Artificial Intelligence thread

[Michael90]

everything to sell more GPUs

https://twitter.com/i/web/status/2080643682408321103

He’s a very good businessman like most Chinese people. Lmao. He’s for open models simply because they are almost all trained on Nvidia moat . The more they spread the better for Nvidia and its ecosystem. If Chinese open models were all trained on Chinese moat/chips with zero reliance on Nvidia, then you can bet he won’t be singing the same song lmao . lol

 

[HighGround]

He’s a very good businessman like most Chinese people. Lmao. He’s for open models simply because they are almost all trained on Nvidia moat . The more they spread the better for Nvidia and its ecosystem. If Chinese open models were all trained on Chinese moat/chips with zero reliance on Nvidia, then you can bet he won’t be singing the same song lmao . lol

In my opinion, it's more about maintaining a competitive landscape.

Huge US labs like Anthropic and OpenAI have been making efforts to corner the market and essentially, outlaw or regulate their competition to death. With only 3-5 major players in the US AI market, they can essentially check Nvidia via procurement. Having healthy competition stops that from happening.

The strongest open source players are all Chinese and they're switching to Chinese hardware anyway. Looks inevitable.

 

[henrik]

UBS notes that China's AI training costs are just 10% of those abroad, yet the sector still achieves healthy gross margins and possesses global competitiveness.

Please, Log in or Register to view URLs content!

How are Chinese AI models going to make a profit? What is their business plan?

 

[subotai1]

He’s a very good businessman like most Chinese people. Lmao. He’s for open models simply because they are almost all trained on Nvidia moat . The more they spread the better for Nvidia and its ecosystem. If Chinese open models were all trained on Chinese moat/chips with zero reliance on Nvidia, then you can bet he won’t be singing the same song lmao . lol

I respectfully disagree. There is a very real danger that political and legal fighting (backed by the US government at the behest of Musk and AI Billionaires) turns AI globally into a nightmare landscape of laws, walled gardens, national and corporate boundaries of who can and cannot build what, access what, install what, move what. That doesn't favor anyone but those that want to preserve what they think of as their current lead and marketshare. You're right that this is not good for Nvidia. But its not about open models, its about AI and compute needs in the long run (and therefore Nvidia).

 

[Kalum Pupeter]

What do people make of this commentary piece from Caixin?

Commentary: America Risks an AI Bubble While China Risks Falling Behind​

By

Please, Log in or Register to view URLs content!

Published: Jul. 24, 2026 5:56 p.m. GMT+8

The United States and China are running neck and neck in the global artificial intelligence race. Yet beneath the surface of this technological rivalry, a stark divergence is emerging: China’s capital expenditure on AI is falling significantly behind that of the U.S. This investment gap reveals fundamental differences in how both economies are funding the next technological revolution, and it carries profound implications for the future of global AI supremacy.

To understand why this matters, we must recognize that the AI era is fundamentally different from the internet boom. The digital economy of the past two decades was defined by light-asset platforms. Companies capitalized on the non-rivalrous nature of data, where user expansion brought marginal costs close to zero. The platform with the strongest network effect won, capturing massive profits.

Generative AI, however, is a heavy-asset endeavor. The scaling laws of large language models demand immense and continuous investments in computing power, data and electricity. Merely maintaining technological parity requires exponentially increasing resources. Furthermore, unlike the traditional internet, AI models incur substantial variable costs with every query. Running these models is hardware-intensive, leading to rapid physical depreciation and relentless upgrade cycles.

Crucially, the value creation of AI far exceeds its value capture. Because large models lack the monopolistic network effects of social media or search platforms, their economic benefits will likely spill over into the broader economy rather than accumulating solely in the hands of the model developers. From a macroeconomic perspective, this positive externality usually leads to underinvestment by private entities, as the societal return outweighs the private return.

This is where the U.S. and China diverge dramatically. In the U.S., an explosion of AI capital expenditure has bypassed traditional cost-benefit constraints. In the first quarter of 2026, AI-related capex accounted for nearly half of U.S. GDP growth and two-thirds of total demand growth.

Despite high interest rates, which typically suppress corporate investment, U.S. markets are gripped by what John Maynard Keynes famously called “animal spirits.” A profoundly optimistic market is subsidizing American AI development. U.S. equity risk premiums have effectively dropped to zero, signaling that investors are willing to fund massive cash-burn operations on the promise of future breakthroughs.

In contrast, China’s AI investment engine is sputtering. By the first quarter of 2026, AI-related capex contributed a mere fraction of a percentage point to China’s GDP growth. Several factors explain this sluggishness.

First, U.S. export controls on advanced graphics processing units have severely restricted hardware procurement, creating a bottleneck for willing Chinese buyers. Second, China’s capital markets remain subdued. Unlike the euphoric American stock market, Chinese equities still carry a noticeable risk premium, reflecting a more cautious investor base.

Third, labor cost differentials play a role. With lower average wages than in the U.S., the immediate economic incentive to replace human cognitive labor with AI is less pressing in China. Finally, China’s broader macroeconomic environment, burdened by a prolonged property slump and weak domestic demand, has compressed corporate profit expectations across the board.

These divergent paths mean Washington and Beijing face vastly different challenges. The U.S. faces pressure at the back end. The risk is that the euphoric expectations driving AI investment fail to materialize quickly enough. If high interest rates eventually exhaust investor patience and cash flows dry up, the resulting burst of the AI bubble could trigger massive capital destruction, similar to the dot-com crash.

