The Great Mispricing: China's AI Has Caught Up. The Money Hasn't.
The Great Mispricing: China's AI Has Caught Up. The Money Hasn't.
Dateline: 13 November 2025
When DeepSeek dropped its reasoning model last January, matching OpenAI's best systems for what they claimed was a fraction of the training cost, Wall Street had a meltdown. A trillion dollars evaporated from U.S. tech stocks in one session. Nvidia? It shed nearly $590 billion in market value. Traders started calling it AI's "Sputnik moment."
Fast forward nine months. Silicon Valley's AI darlings are now worth more than ever. OpenAI's valuation sits somewhere between $150 billion and $300 billion, depending on who you ask. Elon Musk's xAI? It's hit roughly $200 billion.
Meanwhile, over in Beijing and Hangzhou, the companies whose models sparked that panic are worth a pittance by comparison.
DeepSeek's open-weight reasoning model tops the independent rankings for open models right now. Its valuation? Around $15 billion. Fellow Chinese "AI tigers" like Zhipu and Moonshot AI are languishing in the low single-digit billions.
Here's the kicker: by almost any technical measure, China's best models have never been closer to the U.S. frontier. Yet by almost any financial measure, they've rarely been more deeply discounted.
This disconnect between what these systems can do and what investors think they're worth isn't just reshaping the global AI race. It's rewriting the basic playbook for building cutting-edge startups.
From years to months: China closed the gap faster than anyone expected
Most people spent the past decade assuming Chinese AI lagged behind the United States by years. Export controls on advanced Nvidia chips started in 2022. Washington tightened them repeatedly through 2024. The explicit goal? Keep that gap wide open.
That's not what happened.
A U.S. Commerce Department official told Congress in June that Huawei's latest Ascend chips still trail top American GPUs by one to two years. But here's what matters more: China's models were only three to six months behind.
Independent benchmarking backs this up. Analytics firm Artificial Analysis released a Q2 report tracking performance across reasoning, STEM exams and coding. Their finding? The intelligence gap between U.S. and Chinese frontier models "is as narrow now as it has ever been." OpenAI's o3 leads globally (now succeeded by GPT-5), but DeepSeek's R1 and Alibaba's Qwen3 are nipping at its heels.
Translation: at the very top, we're talking months and benchmark points. Not technological epochs.
Images and video: Sora 2 isn't running alone anymore
Language models started the race. Video is where things got visible.
OpenAI launched Sora 2 this autumn. It synthesizes eerily realistic clips with synchronized sound. So convincing that regulators from Tokyo to Washington have been warning about deepfake risks.
But Chinese firms quietly built rivals that creators and platforms worldwide now use interchangeably. ByteDance's Seedance 1.0 arrived mid-2025, generating cinematic, multi-shot 1080p video. It focuses on narrative coherence and runs about ten times faster than earlier models thanks to aggressive distillation and systems optimization.
A recent Time investigation concluded something striking: ByteDance's Seedance and Seedream image tools rival or beat Western tools on realism and character consistency. They're significantly cheaper too. U.S. startups are already integrating them as back-end engines.
Short-video giant Kuaishou rolled out its Kling 2.x family, which generates 1080p clips up to 30 seconds long. Independent reviewers praised its texturing, lighting and motion as "more real" than many Western competitors.
Alibaba turned its Qwen series into a full multimodal stack. The latest Qwen2.5-VL and Qwen3-VL models can parse charts and long documents. They reason over hour-long videos. They output structured JSON for vision tasks. These capabilities are pitched directly against OpenAI's GPT-5 vision and Google's Gemini 2.5.
Image and video generation isn't a one-horse race anymore. OpenAI's Sora 2 might remain the benchmark for raw physics and long-form storytelling. But Seedance, Kling and other Chinese models now sit in the same comparison tables, not some distant second tier.
The open-weights revolution just shifted east
China pulled clearly ahead in one niche with outsized strategic importance: open-weight frontier models.
Artificial Analysis data shows that last November, Alibaba's QwQ-32B preview became the first open-weight model to surpass Meta's Llama 3.1-405B on its composite intelligence index. DeepSeek's R1 and R1-0528 updates consolidated that lead in 2025. China became the locus of the world's most capable open-weight LLMs.
