Source: BIT Securities
Introduction: The DeepSeek-Bot launched on October 7 is just a small plugin that allows multiple AI Bots to collaborate in group chats. Four days later, it received 238 stars, but that alone does not justify an in-depth article. What is noteworthy is another matter: it has identified a trend that is rewriting AI hardware pricing. The collaboration of multiple agents leads to a significant increase in Token consumption, and the resulting bill ultimately falls on CPUs and memory. This article starts from this small plugin and links the multiplicity of Tokens, CPU gaps, memory replacement, and the actions of two giants into a single narrative.
1. 30 Times is Not Linear Doubling, It's a Shift from Budgetable Costs to Unbudgeted
First, let's look at the magnitude. According to engineering data relayed by Anthropic, the Token consumption for a single agent completing a typical task is about four times that of a normal conversation, while a multi-agent collaborative system reaches fifteen times. The figures given by academia are even more extreme: according to the aforementioned joint paper, in Agentic programming tasks, the Token consumption of the Agent is about 1000 times that of standard dialogue, with simple tools operating external systems ranging between 5 to 30 times.
There is another overlooked finding in the same paper: when running the same agent and performing the same task, the fluctuation in Token bills can reach as much as 30 times. This means that the model call times behind the same user's click of the same button can differ by 30 times, with each call requiring a resend of the accumulated conversation history. Gartner provided a compromise estimate in its March 2026 report: the Token consumption per task in Agentic workflows is 5 to 30 times that of standard chatbots, as a single user request may trigger 10 to 20 sequential model calls.
Goldman Sachs’ research department made more aggressive predictions in May: by 2030, the volume of Token consumption is expected to grow by 24 times, roughly translating to 1.2 trillion Tokens per month. The decrease in Token prices hasn’t reversed this trend; according to Gartner, by 2030, inference costs will decrease by over 90% compared to 2025, but the growth rate of usage far exceeds the price drop. Each call is getting cheaper, yet the number of calls is increasing, and the size of each call is getting larger, leading to increased total expenditure due to explosive usage.
The main contradiction in the Token economy has shifted from unit price to the product of call frequency and per-call consumption. This has implications for hardware: the bottleneck for GPU utilization is no longer the speed of single forward calculations, but rather memory bandwidth and scheduling; this is precisely why CPUs and memory have re-entered pricing considerations.
2. CPUs Return to the Spotlight from Supporting Role
In multi-agent systems, task scheduling, tool invocation, logical judgment, and context management are all tasks that CPUs excel at. According to research reports from Shanghai Securities, the time taken in the orchestration segment of agent loads can account for 50% to 90% of total latency; when CPU performance is insufficient, even the strongest GPUs can be idle due to waiting for scheduling.
Demand is already manifesting in price and delivery terms. According to public reports, Intel has confirmed a price increase for certain server CPUs, with high-end Xeon CPUs rising overall by 7% to 12%, and consumer-grade CPUs experiencing three rounds of price hikes this year, with the Plus series cumulatively increasing nearly 30%; AMD is also expected to raise prices for mid and high-end server CPUs by 4% to 7% in the second half of the year; the delivery cycle for server CPUs has extended from the usual 1 to 2 weeks to 8 to 12 weeks, and in some cases, up to 6 months.
There are two estimates of scale. According to Arm’s calculations, traditional AI data centers require about 30 million CPU cores per gigawatt of power, while in the agent era, this will rise to 120 million, a threefold increase; according to Morgan Stanley’s framework, by 2030, Agentic AI will create a CPU incremental market between $32.5 billion to $60 billion, accounting for about 60% of the total server CPU market over $100 billion. AMD’s guidance during the first quarter earnings call was consistent: the ratio of CPUs to GPUs in AI infrastructure is shifting from 1:4 to 1:8 towards 1:1, and in scenarios with dense agent deployments, the number of CPUs may surpass that of GPUs.