China, conversely, faces pressure at the front end. If its current investment slump persists, the compounding nature of AI development means China risks falling permanently behind the U.S. technological frontier. Put simply: America’s risk is overinvestment and a popped bubble; China’s risk is underinvestment and technological obsolescence.

For China, the imperative is clear: it must urgently catch up in AI capital expenditure. The U.S. stock market bubble is currently serving a vital economic function by correcting the natural private underinvestment in a technology with massive social externalities. Because China lacks a comparably hyperactive equity market, it must rely on a different mechanism: industrial policy.

To bridge the gap, Beijing must aggressively mobilize state-backed financial resources, integrating bank credit with targeted industrial policies to fund AI infrastructure, such as advanced computing centers. While direct government investment and policy-driven lending carry their own risks of inefficiency and redundant projects, these are acceptable costs. In the current geopolitical climate, the dangers of technological stagnation far outweigh the costs of potential micro-level inefficiencies.

Recent policy signals suggest China is awakening to this reality, and domestic AI capex is poised for a rebound. But ensuring this investment boom takes root will require more than just targeted subsidies. It demands a broader macroeconomic pivot. Accommodative monetary policy must support fiscal expansion to revive domestic demand, creating the fertile, dynamic economic environment that true innovation requires.

Peng Wensheng is Chief Economist and Head of Research at CICC.

The views expressed in third-party articles are those of the authors and do not necessarily reflect the positions of Caixin.

Please, Log in or Register to view URLs content!

 

[RobertC]

Wednesday the new-formed Little Tech Association comprising 200 Silicon Valley startup companies urged the Trump administration to continue allowing access to open-weight models, including China's, or risk crippling their future.
[

Please, Log in or Register to view URLs content!

]

Today 20 tech companies joined Nvidia to urge the Trump administration to continue allowing access to open-weight models, including China's, or risk undermining America's competitive edge and losing market share.
[

Please, Log in or Register to view URLs content!

]

@Michael90 WRT Nvidia retaining market share, Jensen's losses have been due to market restrictions by the Trump administration not technology or cost. If (when) the Chinese win his market share, that's on the Trump administration not Nvidia.

@HighGround Thank you for your analysis.

@henrik I have confidence in the Chinese companies' ability to make a profit.

@subotai1 Your dystopian future is certainly a possibility but perhaps an AI version of

Please, Log in or Register to view URLs content!

is the more likely future.

@Kalum Pupeter Hopefully their other analyses are more useful.

 

[tokenanalyst]

How are Chinese AI models going to make a profit? What is their business plan?

It seems that Chinese AI companies are scaling according to numbers of users they get. Open sourcing makes sense in term that they can serve a amount of users in their own servers without going into too much in debt.
Also it seems that Chinese companies want to go beyond tokens and they are looking to build ecosystems and sell solutions. The hugging face incident open the door to sell AI cybersecurity specific models trained to protect networks. I think many institutions like banks and governments institutions China and globally are probably looking into this.

In my personal opinion I think the future of these AI models is:

1- Application specific models. Solutions instead of just tokens either locally or by cloud.

2- Flash Hundred of Billions MoE models are going to dominate over multi-trillion parameter models, at least until hardware catch up.
With a human in the loop who knows what he is doing flash models are pretty good, they run many many time faster than frontier models, they are way cheaper to run locally or from a provider. Flash model are productivity machines. I think frontier models are becoming too big, too slow and too expensive to be productive. I think people who think they always need frontiers model is people who want software with the least amount of effort or delusional upper management guys who are sold to the idea of replacing every employee with an LLM. They both get slop results, that looks good at first glance but is crap and worse cannot be repaired because the model output so much slop code that is just better to write the entire codebase from zero.

 

[Michael90]

I respectfully disagree. There is a very real danger that political and legal fighting (backed by the US government at the behest of Musk and AI Billionaires) turns AI globally into a nightmare landscape of laws, walled gardens, national and corporate boundaries of who can and cannot build what, access what, install what, move what. That doesn't favor anyone but those that want to preserve what they think of as their current lead and marketshare. You're right that this is not good for Nvidia. But its not about open models, its about AI and compute needs in the long run (and therefore Nvidia).

It favors Nvidia that’s why they are clamoring for this. Else he won’t be supporting this believe me. lol it just expands his market even more so as a good businessman he supports it. The reason he gives are just excuses . lol. Why can’t he open source all the software his Company produces, so companies like Intel , Huawei , AMD etc et. can ensure their hardware can run it in parity with Nvidia? Lmao since he’s so moral and looking after the greater good ? lol. Seriously , anyone believing this people means they still live in a naive world .

Have we not seen his bullshit shenanigans with PhysX, proprietary FP definitions and CUDA? lol The dude is not different from Dario and Altman et al , since he has locked anyone out of being able to use them and Nvidia even punishes anyone trying to make it run on personal hardware? Look at ZLUDA.

so I call this is support for open source , hypocrisy of the highest order. lol he should lead by example then let’s see. lol. He’s just another greedy entrepreneur out there just for himself and company’s huge profits /valuations

Anyway. Chinese AI startups and companies need to give out their IPs and software codes etc to remain competitive also because by a series of events the US force them into a corner by banning AI chips to China and even equipments making chips so China still lagging behind in this areas has to look for loopholes and other ways to carry on competing and staying in the game. The best way they could do it was through open sourcing their models ironically without the US forcing China into a corner with sanctions and isolation Chinese companies wouldn’t have opted for this model. So the US government only have themselves to blame as well lol