Russia faces severe limitations on its access to AI compute, critical for developing and deploying AI at scale. Russia has no ability to indigenously produce AI accelerators or any of their key inputs, and so is entirely reliant on overseas supply. Sanctions related to the Russia-Ukraine war cut it off from legal access to western AI chips, though it has been actively smuggling AI compute, with reported routes going through India - https://archive.ph/QpU8Z. 

Its other option would be importing Chinese produced AI accelerators from Huawei, Cambricon etc. though to date there has been no reported cases of exports of these chips outside China. One of Russia’s easiest paths to building AI capabilities is in actively seeking Chinese cooperation and access to Chinese chips and models. In December 2024 Russian President Vladimir Putin instructed the Russian state-owned bank and AI leader Sberbank to work with China to further Russia’s AI ecosystem - https://www.cfr.org/article/china-russia-relations-january-2025.

It is far from clear however if China would be willing to sell large amounts of compute to Russia. China is itself highly constrained on its access to AI compute due to US export controls, with its domestic supply almost certainly set to undershoot its domestic demand. Russia would therefore need to offer some substantial advantage to China to make it worth trading off Chinese domestic AI growth for Russian AI gain. China may instead prefer to export only cloud access to AI chips and applications, rather than direct export of chips to Russia.

Ultimately this metric is a key one to track for understanding the AI capabilities available to Russia, and for how the Russia-China relationship is evolving and how China is handling western export controls and secondary sanctions risks. 

This question resolves Yes if, before January 1, 2030, 00:00:00 UTC, credible evidence demonstrates that China has provided Russia with a cumulative total of >100,000 AI accelerators.

1. AI Accelerators Definition

Qualifies as "AI accelerator":

• Graphics Processing Units (GPUs) designed for AI/ML workloads (e.g., Nvidia A100/H100/H200 series, AMD MI series)

• Chinese-manufactured AI accelerators (Huawei Ascend 910/910B/910C, Cambricon MLU series, Biren BR series, Moore Threads, Iluvatar CoreX)

• Tensor Processing Units (TPUs) or similar specialized AI inference/training chips

• Any chip explicitly marketed or documented as designed for AI training or inference workloads

Excludes:

• General-purpose CPUs (even if used for some AI tasks)

• Consumer gaming GPUs not designed for datacenter AI workloads

• Cryptocurrency mining ASICs

• Chips designed primarily for graphics rendering rather than AI compute

Performance threshold: Chips must have performance broadly comparable to or exceeding Nvidia A100. Lower-performance chips count toward the total but are weighted by relative performance (e.g., if a chip has 50% of A100 performance, 2 chips count as 1 toward the threshold). Any chip with >A100 performance only counts as 1 chip with no relative weighting. 

2. "Provided by China" Definition

Qualifies as Chinese provision:

Direct government-to-government:

• Official sales or transfers from Chinese government entities to Russian government

• State-owned enterprise sales to Russian entities

• Officially announced cooperation agreements with chip deliveries

Commercial sales with government approval:

• Sales by Chinese companies (Huawei, Cambricon, etc.) to Russian entities

• Cloud service contracts that involve hardware deployment in Russia

• Leasing arrangements where Chinese firms provide chips for Russian use

Indirect but attributable:

• Third-party transactions where credible evidence shows chips originated from Chinese manufacturers and Chinese government facilitated or permitted the transfer

• Chips transiting through third countries but originating from Chinese supply

Does NOT count:

• Western chips smuggled through China (where China is merely a transit point)

• Chinese chips acquired through smuggling operations without Chinese government knowledge or tacit approval

• Chips stolen or acquired through illicit means that China actively opposed

Attribution standard: Must be reasonable to conclude Chinese entities (government or commercial with government awareness) were the source. Pure black market operations do not count unless evidence suggests Chinese government tolerance or facilitation.

3. "Russia" Definition

Qualifies as provision to Russia:

• Russian government entities, ministries, or agencies

• Russian state-owned enterprises (Sberbank, Rostec, Rosatom, etc.)

