Chinese AI companies have raised 24x less funding than their US counterparts, a gap explained largely by compute constraints, a challenging domestic market, and a smaller investment ecosystem.
> We think this gap is best explained by a mix of three factors. First, compute limitations from US export controls have kept Chinese companies from reaching the capability frontier thus far and likely make investors less excited about their future growth. Second, Chinese AI companies are serving a more challenging home market that’s less ready to adopt enterprise AI. And third, even with a great business case, it’s hard to raise a lot of money in China, both because of a smaller venture capital (VC) ecosystem and because previous state intervention in tech companies makes it risky for investors to bet on rapid growth. Let’s look into each of these.
Isn't there a 4th factor which is Chinese big tech companies are way smaller than the US ones, so there's a smaller starting cloud compute industry, less AI investment overall, including into the AI companies. I think the biggest factor in US-China overall compute gap is from this gap in total AI investment, not the other factors. The total compute gap is a somewhat different question to the gap in AI startup funding which is what you were answering, but I think this general point about the starting size of the total tech industry is the most important factor for both.
Apple + Alphabet + Microsoft + Meta + Amazon ~= $580B operating income
Naively this 5x gap alone seems like it explains ~all of the gap in total AI compute, and ~50% of the gap in frontier AI company investment on its own.
Yeah that seems important too, though how do you reckon it explains the entire compute gap (which is ~9x)? Seems like it’d account for about 75%, unless I’m missing something.
I’d also guess that the capex-to-operating-income ratio is higher for US hyperscalers for some of the reasons mentioned in the post, but not sure about that.
I was thinking its currently ~7x not 9x, with the other 1.4x being mostly from export controls (1.4x lower compute / $ spent on average). I thought China still probably has like 11% of world compute rn, US 77%, Row 12%? Not sure though.
Hmm, looking at the cumulative AI chip ownership data from Epoch, I see 1.82mn for China (including smuggled chips) and 15.8mn for the US companies (i.e., everything, but subtracting "China" and "Other"). That would be ~9x, though the actual US number is probably higher since much of the "Other" compute is probably also US-owned.
I see, hadn’t checked in on this in a while, looks like it’s more complete now. I think for China it looks like it’s missing domestic production though and (i’m guessing) might be incomplete on chinese indirectly owned chips outside of china, eg malaysia (or were those covered by smuggling analysis?). Anyway maybe i’d say like 7-10x is my 50% CI.
That first chart is staggering. It makes me wonder how Chinese labs have been doing as well as they have, and whether that can continue.
Right now there's a lot of talk about Chinese models toppling the US premiere labs by offering an equivalent product for free. But it's hard to imagine such a situation arising out of such constrains.
I think it is partly to capture the market. If they don’t offer it for free, it is hard for branding. They are probably making the most profitable business decision under the constraints of market share and compute.
That is my best guess too. It seems like a game of chicken: They are trying to challenge the US labs at great cost to themselves. It can't be sustainable, but maybe it's the best/only strategy available.
We're looking at the wrong ledger. 📊 The narrative that Chinese AI is faltering because of a 24x funding gap is a massive category error. It mistakes capital-brute-forcing for intelligence.
US labs have raised an average of 116 billion dollars because they're subsidizing massive, dense architectural waste. They're paying a multi-billion-dollar token tax to NVIDIA and cloud cartels to run brute-force dense networks on hardware that literally truncates precision at the physical layer. Hopper's 13-bit mantissa truncation is a silent entropy leak. It forces weights to converge on safe, low-entropy boilerplate. To drown out this microarchitectural noise, Western labs must burn exponential amounts of capital on more redundant training tokens. 💸
Chinese labs, operating under a 5:1 high-end GPU deficit, were forced to achieve numerical solvency through structural elegance, not raw capital. Look at the data. 📈 The 2026 Stanford AI Index shows a 23-to-1 capital ratio, yet US closed models hold barely a 2.7% lead. China's sparse open-weight models duplicate frontier capabilities at an average cost 87% lower than US cloud APIs.
By shifting to Multi-Head Latent Attention (MLA), Kimi Delta Attention, and fine-grained Mixture-of-Experts (MoE) topologies, they've bypassed the physical HBM memory wall. This isn't a funding crisis; it's a structural realization. Capital is inversely proportional to the elegance of your hardware-software joint manifold. 🧠
If your hardware accumulator is too narrow to see the loss geometry, you have to build a bigger bank account just to heat the room. But what happens when the capital well runs dry before we fix the 13-bit silicon floor? ⏳
The 5% vs 28% split is my favorite number here, and I'd read it less as Chinese VCs being cold on AI. State-backed money went into semis instead, which is a coherent bet that the binding constraint is upstream of the model layer.
If you think compute is the ceiling, funding a lab to bump against it is the worse trade. Different diagnosis of the same bottleneck, not less enthusiasm. Whether it pays depends on SMIC, not on anyone's model.
