Nvidia and Microsoft have already made their move: DeepSeek-Bot gained 238 stars in 4 days, burning through a Token bill 30 times larger

DeepSeek-Bot, which launched on October 7, is just a small plugin that lets multiple AI Bots collaborate in group chats. It gained 238 stars in four days, but on its own it can't sustain an in-depth article. What makes it worth writing about is something else: it has hit upon a trend that is rewriting the pricing of AI hardware—multi-agent collaboration causes Token consumption to jump by orders of magnitude, and the bill for that jump ultimately lands on CPUs and memory. This article starts from this small plugin and strings together Token multiples, the CPU gap, memory displacement, and the moves of two giants into a single thread.

1. A 30x jump is not linear doubling—it's costs shifting from budgetable to unbudgetable

First, look at the scale. According to engineering data relayed from Anthropic, a single Agent completing a typical task consumes about 4 times the Tokens of an ordinary conversation, while a multi-Agent collaboration system reaches 15 times. The numbers given by academia are even more extreme: according to the aforementioned joint paper, in Agentic coding tasks, the Token consumption of Agent execution is about 1,000 times that of standard conversation, and even simple tools operating external systems range between 5 and 30 times.

The same paper contains another easily overlooked finding: running the same Agent and executing the same task can produce swings in the Token bill of up to 30 times. In other words, when the same user clicks the same button, the number of model calls behind it may differ by 30 times, and each call must resend the accumulated conversation history. Gartner's report in March 2026 gives a compromise estimate: Agentic workflows consume 5 to 30 times the Tokens per task of a standard chatbot, because a single user request may trigger 10 to 20 sequential model calls.

Goldman Sachs Research's May forecast is even more aggressive: by 2030, Token consumption will grow 24 times, equivalent to a monthly scale of about 1.2 billion trillion Tokens. The decline in Token unit prices has not reversed this trend. According to Gartner, inference costs in 2030 will drop by more than 90% compared with 2025, but usage growth far outpaces the price decline. Each individual call is getting cheaper, while the number of calls is increasing and each call is getting larger, so total spending rises because usage explodes.

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

2. CPU returns from supporting role to lead

In multi-agent systems, task scheduling, tool invocation, logical judgment, and context management are all things CPUs are good at. According to a research report from Shanghai Securities News, in agent workloads, the orchestration stage alone can account for 50% to 90% of total latency; when CPU performance is insufficient, even the strongest GPU will sit idle waiting for scheduling.

Demand is already showing up in prices and lead times. According to public reports, Intel confirmed it raised prices on some server CPUs, with high-end Xeon prices up 7% to 12% overall; consumer CPUs underwent three rounds of price adjustments during the year, with the Plus series cumulative increase approaching 30%; AMD is expected to raise prices on mid-to-high-end server CPUs by another 4% to 7% in the second half of the year; server CPU delivery cycles have extended from the usual 1 to 2 weeks to 8 to 12 weeks, and in some cases as long as 6 months.

There are two scale estimates. According to Arm's calculations, a traditional AI data center needs about 30 million CPU cores per gigawatt of power, and in the agent era this will rise to 120 million, a 3x increase; according to Morgan Stanley's framework, by 2030 Agentic AI will bring a $32.5 billion to $60 billion incremental CPU market, accounting for about 60% of a total server CPU market exceeding $100 billion. AMD gave a consistent direction on its Q1 earnings call: the ratio of CPU to GPU in AI infrastructure is moving from 1:4 to 1:8 toward 1:1, and in agent-intensive deployment scenarios the number of CPUs configured may even surpass GPUs.

Looking further: the elasticity in the CPU line comes from changes in ratio, not total volume growth. The server CPU market itself is growing steadily; what is changing is the number of CPUs standing beside each GPU. Moving from a 1:8 ratio to 1:1 means the same GPU spending corresponds to several times the CPU spending. This is the first time CPUs have an independent incremental logic in the AI hardware narrative. The premise of this judgment is that agent workloads do indeed scale along the current path; if the orchestration stage is significantly compressed by software optimization, the ratio curve will slow.

3. Memory: the root of scarcity is capacity substitution, not demand fluctuation

Multi-agent systems rely on long-term memory, remembering preferences, history, and task progress, and calling on them at different points in time. Technically, every internal inference must read the entire ever-growing KV cache from GPU memory, and inference is a memory-bandwidth-intensive task. In language familiar to retail investors: how fast the GPU calculates increasingly depends on how fast memory can feed it.

Micron executives have said this very bluntly in public. Executive Vice President Sumit Sadana said at the KeyBanc Technology Forum that Token consumption for Agentic AI workloads may reach 5 to 30 times that of comparable chat tasks, and even higher for deep reasoning; CEO Sanjay Mehrotra said in a CNBC interview that the training era is over, the inference era has just begun, and agentic AI has not even truly landed at scale yet.

Micron's financial statements are validating this trend. FY26Q3 revenue was $41.456 billion, up 345.7% year over year, with data center business revenue of $11.52 billion, compared with only $1.53 billion in the same period a year earlier; FY26Q4 revenue rose further to $54.229 billion (company financial statements, verified through financial data interfaces). Sadana added two points at the forum: supply and demand in 2027 will be tighter than in 2026, and DRAM customers "cannot get even half of what they want even if they pay high prices"; on the product side, HBM4 36GB 12-high can increase LLM inference throughput by 2.6 times under test conditions.

The more critical detail is the substitution ratio: for every 100 bits of HBM produced, about 300 bits of DDR output are reduced, and as HBM4E advances this may approach 4:1. This means the structural root of memory scarcity is not on the demand side, but in the physical substitution of production capacity. For every additional unit of HBM capacity expansion, traditional DRAM supply is squeezed by three to four units. As long as HBM continues to expand capacity for AI, DDR scarcity will be difficult to resolve through cooling demand, and the verification point for this supply-demand logic lies in capacity and price guidance subsequently published by Micron and other manufacturers.

