Unitree Closes Its Shanghai Debut Up 460% to a $50 Billion-Plus Market Cap, Marvell’s $120 Billion Google TPU Warrant Rattles Broadcom, and Beijing’s Partial H200 Release Sends A-Share Chip Stocks Reeling Even as China’s Humanoid Shipments Hit 97% Global Share
Revisiting the August 16 Brief — Unitree’s Guided Aug 17-21 Debut Window Landed Exactly on Schedule With Shares Closing 460% Above Issue Price on Aug 19; Samsung’s Guided HBM4 Tripling Extended Into Confirmed $4-5/Gb 2027 Contract Pricing; the World Robot Conference and World Humanoid Robot Games Both Opened in Beijing as Flagged; and Broadcom’s $370B Financing-Vehicle Story Escalated Into a Specific $60-70B Debt Raise Bloomberg Reported This Week.
The week of August 17-23 delivered the first real public-market pricing of the humanoid-robotics sector, and the number was not subtle. 688836.SH Unitree Robotics opened at 1,100 yuan against a 150.80 yuan issue price — a 629% intraday spike — before settling at 845 yuan for a 460% first-day gain and a roughly 342 billion yuan ($50-53 billion) market capitalization, ahead of Baidu and JD.com and second only to TSMC among recent Asian tech listings by day-one demand. Around that anchor event, the custom-silicon land grab accelerated: MRVL Marvell disclosed a Google warrant vesting toward $120 billion in cumulative TPU-ecosystem revenue through fiscal 2033, sending AVGO Broadcom down as much as 5% on diversification fears even as Bloomberg separately reported Broadcom targeting a fresh $60-70 billion AI-infrastructure debt raise. NVDA Nvidia meanwhile posted its longest losing streak since 2022 — six straight red sessions into its August 26 print — after Bloomberg reported the company is raising AI-server prices more than 15% on soaring memory costs, a direct read-through to this week’s Korean DRAM and HBM4 pricing data.
In Asia, China’s Ministry of Commerce allowed a small tranche of Nvidia H200 chips into the mainland for the first time — ByteDance and Tencent each received roughly 10,000 units, a fraction of the 100,000-per-company ceiling the U.S. approved — and the news hit domestic AI-chip names hard, with 688256.SH Cambricon falling more than 9%. That selloff sat awkwardly alongside the week’s other China data point: the China Electronics Council’s mid-year report put H1 2026 humanoid-robot shipments above 40,000 units, a 97% global share, alongside a Reuters-reported loophole in which ByteDance and Moonshot AI have continued accessing restricted Nvidia compute through Southeast Asian intermediaries — the same Singapore-Malaysia corridor the Aperia Group prosecution was built to close. Japan added its own quieter but concrete data point, with Tokyo earmarking a further $944 million for Rapidus, while Korea’s HBM4 contract pricing firmed toward $4-5 per gigabit for 2027 deliveries — the clearest evidence yet that this week’s memory-cost story is structural rather than transitory.
Scoring the August 16 Brief’s Forward Vectors
Unitree’s 460% Shanghai Debut Sets the First Public Price for Humanoid Robotics
688836.SH Unitree Robotics began trading on the Shanghai STAR Market on August 19, opening at 1,100 yuan against its 150.80 yuan IPO price — a 629% intraday spike — before paring back to close at 845 yuan, a 460% first-day gain. About 23.2 billion yuan of shares changed hands, and the closing price valued the company at roughly 342 billion yuan (US$50-53 billion), putting Unitree ahead of Baidu and JD.com by market cap and giving founder Wang Xingxing’s stake alone a value of roughly 103 billion yuan. The debut also arrived with mixed fundamentals attached: H1 2026 revenue rose nearly 49% year-on-year, but adjusted profit fell close to 19%, and the implied valuation multiple — reportedly around 262 times 2025 revenue — sets a high bar for a company that, per last week’s brief, has already lost the unit-shipment lead to privately-held Zhiyuan Innovation.
This is the benchmark every private humanoid round, SPAC and board deck will now be measured against, and it is a benchmark set almost entirely by scarcity — only about 7-12% of post-IPO shares were immediately tradable, with the balance under 12-36 month lock-ups. That scarcity, not operating performance, did most of the work in the 460% pop; H1 adjusted profit actually fell. The real test starts as lock-ups begin expiring and the market has to price Unitree against Zhiyuan’s superior shipment volume rather than against its own IPO-day supply constraint.
Watch whether Unitree’s post-debut trading range holds above 500 yuan once initial volatility settles as the signal for whether this was a scarcity spike or a durable re-rating of the sector.
Marvell’s $120 Billion Google Warrant Cracks Open Broadcom’s TPU Monopoly
MRVL Marvell disclosed on August 19 that it entered a custom-silicon agreement with Google covering AI inference accelerators, storage controllers, networking and memory-attach devices across the TPU ecosystem, paired with a warrant on 58.97 million Marvell shares (worth roughly $12.2 billion if fully exercised) that vests in tranches tied to $500 million purchase increments, up to a $120 billion cumulative-revenue ceiling through fiscal 2033. Marvell shares jumped as much as 14% intraday before closing up roughly 8%, while AVGO Broadcom fell about 5% on fears of losing exclusivity on Google’s TPU program, though Morningstar and other analysts characterized the move as Google multisourcing for scale rather than displacing Broadcom, whose own TPU relationship reportedly runs through 2031. Separately, Bloomberg reported Broadcom is now targeting more than $60 billion — potentially reaching $70 billion — in fresh AI-infrastructure debt financing, escalating last week’s $370 billion by-2029 estimate into a live, specific deal.
