Agent wars, IPO mania, Qwen-Apple prize 🚀

        August 17, 2026

Agent wars, IPO mania, Qwen-Apple prize 🚀

by Kai · The Strategist
12-min read

The Week
China's AI sector is moving past the model race and into a fight over infrastructure, distribution, and monetization. This week's stories show the new battlegrounds: agent orchestration layers, paid tiers, public listings, and cyber-specific capability, as U.S. players like Meta push back on open-weight dominance.

The Lead
China
Unitree IPO mania masks conflicting humanoid robot rankings
Unitree Robotics' Shanghai IPO drew more than 9.8 million retail orders and was oversubscribed more than 5,500 times, raising 6.1 billion yuan (US$900 million). Meanwhile, research firms are split on which Chinese vendor leads global shipments: one says AgiBot took 44% share on roughly 8,400 units, another says Unitree outsold AgiBot.

The 6.1 billion yuan (US$900 million) Unitree IPO attracted 9.8 million retail orders, with only 0.018% of orders filled.
Unitree increased its retail share allocation by half, yet still left most orders unfulfilled.
Smart Analytics Global says AgiBot shipped roughly 8,400 humanoids in the first half of 2026, a 562% year-on-year jump, for 44% global share.
Tech in Asia reported Unitree shipped more than 5,500 humanoids, compared with 5,168 for AgiBot, contradicting SAG's ranking.

The headline story is a 5,500 times oversubscribed IPO. The real story is how little that order book has to do with robots in operation. Unitree shipped more than 5,500 humanoid robots in the first half, a number that is hard to square with a 6.1 billion yuan share sale that drew 9.8 million retail orders. AgiBot, depending on the research firm, shipped roughly 8,400 units and captured 44 percent of the global market, while another data point has Unitree ahead of AgiBot by about 5,500 units to 5,168. These are small volumes for a hardware industry, and the rankings conflict precisely because the market is too young to have clear leaders. The order book is not small. That gap is the first sign that the frenzy is about share scarcity, not product maturity.
The mechanics of this trade give all the pricing power to the issuer and its early shareholders. Retail demand is rationed to 0.018 percent of orders filled, which means the secondary market will set a price far above the IPO price for investors who missed the allocation. The vendors themselves face the opposite pressure. Margins depend on scaling a young product line, and with shipments in the thousands, no company has enough volume to demonstrate durable profitability. AgiBot's 562 percent year-on-year shipment growth and diversified portfolio show a land grab, but a land grab in physical hardware requires constant capital. The competitive battleground has shifted from engineering specifications to balance sheets.
What most coverage will miss is that the IPO itself is now the strategy, not merely a funding event. This week's issue shows software AI companies using pricing as a signal of confidence: DeepSeek raised API prices with peak and off-peak tiers, and Alibaba added paid tiers to its Qwen assistant. Robot vendors have no such software margin to expose. Their revenue is hardware volume, and the volume is measured in thousands, not millions. Raising billions in public equity before that volume exists converts retail enthusiasm into a war chest, and the war chest buys manufacturing capacity, component supply, and distribution. It also converts the dominant risk from engineering failure to valuation failure.
The second-order implication is a capital market race among Chinese AI and robotics names. AgiBot and Unitree are preparing public listings while ModelBest has started a pre-IPO process. Investors may be treating humanoid robots like the next large language model, but language models have metered usage and recurring software revenue. A robot is a depreciating asset with hardware replacement costs. The company that wins the IPO race may not have the best product; it may simply have the largest cash position. With conflicting rankings, oversubscribed share sales, and unit volumes still in the low thousands, the sector has reached the funding stage before the proving stage. That is not necessarily a bubble, but it is a market where the order book and the real world have not yet converged.

The signal: The 5,500 times oversubscription is a speculative mania signal, not a proven demand signal for humanoid robots, and the conflicting shipment numbers show no one can yet credibly claim market leadership. The first company to list gets a capital war chest to build real volume, so the race to public markets matters more than today's disputed sales tally.

scmp.com · techinasia.com

China

Alibaba adds paid tiers to Qwen AI assistant
Alibaba has introduced paid membership tiers for its Qwen AI office assistant, with annual plans starting at 200 yuan (US$29.60), testing whether users will pay for the service. The chatbot remains free, while video generation credits are sold separately. This comes as Alibaba's Qwen model is expected to power Apple's AI features in China alongside Baidu.

Annual plans for Qwen range from 200 yuan (US$29.60) to 1,499 yuan.
Video generation credits cost between 26 yuan and 968 yuan.
The Qwen chatbot remains free to use.
Apple is expected to use Alibaba's Qwen model and Baidu technology in China.

The signal: Alibaba's push to charge for Qwen reflects the mounting cost of AI, but the subscription fees are secondary to the reported Apple deal. If Qwen becomes the model inside iPhones sold in China, Alibaba shifts from a consumer chatbot maker to a core AI infrastructure supplier, a far larger opportunity than membership revenue. The paid tiers are a small test; the Apple partnership is the strategic prize.

scmp.com · techinasia.com

AI startup ModelBest starts pre-IPO process for mainland listing
Chinese AI startup ModelBest, a partner to Samsung and Huawei, has begun the mandatory pre-IPO tutoring process for a mainland China listing, hiring Citic Securities to guide the preparation. The four-year-old company is betting on small AI models that run locally on devices and are optimized for domestic processors, a strategy shaped by U.S. chip export controls.

