What Speirs actually said
Microsoft's reasoning is that agents do much of their work outside the model, on general-purpose CPUs. Speirs said agent workloads act autonomously and produce outputs. He told the paper they therefore need far more infrastructure, computing power and processing capacity than chatbots.
He said that in developer workloads there are cases that use up to 3,000 times the compute of existing generative Copilot or chatbots. He added that this capacity can't simply be built on demand. In his words, it would be several times the size of existing data centers, and out of reach even if every high-performance chip in the world were used.
Read that number carefully:
- It is tied to developer workloads, and the wording is "up to", so it describes a peak case.
- The interview gives no baseline, no measurement method and no unit of comparison. We don't know if it is per task, per user or per system.
- No independent validation has been published.
- The English text is AI-translated from Korean. The paper warns that the quotes may not match the original wording exactly.
So 3,000x is not a rule that applies to every agent or every prompt. It is one Microsoft executive's illustration of how large agent workloads can get. An agent that plans, runs code, searches files, calls tools and retries loops can plausibly burn far more compute than a single chat reply. The size of the multiplier is Microsoft's claim alone.
Why CPUs are part of the story
The Seoul Economic Daily framed this as a shift in AI hardware demand. GPUs dominated the generative AI era because of model training and parallel computation. In the agent era, CPUs matter more because agents spend time orchestrating, executing code and handling tool calls. The paper also cites forecasts that the CPU-to-accelerator ratio could move from roughly 4:1 or 8:1 toward 2:1 or 1:1. That is a forecast reported by the paper, not a measurement.
Arm made a similar argument in its launch announcement. Cobalt 200 demonstrates the impact of purpose-built Arm-based infrastructure. Arm is a design partner on the chip, so treat that as corroboration of the logic rather than independent proof.
The product response: Cobalt 200 VMs
At Microsoft Ignite 2026, Microsoft announced the general availability of Azure Cobalt 200-based virtual machines. Microsoft says they deliver up to 50% better performance than Cobalt 100 and are available in 14 regions, expanding to 23 by the end of 2026. Microsoft's own materials aren't fully consistent on the regional target. Coverage of the Book of News notes a figure of 24, so treat "23 to 24" as the honest range until Microsoft reconciles them.
Key details from Microsoft's Azure announcement:
- Silicon: The chip is built on TSMC's 3nm process and has 132 Arm Neoverse V3 cores. Each vCPU maps to a full physical core, with 3 MB of private L2 cache and 192 MB of system-level cache. It has 12 channels of DDR5, and memory encryption is on by default.
- Offload and security: Azure Boost is included in every VM. Microsoft claims up to 15% more network bandwidth, 20% more remote-storage IOPS and 10% more remote-storage throughput than Cobalt 100. Each server also includes an Azure Integrated HSM.
- Processor modes: Deterministic Mode runs cores at a steady 3.2 GHz. Opportunistic Mode can boost to 3.6 GHz when power and thermal headroom allow, for up to 12% more performance with some frequency variability. You pick between them with the ProcessorMode setting at creation or later, through the Azure portal, Azure CLI, PowerShell, or ARM and Bicep templates.
- VM families: There are five VM series. Two new families are the high-memory Mpsv4, with 16 GiB per vCPU, and the storage-optimized Lpsv5, with up to 23 TB of local NVMe per VM.
- Launch regions: Australia East, Canada Central, Central US, East US, East US 2, Indonesia Central, Italy North, Korea Central, South Central US, Spain Central, Sweden Central, UK South, West US 2 and West US 3.
Vendor benchmarks, and a warning against mixing numbers
Microsoft reports its own benchmark gains over Cobalt 100: up to 1.5x for web serving, 1.6x for Java, 1.5x for AI-agent tool execution and 1.9x for databases. These are vendor figures, not independent tests.
Customer examples are attributed and don't generalize. Amadeus reported flight-search throughput up more than 30% in its benchmark lab. OpenAI said it saw higher throughput and lower latency on the same ChatGPT workloads. Microsoft says Teams video-media processing reached 1.65x.
The 1.5x agent tool-execution figure and the 3,000x claim are unrelated. One compares two Cobalt generations on a benchmark. The other compares how much compute some developer workloads consume against chatbots. A faster CPU doesn't close a 3,000x gap. It only makes each unit of CPU work cheaper in time and power.
What it means for Windows and IT admins
- Arm compatibility comes first. Cobalt is Arm-based, so confirm your workloads and container images have Arm builds before migrating. Microsoft says workloads already on Cobalt 100 can move to Cobalt 200 where they need more performance, larger sizes or the new families.
- Supported platforms. Microsoft says the VMs work with popular Linux distributions, AKS and common container frameworks. They can be deployed from the portal, SDKs, APIs, PowerShell and the Azure CLI, in Virtual Machine Scale Sets and as AKS node pools. Check the deployment documentation for the exact OS support list for your scenario, including Windows guests.
- Choose the processor mode deliberately. Microsoft suggests Deterministic Mode for distributed analytics clusters and services with tight performance targets. It suggests Opportunistic Mode for web and API services, transaction processing and media workloads.
- Plan for capacity. Speirs' point is that compute supply is a real constraint. Teams building agent-heavy developer tooling should expect agent workloads to be costly and test them under realistic loads rather than assuming chatbot-era sizing.
- Maia 300 is a separate announcement. Microsoft's Ignite keynote said Maia 300 will deliver 40% more tokens per watt than leading GPU-based systems. That is a claim about the future, not a statement that it is generally available.
Analysis
Microsoft has a commercial interest in a big compute-demand number. It sells infrastructure, and a scarcity story supports its custom-silicon strategy. That doesn't make the claim wrong, but it does mean the number needs context that the interview doesn't provide. The more defensible takeaway is that agent workloads add CPU-heavy steps such as sandboxes, code compilation and tool calls, and Microsoft is building general-purpose Arm capacity to handle them. Cobalt 200 is a real, shipping product with documented specifications. The 3,000x ratio is a quote.
References
- Azure Cobalt 200 VMs: Faster compute for modern agentic workloads azure.microsoft.com
- Microsoft Says AI Agents Can Use 3,000 Times More Computing Than Chatbots - Seoul Economic Daily Seoul Economic Daily · 2026-11-18T10:19:47+00:00
- Microsoft Azure Cobalt 200 VMs reach general availability, bringing Arm-powered performance to cloud and agentic AI workloads - Arm Newsroom newsroom.arm.com