NVIDIA and Microsoft Respond to Surge in AI Hardware Demand Triggered by DeepSeek-Bot

Introduction: DeepSeek-Bot, launched on October 7, is merely a small plugin enabling multiple AI bots to collaborate within a group chat—gaining 238 stars in four days, a feat too modest to warrant an in-depth article. What makes it worth writing about is something else: it has tapped into a trend reshaping AI hardware pricing. Multi-agent collaboration causes a orders-of-magnitude surge in token consumption, and the resulting bill ultimately lands on CPUs and memory. This article traces a thread connecting this small plugin to token multipliers, CPU shortages, memory swapping, and the moves of two industry giants.

First, consider the scale. According to Anthropic’s engineering data, a single agent consumes approximately four times the tokens of a typical conversation to complete a standard task, while multi-agent collaborative systems reach up to 15 times. Academic figures are even more extreme: according to the aforementioned joint paper, token consumption by agents in agentic programming tasks is about 1,000 times that of standard conversations, and even simple tools for interacting with external systems range from 5 to 30 times.

Another, more easily overlooked finding in the same paper: when running the same agent to perform the same task, the token bill can vary by up to 30 times. This means that when the same user clicks the same button, the number of underlying model calls may differ by a factor of 30, with each call requiring the re-sending of the accumulated conversation history. According to Gartner’s March 2026 report, a conservative estimate suggests that agentic workflows consume 5 to 30 times more tokens per task than standard chatbots, as a single user request may trigger 10 to 20 sequential model calls.

Goldman Sachs Research’s May forecast was even more aggressive: by 2030, token consumption is expected to grow 24-fold, equating to approximately 12 quintillion tokens per month. The decline in token price has not reversed this trend; according to Gartner, inference costs in 2030 are projected to drop more than 90% compared to 2025, but usage growth far outpaces the price decline. Each individual invocation is becoming cheaper, while the number of invocations and their size per call are increasing—total spending is rising due to explosive usage growth.

By integrating usage data and hardware metrics, the primary tension in token economics has shifted from unit price to the product of invocation frequency and per-invocation consumption. This has implications for hardware: the bottleneck for GPU utilization is no longer the speed of individual forward computations, but rather memory bandwidth and scheduling—this is precisely why CPUs and memory are re-entering the pricing equation.

In multi-agent systems, task scheduling, tool invocation, logical reasoning, and context management are all tasks that CPUs excel at. According to a research report by Shanghai Securities News, the orchestration phase alone can account for 50% to 90% of total latency in agent workloads; when CPU performance is insufficient, even the most powerful GPU will remain idle while waiting for scheduling.

Demand has already manifested in price increases and extended lead times. According to public reports, Intel has confirmed price hikes for select server CPUs, with overall increases of 7% to 12% for high-end Xeon processors; consumer-grade CPUs have undergone three rounds of price adjustments this year, with the Plus series累计涨幅接近30%. AMD is expected to raise prices for mid- to high-end server CPUs by an additional 4% to 7% in the second half of the year. Lead times for server CPUs have extended from the usual 1 to 2 weeks to 8 to 12 weeks, with some cases reaching up to six months.

Two estimates have been made regarding scale. According to Arm’s analysis, traditional AI data centers require approximately 30 million CPU cores per gigawatt of power, a figure expected to rise to 120 million cores in the agent era—an increase of threefold. According to Morgan Stanley’s framework, agentic AI will generate an additional CPU market of $32.5 billion to $60 billion by 2030, accounting for approximately 60% of the over $100 billion total server CPU market. AMD’s guidance during its first-quarter earnings call aligns with this trend: the ratio of CPUs to GPUs in AI infrastructure is shifting from 1:4 to 1:8 toward a 1:1 balance, and in agent-dense deployment scenarios, the number of CPUs may surpass that of GPUs.

Looking further: The elasticity in the CPU segment stems from shifts in ratio, not overall growth. The server CPU market itself is growing steadily; the change lies in the number of CPUs accompanying each GPU. A shift from a 1:8 ratio to a 1:1 ratio means that the same GPU expenditure now corresponds to several times the CPU expenditure—this is the first time CPU has developed its own independent growth narrative within the AI hardware story. This assessment assumes that agent workloads continue to scale along the current trajectory; if orchestration is significantly compressed by software optimizations, the ratio curve will slow down.

Multi-agent systems rely on long-term memory to retain preferences, historical data, and task progress, which are accessed at different points in time. Technically, each internal reasoning step requires reading the entire growing KV cache from GPU memory, making inference a memory-bandwidth-intensive task. In simpler terms: how fast the GPU performs is increasingly determined by how quickly its memory can supply data.

Micron’s executives have been straightforward about this in public statements. Sumit Sadana, Executive Vice President, said at the KeyBanc Technology Forum that token consumption for Agentic AI workloads could reach five to 30 times that of comparable chat tasks, with even higher demands for deep reasoning; CEO Sanjay Mehrotra, in an interview with CNBC, stated that the training era has passed and the inference era has just begun, with agent-based AI yet to be truly deployed at scale.

Micron’s financial report validates this trend. In FY26Q3, revenue reached $41.456 billion, a 345.7% year-over-year increase; data center revenue was $11.52 billion, compared to just $1.53 billion in the same period last year. FY26Q4 revenue further rose to $54.229 billion (company financial report, verified via financial data interface). Sadana added two points at the forum: the supply-demand ratio in 2027 will be tighter than in 2026, with DRAM customers unable to secure more than half of their desired volume even at premium prices; on the product side, HBM4 36GB 12-high can increase LLM inference throughput by 2.6x under test conditions.

