The launch of the Kimi K3 model has refocused AI investors on a central question: do cheaper, more efficient models necessarily lead to reduced demand for chips and memory? Analyses from Citigroup and Bank of America Securities lean towards a negative answer, suggesting efficiency improvements could unleash greater usage, ultimately increasing total resource consumption.
According to market analysis, Moonshot AI's Kimi K3 features 2.8 trillion parameters and is designed for a 1 million token context window and long-horizon agent tasks. Citigroup semiconductor analyst Peter Lee believes that even if Kimi K3 sees widespread adoption, demand for general-purpose memory like server DDR5 and eSSD will still increase.
In a July 17th report, Bank of America Securities semiconductor analyst Vivek Arya similarly noted that leading US AI labs are more likely to increase their computing power investments rather than reduce them. If Chinese open-source models continue to close the capability gap, leaders like OpenAI, Anthropic, and Alphabet (ASX: GOOGL) will need to maintain their differentiation through larger training runs, heavier inference loads, and faster product iteration cycles.
This implies that the market's valuation logic for Kimi K3 cannot focus solely on lower per-inference costs. The more critical question is whether a lower-cost model leads to more API calls, longer task chains, and greater token generation. If the answer is yes, then components like GPUs, HBM, DDR5, eSSD, high-speed networking, and inference systems could all continue to benefit.
Low Cost and High Performance Trigger a Paradox
Citigroup views Kimi K3 as a potential case of the "Jevons Paradox" within the AI supply chain. The core of this paradox is that when technological efficiency improvements lower the unit cost of using a resource, consumption can increase so dramatically that total resource use rises instead of falls.
Kimi K3's appeal stems from its combination of low cost and high performance. Its public pricing shows a cost of $0.30 per million tokens for cached input and $15 per million tokens for output. Its full model weights are scheduled for release on July 27th, with the goal of supporting long-horizon agent workflows.
Citigroup's assessment is that a model price reduction does not automatically reduce hardware demand. On the contrary, lower costs could increase the willingness of developers and enterprises to call the model, driving more AI agent deployments. Agent tasks are not one-off queries but involve continuous token generation, reading, and processing across sequential tasks. An increase in call frequency and task length can transform lower unit costs back into higher total resource consumption.
Therefore, the impact of Kimi K3 on the semiconductor chain hinges not on "whether each call is cheaper," but on "whether the total token volume expands." This is precisely why Citigroup remains bullish on demand for server DDR5 and eSSD.
Long-Context Inference Shifts Pressure to Memory
Kimi K3 employs three key technologies: Kimi Delta Attention, Attention Residuals, and Stable LatentMoE. Kimi Delta Attention reduces the cost of the 1 million token context window, Attention Residuals enable selective retrieval of representations across different model depths, and Stable LatentMoE increases sparsity, activating only 16 out of 896 experts per token.
Moonshot AI's technical documentation states this architecture gives Kimi K3 an expansion efficiency 2.5 times that of its K2 predecessor. High sparsity and long-context capability are key foundations for reducing its operational cost.
However, Citigroup emphasizes this does not equate to the disappearance of resource pressure on the inference side. Kimi K3 is still not a lightweight deployment solution; its operation requires multi-node clusters, with super-node configurations exceeding 64 GPUs. More importantly, long-context and agent tasks increase the occupancy of the KV Cache, thereby adding to the memory burden during inference.
KV Cache demand is directly linked to server DDR5 and eSSD. DDR5 handles high-frequency data access, while eSSD benefits from larger cache and data storage needs. For memory manufacturers, the critical variable is not whether a single model is more compute-efficient, but how many times the model is called after price reductions, how many tokens each call generates, and how long the agent task chains become.
Model Convergence Raises the Compute Bar
Bank of America Securities' conclusion echoes Citigroup's but focuses more on GPUs and AI infrastructure. Analyst Vivek Arya argues that stronger Chinese open-source models may not lead US AI giants to cut compute investments. Instead, a narrowing model gap raises the cost for leaders to maintain their advantage.
Bank of America Securities points out that if open-source models continue to converge, OpenAI, Anthropic, and Google will need to rely on larger-scale training, more reinforcement learning and synthetic data loops, heavier inference during testing, and faster product release cycles to defend their differentiation.
