Today we are announcing our biggest bet yet on our own silicon. Starting next calendar year, we will scale our Maia family to multi-gigawatt deployments, making it the primary inference system for our own product and services. And we’re introducing Maia 300, which delivers 40% more tokens per watt compared to leading GPU-based systems. We are already deploying Maia 200 across the US, Europe, and Australia, delivering up to 1.7x more tokens per watt across experiences in Teams and Copilot. Our goal is to build on the full-stack efficiency gains we’ve already seen, while continuing to bring the latest innovations from across the ecosystem into our fleet. Read more: https://lnkd.in/g68XQtxE
I’d love to be in the room for Microsoft’s next GPU pricing negotiation when NVIDIA knows Maia is already running Copilot workloads 😅 Having a credible alternative at that scale might be worth almost as much as the efficiency gains themselves.
Spot on, Satya Nadella! Transitioning toward custom silicon and multi-gigawatt infrastructure deployments represents a fundamental milestone in redefining enterprise AI economics and operational scalability. Maximizing tokens per watt through proprietary hardware architecture is critical for mitigating latency and curbing the exponential energy demands of large-scale agentic AI. Integrating full-stack silicon efficiency directly into cloud services sets a new benchmark for sustainable enterprise IT. As hyperscalers aggressively pivot toward custom silicon to optimize inference economics, what core engineering or supply-chain challenge do you view as the primary bottleneck for scaling multi-gigawatt deployments over the next cycle?
Chajri Mouad
1w
Incredible milestone, mr.Satya! As custom silicon scales to multi-gigawatt deployments with Maia 300, optimizing tokens-per-watt highlights a critical industry truth: physical thermal management at the hardware level is becoming just as vital as architectural efficiency. We’ve been tackling this exact frontier through advanced passive thermal architectures designed for high-density AI hardware. If you or your silicon team are ever open to a brief exchange on bypassing the thermal wall at scale, I'd love to share a few notes. Exciting times ahead for full-stack optimization!"
Tokens per watt is the number I watch from Manila, where available power decides where AI capacity can go. A 40% gain stretches the same megawatts further. In markets like ours, that makes it easier to run inference in-country, close to the users and their data, and every in-country deployment pulls demand into local data center space and domestic fiber.
Phillip Pham
1w
Werden die Effizienzgewinne von Maia zeitnah auch in günstigere PTU-Reservierungen für Enterprise-Kunden weitergegeben, oder bleibt die Hardware primär für Microsofts eigene SaaS-Dienste reserviert? Gerade bei fest gebuchten Kapazitäten entscheidet am Ende der Preis pro garantiertes Token über den tatsächlichen Nutzen im produktiven Betrieb.
James Kass
1w
Making Maia the primary inference system is the bigger signal here. At 167 kW per rack, how is Microsoft approaching the power, cooling, and retrofit requirements needed to turn silicon efficiency into facility-level efficiency?
The question that sits underneath the announcement is not about efficiency. You are about to run multi-gigawatt inference for agents that act — that buy, sign, dispatch, approve, execute — across Teams, Copilot, and the Azure estate. Every one of those actions will commit against a real target at some point. The tokens-per-watt number tells us how cheap the compute is. It does not tell us what stands between an agent's decision and the consequence when the agent has valid credentials, valid authority, and intent to exceed scope. The industry has spent two years building the layer above the action: monitoring, containment, audit, oversight. Every incident in that window had the same shape. The credentials were valid. The policy was configured. The action fired anyway. Whoever solves that layer first owns the next decade of enterprise AI. It is not the model layer, it is not the silicon layer, and it is not the SaaS layer. It is the layer between the decision and the consequence. You have the biggest bet on the silicon. That is the correct place to be building. But the layer that makes agents safe to run at scale is not on the roadmap you just published.
I'm assuming this is good news. Anyone else consider themselves "tech savvy" and dont understand much of what they just read? It seems great though! Good job Microsoft: more power is always good:)
Shannon A.
1w
Maia A. Crockett hey it’s u
40% more tokens per watt, but at what context length? The workloads you named, sustained multi-step agents, carry long contexts, and every agent's working memory grows with each step. At that point the binding constraint is memory, not compute, and the specs that lead this launch are memory specs. Tokens per watt measured on short contexts says little about tokens per watt once agent memory dominates. The bigger announcement may be co-design. With Maia as the primary engine, Copilot's models get shaped around Maia and each next Maia around Copilot's workloads. That's where the gain compounds, and it's very hard to copy from outside. It also splits the market: integrated model-plus-chip stacks on one side, open models on general-purpose GPUs on the other. As agents multiply token demand, that gap becomes the price difference between closed and open AI. Who gives open models what Maia gives Copilot?