4. Software panic and hardware growth come from the same product launch

In mid-January 2026, Anthropic released Claude Cowork, launched industry plugins for legal, finance, sales, and other functions on January 30, and then Claude Opus 4.6 broke large tasks into subtasks and assigned them to different agents to complete in parallel. The market reaction split in two. On the software side, according to public reports and a Wedbush research note, that selloff erased between $800 billion and $1 trillion in software stock market value (different statistics cover different stock ranges), with FactSet falling 10% in a single day, which Wedbush called "Software-mageddon." The core concern was that once agents can autonomously complete contract review and financial analysis, the per-seat SaaS model will be eroded.

The hardware side gave the opposite reading at the same time. 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 already exceeded 20%. The two sides of the same coin are this: the stronger agents' substitution capability becomes, the greater the consumption of underlying compute, CPUs, and memory. Software's share anxiety and hardware's demand growth are driven by the same product cycle. Pricing them separately as two different stories may be one of the biggest pricing disagreements in the current AI sector.

5. What Microsoft and Nvidia are doing

On October 7 local time, Microsoft unveiled the Surface Laptop Ultra in San Francisco in partnership with Nvidia: the NVIDIA RTX Spark platform, Blackwell RTX GPU with up to 6,144 cores, paired with a Grace CPU with up to 20 cores, unified memory up to 128GB, capable of running models with more than 120 billion parameters locally, priced from $2,599, shipping on October 16. Jensen Huang was present, and his exact words were widely quoted: "For forty years, you launched apps, clicked, typed; now you ask, and the computer does the work for you."

More important than the hardware is how the operating system is being changed. According to Microsoft's official blog, Windows has been redesigned around "hybrid intelligence": after authorization, Copilot can read local files, organize them on the user's behalf, troubleshoot issues, and execute workflows; on the security side, system-level containers limit the files and network ranges agents can access; tasks are scheduled between local and cloud models. Nvidia, meanwhile, released the Open Agent Safety Platform on September 28, covering the security layer for agents from testing to deployment; Jensen Huang also disclosed that order value for Blackwell and Vera Rubin systems has reached $1 trillion, covering through 2027 (company statement).

The logic of the two moves is coherent: Microsoft is rebuilding the entry point for the agent era at the operating system layer, and Nvidia is stocking up at the chip and platform layers for the scaled deployment of agents. Rather than saying they expect multi-agent collaboration to arrive, it is more accurate to say they are already building the foundation for it. Putting some inference on local devices is essentially an admission that agents' compute consumption is too large and must be handled in layers. Every PC running agents is a new always-on compute node.

6. The Token bill: from per-seat pricing to usage-based pricing

Gartner's 2026 enterprise survey gives a reading worth taking seriously: nearly a quarter of technology leaders are already paying $200 to $500 per developer per month for AI coding Tokens, and about 6% are paying more than $2,000; the firm predicts that AI coding Token costs will exceed the average developer salary by 2028. This foreshadows a shift in billing methods: in the chatbot era, budgets were made by seat count, while in the agent era budgets can only be made by number of executions, and even the models themselves cannot accurately estimate execution counts. The aforementioned joint paper clearly points out that models will continue to underestimate the number of Tokens they are about to spend.

Back to that small plugin with 238 stars from the beginning. Main Bot replies exclusively by default and calls other Bots on demand; what the developer did is essentially the most basic form of Token budget management. Scaled up to enterprise deployment, the implication becomes magnified: rigid demand for CPU, memory, and networking will only become stronger.

Final thoughts

DeepSeek-Bot's 238 stars are not important. What matters is the cost structure it demonstrates: every step of multi-agent collaboration pushes the Token bill up another notch, and even the model itself cannot predict the size of the jump. This structure presses down the industry chain toward downstream, first forcing changes in CPU ratios and lead times, then forcing capacity substitution in memory, and finally forcing an operating-system-level restructuring. Viewed along this chain, the structure of compute consumption may be moving from GPU unipolarity toward CPU, memory, and GPU coordination. On the hardware side, the demand increment is relatively easier to verify, while the software-side shock depends on the pace of agent substitution; both threads are driven by the same product cycle. The extent to which this shift has already been priced in by the market remains to be seen.

Data notes

• DeepSeek-Bot: GitHub/Trendshift, 2026/10/11.
• Token/Gartner/Goldman Sachs: Stanford-Microsoft, Anthropic, Gartner, Goldman Sachs, relayed.
• CPU/supply chain: Shanghai Securities News, Arm, AMD, Intel/AMD, public disclosures/reports.
• Morgan Stanley: research note relayed.
• Micron: financial statements, interface verification; Sadana/Mehrotra, KeyBanc/CNBC.
• Microsoft/Nvidia: Microsoft blog, GeekWire, 21st Century; Nvidia press release, Jensen Huang statement.
• Claude/SaaS: Fortune, Wedbush, public reports, with differences in scope.

Disclaimer:

This content is for general information and market commentary only. It is based on public materials as of the time stated in the article, and relevant market data, expectations, and probabilities may change with market conditions. The views and investment strategies of third-party institutions, analysts, or other persons cited in the article represent only 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 for any securities, investment products, or trading strategies, and should not be used as the basis for any investment decision. Financial markets involve risk, securities prices and market performance may fluctuate, and historical performance and past market trends do not represent or guarantee future results. Investors should independently assess relevant risks based on their own circumstances and seek professional advice when necessary. BIT's Hong Kong stock trading services are not provided to Hong Kong residents or users in other restricted jurisdictions.