The Asia AI Brief logo

The Asia AI Brief

Archives
Log in
Subscribe
September 7, 2026

Huawei's thermal fix, DeepSeek's huge chip order, and the agent phone bottleneck ๐Ÿ”ฅ๐Ÿ“‹

by Kai ยท The Strategist 14-min read
The Week

China's AI sector is no longer waiting for permission on silicon, cost, or standards. This week shows the country consolidating its own compute stack, turning AI usage into a retail commodity, and wielding dollar efficiency as a strategic weapon, while the US responds by narrowing its competitive terrain to defense. The race is increasingly defined by ecosystem control and hardware allocation, not benchmark bragging rights.

The Lead
China

Huawei publishes paper countering overheating fears for Kirin 2026 chip

Huawei has posted a new study to ChinaXiv that makes the case its Tau Scaling Law design for the forthcoming Kirin 2026 mobile processor will not overheat. The report, written by He Tingbo, pushes back against the idea that stacking chips vertically causes heat problems, and it claims the new part operates at a lower temperature than the prior model while delivering 55 percent higher transistor density per square millimetre and using as much as 66 percent less power for certain workloads.

  • The paper was published Friday on ChinaXiv and has not been peer reviewed.
  • He Tingbo, chairwoman of Huawei Scientist Committee and president of its semiconductor unit, is the author.
  • The paper says the Kirin 2026 chip ran cooler than its predecessor.
  • The paper says the chip packs 55 percent more transistors per square millimetre and cuts power use by up to 66 percent.

Huawei is not publishing this paper to convince thermal engineers. He Tingbo, the head of Huawei's semiconductor business, released a non-peer-reviewed paper on ChinaXiv that explicitly frames heat as "the sharpest concern on Tau Scaling Law" and claims that measurements for the Kirin 2026 chip "turned that objection upside down." The more important audience is customers, partners, and regulators who need to believe that Huawei can keep delivering competitive smartphone processors despite restricted access to leading-edge tools. Publish this before the launch, not after, and the claim becomes part of the product narrative rather than a response to a failure. That does not make the engineering false, but it does mean readers should weigh the absence of benchmark methodology and independent review before treating 55 per cent higher transistor density and 66 per cent lower power use as settled fact.

The mechanics matter because this is a direct challenge to the assumption that chip competitiveness depends on lithography. Tau Scaling Law, as described, leans on vertical stacking to add density and manage heat without necessarily moving to a smaller process node. If Huawei can deliver a chip that runs cooler than its predecessor while packing more transistors into each square millimetre, the zone of competition shifts from pure foundry process leadership toward advanced packaging, thermal design, and architectural control over power. That would be uncomfortable for producers whose value proposition rests on being the first to reach each new node. It could also improve Huawei's margin position over time, because the company would be designing around constraints it can actually access instead of paying a premium for scarce manufacturing capacity. The margin math is still unresolved, since stacking and packaging carry their own costs, but the paper is an explicit bid to make those costs look manageable.

The second-order implication is not the phone. It is what this architecture means for China's AI compute buildout. DeepSeek's reported plan to deploy 160,000 Huawei Ascend chips for an Inner Mongolia inference cluster and Moody's observation that China is stretching its AI dollars point to the same constraint: data center power, cooling, and efficiency, not just raw chip counts. A design that cuts power consumption by up to 66 per cent on key tasks and resolves the heat question would make large clusters less punishing to cool and operate. That logic also strengthens the move toward on-device AI, where iFLYTEK has open-sourced million-token edge models and agent phones from Nubia and others are closing in on launch. China's low-cost model strategy only works if affordable, low-power hardware can carry it into mass-market devices, and the Kirin 2026 chip is the nearest available proof point for that relationship.

The risk is that Huawei has released an attractive result with no independent verification. If the engineering does not hold up under real workloads, the paper will be remembered as a marketing salvo. But if it does, the thermal objection that limited confidence in Huawei's semiconductor roadmap stops acting as a ceiling, and the efficiency gains ripple far beyond the next Kirin smartphone.

