Artificial Intelligence thread

tphuang

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
you just need to start moving to Ascend NPUs for inference. It's a pain to switch your tech stack, but companies should be working on this.

Besides, they all have accumulated a whole bunch of AI chips already and smuggling will continue
 
9dashline

9dashline

Captain
Registered Member
Hyper said:
I am sure inferencing can be done on new server cpu. They are dirt cheap compared to GPU. They might be bandwidth constrained but the raw storage can ease the load.
Click to expand...

1. BACKDROP: THE NEW COLD WAR OVER AI AND SEMICONDUCTORS
The U.S. has launched its most sweeping export controls yet under the Framework for Artificial Intelligence Diffusion to obstruct Chinese access to advanced GPUs. These rules divide the world into three tiers (Tier 1: top allies; Tier 2: everyone else; Tier 3: arms-embargoed states, including China) and impose strict licensing on both AI chips and large-scale training. America’s rationale: AI progress is accelerating so fast that even a few months’ lead could cement global supremacy—essentially forming a new “AI OPEC.”


2. CORE OF THE NEW RULES


  • Tiers and Validated End Users (VEUs):
    • Tier 1 (U.S. + 18 aligned countries) can import frontier GPUs under relaxed constraints if run by “UVEUs” (like Microsoft, Amazon, Google).
    • Tier 2 faces country-level GPU caps (49k–99k “H100 equivalents”) unless an entity becomes an NVEU with special approval.
    • Tier 3 (China, Russia, etc.) is largely banned.
  • Model Training Limits: Training any model that consumes ≥1e26 FLOPs (or the largest open-model threshold) in Tier 2 or Tier 3 is blocked. Even open-source can be restricted if fine-tuning uses more than 2e25 FLOPs.
  • Model Weights Export: Frontier closed-source model weights above 1e26 FLOPs need special licenses for non–Tier 1 locations.
  • Small Loopholes: Up to 1,699 H100-equivalent GPUs can be exported per Tier 2 customer per year without a license, but that’s overshadowed by the broader clampdown.

3. MOTIVATION: STOPPING CHINA’S AI UPSURGE
Past GPU restrictions were riddled with loopholes: re-exports, PRC SKUs (H20, MI308X), or data centers in “unrestricted” Tier 2 countries. The new rules aim to end these via:


  • Global Licensing: Tying hardware usage to a VEU system that requires rigorous security, supply-chain checks, and location quotas.
  • Frontier Model Policing: If a Tier 3 or suspect Tier 2 entity tries to train an ultra-large model on Tier 1 hardware, providers must raise red flags.
  • Geopolitical Pressure: Tier 2 governments can double their national AI chip cap (49k→99k) only if they align with Washington’s export regime.

4. NEAR-TERM IMPACT ON THE GLOBAL AI RACE


  1. Concentrating AI in Tier 1: U.S. hyperscalers get streamlined approvals to build massive HPC on home soil or in a few allied countries, boosting “onshore” AI.
  2. Shrinking Tier 2 Mega-Clusters: Plans for colossal data centers in Malaysia, India, Brazil may stall unless anchored by a U.S.-approved VEU. This hits colocation operators who expected to serve Chinese or unaligned players.
  3. China’s Workarounds:
    • Exploit small shipments under 1,699 GPUs each, aggregated via shell companies.
    • Accelerate domestic DUV-based chips for AI clusters in China.
    • Sabotage or co-opt TSMC to degrade U.S. supply.

5. AGI TIMELINES AND TRANSFORMER CEILINGS
These restrictions underscore Washington’s belief that AGI could be near. OpenAI’s o3, DeepSeek v3, and other leaps challenge the notion that “transformers have peaked.” The U.S. is betting that controlling sub-2 nm GPUs for a brief period might lock in an irreversible lead. Yet if hardware-based scaling hits diminishing returns at 1.5 nm, or if algorithmic breakthroughs let rivals do more with mid-tier chips, U.S. controls might become less decisive.


6. DEDOLLARIZATION VS. “AI OPEC”
The U.S. seeks to replace petrodollar influence with “AI-dollar” hegemony, charging the world for advanced AI services. Tier 2 or Tier 3 states hoping to escape the dollar system may find themselves tethered to U.S.-based clouds if they want frontier AI. On the other hand, if China can sustain data center expansions (even with second-tier chips) and rally the Global South to dump the dollar, it may undermine the financial leverage behind America’s new AI monarchy.


