Microsoft Now Has 3 Generations of Its Own AI Chips. Nvidia Is Still the Chip Stock to Buy.

Microsoft announced Maia 300 and plans to ramp up its custom chips in 2027, aiming to make Maia the primary inference system for its own products; Microsoft estimates 40% more tokens per megawatt than top GPU-based systems. Nvidia's revenue rose 106% year over year to $96.2 billion in fiscal Q2 2027, hyperscaler spending on its chips more than doubled to $48.7 billion, and its fiscal Q3 revenue guidance is about $108 billion. The article argues Maia poses a limited near-term threat, noting Nvidia also expects roughly 70% fiscal 2028 revenue growth and trades at about 14 times expected fiscal 2028 earnings.

Analysis

Inference is a mix shift, not yet a demand signal. Microsoft’s custom silicon could take share in predictable, internally controlled inference workloads while leaving Nvidia better positioned for training, rapidly changing models, and customers that value portability. If cheaper inference expands token usage, lower cost per token may partly offset lost accelerator share; the key variable is Nvidia’s absolute shipments, not custom-chip announcements. Microsoft’s efficiency claim is unverified until production deployment and workload-level results arrive.

Horizon and risks: Over the next 1–3 months, Nvidia guidance and hyperscaler capex are the more actionable checks on demand than Maia milestones. Over 6–18 months, watch whether Maia reaches meaningful production scale, whether Microsoft shifts workloads rather than merely adding capacity, and whether other large buyers accelerate proprietary silicon. Nvidia’s customer concentration makes a change in one buyer’s purchasing consequential; continued growth among non-hyperscale customers offers a potential offset. The cited low forward multiple is not a sufficient margin of safety if earnings estimates depend on constrained supply and sustained growth.

Contrarian read: Custom chips could be bullish for total AI compute by improving economics and expanding inference, even as they pressure Nvidia’s share. Conversely, strong Nvidia revenue can conceal future mix risk because existing deployments and supply constraints delay the visible impact of substitution. No basis yet to short Microsoft: successful internal silicon could improve its product economics even if it changes supplier mix.

Key Decisions for Investors

• Consider a modest, staged long in NVDA rather than chasing strength; the thesis is continued absolute demand growth, not immunity from custom-chip share loss. Reassess after the next results and guidance. Falsify if hyperscaler purchases weaken alongside slowing non-hyperscaler data-center demand, or if earnings expectations are cut materially.
• Do not short MSFT on the Maia announcement alone. Treat production-scale deployment, independently demonstrated tokens per megawatt, and evidence of workload migration from third-party GPUs as prerequisites for a stronger displacement thesis.
• Set an alert for the next 1–3 months on Nvidia guidance, customer concentration disclosures, and hyperscaler capex; over 6–18 months track Maia deployment scale and the split between inference workloads on proprietary chips versus Nvidia systems. Those data are missing and matter more than announced chip specifications.
• Avoid treating the forward earnings multiple as proof NVDA is cheap: stress-test the position against slower growth, supply normalization, and customer insourcing. If those risks worsen while the stock rerates upward, trim rather than relying on the historical growth trajectory.