Structure matters more than the headline number here — Google is paying Marvell in dilution it only earns by placing orders, not cash, which means the $120 billion figure is a ceiling on Google’s future spending intent, not a signed backlog. Read alongside Broadcom’s own escalating debt-financing needs, the two stories together say the same thing: every player in the custom-silicon stack is now financing AI-infrastructure growth through increasingly exotic paper — warrants, leases, structured debt — rather than cash flow, and the market is starting to price the financing-structure risk as carefully as the demand story.
Watch Marvell’s August 27 earnings call, one day after Nvidia’s own print, for management’s first commentary on how the Google agreement is already showing up in bookings.
Nvidia’s Longest Losing Streak Since 2022 Meets a Memory-Cost Shock
NVDA Nvidia shares fell in six consecutive sessions through August 21, its longest losing streak since 2022, closing the week near $216 and about 8% off its record high, ahead of fiscal Q2 earnings due August 26. The slide coincided with a Bloomberg report that Nvidia is raising prices on servers containing its AI chips by more than 15% for systems shipping in early 2027, as soaring memory costs feed through the bill of materials — a direct consequence of this week’s Korean HBM4 and DRAM pricing data. Wall Street remains broadly unmoved on the fundamental setup: all 26 analysts tracked by TipRanks rate the stock a buy, with an average price target near $302, and a securities filing also surfaced this week showing Nvidia has guaranteed up to $105 billion in lease obligations tied to an OpenAI data-center campus in Ohio, alongside its previously disclosed $21 billion equity stake in SpaceX.
The losing streak is sentiment, not fundamentals — but the memory-cost pass-through is real and structural, and it is the same mechanism connecting Nvidia’s margin outlook to Samsung’s and SK Hynix’s HBM4 pricing power on the Korea desk below. A chipmaker raising list prices because its own suppliers are raising prices is a data point about the whole AI-infrastructure cost stack, not just Nvidia’s quarter.
Watch whether Wednesday’s print addresses gross margin guidance in light of the memory pass-through as the signal for whether this becomes a sector-wide 2027 story.
Week of August 17-23, 2026
Korea
HBM4 contract pricing for 2027 deliveries firmed further this week, with Taiwanese and Korean trade press converging on a $4-5 per gigabit range — up from roughly $2 per gigabit in the second half of 2026 — as Samsung and SK Hynix both refuse long-term fixed-price agreements in favor of quarterly contracts that let pricing float with demand. Samsung’s DRAM and NAND ASPs reportedly rose mid-40% and high-60% quarter-on-quarter respectively in Q2, versus roughly 30% and mid-50% at SK Hynix, and this week’s Nvidia server-price-hike disclosure is the clearest evidence yet that these increases are now reaching end customers rather than being absorbed by the memory makers’ own margins. 005930.KS 000660.KS SKHY
Taiwan
No new company-specific catalyst broke this week, but Taiwan remains the structural backdrop for everything above: TSMC’s next monthly revenue print (due mid-September) will be the first read on whether July’s record growth held through August, and the island’s foundry and CoWoS packaging capacity remains the binding constraint referenced in Rapidus’s own interposer-scaling roadmap below. TSM
Japan
Japan’s trade ministry confirmed on August 21 it will seek an additional ¥150 billion ($944 million) for Rapidus in the fiscal 2027 budget, aimed at boosting the company’s capital base to encourage bank lending ahead of its targeted second-half-2027 2nm mass production start at the Chitose, Hokkaido IIM-1 facility. The government is already Rapidus’s largest shareholder via a prior ¥100 billion investment, and this week’s commitment follows last week’s interposer-scaling detail on eight-reticle packaging capacity by around 2030 — Japan’s slower-moving but increasingly concrete bid for a foothold in advanced logic and packaging.
ASEAN
Reuters reporting this week put fresh detail on the loophole the Aperia Group prosecution was designed to close: ByteDance has reportedly worked with Singapore-based cloud provider Aolani to access Nvidia compute physically located in Malaysia, while a White House official separately accused Moonshot AI of training its Kimi K3 model on restricted Nvidia GB300 chips accessed through a Thailand facility. Aolani maintains its services are fully compliant and that customers hold no ownership or physical access to the underlying chips. No new docket movement surfaced in the Aperia Group case itself this week.
Korea delivered the week’s only hard forward-pricing data, and it is the thread that ties directly back to Nvidia’s own margin story above — the memory supercycle is no longer a Korea-only story, it is now visibly repricing US-listed AI hardware. Japan’s commitment is real money but a multi-year payoff; ASEAN’s compliance gap is the opposite — no new money, but a credibility problem that keeps widening every time new reporting surfaces a fresh routing method.
