By Dongyan Zheng in AI — 30 Jun 2026

ChinaBiz Briefing | Cambricon ¥1T, Kimi $31.5B, DeepSeek V4 Pricing, BYD Chip, 15 Embodied AI Unicorns

China's technology and capital markets closed the first half of 2026 with a cluster of milestones that collectively signal a sector in structural acceleration, not speculative drift. A domestic AI chipmaker crossed a valuation threshold that would have seemed implausible two years ago. A large language model startup repriced itself upward by 57.5% in a matter of weeks. A price war is being prosecuted not through subsidies but through genuine architectural efficiency. And a robotics investment frenzy is minting unicorns at a pace that is simultaneously exciting and arithmetically unsustainable. Taken together, June 30 reads less like a single trading day and more like a half-year report card for China's technology ambition.

Cambricon Becomes China's First AI Chip Company Worth ¥1 Trillion

Cambricon's market capitalization breached RMB 1 trillion (US$138.9 billion) on June 30, rising as much as 8.84% intraday after China's State Council on June 29 issued a directive to accelerate national-scale intelligent computing cluster construction and expand "AI+" deployment across manufacturing, healthcare, and automotive sectors. The stock has gained more than 80% since the start of 2026. In Q1 alone, the company posted revenue of RMB 2.825 billion — a 159.56% year-on-year increase — and generated positive operating cash flow for the first time on a quarterly basis.

The milestone matters for reasons that extend beyond the headline number. Cambricon is the only publicly listed pure-play general-purpose AI chip designer in China, which makes it the default benchmark for the country's semiconductor self-sufficiency narrative. Its forward P/E of 250x is elevated, but not the most extreme in its peer group — Haiguang trades at 316x. The more important data point is market share reality: Cambricon held roughly 2.9% of China's AI accelerator card shipments in 2025 by unit volume, while Nvidia retained 55% despite export controls. The bull case rests on a structural shift already visible in aggregate data: domestically produced AI accelerators captured 41% of China's total market in 2025, up sharply year-on-year, and Cambricon reportedly holds more than 35% share of compute deployed for large language model training and inference within that domestic cohort. Whether Q1's 159% growth rate is sustainable as base effects steepen will be the defining question when Q2 results arrive.

Kimi Raises at $31.5B as ARR Crosses $300M — China's Anthropic Moment

Moonshot AI, the startup behind the Kimi large language model, has launched a new funding round at a pre-money valuation of $31.5 billion — up sharply from the $20 billion valuation of its previous round, which only recently closed. The catalyst is a revenue inflection: Kimi's annualized recurring revenue crossed $300 million in mid-June 2026, driven primarily by developer adoption and API consumption. API revenue now accounts for more than 70% of total revenue.

The revenue mix is the signal investors should focus on. A business generating over 70% of revenue from API consumption carries stronger scalability and retention characteristics than one dependent on advertising or consumer subscriptions. Observers have drawn explicit comparisons to Anthropic's early commercial expansion — expanding developer call volumes, rising API revenue share, growing overseas paying users, and pricing power that moves upward alongside model improvements. No lead investor or target raise size has been disclosed, but the valuation trajectory — from $3 billion to $20 billion to $31.5 billion within roughly 18 months — reflects a market that is pricing Kimi's commercial momentum, not merely its model benchmarks.

DeepSeek V4 Doubles Peak-Hour Prices — and Is Still 17× Cheaper Than GPT-5.5

DeepSeek announced a peak-valley pricing structure for its V4 API on June 29, with output rates doubling during Beijing business hours to RMB 12 per million tokens (approximately US$1.71). Off-peak rates remain at RMB 6 per million tokens. By comparison, GPT-5.5’s equivalent output rate stands at US$30 per million tokens — roughly RMB 210, making DeepSeek’s pricing significantly lower. The announcement triggered 2.91 million views on Zhihu within hours. DeepSeek V4 Flash has held the top spot on OpenRouter’s global single-model usage ranking for six consecutive weeks, recording 4.66 trillion weekly token calls.

