Microsoft’s public roadmap for AI hardware just shifted from “partner-first” pragmatism toward a clear, long-term bet on in‑house silicon — and that matters for every Windows admin, Azure customer, and datacenter architect who pays the GPU bill. At Italian Tech Week, Microsoft CTO Kevin Scott said the company wants to run “mainly Microsoft chips” in its AI data centers, while also acknowledging Microsoft will continue to use NVIDIA and AMD where they deliver the best price-performance today. That declaration crystallizes a multi-year strategy that ties chip design to racks, cooling, networking, and the models that run on them — a systems play aimed at controlling cost, latency, and supply risk as generative AI explodes demand for compute.
Background / Overview
Microsoft’s public push into custom silicon is not new — the company announced the Azure Maia AI accelerator and the Arm-based Azure Cobalt 100 CPU in 2023–2024 as part of an integrated systems strategy. Those projects were presented as components of a broader goal: to optimize the entire stack from silicon through server and rack to software and services. The recent comments from Kevin Scott and Microsoft AI CEO Mustafa Suleyman make the ambition explicit: Microsoft seeks self-sufficiency in parts of its AI stack while remaining pragmatic about using external chips when appropriate.This article unpacks what Microsoft’s intention to “run mainly Microsoft chips” really means: the technical building blocks, the competitive context, the economic rationale, the practical risks, and the immediate implications for enterprise customers and IT decision-makers. Where public details are thin—particularly numeric performance claims—those will be flagged and treated cautiously until independent benchmarks appear.
Why Microsoft is Doubling Down on First‑Party Silicon
The strategic drivers
- Control of cost at hyperscale. Owning silicon and the systems that house it changes how unit economics are amortized; at very large scale, a custom accelerator can reduce cost per useful operation compared with renting or buying third‑party GPUs.
- Latency and tight integration. Custom accelerators paired with bespoke interconnects and networking can reduce end‑to‑end latency for real‑time and inference workloads (Copilot, Teams voice, live transcription).
- Supply‑chain and concentration risk. Relying heavily on a small set of external vendors creates strategic exposure; internal silicon hedges that risk while freeing capacity to purchase vendor GPUs for other customers.
- Systems optimization. Microsoft frames Maia and Cobalt as parts of a co‑design effort: chip, rack, cooling (microfluidic cooling experiments), and software runtimes tuned for model formats and sharding strategies. This vertical integration can unlock higher rack density and efficiency if engineering execution succeeds.
Where the pragmatism shows
Kevin Scott emphasized that Microsoft is “not religious” about silicon: the company will use NVIDIA and AMD where they are currently best and will “entertain anything” to secure compute capacity during shortages. That pragmatic stance explains why Microsoft continues to buy GPUs even as it scales its Maia and Cobalt deployments. In short: it’s a hybrid trajectory, not an overnight replacement.The Hardware Stack: What Microsoft Already Has (and What’s Next)
Azure Maia 100 — Microsoft’s AI accelerator
Microsoft introduced the Azure Maia 100 as a custom AI accelerator designed for cloud AI workloads. Public materials describe the chip as a large, 5 nm monolithic device, purpose‑built for low‑precision tensor math and large model sharding at rack scale. Microsoft positions Maia as a systems product — silicon plus server board, custom racks, and software (MX formats, Triton compatibility) — with the aim of improving cost per useful token for typical inference patterns.Important caveats: Microsoft and trade press have quoted ambitious TFLOPS and architecture figures for Maia, but those vendor numbers require independent verification. Until third‑party benchmarks and reproducible test methodologies appear, treat raw vendor TFLOPS claims as indicative but not definitive.