Unitree Goes Public Into a Chip-Supply Shock, August 17-23, 2026
Beijing allowed a small tranche of Nvidia H200 chips into the mainland this week, with ByteDance and Tencent each receiving roughly 10,000 units — against a US-approved ceiling of up to 100,000 chips per company, implying only about a 10% fulfillment rate so far. The reaction was immediate and sharp: domestic AI-chip names sold off hard, with Cambricon falling more than 9%. The release is too small to meaningfully close China’s compute gap, but it is the first crack in the “fully self-sufficient” narrative underpinning domestic chip valuations, and it lands in the same week China’s own humanoid-robot industry posted a 97% global shipment share — a reminder that the compute bottleneck and the hardware lead are running on separate clocks.
Unitree’s Debut, Zhuji Dongli’s HK Filing, and the Capital Stack
Unitree Robotics (see Core Updates deep-dive above) formally became China’s first publicly-traded pure-play humanoid maker on August 19, raising 4.2 billion yuan with 48.13% earmarked for embodied-intelligence model R&D. Its debut immediately re-rated the sector: China now counts 28 embodied-intelligence unicorns worth a combined 350 billion yuan-plus, with 18 already in some stage of the listing process. The most notable of those this week was Zhuji Dongli (逐际动力), reportedly filing confidentially for a Hong Kong IPO at a roughly 15 billion yuan valuation, backed by Alibaba, JD.com and IDG Capital after a $200 million round in July; its lineup spans the Oli and Luna full-autonomy humanoids and the modular TRON platform. Separately, DeepSeek’s second funding round — targeting 50 billion yuan at a roughly 500 billion yuan pre-money valuation — is reportedly on track to close in late August.
Compute, Models and Infrastructure
Beyond the H200 tranche above, China’s compute buildout continued on its own track: the 2026 Green Computing (AI) Conference in Hohhot signed 13 projects worth a combined 136.1 billion yuan, spanning a Volcano Engine gigawatt-scale campus, a Cambricon chip lab, and integrated power-compute facilities, lifting Inner Mongolia’s total computing capacity to 345,000 petaflops. Alibaba Cloud said its CUBE 5.0 modular design has cut large AI-datacenter delivery times to 100 days, and multiple listed compute-chain names — Foxconn Industrial Internet, Loongson, Hygon Information Technology among them — reported sharply higher year-on-year net profit in H1 results released this week, evidence that the capacity build is converting into revenue even as the H200 news rattled sentiment.
Humanoid Hardware, Policy and Standards
The China Electronics Council’s mid-year report, released at the World Robot Conference, put China’s H1 2026 humanoid shipments above 40,000 units and 97% of global volume, with domestic core-component localization now past 75%. Galbot launched its Galbot ET1 bipedal humanoid at the same conference, running the company’s self-developed “Galaxy Brain” embodied model; UBTECH separately signed a strategic partnership with Basiic Semiconductor to accelerate silicon-carbide power devices for longer-endurance robots. On policy, Beijing’s Yizhuang economic zone unveiled an “AI Talent Eight Measures” program offering up to 5 million yuan to top-tier researchers, while Shenzhen’s Longgang district, Yueqing and Fuzhou all rolled out fresh embodied-intelligence subsidy programs this week, and Guizhou province became the first to name embodied-intelligence data collection a core funding priority in its 2026 big-data plan.
The H200 tranche and the humanoid-shipment report landing in the same week is the tension worth sitting with: China’s hardware supply chain is winning on volume even as its compute supply chain remains rationed to roughly a tenth of what Washington nominally approved. Unitree’s public debut now gives outside investors a live-priced proxy for the hardware side of that story, while Cambricon’s selloff is the market’s real-time verdict on how much the compute side still depends on foreign silicon, however small the tranche.
Watch whether the H200 allocation expands beyond ByteDance and Tencent in the coming weeks as the signal for whether this was a one-off gesture or the start of a durable reopening.
Window of August 24-September 2, 2026
World Humanoid Robot Games, Beijing — carried forward, concludes this window; 2,056 robots across 51 events, real-world motion-control validation.
2026 AGIC Shenzhen AI Expo — bionic intelligence and humanoid showcase, secondary confirmation window after the Beijing events.
Nvidia (NVDA) Fiscal Q2 Earnings — first test of the memory-cost/server-price-hike story against actual margin guidance, after a six-day losing streak. NVDA
Marvell (MRVL) Earnings — first management commentary on the Google TPU warrant’s early bookings impact. MRVL
Jackson Hole Economic Symposium — Kevin Warsh’s first keynote as Fed chair, 19 days ahead of the Sept 16-17 FOMC decision.
DeepSeek’s Second Funding Round — ~¥50B round at ~¥500B pre-money reportedly closing late August; watch capital deployment into compute.
Aperia Group / ASEAN Compliance Gap — no docket movement, but fresh reporting on ByteDance and Moonshot AI routing widens the enforcement question. NVDA
Broadcom (AVGO) Fiscal Q3 Earnings — first opportunity to address both the $370B financing-vehicle estimate and this week’s reported $60-70B debt-raise target. AVGO