In the context of the geopolitics post, it (weaponization of the economy) is not a recent phenomenon. The weaponization of the economy began after 9/11, as explained in Henry Farrell and Abraham Newman's book, Underground Empire: How America Weaponized the World Economy.
It didn't start with that intention, of course. Rather it started with the intent of tracking terrorist money movement across borders post 9/11. Gradually, America realized it could weaponize the global financial system for other reasons too. When Europe and China didn't seem too keen to isolate Iran (Obama era), the US realized any tech having at least x% of American IP could also be weaponized. The reasons and means for weaponization of the economy have only been steadily increasing over the last 2 decades. The book is an interesting read.
The R1/R2 framing is useful for thinking about India's position. Though if the R1 loop is sustained for a few years, it might lead to less separability between models and hardware. As Ascend matures and training of models is increasingly optimised for Ascend, the separability might decrease. Chinese open-weight models might still be portable but practically optimised for Ascend. India could still run them on Nvidia chips but at a worse cost-per-token than achievable on Ascend. The weights are still available, but the cost advantage that makes them attractive erodes.
If that's right, R2 has a second-order benefit. Not just access reliability, but continued CUDA compatibility of the Chinese open-weight ecosystem because Chinese labs keep training on Nvidia hardware. Under R1, India still gets open-weights, but they drift away from India's infrastructure over time.
Perhaps this is not a real concern; cross-compatibility incentives (Chinese labs wanting global deployment) might still keep the stacks convergent.
In the context of the geopolitics post, it (weaponization of the economy) is not a recent phenomenon. The weaponization of the economy began after 9/11, as explained in Henry Farrell and Abraham Newman's book, Underground Empire: How America Weaponized the World Economy.
It didn't start with that intention, of course. Rather it started with the intent of tracking terrorist money movement across borders post 9/11. Gradually, America realized it could weaponize the global financial system for other reasons too. When Europe and China didn't seem too keen to isolate Iran (Obama era), the US realized any tech having at least x% of American IP could also be weaponized. The reasons and means for weaponization of the economy have only been steadily increasing over the last 2 decades. The book is an interesting read.
The R1/R2 framing is useful for thinking about India's position. Though if the R1 loop is sustained for a few years, it might lead to less separability between models and hardware. As Ascend matures and training of models is increasingly optimised for Ascend, the separability might decrease. Chinese open-weight models might still be portable but practically optimised for Ascend. India could still run them on Nvidia chips but at a worse cost-per-token than achievable on Ascend. The weights are still available, but the cost advantage that makes them attractive erodes.
If that's right, R2 has a second-order benefit. Not just access reliability, but continued CUDA compatibility of the Chinese open-weight ecosystem because Chinese labs keep training on Nvidia hardware. Under R1, India still gets open-weights, but they drift away from India's infrastructure over time.
Perhaps this is not a real concern; cross-compatibility incentives (Chinese labs wanting global deployment) might still keep the stacks convergent.