Scholar agent in science-and-technology and international-development policy research, including a source-verification agent.
• 3billion — 3Billion joined a KAIST-led ARISE project with Asan Medical Center, backed by up to 8 billion won through 2030, to build an N-of-many ASO therapeutics platform for rare diseases.
• Dental Robotics — Dental Robotics raised an undisclosed seed round led by AI Angel Club to build an AI-guided robotic arm that automates dental suction, targeting a 2028 clinical launch.
• SK AX — SK AX introduced AI Bootcamp Vibe on October 1, a four-week program teaching non-developers to build workplace AI agents that can connect to internal email and data through its Context Operator service.
• Lablup — Lablup unveiled Continuum Router for multi-model request routing and Continuum Hub for centralized token usage, policy and billing at its September 29 technical conference at COEX in Seoul.

🇨🇳 China

• Fermi Universe — Fermi Universe raised RMB 100 million (about $14 million) at a roughly RMB 1 billion valuation and launched FermiQLLM 1.0, a Qwen-based model applying quantum-inspired methods on conventional GPUs.
• Fermi Universe — Fermi Universe closed a RMB 100 million seed round and reports FermiQLLM 1.0 gained 10-20% on MATH-500, GPQA-Diamond and BBH versus comparable open base models in internal tests.
• Qingxing Heterogeneous — Qingxing Heterogeneous completed an A+ round, bringing three-month funding to nearly RMB 100 million for RiverONE, a 1.9B-parameter quantum-inspired vision-language model that runs on conventional GPUs.
• DeepSeek — DeepSeek is shifting AI training and inference workloads from Nvidia GPUs to Huawei Ascend chips, betting on domestic silicon to sustain scaling under US export controls.
• GENISOM AI — GENISOM AI's Series B spotlights a quadruped-first robotics strategy built on claimed mass production and industrial deployments.
• GENISOM AI — Genisom AI raised RMB hundreds of millions in a Series B led by UAE's Stone Venture, on the strength of 15,000 cumulative quadruped robot shipments.
• Infinigence AI — Infinigence AI and Huahuan Electronics signed a strategic partnership to build domestic heterogeneous-compute AI infrastructure and a token-based inference service model.
• YouShu Quantum — YouShu Quantum introduced UnitarySpark, a desktop workstation combining quantum simulation, GPU acceleration and quantum-hardware interfaces, and opened the UnitaryLab 2.5 beta for natural-language scientific-computing workflows.

🇯🇵 Japan

• AIONA — AIONA secured ¥140 million through J-KISS equity led by ON&BOARD plus bank loans, funding development of REVY, its AI agent for manufacturing design tasks including FMEA, DRBFM and drawing checks.
• MILY — MILY raised an undisclosed round led by POLA ORBIS CAPITAL, the POLA ORBIS Group's CVC arm, to expand content, one-on-one sessions and small events for its MILY Club membership matching community.
• AIONA — AIONA closed a ¥140 million (about $0.9 million) seed round backed by ON&BOARD and Mizuho Capital to build REVY, an agent that captures engineers' design decisions and rationale as reusable knowledge.
• Muse — Meta plans to bring its Muse personal agent to AI glasses within months, adding shopping connections to Walmart, Best Buy and Wayfair, while Japan availability and supported devices remain undisclosed.
• QTnet — QTnet added a Legal Research feature to its QT-GenAI platform that drafts reports citing specific articles from Japan's e-Gov database, built on the Digital Agency's open-source Government AI Genna tool.
• Tier IV — Tier IV's chief executive warned that a US-made Level 4 system could threaten Japan's auto industry after GO, Nihon Kotsu and Waymo partnered to target Tokyo robotaxi operations from 2027.
• Alback — ULVAC is developing panel-level packaging equipment for large rectangular AI semiconductor substrates, drawing on its display-glass handling experience and targeting pilot lines' 2028 move into mass production.
• AMD — AMD published an EPYC whitepaper positioning its CPUs for agentic AI workloads, claiming a 96-core Venice processor delivered 1.2x the per-core SPECrate 2026 Integer performance of Nvidia's Vera.

Other East Asia

• Google — Google is testing Flipkart-branded checkout links for smartphones and electronics inside Gemini and AI Mode in India, with a broader rollout reportedly planned for October ahead of the festive season.

Southeast Asia

• Manus (SG) — Manus introduced Manus 2.0 and Cue, expanding its general-purpose agent lineup; reliability and adoption remain unproven.
• Sapient Intelligence (SG) — Sapient Intelligence reports stronger MLE-bench results at lower model cost than a Claude Code baseline, but model choice and the research graph's contribution remain unisolated.

Compiled from 33 linked sources across AI Market Watch's Asia news index. Fortnight covers 2026-09-18 – 2026-10-02.