This logic does not depend on which model leads in the short term. Leaderboard rankings may change rapidly, but what enterprises actually buy is stable, low-latency, highly available AI output with a lower "cost per unit of effective output." As model capabilities converge, competitive pressure transfers to the underlying infrastructure, including GPUs, HBM, high-speed networks, and inference systems.
Therefore, Bank of America Securities believes the emergence of models like Kimi K3 should not be simplistically interpreted as "more efficient models, fewer chips." A more realistic path is that intensified model competition forces leaders to continue ramping up compute investments to stay ahead.
MoE Architecture Shifts Bottlenecks to Memory and Interconnect
Kimi K3's use of a Mixture-of-Experts (MoE) architecture means the total parameter count and the parameters activated per inference can be separated. This change reduces some computational pressure but shifts infrastructure bottlenecks from pure compute power to memory access, expert routing, response latency, and interconnect capabilities.
Bank of America Securities stresses that MoE inference requires systems to move data, schedule expert modules, and connect larger compute clusters more efficiently. NVIDIA (ASX: NVDA) has indicated that modern MoE inference requires larger GPU domains. Its estimates show the GB300 NVL72 can deliver up to 25 times the performance per watt of the Hopper platform on leading open-source models.
CoreWeave's testing around Kimi K2.6 also illustrates that even though open-source MoE models activate only a fraction of parameters per inference, maintaining leadership in speed and cost-effectiveness still requires optimized NVIDIA GB300/GB200 NVL72 infrastructure.
This implies that model weights may gradually become commoditized, but the GPUs, high-bandwidth memory, network interconnects, and inference systems required to run them will not lose value. On the contrary, as model call volumes expand, these areas could become the focus of even more intense competition.
Token Usage Continues to Expand, Demand Has Not Peaked
Bank of America Securities also cites OpenRouter data indicating that model call volumes are still growing rapidly. As a third-party API platform connecting multiple model labs, token usage on OpenRouter continues to rise, with token usage for models from Chinese AI labs now exceeding that of non-Chinese lab models.
This data does not directly represent all enterprise and consumer scenarios, but it at least shows that developer choices within the open model ecosystem are changing. Falling model prices and improving capabilities may drive continued expansion in call volume, rather than being offset by the efficiency gains of any single model.
Enterprise paid usage is also increasing. Ramp data shows that as of June 2026, approximately 55% of US businesses have paid subscriptions for AI models, platforms, or tools, higher than the 21% estimate from the US Census Bureau's BTOS survey. By model, enterprise adoption rates are 42.4% for Anthropic and 39.5% for OpenAI.
However, AI spending remains highly concentrated. The top 1% of enterprise users spend an average of about $4,833 per employee per month on AI, the top 10% spend $516, while the overall median is just $11. Subscription rates reach 79.8% in the technology and media sectors, and adoption is 65.5% among large enterprises, higher than 61.3% for medium-sized and 48.7% for small businesses.
This data supports the judgment that AI usage is still in a diffusion phase. If lower-cost models lower the barrier to entry, future incremental demand may come from more enterprises, more developers, and more agent applications.
From "Less Compute" to "More Workload"
The shared conclusion from Citigroup and Bank of America Securities is that the significance of Kimi K3 lies not in whether a single model reduces unit inference cost, but in whether lower costs unleash a larger scale of workloads.
For Citigroup, the most direct beneficiaries are server DDR5 and eSSD. Long-context processing, KV Cache, and agent tasks will increase memory access and storage demands. For Bank of America Securities, the more important areas are GPUs, HBM, high-speed networks, and inference systems, as model competition will compel leaders to continue investing in infrastructure.
Bank of America Securities also mentions that Kimi K3's involvement with "45nm open-source EDA" should not be simplistically read as the replacement of commercial EDA. Instead, it indicates that chip design still relies on EDA tools. For advanced process nodes, commercial EDA vendors like Cadence and Synopsys remain in a critical position.
The risk lies in an alternative scenario: if model compression, inference optimization, and hardware efficiency gains outpace new workload growth over the long term, AI infrastructure expansion could cool off periodically. However, within the current framework of these two investment banks, Kimi K3 appears more like a catalyst for increased usage rather than a signal weakening chip demand. Following efficiency gains, the market may instead need to focus on more tokens, more inference, and greater consumption of underlying hardware.
Comments