Citi: Citi AI Weekly: Gemini retakes top model ranking as locked-in memory supply drives up compute costs

Gemini returns to the top: why is the commercial impact limited?

Google’s newly released Gemini 4 Carbon scored 59 on the Artificial Analysis Intelligence Index, ranking first; Google last held the top spot in February with Gemini 3.1 Pro. The lead of the top closed-source model over the leading open model widened to 27 points, the largest since June.
The evaluation followed the 27 October index update, and the new index relies mainly on private datasets. Carbon is currently available only to cyber defenders, so Citi believes its direct commercial impact may be limited.
As measured by Artificial Analysis tasks, Carbon uses 2.4 times as many output tokens as GPT-6 Astra, but its cost per task calculated at listed prices is only 24% higher, partly because Carbon’s output tokens are cheaper.

Memory prices are rising while order backlogs continue to grow. Citi expects the price of Micron’s HBM4 12-layer product to more than double year over year; SK hynix has sold more than half of its capacity for 2027–2028, Micron has committed more than 75% of its fiscal 2027 shipments, and Samsung plans to allocate 60%–70% of its capacity to long-term agreements.
Citi notes that cached context accounts for about 80% of leading models’ costs on Artificial Analysis tasks. As compute design shifts toward full racks and shared memory pools, memory becomes harder to substitute: smaller AI cloud providers without long-term contracts may have to buy memory at spot prices, offsetting revenue from higher GPU rental rates.
Rental rates have already diverged: since November 2025, rental rates for the memory-rich B300 have risen 69%, while H100 rates have risen 33% over the same period.

Nvidia’s Minerva uses AI to predict component failures and address problems before they slow jobs, freeing up more compute from existing GPUs. This capability becomes more valuable as tasks get longer: in April this year, the leading model handled tasks that take humans 17 hours with a 50% success rate, versus a corresponding duration of 2 hours a year earlier.
Citi judges that efficiency gains are unlikely to affect near-term infrastructure demand as long as GPUs remain scarce. Test-equipment maker Teradyne says its AI demand now spans areas including memory, interconnects and integrated systems, driving related revenue up 70% year over year.

Google DeepMind’s Genie 4 makes simulation training for robots more realistic, but Citi believes only real-world data can demonstrate that the operational policies learned through training are effective, so early adopters may build a data advantage.
Manufacturing still lags finance in software adoption: among companies using Ramp, 61% of manufacturers paid for AI software subscriptions in September, compared with 74% of financial firms. SAP has begun dispatching tasks to robots from order, inventory and maintenance systems, giving it a software entry point to participate in robot adoption.
The robot learning conference from 9 November to the 12th may show whether simulation technology is shortening the time from pilots to robot fleet deployment; if it is, Citi believes physical AI’s contribution to hardware and inference demand could accelerate.