A separate survey of open-source LLMs this year describes China as a "global leader in open-source AI." The core stack? DeepSeek, Qwen, Yi, Baichuan and GLM. Most come under permissive licenses like Apache 2.0 that allow commercial use.
Meta's long-awaited Llama 4 family struggled by comparison. Scout, Maverick and the still-delayed Behemoth haven't lived up to their billing as the flagship open alternative to GPT-5.
The vanilla Maverick model ranked noticeably below top closed models on popular chat benchmarks. Meta initially touted parity with GPT-4o. Researchers accused Meta of "benchmark gaming" after a tuned experimental Maverick variant quietly climbed to the top of the LMSYS Arena leaderboard, misrepresenting typical performance. Internal reports and independent analysts described Llama 4 as "lost," over-engineered and late. It became a victim of escalating compute costs and an unclear product narrative.
Don't get me wrong: Llama 4 isn't a failure. It remains a strong open-weight model in many tasks. Some benchmarks even show it slightly ahead of DeepSeek on factual accuracy while trailing on complex reasoning and coding.
But in the symbolic race for open-weight leadership? Meta's stumbles proved costly. DeepSeek R1-0528 is now rated the world's most intelligent open-weight model. Chinese labs like Alibaba and Z.ai routinely open-source their flagships. The center of gravity for open models has clearly shifted east.
Building frontier AI on a hardware diet
China's progress looks most striking when you consider the hardware it doesn't have.
Washington has repeatedly tightened export controls on advanced GPUs since 2022. Nvidia's A100, H100 and even "China-only" variants like the A800 and H800 got blocked. Later rules extended limits to the H20. U.S. allies aligned with many of these measures. American citizens can't support some Chinese fabs anymore.
A new round of rules this year even targeted cloud access and the sharing of model weights with Chinese actors.
These controls bite hard. Huawei expects to produce no more than 200,000 of its advanced Ascend AI chips in 2025. That falls far short of domestic demand. U.S. guidance now warns allies against using those chips anywhere in the world.
Yet instead of freezing China's AI ambitions, the sanctions catalyzed a wave of frugal engineering.
DeepSeek's V3 and R1 models reportedly match GPT-4-class performance while costing around $5.5 million to train. That's orders of magnitude less than the hundreds of millions often cited for U.S. frontier models. Z.ai's GLM-4.5V vision-language model gets priced at roughly 13% of DeepSeek's serving costs. It can run on as few as eight Nvidia H20 chips. Its follow-on GLM-4.6 is optimized for domestic accelerators from Cambricon and Moore Threads.
DeepSeek's latest releases are tuned from day one for Chinese-made chips and software stacks like Huawei's Ascend/CANN and Cambricon's MLUs. This reduces their dependence on Nvidia's CUDA ecosystem. Alibaba and Baidu have begun training large models on their own in-house chips, partly replacing Nvidia hardware.
A growing body of analysis now argues that export controls unintentionally accelerated Chinese efforts to build a full indigenous AI stack. Meanwhile, they've hurt U.S. chipmakers' revenues and R&D budgets.
If China can achieve near-frontier performance using weakened H20 GPUs, domestic accelerators and ruthlessly optimized training schemes, the marginal advantage of the very latest U.S. chips starts looking less decisive than Washington hoped.
Follow the money: why valuations diverged
Technologically, the gap has narrowed. Financially? The chasm is immense.
Stanford's 2025 AI Index says U.S. private AI investment hit about $109 billion in 2024. That's nearly twelve times China's $9.3 billion. A more recent industry survey estimates the U.S. still leads with roughly $67.2 billion of AI investment in 2025 to China's $43.8 billion. It counts 47 AI unicorns in the U.S. versus far fewer in China.
Look at the valuations. OpenAI sits at $150 to $300 billion, depending on which late-stage deal you believe. Anthropic? Upwards of $50 billion. xAI hit $200 billion after its latest $10 billion raise. CoreWeave and other AI compute providers? North of $20 billion.
