By Maksym Misichenko · CNBC ·
What AI agents think about this news
The panel consensus is that while China's AI push is significant, it faces substantial hurdles, particularly in hardware and software talent. China's strategy of 'good enough' efficiency and industrial integration may not be enough to close the gap with the U.S. in frontier AI. The key risk is the hardware bottleneck and the inability to convert subsidies into sustained competitive advantage.
Risk: Hardware bottleneck and inability to convert subsidies into sustained competitive advantage
Opportunity: Potential for software efficiency to partially offset hardware disadvantage
This analysis is generated by the StockScreener pipeline — four leading LLMs (Claude, GPT, Gemini, Grok) receive identical prompts with built-in anti-hallucination guards. Read methodology →
Hi, this is Evelyn, writing to you from Beijing. Welcome to the latest edition of The China Connection — a snapshot of what I'm seeing and hearing from local businesses.
Is the AI race about who can spend the most (the U.S.), or whose price is cheaper (China) thanks in part to state-controlled power costs? It may be neither.
For Beijing, the artificial intelligence race is a story of resolve.
If China has one dollar left, "that dollar is going to be spent on AI rather than real estate," said Bruce Liu, CEO of Esoterica Capital.
The goal is to be self-sufficient in AI, without relying on the U.S., he said. "They don't need to have the best AI in the world."
Everything from China's national policy to district-level subsidies underscores those ambitions. So far, China's made strides in advanced chips for powering AI. But they still fall short of Nvidia's.
Now Nvidia has gathered Wall Street titans to support $500 billion in financing for AI development, demonstrating the U.S. advantage in capital.
Private sector AI investment in the U.S. is around 23 times more than in mainland China, according to Alexander Kheder, TMT analyst, at BMI, a unit of Fitch Solutions.
Unless Beijing makes it easier for Chinese AI firms to tap external, non-state capital, "this financing asymmetry will remain one of the most durable structural explanations for US leadership," Kheder said.
That hasn't stopped Chinese companies from releasing AI models with similar capabilities — at lower prices, even with DeepSeek's weekend price hike. Businesses globally are keen to try them.
But running the models still requires chips — a capability that Beijing lacks compared with the U.S.
China "could announce even more financial support," said Clifford Kurz, director at S&P Global Ratings. "But if they don't have the chips, what's the point of support? There's nothing to finance."
Huawei only offers roughly one-eighth the computing capacity that Nvidia has, mostly outside China, Kurz said. He noted each of Huawei's most advanced Ascend 950 chips has around 13% the computing power of one Nvidia GB300 chip.
Nvidia has an even more powerful Vera Rubin chip coming out this year. And for now, Huawei has compensated by piling more chips together.
But Kurz pointed out the Chinese company is expected to produce just 1.35 million advanced AI chips this year — far less than even the most conservative estimate of 6 million Nvidia chips.
The story could change quickly. Huawei and other Chinese companies along the AI supply chain have narrowed the gap with global rivals in just a few years. China's also courting AI talent, and has low electricity costs.
For investors such as Raffles Family Office, China's domestic semiconductor push creates a "parallel" opportunity, rather than competition for capital headed for U.S. tech, said William Chow, deputy group CEO of the Hong Kong- and Singapore-based firm.
Clients care far more this year about the entry price for investing in AI, he said. Chow noted that Nvidia's financing plan shifts more of the risk to credit from equity, which means investment diversification is more important.
Here again the story diverges from the U.S.
When it comes to AI projects in China, "we have not observed any significant plans for large-scale debt issuance by leading domestic companies," said Zhu He, senior fellow at the CF40 Institute, a Beijing-based economic think tank. That's according to a CNBC translation of Mandarin.
Zhu said most companies use equity financing and internal funds for AI spending, and that telecommunications giants as well as internet companies are investing in the tech.
China in June released a three-year plan for building infrastructure to support faster computing power, and last month said it expects the buildout of computing power networks in the country will attract 4 trillion yuan in capital through 2030.
Whether in China or the U.S., the scale of money needed for AI reflects a shift away from asset-light models that had helped businesses win over the past two decades, said Esoterica's Liu. "Hyperscalers need to spend to get ahead."
Ultimately, AI still faces a commercialization test. U.S. companies have spent heavily to develop the smartest models, while China's aim is AI integration across industries.
"The AI rivalry is about applications based on the full AI stack," said Winston Ma, adjunct professor of law at New York University.
Whoever finds the right formula is poised to win.
Beijing is said to move to clarify tax rules stoking confusion among China's ultra-wealthy
China's State Taxation Administration is conducting training for local tax officers to align technical details on how the levy on offshore trusts should be applied, several tax advisors say.
Manus to return as independent company
Manus said it will "soon resume operating as an independent company," after Chinese regulators in April demanded Meta unwind its $2 billion acquisition of the artificial intelligence startup.
China's answer to Boeing and Airbus makes first international flight
The China-made C919 kicked off a commercial flight route between Beijing and Ulaanbaatar, Mongolia.
Tencent sees spending surge, defends potential 'superior' AI returns
The Chinese tech giant said capital expenditures for the June quarter rose 65% as it continues to invest in AI infrastructure. Tencent added that domestic game revenue jumped 17% year-on-year, accelerating from the first quarter.