• Russian private companies operating within Russia

• Russian research institutions or universities

• Chips physically located in Russia or Russian-controlled territory

Does NOT count:

• Russian individuals or companies using Chinese cloud services physically located in China

• Russian entities accessing chips via standard commercial cloud APIs without physical hardware transfer

4. Evidence Standards

Tier 1 - Official announcements (strongest):

• Chinese or Russian government statements specifying chip quantities

• Company announcements from Chinese chip manufacturers

• Regulatory filings or customs data

• Official cooperation agreements with implementation details

Tier 2 - Investigative reporting (requires corroboration):

• Reports from major news outlets (Reuters, Bloomberg, FT, WSJ, NYT, AP) based on:

  • Multiple independent sources (officials, industry sources, etc.)

  • Shipping manifests, customs records, or commercial documents

  • On-the-record statements from involved parties

• Analysis by defense/technology research institutions (CSIS, RUSI, CNAS, IISS) with documented methodology

Corroboration requirements:

• Tier 1 evidence alone is sufficient

• Tier 2 requires at least 2 independent sources OR 1 source + supporting technical evidence

Cumulative counting: The 100,000 threshold is cumulative from January 1, 2025 through December 31, 2029. Evidence can come from multiple reports covering different time periods or shipments.

Timing: All chip transfers must occur between January 1, 2025 and December 31, 2029. Evidence of these transfers can emerge after the deadline (grace period until July 1, 2030 for evidence to emerge).

Cumulative counting: Multiple shipments, contracts, or transfers can be aggregated to reach the 100k threshold. If evidence shows 40k chips in 2026, 35k in 2027, and 30k in 2028, this cumulates to 105k and resolves Yes.

Performance weighting: If chips have significantly different performance levels, moderators may normalize to "A100-equivalent" units. For example:

• Huawei Ascend 910B ≈ 1.0 A100-equivalent

• Older chips with 50% performance ≈ 0.5 A100-equivalent

• More powerful chips do not count as >1.0 equivalent

This weighting only applies if evidence clearly indicates mixed generations. If uncertain, raw chip counts are used.

Cloud vs. hardware: Physical chip transfers or deployment of hardware in Russia counts. Pure cloud service access to chips physically in China does NOT count unless the contract explicitly involves hardware transfer or Russian control of specific hardware.

Leasing and time-limited access: Chips leased or provided for limited time periods count toward the total based on physical units provided, not the duration of access.

Dual-use chips: If chips are marketed as both AI and general HPC but meet the AI accelerator definition, they count.

Chips in transit: Chips that are ordered/paid for but not yet delivered by December 31, 2029 do NOT count. Physical delivery must occur.

Failed deliveries: Chips that are ordered but intercepted, sanctioned, or otherwise prevented from reaching Russia do NOT count.

Returned or non-functional chips: If credible evidence shows significant quantities were returned as defective or never became operational, moderators may adjust counts. However, burden of proof is high—delivery presumed successful unless clear evidence otherwise.

Chinese chips in Western facilities: If China provides chips that are physically located in third countries for Russian use (e.g., Chinese datacenter in Kazakhstan serving Russian customers), this counts if Russian entities have meaningful control/access.

Smuggled Western chips: Western chips (Nvidia, AMD) that transit through China but originate from Western manufacturers do NOT count, even if China facilitates the smuggling. Only Chinese-manufactured chips or chips where China is the actual provider count.

Evidence uncertainty: If evidence provides ranges (e.g., "between 80,000 and 150,000 chips"), moderators use the midpoint. If evidence suggests "at least X" or "more than Y," and the lower bound exceeds 100k, this resolves Yes.

Ambiguous sourcing: If unclear whether chips are Chinese-manufactured vs. Western chips transiting through China, moderators require positive evidence of Chinese origin. Ambiguity resolves toward No. 

Military vs. civilian: No distinction between military and civilian end-users in Russia. All Russian entities count equally.

Verification timing: Evidence can be published/discovered up to 6 months after the December 31, 2029 deadline (until July 1, 2030) as long as it documents transfers occurring before the deadline.

"Provide" vs. "sell": No distinction between gifts, sales, leases, or other forms of transfer. All forms of provision count.