Azure Cobalt 100 — a first‑party cloud CPU
The Azure Cobalt 100 is Microsoft’s in‑house Arm‑based cloud CPU. Microsoft announced Cobalt‑powered VM families with claims of improved price‑performance over prior Arm instances and broad region availability. Cobalt is part of the same systems story: a custom host CPU co‑designed to support Maia accelerators and to reduce dependence on commodity x86 servers for some cloud services.Cooling innovation: microfluidics and Corintis
Power density is now a gating factor for AI racks. Microsoft has publicly discussed prototype in‑chip microfluidic cooling and a partnership with Corintis to bring microchannel coolant pathways into production. Lab results reported by Microsoft suggest significant heat removal improvements, but large‑scale reliability, leak detection, and supply‑chain implications remain open questions. These cooling advances are central to Microsoft’s pitch: higher sustained TDPs per accelerator and denser racks unlock the economics of first‑party silicon.Timeline, Reality Check, and Execution Risks
Delays, talent churn, and production schedules
Custom silicon is hard. Public reporting in mid‑2025 indicated the next‑generation Maia (often referenced as Maia 200 or “Braga” internally) slipped into 2026 for mass production due to design revisions, staff turnover, and integration issues. That timeline slippage illustrates the execution risk inherent to attempting to close the raw compute gap with entrenched GPU vendors rapidly. Microsoft’s aspiration to “mainly” run its own silicon is conditional on future Maia generations meeting yield, bandwidth, and interconnect targets.Software and ecosystem friction
NVIDIA’s dominance rests not only on FLOPS but on decades of software investment: CUDA, cuDNN, tuned kernels, profiling, and a vast third‑party ecosystem. Microsoft is mitigating this with SDKs, Triton compatibility, and the MX data format to ease model portability — but porting and validating large models across new ISAs and runtimes is nontrivial. Expect incremental adoption of Maia for customers and ISVs until tooling maturity and performance-per-dollar are proven in production.Production economics and yield risk
A large, monolithic 5 nm die with high transistor counts has inherent manufacturing yield risk. If yields are poor, per‑unit costs rise and the projected total cost of ownership advantage may evaporate. Microsoft’s ability to amortize design and NRE across massive internal volume is critical — but that’s a forecast, not a guarantee.Market Impact: What This Means for NVIDIA, AMD, and Hyperscalers
Near term: GPUs remain indispensable
For high‑end training and GPU‑tuned stacks, NVIDIA and AMD GPUs still lead in raw tensor throughput and ecosystem support. Microsoft’s custom silicon will not immediately replace GPUs across Azure; instead, expect a mixed datacenter that includes GPUs, Maia accelerators, and Cobalt hosts. That balance preserves customer choice while gradually shifting internal workloads to proprietary stacks as they prove out.Strategic bargaining and vendor dynamics
Microsoft’s growing self‑sufficiency provides leverage in vendor negotiations and reduces exposure to supply scarcity. It could free up GPU capacity for paying customers while moving more internal inference onto Maia, subtly reshaping cloud procurement dynamics. That said, Microsoft remains a major GPU buyer; any vendor relations will be complex and interdependent rather than simply adversarial.The broader hyperscaler trend
Microsoft is following a pattern set by Google (TPUs) and AWS (Trainium/Inferentia). Hyperscalers are optimizing vertically to reduce per‑operation costs and gain specialized features. The likely long‑term market structure is heterogeneous: GPUs for some workloads, proprietary ASICs for others, and interoperability layers (Triton, ONNX, MX) as the path to portability. This makes cloud selection and workload portability strategic priorities for enterprises.Partnership Tensions: OpenAI and the “Right of First Refusal”
Microsoft remains OpenAI’s largest backer and strategic cloud partner, but the relationship has shown strain as OpenAI pursues multi‑cloud arrangements (Stargate) and expanded compute partnerships. Microsoft has reportedly re‑negotiated or declined additional terms around training support, and OpenAI’s own data center plans have reduced Microsoft’s exclusivity. Microsoft’s push for its own silicon and model fleet is partly a hedge against partner uncertainty — an attempt to retain product roadmaps and feature timelines even if partner dynamics shift. This is consistent with Mustafa Suleyman’s public comments about building “off‑frontier” models and becoming more self‑sufficient.Strengths of Microsoft’s Approach
- Systems engineering mindset: co‑design of silicon, racks, cooling, and runtime is the right architectural move for hyperscale optimization. When executed well, it reduces wasted headroom and unlocks price/performance improvements unavailable to general‑purpose hardware.
- Financial firepower and scale: Microsoft’s capital commitments to AI data centers (large multi‑billion dollar spends) give it the runway to iterate on silicon and amortize R&D at hyperscale. That scale is a genuine advantage compared with smaller firms.
- Pragmatic hybrid strategy: continuing to use best‑in‑class GPUs while building in‑house options reduces short‑term risk and keeps Microsoft competitive across customer needs.
Material Risks and Uncertainties
- Execution and timeline risk: Maia 200 delays show how quickly planned advantages can slip. Missed schedules allow competitors to extend performance leads or capture customer mindshare.