In China, DeepSeek is valued at roughly $15 billion. Zhipu, also known as Z.ai, stood at about 20 billion yuan in its last disclosed round. That's around $2.7 billion, despite topping multiple open benchmarks with its GLM models. Moonshot AI, maker of the long-context chatbot Kimi, is reportedly approaching a valuation below $4 billion.
Public markets tell a similar story. The Hang Seng Tech Index trades at around 21 times forward earnings. That's below its five-year average and at a discount to the Nasdaq-100, even after a 40 to 50 percent rally this year. The index includes Alibaba, Baidu, Xiaomi and other AI-heavy firms.
U.S. bank CEOs now openly warn of an AI-driven valuation bubble in American tech stocks. They're predicting 10 to 20 percent corrections after two years of manic inflows.
The worrying question for investors isn't simply whether there is an AI bubble in the U.S. It's whether they're mispricing the geography of the next wave of value creation.
When venture capital becomes optional
Part of the answer lies in how Chinese AI firms get funded.
Venture capital for Chinese AI startups has fallen sharply. One recent analysis notes that VC funding for the sector dropped nearly 50 percent year-on-year in early 2025. It hit just $4.7 billion in Q2, the lowest in a decade. Broader economic slowdown and regulatory uncertainty played their roles.
But that doesn't mean there's no money. It's just coming from somewhere else.
Beijing is assembling a one trillion yuan national guidance fund to back "hard tech." That's about $138 billion aimed at AI, quantum and semiconductors. It's one of the largest state-backed pools of risk capital in the world. Sub-national "government guidance funds" and city-level programs offer subsidies, cheap office space and tax breaks to AI firms. The line between VC and industrial policy gets pretty blurry.
Z.ai's latest rounds were led by municipal and provincial state-owned capital. This happened even after the U.S. Commerce Department blacklisted the company.
Then there's DeepSeek, the outlier. A popular data-driven profile this year noted something remarkable: the company has zero conventional VC funding. It's 84 percent founder-owned. It became a multi-billion-dollar "open-source unicorn" and reached over 30 million monthly active users largely on product merit and word-of-mouth.
In this ecosystem, capital isn't the scarce resource. Compute and talent are. Once those are secured through state-backed cloud credits, local chipmakers or big-tech partnerships, the need for traditional Sand Hill Road-style venture capital diminishes.
For Western founders, that raises a provocative possibility: maybe high-growth AI startups don't always need conventional VC to scale. For Chinese founders, it's increasingly just reality.
Can anything still stop China from becoming an AI superpower?
None of this means China has "won" the AI race. U.S. labs still dominate closed frontier models. GPT-5, Claude Sonnet 4.5 and Gemini 2.5 remain the systems to beat in many high-stakes domains, from legal reasoning to complex coding and agentic workflows.
China also faces serious headwinds.
Chips remain a problem. Despite workarounds, domestic accelerators still trail Nvidia's best GPUs by a generation or more. Export caps on AI chips and manufacturing equipment remain firmly in place.
Capital controls and trust issues complicate things. Geopolitical tensions, sanctions and restrictions on Chinese tech listings abroad depress valuations. They limit global adoption of Chinese AI in sensitive sectors.
Regulation at home adds another layer. Chinese platforms face their own tightening rules around AI content, data security and algorithm governance. These constraints may yet slow monetization.
But the past two years clarified something important: export controls and starved VC pipelines haven't prevented China from reaching near-parity in core AI technologies. They've made that progress more resource-efficient at best. Perhaps more domestically rooted too.
Consider the combination: world-class open-weight models from DeepSeek, Alibaba and Z.ai. Highly competitive video and image systems like Seedance and Kling. Aggressive state-backed capital replacing traditional VC. A vast internal market hungry for AI-enabled automation.
It's difficult to see how hardware throttles alone could permanently lock China out of the next industrial revolution.
The bigger question may not be whether China can catch up. That's already happening in real time. The question is whether global investors, regulators and rivals are prepared for a world where the most important open models, and a growing share of cutting-edge applied AI, get built in a system that doesn't depend on Silicon Valley's money or Washington's permission.
NOT INVESTMENT ADVICE