7. EROEI CONSTRAINTS AND DATACENTER ENERGY
Declining net-energy returns complicate HPC expansions. The new rules ironically concentrate HPC in Tier 1, which might strain local energy grids unless AI-driven breakthroughs in fusion or advanced renewables arrive. Meanwhile, countries with abundant cheap power (e.g., Middle East) find themselves in Tier 2. They can host new HPC only if a major U.S. firm has a green light. This might hamper China’s ability to exploit cheap overseas energy for giant GPU clusters.


8. MASSIVE JOB DISPLACEMENT AS A SIDE EFFECT
The tighter distribution of top GPUs in Tier 1 may lead to faster rollout of advanced AI in the U.S. and allied nations, automating sectors and displacing white-collar and blue-collar workers. Tier 2 countries lose potential AI expansions, meaning they lose both the tech opportunities and some local job creation in HPC centers. The U.S. can then require an “AI service tax,” funneling global profits back home.


9. STRATEGIC FORECAST: CATASTROPHIC OR DOMINANT?


  1. Effective Enforcement: If these controls are fully enforced, China faces a hardware drought for frontier models. U.S. hyperscalers dominate “AGI first,” forging an AI OPEC.
  2. Backlash & Loopholes: If black markets flourish or if Tier 2 states balk at the U.S. regime, China might still amass GPU clusters—just more covertly.
  3. Algorithmic Wildcards: Should Chinese researchers or open-source communities discover radical training efficiencies, the hardware blockade might mean less than the U.S. hopes.
  4. Potential War Flashpoint: If Beijing deems it impossible to catch up under these controls, it may gamble on a TSMC “rugpull,” forcibly leveling the field by destroying the West’s advanced GPU supply chain.

10. CONCLUSION: A FORCED RESTRUCTURING OF GLOBAL AI
The AI Diffusion Framework marks a profound realignment. While it centralizes HPC under vetted American or allied hands, it also triggers new black markets, fosters local DUV innovations, and intensifies the arms race for more efficient AI architectures. In the U.S. view, only a strict, worldwide licensing net can block China from achieving frontier AI. In China’s view, they have to circumvent or smash that net, lest they be locked out of the greatest technological revolution of our era.

Whichever side prevails, the next few years look pivotal. The question isn’t whether the AI race intensifies, but how far each superpower will go to keep or sabotage the other’s ascent—and whether a single slip might abruptly transform the entire global order for decades.
 
tphuang

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
9dashline said:
1. BACKDROP: THE NEW COLD WAR OVER AI AND SEMICONDUCTORS
The U.S. has launched its most sweeping export controls yet under the Framework for Artificial Intelligence Diffusion to obstruct Chinese access to advanced GPUs. These rules divide the world into three tiers (Tier 1: top allies; Tier 2: everyone else; Tier 3: arms-embargoed states, including China) and impose strict licensing on both AI chips and large-scale training. America’s rationale: AI progress is accelerating so fast that even a few months’ lead could cement global supremacy—essentially forming a new “AI OPEC.”


2. CORE OF THE NEW RULES


  • Tiers and Validated End Users (VEUs):
    • Tier 1 (U.S. + 18 aligned countries) can import frontier GPUs under relaxed constraints if run by “UVEUs” (like Microsoft, Amazon, Google).
    • Tier 2 faces country-level GPU caps (49k–99k “H100 equivalents”) unless an entity becomes an NVEU with special approval.
    • Tier 3 (China, Russia, etc.) is largely banned.
  • Model Training Limits: Training any model that consumes ≥1e26 FLOPs (or the largest open-model threshold) in Tier 2 or Tier 3 is blocked. Even open-source can be restricted if fine-tuning uses more than 2e25 FLOPs.
  • Model Weights Export: Frontier closed-source model weights above 1e26 FLOPs need special licenses for non–Tier 1 locations.
  • Small Loopholes: Up to 1,699 H100-equivalent GPUs can be exported per Tier 2 customer per year without a license, but that’s overshadowed by the broader clampdown.

3. MOTIVATION: STOPPING CHINA’S AI UPSURGE
Past GPU restrictions were riddled with loopholes: re-exports, PRC SKUs (H20, MI308X), or data centers in “unrestricted” Tier 2 countries. The new rules aim to end these via:


  • Global Licensing: Tying hardware usage to a VEU system that requires rigorous security, supply-chain checks, and location quotas.
  • Frontier Model Policing: If a Tier 3 or suspect Tier 2 entity tries to train an ultra-large model on Tier 1 hardware, providers must raise red flags.
  • Geopolitical Pressure: Tier 2 governments can double their national AI chip cap (49k→99k) only if they align with Washington’s export regime.