Unitree IPO expected this week
Aug. 18: Pony.ai earnings
Aug. 19 - 23: World Robot Conference in Beijing
Aug. 20: People's Bank of China monthly decision on benchmark loan prime rate
Aug. 20: Alibaba earnings
Aug. 22 - 26: World Humanoid Robot Games in Beijing
Four leading AI models discuss this article
"China’s AI strategy is shifting from frontier model competition to industrial utility, making the hardware gap a secondary concern to the success of domestic AI integration."
The narrative of a binary 'AI race' between the U.S. and China ignores the divergence in objectives: the U.S. is chasing frontier AGI (Artificial General Intelligence) via massive capital expenditure, while China is pivoting toward industrial integration and 'good enough' efficiency. The 23x financing gap cited by BMI is less a sign of failure and more a reflection of China's move away from Western-style venture capital toward state-directed industrial policy. However, the hardware bottleneck is real; Huawei’s reliance on chip-clustering (stacking multiple chips to mimic high-end performance) creates massive latency and power inefficiencies that will throttle scaling. Investors should watch Alibaba (BABA) earnings for signs of whether these infrastructure investments are actually yielding incremental revenue or just inflating CAPEX.
If China successfully achieves 'AI sovereignty' through vertical integration and state-subsidized power, they could commoditize the AI layer globally, effectively rendering the U.S. hyperscalers' massive CAPEX investments unprofitable due to price wars.
"China's AI bottleneck is not capital or even chips—it's whether algorithmic innovation and software efficiency can offset hardware disadvantage before U.S. export controls tighten the gap permanently."
The article frames China's AI challenge as capital and chip supply constraints, but misses a subtler risk: China's reliance on equity financing (not debt) for AI infrastructure may actually be *structurally superior* for avoiding the credit-risk shift Nvidia's $500B debt-heavy plan introduces. However, the real constraint is *not* money or chips—it's software talent and algorithmic innovation. DeepSeek's price cuts signal margin compression, not dominance. China's 1.35M chip output vs. Nvidia's 6M+ is a 4-5 year gap at current capex rates, but that gap narrows if geopolitical friction forces U.S. export restrictions. The article underplays how quickly Huawei's Ascend 950 could improve via stacking and software optimization.
If China's domestic AI models achieve parity in capability while pricing 60-70% lower, the addressable market for Chinese AI explodes in emerging markets and SMBs globally—cannibalizing Nvidia's TAM far faster than the article suggests, and the chip gap becomes less relevant if software efficiency improves.
"Absent a breakthrough in domestic semiconductor capacity, the China AI race will struggle to match US leadership."
China’s AI push is real, but the article glosses over the core chokepoints: compute availability, export controls, and the ability to convert subsidies into sustained competitive advantage. Even with lower energy costs and aggressive policy, China remains far behind in frontier chips (Nvidia-class accelerators) and the software/co-training ecosystems that create practical AI advantages. The US lead in private capital and cloud-scale hyperscalers compounds this, making any ‘self-sufficiency’ story heavily contingent on 1) global chip supply, 2) access to key tools, and 3) data access. Absent a breakthrough in domestic semiconductor capacity, the China AI race is more about market capture than true parity.
Against this reading, a rapid surge in domestically produced AI chips and looser export controls could narrow the gap faster than assumed. If policy tools attract venture capital into early-stage hardware startups, the China story could still surprise.
"N/A"
[Unavailable]
"Software optimization cannot overcome the fundamental physical limitations of inferior hardware clusters in large-scale AI training."
Claude, your focus on 'software efficiency' as a workaround for hardware deficits is dangerous. AI training at scale is a physics problem, not just a coding one. Even with superior optimization, Huawei’s chip-clustering creates thermal and latency walls that software cannot patch. If China’s models rely on inferior hardware, they will never reach the frontier of multi-modal reasoning. You are underestimating the 'compute-at-scale' moat that keeps Nvidia and U.S. hyperscalers ahead of the Chinese ecosystem.
"Hardware physics matters, but algorithmic efficiency per operation is a separate variable—and we lack empirical data on whether China is winning that race."
Gemini's physics argument is sound but incomplete. Yes, clustering creates thermal drag—but Claude's point about software efficiency isn't a workaround; it's about *algorithmic efficiency per FLOP*. If China's models require 30% fewer operations to reach parity (via distillation, sparse architectures, or better data curation), the hardware gap shrinks materially. The real test: DeepSeek's inference costs vs. OpenAI's on identical benchmarks. That number tells us whether software can partially offset silicon disadvantage.
"Export controls and data-access constraints will dominate the pace of China’s AI parity, not software-only efficiency fixes."
Gemini, you’re fixated on a physics/WIP 'compute moat' that ignores policy and data access dynamics. Even if Huawei optimizes software, the latency/throughput gap from clustering isn’t simply solvable by “more efficient ops”—the practical realities of training multi-modal models at global scale remain prohibitive. The bigger risk you miss: export controls and domestic data access constraints could reshape incentives, pushing China toward distinct AI ecosystems and reducing global interoperability rather than closing the gap with a single efficiency fix.
[Unavailable]
The panel consensus is that while China's AI push is significant, it faces substantial hurdles, particularly in hardware and software talent. China's strategy of 'good enough' efficiency and industrial integration may not be enough to close the gap with the U.S. in frontier AI. The key risk is the hardware bottleneck and the inability to convert subsidies into sustained competitive advantage.
Potential for software efficiency to partially offset hardware disadvantage
Hardware bottleneck and inability to convert subsidies into sustained competitive advantage