- Ecosystem and portability friction: Without mature tooling and widespread validation, customers and partners may be slow to migrate. The software handoff (Triton, MX format) must prove frictionless.
- Yield and cost risk: manufacturing large, dense dies at scale is expensive and fragile; poor yields or higher-than-expected per‑unit costs would weaken the TCO case.
- Operational risk for microfluidics: novel cooling introduces new failure modes and maintenance paradigms. Prototypes look promising in labs; fleet‑level reliability remains unproven.
- Competitive response: NVIDIA’s extraordinary valuation and cash flow (the company briefly reached the $4 trillion market‑cap milestone) mean the market leader can accelerate product cycles, software investment, and pricing responses — making the gap harder to close. Microsoft must out‑engineer and out‑scale incumbents to meet its stated goal.
What Enterprises and IT Leaders Should Do Now
- Reassess cloud procurement assumptions. Expect heterogeneous compute offerings from cloud providers and measure TCO by useful work (latency, throughput, cost per inference), not raw FLOPS.
- Invest in portability now. Standardize on interoperable runtimes and model packaging (Triton, ONNX, containerized runtimes) so workloads can be moved between GPUs and proprietary accelerators with less friction.
- Pilot latency‑sensitive services on in‑house accelerators as they become available. Services with millisecond SLAs (voice, live transcription) are the likeliest early winners for Maia‑class hosts.
- Require transparent benchmarks and SLAs. Negotiate capacity and performance guarantees where possible; treat vendor performance figures as marketing until validated by independent benchmarks.
- Monitor partner model routing and governance. Multi‑model stacks will be the norm; policies that decide what model processes regulated or sensitive data will be essential for compliance.
Verifiable Claims vs. Unverified Numbers — A Clearer View
- Verifiable: Kevin Scott publicly stated Microsoft aims to run mainly Microsoft chips in the future, while continuing to use NVIDIA and AMD where appropriate. That comment was reported in CNBC.
- Verifiable: Microsoft announced Azure Maia 100 and Azure Cobalt 100 and has published blog posts describing the co‑design approach and target workloads. These product announcements and blog posts are public.
- Verifiable: Reporting in mid‑2025 documented delays to the next‑generation Maia ramp into 2026; this delay was covered by Reuters and other outlets. Execution timelines have shifted.
- Unverified / vendor‑provided: Specific TFLOPS figures, single‑GPU throughput claims, and per‑token cost targets for Maia/MAI models should be treated cautiously until independent benchmarks and reproducible test methodologies are published. Microsoft and press reporting include many performance claims that are persuasive but not yet peer‑verified.
The Competitive Landscape and What Comes Next
Microsoft’s move is part of a larger industry dynamic: Google (TPUs), AWS (Trainium/Inferentia), and hyperscalers generally are leaning into custom silicon and systems engineering to extract edge‑case efficiencies for AI workloads. That trend reduces the near‑term fungibility of cloud compute and raises the importance of interoperability layers and vendor‑agnostic tooling.If Microsoft can execute on Maia/Cobalt and pairing innovations such as microfluidic cooling, it will gain persistent operational advantages for certain classes of workloads. If it misses timelines, vendor ecosystems sharpened by NVIDIA and AMD will maintain their lead, and Microsoft’s multi‑vendor approach will remain necessary for many customers. Either way, the era of treating the GPU as the only relevant cloud commodity is over; architecture decisions now include ASICs, DPUs, custom Arm hosts, and thermal systems as first‑order considerations.
Conclusion
Kevin Scott’s blunt statement — that Microsoft “wants to run mainly Microsoft chips” — is not a marketing flourish. It is the public articulation of a deliberate, systems‑level strategy: design silicon for the company’s most important workloads, pair it with purpose‑built racks and cooling, and capture the operational benefits of vertical integration. The move is strategically rational given Microsoft’s scale, capital, and product footprint, but it is not without real engineering and business risks: manufacturing yield, software ecosystem maturity, timeline slippage, and operational unknowns around novel cooling solutions.For enterprises and IT teams, the practical takeaway is to prepare for a heterogeneous compute world: optimize for portability, demand transparent performance claims, and pilot the new accelerators where latency and TCO matter most. Microsoft’s aspiration to reduce dependence on NVIDIA and AMD will remap vendor relationships and datacenter economics over years — not months — but the direction is now unmistakable. The race for compute at scale has become a race to own more of the full stack.
Source: Windows Central Microsoft’s AI future may not need NVIDIA or AMD anymore