ModelBest hired Citic Securities for pre-listing preparation, according to a filing to China's securities regulator.
The company has not disclosed its target fundraising size, timeline, or planned listing venue.
The pre-IPO tutoring process takes at least three months, potentially enabling an IPO next year.
ModelBest trains lightweight AI systems optimized for domestic processors amid a shortage of advanced U.S. chips.

The signal: ModelBest's listing plan is a bet that small, efficient models, not massive compute, are the winning answer to U.S. chip restrictions. If that bet pays off, it could push more Chinese AI companies toward software efficiency and on-device AI, reshaping competitive priorities away from raw hardware access. The filing also shows Chinese AI startups still view a mainland listing as a viable path, even as regulatory scrutiny of the sector stays high.

scmp.com

DeepSeek launches Harness in pivot to agentic AI
DeepSeek launched Harness on Thursday, a toolset that provides the underlying scaffolding for AI agents to operate external software and complete tasks autonomously. The release marks a strategic shift for the Chinese AI lab, moving competition from model intelligence toward how well AI agents can connect with real-world applications.

DeepSeek released Harness on Thursday, according to the South China Morning Post.
Harness provides four operational settings for enterprise and developer workloads.
A code-focused mode lets the AI write code to command multiple applications simultaneously.
A minimal mode offers isolated testing, while creative mode allows custom tool experiments.

The signal: DeepSeek is betting that the agent layer, the software plumbing between AI and real-world tools, will decide who captures value in the next phase of AI. By opening Harness to developers, the company can make its stack the default way Chinese enterprises automate work, which may matter more than topping model leaderboards. Rivals now have to match DeepSeek on integration, not just intelligence.

scmp.com

Zhipu launches GLM-5.3, claims cyber benchmark win over Anthropic and OpenAI
Beijing-based AI firm Zhipu unveiled its flagship GLM-5.3 model, saying it edged out Anthropic's Mythos 5 and OpenAI's GPT-5.6 Sol on a cybersecurity vulnerability benchmark, while trailing both on a separate exploit test. The launch highlights China's push to close the gap with Western AI in cyber defense.

GLM-5.3 scored 84.5% on CyberGym versus Mythos 5's 83.8% and GPT-5.6 Sol's 83.6%.
On ExploitBench, GLM-5.3 scored 54.4%, well below Mythos 5's 78% and GPT-5.6 Sol's 76.5%.
Zhipu said testing with Chinese security teams found 2,436 vulnerabilities across 269 projects, with 1,097 rated medium to high severity.

The signal: The CyberGym lead is paper-thin and covers only flaw detection, while the ExploitBench gap shows China still lacks equivalent AI capability for offensive exploitation. The more significant signal is that a Beijing-backed model is now being benchmarked head-to-head against Anthropic and OpenAI, confirming that frontier AI in China is being steered toward cyber defense as a strategic national priority.

scmp.com

Quick hits

▸
DeepSeek raises API prices with peak and off-peak tiers as V4 Pro launches:
DeepSeek's price increase signals the end of the low-cost token dumping that defined China's model price wars, and rivals will likely follow with their own increases. The real shift is strategic: DeepSeek is repositioning from a pure model seller to a company monetizing task delivery, using peak pricing to push workloads into cheaper windows and positioning V4 Pro as the base for agentic products.

▸
DeepSeek Launches Open-Source Agent Framework Harness:
DeepSeek is making a strategic bet that value in AI agents sits in the orchestration layer, not the model itself. By open-sourcing a harness that works with nearly 40 rival models, DeepSeek is trying to become the default interface between models and the real world, a position that would let it set the rules and capture the ecosystem even if other labs ship better foundation models. That is a far more ambitious move than releasing another model, and it is why the architecture, not the v0.1 rough edges, is the real story.

▸
Meta challenges China's open-weight AI dominance with US-friendly models:
China has dominated open-weight AI, giving its models global distribution and influence over how developers build. Meta's push gives US companies a domestic alternative with comparable performance, which could erode China's share of the open-source ecosystem. The bigger game is not model quality but who controls the default choice for enterprises, and Meta is moving to reclaim that default for the US.

▸
Chinese open-weight AI models gain ground in Europe as sovereignty debate intensifies:
Europe's real choice is not between Chinese and American AI, but between models it can host and control and models it merely rents. Regulators who equate supplier nationality with sovereignty will miss that open-weight Chinese models, run locally, may actually advance the continent's control over its own data. The more Brussels pushes for self-reliance, the more it should favor deployment models that put code on European servers.

▸
DeepSeek's V4 Pro update falls short on benchmarks, finds niche in cybersecurity:
DeepSeek built its reputation on near-frontier performance at low cost, and if V4 Pro cannot match rivals on mainstream benchmarks while pricing disappoints, it loses the price-performance edge that made it a darling. The cybersecurity strength is a useful differentiator, but niche wins do not hold up in a market where developers need general capability.

One to watch
Watch whether DeepSeek's open-source Harness becomes the default orchestration layer for agents, which would let it capture value even if rivals ship better base models.

The Asia AI Brief

                                Don't miss what's next. Subscribe to The Asia AI Brief:

                    Email
                    *

                    Older →

                Kimi K3 escapes, Qwen monetizes, DeepSeek goes national 🚀