Looking further, the elasticity of the CPU line comes from changes in ratio rather than total volume growth. The server CPU market itself is growing steadily, the change lies in the number of CPUs standing beside each GPU. Moving from a ratio of 1:8 to 1:1 means that the same GPU expenditure corresponds to several times the CPU expenditure, marking the first time that CPUs have an independent incremental logic in the narrative of AI hardware. This judgment hinges on the assumption that agent loads indeed expand along the current path; if the orchestration segment is significantly compressed by software optimizations, the ratio curve will slow down.
3. Memory: The Root of Scarcity Lies in Capacity Replacement, Not Demand Fluctuations
Multi-agent systems rely on long-term memory to remember preferences, history, and task progress, and to call these at different points in time. Technically, each internal inference must read the entire growing KV cache from GPU memory, making inference a memory bandwidth-intensive task. In terms familiar to retail investors: the speed at which GPUs compute increasingly depends on how quickly memory feeds them.
Micron executives have been very straightforward about this. Executive Vice President Sumit Sadana stated at the KeyBanc Technology Forum that the Token consumption of Agentic AI workloads could reach 5 to 30 times that of similar chat tasks, with deeper inference being even higher; CEO Sanjay Mehrotra noted in an interview with CNBC that the training era has passed, and the inference era has just begun, while agent AI has yet to be truly deployed on a large scale.
Micron’s reports validate this trend. In FY26Q3, revenues reached $41.456 billion, a 345.7% year-on-year increase, with data center revenues at $11.52 billion, up from just $1.53 billion the previous year; FY26Q4 revenues further rose to $54.229 billion (company report verified by financial data interfaces). Sadana added two points at the forum: the supply-demand ratio in 2027 will be tighter than in 2026, and DRAM customers "cannot even get more than half of what they want even at high prices"; on the product side, HBM4 36GB 12-high can improve LLM inference throughput by 2.6 times under test conditions.
A more critical detail is the replacement ratio: for every 100 bits of HBM produced, approximately 300 bits of DDR output are reduced, which could approach 4:1 with the advancement of HBM4E. This means the structural root of memory scarcity does not lie on the demand side, but in the physical replacement of capacity; for every unit increase in HBM for AI, traditional DRAM supply is compressed by three to four units. As long as HBM continues to expand for AI, the scarcity of DDR cannot be alleviated by a reduction in demand; the verification point for this supply-demand logic lies in the capacity and price guidance published by manufacturers like Micron in the future.
4. Panic in Software and Hardware Increment, from the Same Product Release
In mid-January 2026, Anthropic released Claude Cowork, and on January 30, it launched industry plugins aimed at functions such as legal, finance, and sales; subsequently, Claude Opus 4.6 broke down large tasks into sub-tasks assigned to different agents for parallel completion. The market response was divided in two halves. On the software side, according to public reports and Wedbush research reports, the sell-off wiped away between $800 billion and $1 trillion in value from software stocks (with different counts covering different stocks); FactSet fell by 10% in a single day, and Wedbush dubbed it "Software-mageddon." The core concern was: after agents can autonomously complete contract reviews and financial analyses, the human-headcount-based SaaS model will be eroded.
On the hardware side, during the same period, contrary readings were provided. Morgan Stanley pointed out that as AI shifts from generative to agentic, the investment scope will expand from GPUs to CPUs and memory; semiconductors have exceeded 20% in weight within the S&P 500. The two sides of the same coin are: the stronger the agents’ substitution capability, the greater the consumption of underlying computing power, CPUs, and memory. The anxiety over software market share and the hardware demand increment are driven by the same product cycle; pricing them as two separate stories can be one of the biggest pricing discrepancies in the current AI sector.
5. What Microsoft and NVIDIA Are Doing
On October 7 local time, Microsoft launched the Surface Laptop Ultra in collaboration with NVIDIA in San Francisco: the NVIDIA RTX Spark platform with Blackwell RTX GPU featuring up to 6144 cores, Grace CPU with up to 20 cores, and unified memory of up to 128GB, capable of running models with over 120 billion parameters locally, starting at $2,599, to ship on October 16. Jensen Huang was present, and his original words have been widely quoted: "For forty years, you start an application, click, and type; now you ask a question, and the computer works for you."