The pricing move is best understood as demand management, not a revenue patch. The off-peak baseline already represents a 75% reduction from V4's April 24 launch price — the peak surcharge is an increment above a permanently discounted floor, not a rollback. Two days before the announcement, DeepSeek published the DSpark inference framework, co-developed with Peking University, which lifts single-user inference speed by 57–85% depending on model variant. The sequencing — efficiency paper on June 27, pricing announcement on June 29 — was almost certainly deliberate: DSpark expands the cost elasticity buffer that makes the price floor structurally defensible. For North American developers, Beijing peak hours correspond to 21:00–06:00 U.S. Eastern Time, meaning American enterprise users access DeepSeek APIs at off-peak rates by default — an unintended but structurally durable subsidy. DeepSeek's mid-June funding round of more than RMB 50 billion at a post-money valuation exceeding RMB 338 billion (US$46.9 billion) funds infrastructure expansion to sustain that cost advantage, not to close a cash shortfall.

BYD's Xuanji A3 Chip Targets 2027 Production Debut on Denza

BYD is on track to deploy its proprietary Xuanji A3 autonomous driving chip in a mass-production vehicle for the first time in 2027, with the Denza brand expected to serve as the launch platform, according to an exclusive LatePost Auto report. The Xuanji A3 is a 4-nanometer chip delivering over 700 TOPS per unit; a three-chip configuration reaches 2,100 TOPS — sufficient for L3 and L4 autonomous driving. BYD claims 20% lower power consumption per compute unit than comparable products, and proprietary algorithm optimization that doubles effective compute utilization. The chip was officially unveiled in May 2026, with Chairman Wang Chuanfu declaring: "The first half of the EV era was defined by batteries; the second half of the intelligent era will be defined by chips."

The strategic implication is significant. BYD has historically dominated through vertical integration of battery and powertrain technology; the Xuanji A3 signals an extension of that model into the semiconductor layer. If the 2027 deployment is realized, BYD transitions from a battery-centric hardware supplier to a full-stack intelligent vehicle platform — a shift that would intensify competitive pressure on both domestic rivals and established chip vendors, including Nvidia and Horizon Robotics, that currently supply China's EV market. The timeline carries inherent execution risk: chip-to-vehicle integration typically requires at least one year of independent validation across the chip, algorithm, and full-vehicle compatibility layers.

15 Embodied AI Unicorns in Six Months — and a Cash Runway Problem

At least 25 Chinese embodied intelligence startups now carry valuations above RMB 10 billion (US$1.39 billion), with 15 crossing that threshold in H1 2026 alone. The sector absorbed more than RMB 46 billion (US$6.39 billion) in disclosed funding during the period — a pace that eclipses even the 2023 large language model funding frenzy. The latest entrants, AI² Robotics and X Square Robot, each disclosed valuations exceeding RMB 20 billion on June 29.

Notable transactions include:

Sudo AI reaching a RMB 13.6 billion valuation just 11 months after incorporation;
TARS securing a US$455 million Pre-A round co-led by Hillhouse and Sequoia China;
COOWA — one of the few profitable players in the sector — raising US$600 million-plus at a US$3 billion valuation.

The structural tension is explicit and quantified: most startups in the cohort carry cash runways of only 18 to 24 months, according to industry analysts, meaning a consolidation or elimination event is likely between 2027 and 2028. Spirit AI founder Han Fengtao offered the sector's most candid self-assessment: current model capability is equivalent to "a one-to-two-year-old child," and meaningful scaled deployment is "at least two years away." Jensen Huang called 2026 "the commercialization year for humanoid robots" at GTC in March; several Chinese founders privately call it "the year of elimination." The investment thesis animating the cohort is explicitly an options bet — pricing in the possibility of a "GPT moment for the physical world" rather than near-term earnings. The companies that exit the 2027–2028 window with defensible unit economics will define the next credible cohort; those that do not will validate the more cautious reading of today's valuations.