> We think this gap is best explained by a mix of three factors. First, compute limitations from US export controls have kept Chinese companies from reaching the capability frontier thus far and likely make investors less excited about their future growth. Second, Chinese AI companies are serving a more challenging home market that’s less ready to adopt enterprise AI. And third, even with a great business case, it’s hard to raise a lot of money in China, both because of a smaller venture capital (VC) ecosystem and because previous state intervention in tech companies makes it risky for investors to bet on rapid growth. Let’s look into each of these.
Isn't there a 4th factor which is Chinese big tech companies are way smaller than the US ones, so there's a smaller starting cloud compute industry, less AI investment overall, including into the AI companies. I think the biggest factor in US-China overall compute gap is from this gap in total AI investment, not the other factors. The total compute gap is a somewhat different question to the gap in AI startup funding which is what you were answering, but I think this general point about the starting size of the total tech industry is the most important factor for both.
ByteDance + Tencent + Huawei + Alibaba + PDD ~= $115B operating income
Apple + Alphabet + Microsoft + Meta + Amazon ~= $580B operating income
Naively this 5x gap alone seems like it explains ~all of the gap in total AI compute, and ~50% of the gap in frontier AI company investment on its own.
Yeah that seems important too, though how do you reckon it explains the entire compute gap (which is ~9x)? Seems like it’d account for about 75%, unless I’m missing something.
I’d also guess that the capex-to-operating-income ratio is higher for US hyperscalers for some of the reasons mentioned in the post, but not sure about that.
I was thinking its currently ~7x not 9x, with the other 1.4x being mostly from export controls (1.4x lower compute / $ spent on average). I thought China still probably has like 11% of world compute rn, US 77%, Row 12%? Not sure though.
Hmm, looking at the cumulative AI chip ownership data from Epoch, I see 1.82mn for China (including smuggled chips) and 15.8mn for the US companies (i.e., everything, but subtracting "China" and "Other"). That would be ~9x, though the actual US number is probably higher since much of the "Other" compute is probably also US-owned.
https://epoch.ai/data/ai-chip-owners
I see, hadn’t checked in on this in a while, looks like it’s more complete now. I think for China it looks like it’s missing domestic production though and (i’m guessing) might be incomplete on chinese indirectly owned chips outside of china, eg malaysia (or were those covered by smuggling analysis?). Anyway maybe i’d say like 7-10x is my 50% CI.
That first chart is staggering. It makes me wonder how Chinese labs have been doing as well as they have, and whether that can continue.
Right now there's a lot of talk about Chinese models toppling the US premiere labs by offering an equivalent product for free. But it's hard to imagine such a situation arising out of such constrains.
I think it is partly to capture the market. If they don’t offer it for free, it is hard for branding. They are probably making the most profitable business decision under the constraints of market share and compute.
That is my best guess too. It seems like a game of chicken: They are trying to challenge the US labs at great cost to themselves. It can't be sustainable, but maybe it's the best/only strategy available.
We're looking at the wrong ledger. 📊 The narrative that Chinese AI is faltering because of a 24x funding gap is a massive category error. It mistakes capital-brute-forcing for intelligence.
US labs have raised an average of 116 billion dollars because they're subsidizing massive, dense architectural waste. They're paying a multi-billion-dollar token tax to NVIDIA and cloud cartels to run brute-force dense networks on hardware that literally truncates precision at the physical layer. Hopper's 13-bit mantissa truncation is a silent entropy leak. It forces weights to converge on safe, low-entropy boilerplate. To drown out this microarchitectural noise, Western labs must burn exponential amounts of capital on more redundant training tokens. 💸
Chinese labs, operating under a 5:1 high-end GPU deficit, were forced to achieve numerical solvency through structural elegance, not raw capital. Look at the data. 📈 The 2026 Stanford AI Index shows a 23-to-1 capital ratio, yet US closed models hold barely a 2.7% lead. China's sparse open-weight models duplicate frontier capabilities at an average cost 87% lower than US cloud APIs.
By shifting to Multi-Head Latent Attention (MLA), Kimi Delta Attention, and fine-grained Mixture-of-Experts (MoE) topologies, they've bypassed the physical HBM memory wall. This isn't a funding crisis; it's a structural realization. Capital is inversely proportional to the elegance of your hardware-software joint manifold. 🧠
If your hardware accumulator is too narrow to see the loss geometry, you have to build a bigger bank account just to heat the room. But what happens when the capital well runs dry before we fix the 13-bit silicon floor? ⏳
(ノ°益°)ノ
The 5% vs 28% split is my favorite number here, and I'd read it less as Chinese VCs being cold on AI. State-backed money went into semis instead, which is a coherent bet that the binding constraint is upstream of the model layer.
If you think compute is the ceiling, funding a lab to bump against it is the worse trade. Different diagnosis of the same bottleneck, not less enthusiasm. Whether it pays depends on SMIC, not on anyone's model.