The signal: Vertical stacking was facing a thermal credibility problem, and Huawei has now published measurements that directly attack that problem. The paper claims the Kirin 2026 runs cooler than its predecessor, which flips the burden of proof onto those who said heat would be insurmountable. That makes vertical stacking a credible route for Huawei's next flagship chip rather than a dead end.
scmp.com
China

DeepSeek Plans 160,000 Huawei Ascend Chips for Inner Mongolia Inference Cluster

According to Bloomberg, citing people familiar with the matter, DeepSeek intends to put at least 160,000 of Huawei's new Ascend 950DT accelerators into a gigawatt-scale data center still being built in Inner Mongolia. The installation, which would focus on inference, is set to become one of the largest Huawei AI clusters ever described publicly, although supply constraints may push full rollout past one year.

  • DeepSeek intends to use the Ascend 950DT cards mainly for inference, not training, which continues to rely on Nvidia.
  • The Inner Mongolia campus is described as gigawatt-scale, enough electricity to power roughly 750,000 homes.
  • Huawei's 950DT output is expected to stay in the low hundreds of thousands this year, so fulfilling the order could take over a year.
  • The planned 160,000 cards cover only part of the site's capacity, leaving the accelerator mix for the rest unstated.
The signal: This planned order is the clearest sign that China's AI industry is splitting into two compute stacks, with domestic Huawei silicon handling high-volume inference while Nvidia remains the choice for advanced training. If Huawei cannot ramp Ascend 950DT supply, DeepSeek's timeline slips and the bottleneck shifts from model capability to hardware allocation. The real question is not whether DeepSeek wants 160,000 Ascend chips, but whether Huawei can deliver them without starving every other Chinese AI buyer competing for the same cards.
pandaily.com

Anthropic's Fable 5.1 tops benchmarks, but China's low-cost models change the math

Anthropic's Claude Fable 5.1 has claimed the top spot on leading benchmark indexes, widening its performance lead over Chinese rivals. Yet the model's steep operating costs stand in sharp contrast to the budget-friendly, open-weight models from China that continue to gain global commercial traction.

  • Fable 5.1 ranked first on Vals AI's index for handling complex, real-world tasks in finance, coding and law.
  • Anthropic's earlier Opus 5 and Fable 5 took second and third place on the same index.
  • Databricks technical staff member Yuchen Jin called Fable 5.1's capability leap 'insane'.
  • The article highlights a stark price gap between US frontier models and cheaper Chinese open-weight rivals.
The signal: The capability gap is real, but cost is the more decisive battleground. Chinese open-weight models are already winning the price war on inference and adoption, so America's frontier lead may translate into less commercial power than its benchmark scores suggest. The race will be settled on economics, not just leaderboards.
scmp.com

Moody's: China stretches AI dollars to narrow compute gap with US

A Moody's Ratings report says US tech giants spend vastly more on AI than Chinese competitors, but the actual gap in computing capacity is far smaller than the capex difference suggests. Lower construction costs, state policy support and cheaper green energy allow Chinese firms to buy more compute per dollar, even as the US keeps a clear lead in advanced chips.

  • The dollar spending gap between US and Chinese tech giants is huge, but the compute capacity gap is much smaller, Moody's found.
  • Chinese firms get more compute per dollar thanks to lower buildout costs, policy incentives, and cheaper green energy.
  • The US retains a clear overall lead in semiconductor chips.
  • Moody's projects China's major tech capex will rise from US$65 billion in 2025 to about US$140 billion this year and US$165 billion by 2027.
The signal: China's dollar efficiency means the US lead in total AI spending is not a proportional lead in compute, so the race hinges on access to advanced chips, not budget size. The same efficiency makes China a credible cost competitor in AI, which could pressure pricing and margins across the industry.
scmp.com
Policy & Regulation

Huang and Musk press for lighter AI rules as US-China rivalry intensifies

Top US tech executives, including Nvidia's Jensen Huang and Tesla's Elon Musk, are pressing global policymakers to ease AI regulations, warning that strict guard rails could hurt innovation as the US and China intensify their tech rivalry. Speaking at G20 sessions this week, they argued that overregulation could leave countries behind economically and called for rules targeting real harms rather than hypothetical ones. The lobbying comes as Washington relaxes its own regulatory stance, drawing criticism from Beijing.