More critically, the replacement ratio is key: for every 100 bits of HBM produced, approximately 300 bits of DDR output are displaced—a ratio that may approach 4:1 as HBM4E advances. This means the structural root of memory shortages lies not in demand, but in physical production substitution: each incremental increase in HBM production reduces traditional DRAM supply by three to four times. As long as HBM continues to expand for AI, DDR shortages cannot be resolved by cooling demand alone. The validation of this supply-demand dynamic will be evident in future capacity and pricing guidance from manufacturers such as Micron.

In mid-January 2026, Anthropic released Claude Cowork, and on January 30, it launched industry-specific plugins tailored for legal, finance, sales, and other functions. Shortly after, Claude Opus 4.6 began breaking down complex tasks into subtasks and assigning them to multiple agents for parallel execution. Market reactions were divided. On the software side, according to public reports and Wedbush research, the sell-off erased between $800 billion and $1 trillion in software market capitalization (with variations depending on the scope of stocks included); FactSet dropped 10% in a single day, which Wedbush termed "Software-mageddon." The core concern was that once agents could autonomously handle contract reviews and financial analyses, the per-user SaaS pricing model would be eroded.

The hardware side is providing contradictory readings simultaneously. Morgan Stanley noted that as AI shifts from generative to agent-based models, investment focus will expand from GPUs to CPUs and memory; the weight of semiconductors in the S&P 500 has surpassed 20%. Two sides of the same coin: the stronger an agent’s substitution capability, the greater the demand for underlying compute power, CPUs, and memory. The software’s share anxiety and the hardware’s incremental demand are driven by the same product cycle—treating them as two separate stories for pricing may be one of the largest pricing divergences in the current AI sector.

On October 7 local time, Microsoft unveiled the Surface Laptop Ultra with NVIDIA in San Francisco: the NVIDIA RTX Spark platform, featuring up to 6,144 cores on the Blackwell RTX GPU and up to 20 cores on the Grace CPU, with up to 128 GB of unified memory capable of running models with over 120 billion parameters locally. Priced from $2,599, it ships on October 16. Jensen Huang attended the event, and his exact words were widely quoted: “For forty years, you launched applications, clicked, and typed; now you ask questions, and your computer does the work.”

More important than hardware is the redesign of the operating system. According to Microsoft’s official blog, Windows has been rearchitected around “hybrid intelligence”: with proper authorization, Copilot can access local files, organize them, troubleshoot issues, and execute workflows; security is enforced through system-level containers that limit agents’ access to files and network resources; tasks are dynamically scheduled between local and cloud-based models. NVIDIA, meanwhile, launched the Open Agent Safety Platform on September 28, covering security layers from testing to deployment. Jensen Huang also revealed that orders for the Blackwell and Vera Rubin systems have reached $1 trillion, extending through 2027 (according to company figures).

The two actions are logically connected: Microsoft is rebuilding the entry point at the operating system level for the agent era, while NVIDIA is stockpiling at the chip and platform level for scalable agent deployment. Rather than merely anticipating the arrival of multi-agent collaboration, they are already building its foundation. Offloading some reasoning to local devices essentially acknowledges that agents consume excessive computational power and must be processed in layers—each PC running an agent becomes a new persistent computational node.

A Gartner 2026 enterprise survey presents a compelling figure: nearly a quarter of technology leaders are already paying $200 to $500 per month per developer for AI coding tokens, with about 6% paying over $2,000; the firm predicts that AI coding token costs will exceed the average developer salary by 2028. This signals a shift in billing models: while the chatbot era relied on seat-based budgeting, the agent era requires budgeting by execution count—a metric even the models themselves cannot accurately predict. The aforementioned joint paper explicitly notes that models consistently underestimate the number of tokens they will consume.

Back to the small plugin with 238 stars: the Main Bot defaults to exclusive replies and calls other bots on demand—what the developer essentially did was the most basic form of token budgeting. When scaled to enterprise deployment, this micro-level action becomes significantly more impactful: the rigid demands on CPU, memory, and network resources will only intensify.

The 238 stars on DeepSeek-Bot are not important; what matters is the cost structure it demonstrates: with each collaborative step among multi-agents, the token bill jumps a tier, and the magnitude of these jumps is unpredictable even for the model itself. This structure pushes downstream along the supply chain, first altering CPU ratios and lead times, then driving memory capacity substitution, and ultimately triggering OS-level rearchitecting. From this chain, it appears that compute consumption is shifting from a GPU-centric model toward a coordinated system of CPU, memory, and GPU—hardware demand increments are relatively verifiable, while software impacts depend on the pace of agent substitution; both trends are driven by the same product cycle. To what extent this shift has already been priced in by the market remains to be seen.

• DeepSeek-Bot: GitHub/Trendshift, 10/11/2026.
• Token/Gartner/Goldman Sachs: Stanford-Microsoft, Anthropic, Gartner, Goldman Sachs, paraphrased.
• CPU/Supply Chain: Shanghai Securities News, Arm, AMD, Intel/AMD, publicly disclosed/reported.
• Morgan Stanley: Report summary.
• Micron: Earnings report, interface verification; Sadana/Mehrotra, KeyBanc/CNBC.
• Microsoft/NVIDIA: Microsoft blog, GeekWire, 21st Century; NVIDIA press release, Jensen Huang's statements.
• Claude/SaaS: Fortune, Wedbush, publicly reported, differing statements.

Disclaimer: The content of this article is provided solely for general information and market commentary, based on publicly available data as of the time mentioned herein. Market data, expectations, and probabilities referenced herein may change due to market conditions. Opinions and investment strategies cited from third-party institutions, analysts, or other individuals reflect only their views at a specific point in time and do not represent BIT’s views or recommendations. 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 risks; security prices and market performance may fluctuate, and historical performance or past market trends do not indicate 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 available to users residing in Hong Kong or other restricted jurisdictions.