4. NEAR-TERM IMPACT ON THE GLOBAL AI RACE


  1. Concentrating AI in Tier 1: U.S. hyperscalers get streamlined approvals to build massive HPC on home soil or in a few allied countries, boosting “onshore” AI.
  2. Shrinking Tier 2 Mega-Clusters: Plans for colossal data centers in Malaysia, India, Brazil may stall unless anchored by a U.S.-approved VEU. This hits colocation operators who expected to serve Chinese or unaligned players.
  3. China’s Workarounds:
    • Exploit small shipments under 1,699 GPUs each, aggregated via shell companies.
    • Accelerate domestic DUV-based chips for AI clusters in China.
    • Sabotage or co-opt TSMC to degrade U.S. supply.

5. AGI TIMELINES AND TRANSFORMER CEILINGS
These restrictions underscore Washington’s belief that AGI could be near. OpenAI’s o3, DeepSeek v3, and other leaps challenge the notion that “transformers have peaked.” The U.S. is betting that controlling sub-2 nm GPUs for a brief period might lock in an irreversible lead. Yet if hardware-based scaling hits diminishing returns at 1.5 nm, or if algorithmic breakthroughs let rivals do more with mid-tier chips, U.S. controls might become less decisive.


6. DEDOLLARIZATION VS. “AI OPEC”
The U.S. seeks to replace petrodollar influence with “AI-dollar” hegemony, charging the world for advanced AI services. Tier 2 or Tier 3 states hoping to escape the dollar system may find themselves tethered to U.S.-based clouds if they want frontier AI. On the other hand, if China can sustain data center expansions (even with second-tier chips) and rally the Global South to dump the dollar, it may undermine the financial leverage behind America’s new AI monarchy.


7. EROEI CONSTRAINTS AND DATACENTER ENERGY
Declining net-energy returns complicate HPC expansions. The new rules ironically concentrate HPC in Tier 1, which might strain local energy grids unless AI-driven breakthroughs in fusion or advanced renewables arrive. Meanwhile, countries with abundant cheap power (e.g., Middle East) find themselves in Tier 2. They can host new HPC only if a major U.S. firm has a green light. This might hamper China’s ability to exploit cheap overseas energy for giant GPU clusters.


8. MASSIVE JOB DISPLACEMENT AS A SIDE EFFECT
The tighter distribution of top GPUs in Tier 1 may lead to faster rollout of advanced AI in the U.S. and allied nations, automating sectors and displacing white-collar and blue-collar workers. Tier 2 countries lose potential AI expansions, meaning they lose both the tech opportunities and some local job creation in HPC centers. The U.S. can then require an “AI service tax,” funneling global profits back home.


9. STRATEGIC FORECAST: CATASTROPHIC OR DOMINANT?


  1. Effective Enforcement: If these controls are fully enforced, China faces a hardware drought for frontier models. U.S. hyperscalers dominate “AGI first,” forging an AI OPEC.
  2. Backlash & Loopholes: If black markets flourish or if Tier 2 states balk at the U.S. regime, China might still amass GPU clusters—just more covertly.
  3. Algorithmic Wildcards: Should Chinese researchers or open-source communities discover radical training efficiencies, the hardware blockade might mean less than the U.S. hopes.
  4. Potential War Flashpoint: If Beijing deems it impossible to catch up under these controls, it may gamble on a TSMC “rugpull,” forcibly leveling the field by destroying the West’s advanced GPU supply chain.

10. CONCLUSION: A FORCED RESTRUCTURING OF GLOBAL AI
The AI Diffusion Framework marks a profound realignment. While it centralizes HPC under vetted American or allied hands, it also triggers new black markets, fosters local DUV innovations, and intensifies the arms race for more efficient AI architectures. In the U.S. view, only a strict, worldwide licensing net can block China from achieving frontier AI. In China’s view, they have to circumvent or smash that net, lest they be locked out of the greatest technological revolution of our era.