LandSpace Clears Final Ground Test for Reusable Zhuque-3, July Launch Expected

LandSpace completed a full-stack static fire test of its reusable Zhuque-3 Y2 launch vehicle on June 29, closing out all key ground verification milestones. The test validated thrust stability during simultaneous multi-engine ignition, structural load-bearing capacity, and rocket-to-ground system coordination. A June launch is no longer feasible given remaining pre-launch sequence requirements; based on standard post-static-fire procedures requiring one to two weeks for data consolidation, the launch window is expected in July 2026.

The mission is designed to validate Zhuque-3's reusability and recovery capabilities — the technical and commercial cornerstone of LandSpace's ambition to compete in the reusable launch vehicle market. China's commercial space sector has attracted significant capital in 2025–2026, and a successful Zhuque-3 Y2 flight and recovery would represent a meaningful proof point for domestic reusable launch economics, directly relevant to investors tracking the sector's long-term infrastructure buildout.

XPeng Launches X-Mind Predictive World Model Framework for Autonomous Driving

XPeng has unveiled X-Mind, a technical framework that embeds a predictive world model directly into its onboard driving intelligence system. The architecture's core efficiency claim centers on a deep compression autoencoder that compresses 12 frames of future world inference into 96 tokens, stripping planning-irrelevant visual noise and retaining only core semantic priors — road topology, traffic light states, and navigation intent. The framework was trained on hundreds of millions of real-world frames and addresses the computational bottleneck associated with long-context processing on vehicle hardware. XPeng CEO He Xiaopeng separately confirmed on June 26 that the company's VLA 2.0 system is on a confirmed path toward global deployment, citing UN WP29 approval of two autonomous driving regulations, with the urban NGP standard set to become mandatory EU regulation within six months.

The X-Mind release signals XPeng's strategic differentiation on AI architecture rather than hardware specifications — a positioning increasingly important as Chinese automakers seek to convert domestic technological development into international commercial traction. The regulatory clarity provided by WP29 approval creates a concrete timeline for legal autonomous driving operation in global markets by end-2026, potentially accelerating the international expansion plans of multiple Chinese intelligent vehicle manufacturers simultaneously.

What to Watch in H2 2026

The second half of the year will stress-test several of today's narratives simultaneously. Cambricon's Q2 results will determine whether its 159% revenue growth rate is a durable trajectory or a policy-accelerated spike. DeepSeek's V4 full commercial release in mid-July — backed by Ascend supernode infrastructure scaling and a freshly capitalized balance sheet — will reveal whether its cost structure advantage survives a new architecture cycle. BYD's Xuanji A3 integration timeline will face the inherent complexity of chip-to-vehicle validation. The embodied AI cohort's 18-to-24-month cash runway begins its countdown now. And LandSpace's Zhuque-3 Y2 recovery attempt will serve as a public test of China's commercial reusable launch ambitions. The half-year report card is strong; the second-half exam is harder.

Related Coverage:

DeepSeek Doubles Peak-Hour API Prices — Still 17× Cheaper Than OpenAI XPeng’s X-Mind AI Framework: Autonomous Driving That “Sees the Future”BYD's In-House Autonomous Driving Chip Set for 2027 Production Debut on Denza Models Geely's Xingyuan Tops China's Auto Market, but Structural Shifts Are Already Eroding Its Throne LandSpace’s Zhuque-3 Y2 Passes Full-Stack Test, Eyes July Launch Window 15 Embodied AI Unicorns in 6 Months: China’s Robot Race Hits a Reality Check Cambricon Crosses RMB 1 Trillion Threshold, Cementing China's Domestic AI Chip Crown Kimi's Valuation Surges to $31.5 Billion as ARR Tops $300 Million