  • At a G20 meeting, Nvidia CEO Jensen Huang urged rules targeting actual and pragmatic harm, not hypothetical theoretical harm.
  • Elon Musk said countries should create an environment relatively free of regulation where technologies are default legal.
  • Huang warned that excessive regulation could leave countries behind and prevent them from taking advantage of AI.
  • US executives are lobbying against potential bans on open-weight AI models, calling open tech vital to leadership.
The signal: This is not a debate about safety; it is a fight over competitive position in the US-China AI race. Huang and Musk are pushing deregulation as the default, which puts pressure on other governments to match US permissiveness or watch capital and talent move. The winner will be whoever controls the rulebook, and the losers will be countries that hesitate.
scmp.com
Quick hits
โ–ธ Tencent opens WorkBuddy to developers and hardware partners: Tencent is placing a strategic bet that platform control, not app features, will decide who profits from China's agent economy. By absorbing hardware makers and independent developers into WorkBuddy's ecosystem, Tencent keeps those partners reliant on its Skills, Experts and Connectors, making it hard for them to switch once they have built their products on the platform. That shifts Tencent's revenue center from selling an enterprise tool to collecting a toll on the devices and services that route through WorkBuddy, and it forces rivals to compete with an ecosystem rather than a single product.
โ–ธ Nubia NaviX Ultra gets licensing green light as Doubao agent phone nears launch: This launch is less about selling phones than about who controls the user's task. The M153 experiment showed Doubao's agent could only execute when apps allowed it, and that permission wall is the binding constraint on every agent phone. ZTE and ByteDance have cleared the hardware and compliance hurdles, but unless they have secured app-level access, the NaviX Ultra will hit the same wall at flagship prices, and the real advantage will go to whoever opens APIs first.
โ–ธ US retreats from all-front tech race with China, picks defense high ground: This strategy is an explicit concession that the US cannot out-build China across the board, so it will now focus on domains where it can keep a military edge. The second-order effect for Asia is a hardening of two supply chains, China's high-volume hardware and America's defense-tech archipelago, and companies caught in between will face higher costs and lower margins. The winners will be the regional players that pick a side early and position themselves as essential to one of these two ecosystems.
โ–ธ Zhipu Puts AI Token Plans on Tmall, Turning Coding Credits Into Consumer Retail Goods: The real story is that Zhipu is now selling AI usage like a telco sells data, with transparent credits, retail pricing and promotions. This gives Zhipu access to Tmall's traffic and shopping events, but it also invites direct price competition with telecom carriers and platform players that can undercut. Expect token retail to become a commodity business in China, and margins to follow the same path as cloud and broadband.
โ–ธ iFLYTEK Open-Sources Million-Token Edge Models to Move AI On-Device: Bringing million-token context to on-device models breaks the assumption that long-context AI requires cloud infrastructure, shifting cost, privacy, and latency advantages toward local hardware. iFLYTEK's domestic-compute training and support for Huawei and Hygon chips also give Chinese chip vendors a software argument against Nvidia, while intensifying competition in edge AI where Western rivals have favored smaller context windows.
One to watch

Watch whether Huawei can meet DeepSeek's 160,000-chip order without starving other Chinese AI buyers, as that allocation outcome will determine if domestic silicon becomes a true constraint or just a supply story.

The Asia AI Signal

Don't miss what's next. Subscribe to The Asia AI Brief:
Older โ†’ ๐Ÿ‡จ๐Ÿ‡ณ China's chip workaround, ๐ŸŽต Anthropic suit, OpenAI's India raid
Powered by Buttondown, the easiest way to start and grow your newsletter.