NVIDIA and Microsoft have taken action: in 4 days, 238 stars of DeepSeek-Bot burned a 30-fold Token bill

Introduction: The DeepSeek-Bot launched on October 7 is merely a small plugin that allows multiple AI Bots to collaborate in group chats. With 238 stars in four days, it alone cannot support an in-depth article. What is worth writing about is another matter: it has hit upon a trend that is rewriting AI hardware pricing. Multi-agent collaboration has caused a significant jump in Token consumption, and the resulting bills ultimately fall on CPUs and memory. This article starts from this small plugin and connects Token multiples, CPU gaps, memory swapping, and the actions of two giants into a single narrative.

1. 30 times is not a linear doubling; it is a shift from budgetable to unbudgetable costs

First, let's look at the scale. According to engineering data from Anthropic, the Token consumption for a single Agent completing a typical task is about 4 times that of a regular conversation, while a multi-Agent collaborative system reaches 15 times. The figures provided by academia are even more extreme: according to the aforementioned joint paper, the Token consumption for Agents executing Agentic programming tasks is about 1000 times that of standard conversations, with simple tools operating external systems falling between 5 to 30 times.

Another easily overlooked finding in the same paper is that when running the same Agent and executing the same task, the fluctuation in Token bills can reach 30 times. This means that when the same user clicks the same button, the number of model calls behind it may differ by 30 times, with each call needing to resend the accumulated conversation history. Gartner's report in March 2026 provides a compromise estimate: 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 a more aggressive prediction in May: by 2030, Token consumption will increase 24 times, with monthly usage estimated at about 12 trillion Tokens. The decline in Token unit price has not reversed this trend; according to Gartner, it is expected that reasoning costs will drop by more than 90% by 2030 compared to 2025, but the growth rate of usage far exceeds the rate of price decline. Each call is becoming cheaper, but the number of calls is increasing, and each call is getting larger, leading to an increase in total expenditure due to explosive usage.

Combining call data with hardware readings, the main contradiction in the Token economy has shifted from unit price to the product of call frequency and per-call consumption. The implication for hardware is that the bottleneck in GPU utilization is no longer the speed of a single forward computation, but rather memory bandwidth and scheduling; this is precisely why CPUs and memory have re-entered the pricing spotlight.

2. CPUs return from supporting roles to leading roles

In multi-agent systems, task scheduling, tool invocation, logical judgment, and context management are all areas where CPUs excel. According to research from Shanghai Securities Journal, the time spent on orchestration alone can account for 50% to 90% of total latency in agent workloads; when CPU performance is insufficient, even the strongest GPU can remain idle due to waiting for scheduling.

Demand has manifested in pricing and delivery times. According to public reports, Intel has confirmed price increases for some server CPUs, with high-end Xeon prices rising by 7% to 12%, and consumer-grade CPUs experiencing three rounds of price adjustments this year, with the Plus series seeing a cumulative increase of nearly 30%; AMD is expected to raise prices for mid-to-high-end server CPUs by another 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, with some cases extending 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, which will rise to 120 million in the agent era, a threefold increase; according to Morgan Stanley's framework, by 2030, Agentic AI will create a CPU incremental market of $32.5 billion to $60 billion, accounting for about 60% of the total server CPU market exceeding $100 billion. AMD's guidance during its 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 deployment, the number of CPU configurations may surpass that of GPUs.

Looking further: the elasticity of this CPU line comes from changes in ratios rather than total growth. The server CPU market itself is growing steadily; the change lies in the number of CPUs standing next to each GPU. A shift from 1:8 to 1:1 means that the same GPU expenditure corresponds to several times the CPU expenditure, marking the first time CPUs have an independent incremental logic in the AI hardware narrative. This judgment is predicated on the assumption that agent workloads indeed scale along the current path; if the orchestration phase is significantly compressed by software optimization, the ratio curve will slow down.

3. Memory: The root of scarcity lies in capacity replacement, not in demand fluctuations

Multi-agent systems rely on long-term memory to remember preferences, history, and task progress, and to invoke them 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: whether a GPU computes quickly increasingly depends on how quickly memory feeds it.

Micron executives have been very straightforward about this matter in public. Executive Vice President Sumit Sadana stated at the KeyBanc Technology Forum that Token consumption for Agentic AI workloads could reach 5 to 30 times that of similar chat tasks, with deeper inferences being even higher; CEO Sanjay Mehrotra mentioned in an interview with CNBC that the training era is over, and the inference era has just begun, while agent AI has not yet truly scaled.

Micron's reports validate this trend. In FY26Q3, revenue reached $41.456 billion, a year-on-year increase of 345.7%, with data center business revenue at $11.52 billion, compared to only $1.53 billion in the same period last year; FY26Q4 revenue 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 testing conditions.

A more critical detail is the replacement ratio: for every 100 bits of HBM produced, about 300 bits of DDR output are reduced, and with the advancement of HBM4E, this could approach 4:1. This means that the structural root of memory scarcity lies not in demand but in the physical replacement of capacity; for every unit of HBM expanded for AI, the supply of traditional DRAM is compressed by three to four units. As long as HBM continues to expand for AI, the scarcity of DDR cannot be alleviated by cooling demand; the verification point for this supply-demand logic lies in the subsequent capacity and price guidance released by manufacturers like Micron.