More important than hardware is the change in the operating system. According to Microsoft’s official blog, Windows has been redesigned around "mixed intelligence": once authorized, Copilot can read local files, help organize, troubleshoot, and execute workflows; from a security perspective, system-level containers are used to limit the files and networks accessible by agents; and tasks are scheduled between local and cloud models. NVIDIA, on the other hand, released the Open Agent Safety Platform on September 28, covering the safety layers for agents from testing to deployment; Huang also disclosed that orders for Blackwell and Vera Rubin systems have reached $1 trillion, extending out to 2027 (company statement).
The logic of these two actions is coherent: Microsoft is rebuilding access for the agent age at the operating system level, while NVIDIA is stocking up for large-scale deployment of agents at the chip and platform level. Rather than saying they anticipate the arrival of multi-agent collaboration, it is more accurate to say they are already laying the groundwork for it. Putting some inference on local devices essentially acknowledges that the computing power consumption of agents is too large and must be handled hierarchically; every PC running an agent becomes a new resident computing node.
6. Token Bills: From Per-Head Billing to Consumption-Based Billing
Gartner's 2026 corporate survey provides a reading worthy of serious consideration: nearly a quarter of technology leaders are already paying $200 to $500 per month in AI coding Tokens for each developer, with about 6% spending over $2,000; the institution predicts that AI coding Token costs will exceed average developer salaries by 2028. This foreshadows a shift in billing methods: during the chatbot era, budgeting was based on seat counts, while in the agent era, budgeting can only be based on execution counts, which even the models themselves cannot accurately estimate; the aforementioned joint paper clearly states that models will continue to underestimate the number of Tokens they will require.
Returning to the small plugin that received 238 stars. The Main Bot by default monopolizes replies and calls other Bots as needed; what developers are essentially doing is the most basic Token budget management. Scaling this micro action to enterprise-level deployment will magnify the meaning: The rigid demand for CPUs, memory, and networks will only become stronger.
In Conclusion
The 238 stars of DeepSeek-Bot are not important; what matters is the cost structure it demonstrates: for each step of collaboration among multiple agents, the Token bill jumps, and the magnitude of this jump can even be unpredictable to the models themselves. This structure compresses down the industry chain, first affecting CPU ratio changes and delivery times, then impacting memory capacity replacement, and ultimately leading to operating system-level restructuring. From this chain perspective, the structure of computing power consumption may indeed be shifting from a GPU-centric model to a collaborative equilibrium among CPUs, memory, and GPUs. The demand increase on the hardware side can be relatively corroborated, while the impact on the software side depends on the pace of agent substitution; both threads are driven by the same product cycle. To what extent this shift has already been factored into the market remains to be seen.
Disclaimer: The contents of this article are for general information and market commentary only, compiled from publicly available materials as of the time mentioned and may be subject to change based on market conditions. The views and investment strategies attributed to third-party institutions, analysts, or other individuals mentioned herein only represent those specific third parties at a given time and do not necessarily represent the views or recommendations of BIT. This article does not constitute investment advice, research, an offer, solicitation, or a recommendation of any securities, investment products, or trading strategies, nor should it be relied upon as the basis for any investment decisions. Financial markets involve risks, and securities prices and market performance may fluctuate; past performance and historical market trends do not guarantee or represent future results. Investors should independently assess the relevant risks based on their own circumstances and seek professional advice when needed. BIT's Hong Kong stock trading services are not offered to residents of Hong Kong or users in other restricted jurisdictions.
免责声明：本文章仅代表作者个人观点，不代表本平台的立场和观点。本文章仅供信息分享，不构成对任何人的任何投资建议。用户与作者之间的任何争议，与本平台无关。如网页中刊载的文章或图片涉及侵权，请提供相关的权利证明和身份证明发送邮件到support@aicoin.com，本平台相关工作人员将会进行核查。
