#openai据报完成新一代bel模型预训练
1. OpenAI re-enters the competition for large-scale pretraining
A 10T-parameter model means the model scale continues to expand to extremes. However, the number of parameters alone does not equal model capability. Training data quality, compute budget, architecture, MoE design, and post-training are just as important.
2. Bel may not be a consumer-facing product directly
Based on the leaked description, it looks more like a **“foundation model”**:
Bel → post-training/inference capability → Astra / GPT-6 and subsequent products
Therefore, even if Bel has already completed pretraining, it doesn’t mean users can immediately use a “10T-parameter model.”
3. What’s truly worth paying attention to is Astra
In currently available public information, OpenAI has been more explicit about Astra rather than Bel. That suggests the market may be linking an unverified internal training project to the next generation product roadmap.
4. The AI compute supply chain may be stimulated again
If a 10T-class model really enters large-scale training, in the long run it would mean:
AI GPU demand continues to increase data center capex stays elevated HBM demand continues to grow high-speed networking/switches demand increases liquid cooling and power data center infrastructure benefits
Impact on the AI industry / U.S. stocks
My view:
Short term: positive for the AI compute supply chain, but the truthfulness of the news remains the biggest discount factor.
If OpenAI later officially confirms Bel, and it truly reaches a 10T+ scale, then the market may once again raise expectations for the next round of AI compute capex.
Beneficiary directions roughly are:
GPU → HBM → high-speed network → optical modules → servers → liquid cooling → data centers → electricity
Among these, what’s really worth watching isn’t the “10T” number, but whether OpenAI starts deploying the corresponding compute at large scale—and whether inference costs for the next-generation models drop significantly.
---------------------
$SNDK Trading strategy
After 1,468 stabilizes, go long
Stop loss: 1,451
Target: around 1,560 to reduce positions or take profit
【👇👇Click the trading link below👇👇】🚀🚀🚀🎉🎉🎉🎉🎉
1. OpenAI re-enters the competition for large-scale pretraining
A 10T-parameter model means the model scale continues to expand to extremes. However, the number of parameters alone does not equal model capability. Training data quality, compute budget, architecture, MoE design, and post-training are just as important.
2. Bel may not be a consumer-facing product directly
Based on the leaked description, it looks more like a **“foundation model”**:
Bel → post-training/inference capability → Astra / GPT-6 and subsequent products
Therefore, even if Bel has already completed pretraining, it doesn’t mean users can immediately use a “10T-parameter model.”
3. What’s truly worth paying attention to is Astra
In currently available public information, OpenAI has been more explicit about Astra rather than Bel. That suggests the market may be linking an unverified internal training project to the next generation product roadmap.
4. The AI compute supply chain may be stimulated again
If a 10T-class model really enters large-scale training, in the long run it would mean:
AI GPU demand continues to increase data center capex stays elevated HBM demand continues to grow high-speed networking/switches demand increases liquid cooling and power data center infrastructure benefits
Impact on the AI industry / U.S. stocks
My view:
Short term: positive for the AI compute supply chain, but the truthfulness of the news remains the biggest discount factor.
If OpenAI later officially confirms Bel, and it truly reaches a 10T+ scale, then the market may once again raise expectations for the next round of AI compute capex.
Beneficiary directions roughly are:
GPU → HBM → high-speed network → optical modules → servers → liquid cooling → data centers → electricity
Among these, what’s really worth watching isn’t the “10T” number, but whether OpenAI starts deploying the corresponding compute at large scale—and whether inference costs for the next-generation models drop significantly.
---------------------
$SNDK Trading strategy
After 1,468 stabilizes, go long
Stop loss: 1,451
Target: around 1,560 to reduce positions or take profit
【👇👇Click the trading link below👇👇】🚀🚀🚀🎉🎉🎉🎉🎉