4. The panic in software and the increment in hardware stem from the same product release

In mid-January 2026, Anthropic released Claude Cowork, and on January 30, it launched industry plugins for legal, financial, sales, and other functions, followed by Claude Opus 4.6, which breaks down large tasks into sub-tasks and assigns them to different agents for parallel completion. The market's reaction was split in two. On the software side, according to public reports and Wedbush research, the software market value wiped out during that round of sell-off was between $800 billion and $1 trillion (different statistics cover different ranges of stocks), with FactSet dropping 10% in a single day, and Wedbush termed it "Software-mageddon." The core concern is that after agents can autonomously complete contract reviews and financial analyses, the per-head charging SaaS model will be eroded.

On the hardware side, at the same time, contrary readings were given. Morgan Stanley pointed out that as AI shifts from generative to agentic, the investment scope will expand from GPUs to CPUs and memory; semiconductors' weight in the S&P 500 has surpassed 20%. The two sides of the same coin are: the stronger the agents' substitution ability, the greater the consumption of underlying computing power, CPUs, and memory. The anxiety over software's share and the incremental demand for hardware are driven by the same product cycle; treating them as two separate stories for pricing may be one of the biggest pricing discrepancies in the current AI sector.

5. What are Microsoft and NVIDIA 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 the Blackwell RTX GPU featuring up to 6144 cores, paired with the Grace CPU of up to 20 cores, and unified memory of up to 128GB, capable of running models with over 120 billion parameters locally, starting at $2599, with shipments on October 16. Jensen Huang was present, and his exact words were widely quoted: "For forty years, you start an application, click, and type; now you ask a question, and the computer does the work for you."

More important than hardware is the change in the operating system. According to Microsoft's official blog, Windows has been redesigned around "hybrid intelligence": once authorized, Copilot can read local files, organize them, troubleshoot, and execute workflows; on the security front, system-level containers limit the files and network scope accessible to agents; tasks are scheduled between local and cloud models. NVIDIA also released the Open Agent Safety Platform on September 28, covering the safety layers for agents from testing to deployment; Huang simultaneously disclosed that orders for the Blackwell and Vera Rubin systems have reached $1 trillion, extending to 2027 (company statement).

The logic behind these two actions is coherent: Microsoft is rebuilding the entry point for the agent era at the operating system level, while NVIDIA is stocking up for the 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 building the foundation for it. By placing part of the inference on local devices, it essentially acknowledges that the power consumption of agents is too large and must be processed in layers; every PC running agents becomes a new permanent computing node.

6. Token bills: From per-head charging to consumption-based charging

Gartner's 2026 enterprise survey provides a reading that deserves serious attention: nearly a quarter of technology leaders are already paying $200 to $500 per month in AI coding Tokens for each developer, with about 6% exceeding $2000; the agency predicts that the cost of AI coding Tokens will exceed the average salary of developers by 2028. This indicates a shift in billing methods: during the chatbot era, budgets were made based on the number of seats, while in the agent era, budgets can only be made based on execution counts, and execution counts are something the model itself cannot accurately estimate. The aforementioned joint paper clearly states that models will continue to underestimate the number of Tokens they will spend.

Returning to the small plugin with 238 stars. The Main Bot defaults to exclusive replies and calls other Bots on demand; what the developer is doing is essentially the most basic Token budget management. When this micro action is scaled to enterprise-level deployment, its implications will be amplified: the rigid demand for CPUs, memory, and networks will only become stronger.

Conclusion

The 238 stars of DeepSeek-Bot are not important; what matters is the cost structure it demonstrates: with each collaborative step of multiple agents, the Token bill jumps a tier, and the magnitude of the jump is something even the model cannot predict. This structure compresses downstream along the industry chain, first squeezing out changes in CPU ratios and delivery times, then squeezing out capacity replacement in memory, and finally squeezing out operating system-level reconstruction. From this chain, the structure of computing power consumption may be shifting from a GPU-centric model to a collaboration of CPU, memory, and GPU, where the incremental demand on the hardware side is relatively more verifiable, while the impact on the software side depends on the pace of agent substitution; both clues 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 content of this article is for general information and market commentary purposes only, compiled from publicly available information as of the time mentioned in the text. Relevant market data, expectations, and probabilities may change with market conditions. The views and investment strategies cited from third-party institutions, analysts, or other individuals represent the views of the relevant third parties at a specific time and do not represent the views or recommendations of BIT. This article does not constitute investment advice, investment research, an offer, solicitation, or recommendation of any securities, investment products, or trading strategies, nor should it be the basis for any investment decision. Financial markets carry risks, and securities prices and market performance may fluctuate; past performance and historical market trends do not represent or guarantee future results. Investors should independently assess the relevant risks based on their own circumstances and seek professional advice when necessary. BIT's Hong Kong stock trading services are not provided to residents of Hong Kong or users in other restricted jurisdictions.