Whichever side prevails, the next few years look pivotal. The question isn’t whether the AI race intensifies, but how far each superpower will go to keep or sabotage the other’s ascent—and whether a single slip might abruptly transform the entire global order for decades.
Click to expand...

your quality of posts on this thread is really poor and it doesn't seem like you have gotten the message even after I deleted your recent posts. This above post looks like you either copied and pasted from a link that you are not referencing or AI generated. If you are posting like this on an AI thread, you have to specify that. One more post like this and you will be banned from this thread.
 
tphuang

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
btw, deepseek app is available for download on google store and also on iOS.

Please, Log in or Register to view URLs content!

I have already downloaded it.

It already has 24 reviews and now 30th on iOS App Store for productivity

Kling AI also continuing to move in the apps, now #83 in graphics and design.
 
T

tokenanalyst

Brigadier
Registered Member
9dashline said:
1. BACKDROP: THE NEW COLD WAR OVER AI AND SEMICONDUCTORS
The U.S. has launched its most sweeping export controls yet under the Framework for Artificial Intelligence Diffusion to obstruct Chinese access to advanced GPUs. These rules divide the world into three tiers (Tier 1: top allies; Tier 2: everyone else; Tier 3: arms-embargoed states, including China) and impose strict licensing on both AI chips and large-scale training. America’s rationale: AI progress is accelerating so fast that even a few months’ lead could cement global supremacy—essentially forming a new “AI OPEC.”


2. CORE OF THE NEW RULES


  • Tiers and Validated End Users (VEUs):
    • Tier 1 (U.S. + 18 aligned countries) can import frontier GPUs under relaxed constraints if run by “UVEUs” (like Microsoft, Amazon, Google).
    • Tier 2 faces country-level GPU caps (49k–99k “H100 equivalents”) unless an entity becomes an NVEU with special approval.
    • Tier 3 (China, Russia, etc.) is largely banned.
  • Model Training Limits: Training any model that consumes ≥1e26 FLOPs (or the largest open-model threshold) in Tier 2 or Tier 3 is blocked. Even open-source can be restricted if fine-tuning uses more than 2e25 FLOPs.
  • Model Weights Export: Frontier closed-source model weights above 1e26 FLOPs need special licenses for non–Tier 1 locations.
  • Small Loopholes: Up to 1,699 H100-equivalent GPUs can be exported per Tier 2 customer per year without a license, but that’s overshadowed by the broader clampdown.

3. MOTIVATION: STOPPING CHINA’S AI UPSURGE
Past GPU restrictions were riddled with loopholes: re-exports, PRC SKUs (H20, MI308X), or data centers in “unrestricted” Tier 2 countries. The new rules aim to end these via:


  • Global Licensing: Tying hardware usage to a VEU system that requires rigorous security, supply-chain checks, and location quotas.
  • Frontier Model Policing: If a Tier 3 or suspect Tier 2 entity tries to train an ultra-large model on Tier 1 hardware, providers must raise red flags.
  • Geopolitical Pressure: Tier 2 governments can double their national AI chip cap (49k→99k) only if they align with Washington’s export regime.

4. NEAR-TERM IMPACT ON THE GLOBAL AI RACE


  1. Concentrating AI in Tier 1: U.S. hyperscalers get streamlined approvals to build massive HPC on home soil or in a few allied countries, boosting “onshore” AI.
  2. Shrinking Tier 2 Mega-Clusters: Plans for colossal data centers in Malaysia, India, Brazil may stall unless anchored by a U.S.-approved VEU. This hits colocation operators who expected to serve Chinese or unaligned players.
  3. China’s Workarounds:
    • Exploit small shipments under 1,699 GPUs each, aggregated via shell companies.
    • Accelerate domestic DUV-based chips for AI clusters in China.
    • Sabotage or co-opt TSMC to degrade U.S. supply.

5. AGI TIMELINES AND TRANSFORMER CEILINGS
These restrictions underscore Washington’s belief that AGI could be near. OpenAI’s o3, DeepSeek v3, and other leaps challenge the notion that “transformers have peaked.” The U.S. is betting that controlling sub-2 nm GPUs for a brief period might lock in an irreversible lead. Yet if hardware-based scaling hits diminishing returns at 1.5 nm, or if algorithmic breakthroughs let rivals do more with mid-tier chips, U.S. controls might become less decisive.


6. DEDOLLARIZATION VS. “AI OPEC”
The U.S. seeks to replace petrodollar influence with “AI-dollar” hegemony, charging the world for advanced AI services. Tier 2 or Tier 3 states hoping to escape the dollar system may find themselves tethered to U.S.-based clouds if they want frontier AI. On the other hand, if China can sustain data center expansions (even with second-tier chips) and rally the Global South to dump the dollar, it may undermine the financial leverage behind America’s new AI monarchy.


7. EROEI CONSTRAINTS AND DATACENTER ENERGY
Declining net-energy returns complicate HPC expansions. The new rules ironically concentrate HPC in Tier 1, which might strain local energy grids unless AI-driven breakthroughs in fusion or advanced renewables arrive. Meanwhile, countries with abundant cheap power (e.g., Middle East) find themselves in Tier 2. They can host new HPC only if a major U.S. firm has a green light. This might hamper China’s ability to exploit cheap overseas energy for giant GPU clusters.


8. MASSIVE JOB DISPLACEMENT AS A SIDE EFFECT
The tighter distribution of top GPUs in Tier 1 may lead to faster rollout of advanced AI in the U.S. and allied nations, automating sectors and displacing white-collar and blue-collar workers. Tier 2 countries lose potential AI expansions, meaning they lose both the tech opportunities and some local job creation in HPC centers. The U.S. can then require an “AI service tax,” funneling global profits back home.


9. STRATEGIC FORECAST: CATASTROPHIC OR DOMINANT?


  1. Effective Enforcement: If these controls are fully enforced, China faces a hardware drought for frontier models. U.S. hyperscalers dominate “AGI first,” forging an AI OPEC.
  2. Backlash & Loopholes: If black markets flourish or if Tier 2 states balk at the U.S. regime, China might still amass GPU clusters—just more covertly.
  3. Algorithmic Wildcards: Should Chinese researchers or open-source communities discover radical training efficiencies, the hardware blockade might mean less than the U.S. hopes.
  4. Potential War Flashpoint: If Beijing deems it impossible to catch up under these controls, it may gamble on a TSMC “rugpull,” forcibly leveling the field by destroying the West’s advanced GPU supply chain.

10. CONCLUSION: A FORCED RESTRUCTURING OF GLOBAL AI
The AI Diffusion Framework marks a profound realignment. While it centralizes HPC under vetted American or allied hands, it also triggers new black markets, fosters local DUV innovations, and intensifies the arms race for more efficient AI architectures. In the U.S. view, only a strict, worldwide licensing net can block China from achieving frontier AI. In China’s view, they have to circumvent or smash that net, lest they be locked out of the greatest technological revolution of our era.

Whichever side prevails, the next few years look pivotal. The question isn’t whether the AI race intensifies, but how far each superpower will go to keep or sabotage the other’s ascent—and whether a single slip might abruptly transform the entire global order for decades.
Click to expand...
1- Unenforceable crap, US Companies are not going to spend resources enforcing Biden crappy rules, so it will be to Tiers countries to do that Job, what Americans companies are going to do is to spend huge amount of money lobbying the hell out of the next money/business hungry administration to remove these export controls.

2- I think between SMIC and prior 2024 TSMC manufactured Huawei GPUs is more than enough for a few years at least.

3-I think the next fab boom in the coming years in China is not going to be mature nodes but advanced nodes 14nm/7nm/5nm rendering these controls useless and even counterproductive.
 
L

luminary

Senior Member
Registered Member
I'm stumped at how they come to this conclusion about the chinese chip ecosystem.

I was also wondering why Israel is in tier 2 (as they are building supercomputers there), but then realized it might be a useful negotiation lever for Trump.
 
Last edited:
tphuang

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
luminary said:
I'm stumped at how they come to this conclusion about the chinese chip ecosystem.

I was also wondering why Israel is in tier 2 (as they are building supercomputers there), but then realized it might be a useful negotiation lever for Trump.
View attachment 143574
Click to expand...
Did I not tell you in the other thread to post the link and quote the section that you want to display?

are you just trying to annoy me?
 
9dashline

9dashline

Captain
Registered Member
luminary said:
I'm stumped at how they come to this conclusion about the chinese chip ecosystem.

I was also wondering why Israel is in tier 2 (as they are building supercomputers there), but then realized it might be a useful negotiation lever for Trump.
View attachment 143574
Please, Log in or Register to view URLs content!
Click to expand...
China can just kill 2 birds with one stone and bump up AR timeline to this year. Obvious move

And for the record, Patel obviously is Jia Hind, but he is saying exactly what I been saying for long time now. except his writing is behind a $500 pay wall and I gave my insights for free

And, for various reasons, I dont think Americas grand plan will work out the way it hoped
 
Last edited:
You must log in or register to reply here.
Top