Microsoft Maia 300 AI Chip

Maia 300's headline spec isn't a FLOPS number - it's memory. The chip splits its work across multiple chiplets: a main compute die manufactured on TSMC's N2P process, paired with separate memory-controller and I/O chiplets built on a more mature 3nm node ^[1]. Wrapped around that compute core are eight stacks of HBM4 memory totaling 288GB, delivering up to 18 TB/s of peak bandwidth, all inside a 950-watt power budget per accelerator ^[2]. That's a 2.5x jump in memory bandwidth and 33% more capacity than the Maia 200 that shipped earlier in 2026, even though raw compute only grew a comparatively modest 36% to roughly 13.8 petaflops at FP4 precision and 6.9 petaflops at FP8 ^[2][3].

The reason memory gets the bigger upgrade is that inference, especially the kind involved in long-context reasoning and multi-step agent workflows, is bottlenecked less by how fast a chip can multiply matrices and more by how quickly it can move a model's parameters and conversation context in and out of memory during the token-by-token 'decode' phase. Futurum Research's Brendan Burke describes Maia 300 as built from the ground up for exactly that profile: 'long-context, reasoning-intensive, decode-heavy inference' ^[4]. In other words, Microsoft isn't chasing Nvidia on a generic compute benchmark - it's optimizing for the specific shape of workload that Copilot, GitHub Copilot, and OpenAI's models actually generate in production.

The most important nuance in Microsoft's announcement is also the easiest to miss: Maia 300 is an inference chip, not a training chip. Microsoft has been explicit that it will keep using Nvidia and AMD GPUs to train its own and OpenAI's models, while shifting the work of actually running those trained models - answering a Copilot prompt, generating GitHub Copilot's code suggestions - onto its own silicon ^[5]. Maia 200 already does this today for Microsoft 365 Copilot, GitHub Copilot, Microsoft Foundry, OpenAI models, and Microsoft's own MAI models, and Maia 300 is positioned to take over that role at far greater scale, with 96 chips per liquid-cooled rack drawing about 167kW - double Maia 200's density ^[6][7].

Analyst Kieron Allen's framing cuts through the 'Nvidia killer' narrative that circulated after the reveal: 'By investing in these chips, Microsoft isn't necessarily trying to compete with NVIDIA. Rather, it aims to ensure that Azure doesn't rely entirely on NVIDIA infrastructure' ^[8]. That's a hedge, not a coup. Nvidia and AMD still get Microsoft's training dollars for both Microsoft's own models and OpenAI's.

Every performance claim attached to Maia 300 - the 40% tokens-per-megawatt improvement, the 7x rack-scale throughput gain - comes from Microsoft itself. As of the announcement, no third-party lab had independently tested the chip ^[2]. Moor Insights and Strategy's Matt Kimball captured the measured version of this skepticism, calling Maia 300 'a significant step forward in Microsoft's custom silicon strategy' while explicitly declining to predict it will slow Nvidia or AMD GPU purchases across Azure, and flagging that it still doesn't answer whether Microsoft's heavy AI infrastructure spending will keep squeezing its margins ^[4].

That caution contrasts with Wall Street's reaction: Deutsche Bank raised its Microsoft price target to $600, RBC held an Outperform rating with a $640 target, and Evercore ISI lifted its target to $635 from $528 - all citing progress on Microsoft's vertical-integration story ^[4]. The gap between enthusiastic stock-price reactions and the absence of independent benchmarks is the real story here - investors are pricing in a strategic narrative (less Nvidia dependence, more margin control) well before anyone outside Microsoft has verified the chip delivers what's claimed. That same skepticism has surfaced in public discussion of the launch, where commenters have pushed back on the 40%-tokens-per-megawatt figure as narrow or cherry-picked without a disclosed benchmark methodology - echoing the same gap analysts identified.

While headlines focused on the Nvidia angle, Maia 300's design choices are sending a louder signal to the memory industry. Packing eight HBM4 stacks into a single accelerator package, then multiplying that across a 96-chip rack, has drawn comparisons from financial commentary covering the memory sector to some of the most memory-intensive AI accelerators built to date. Industry trackers separately point to it as a direct driver of HBM4 demand - in contrast to Intel's parallel strategy of leaning on cheaper LPDDR5X memory for its own AI silicon ^[9].

That divergence matters because it splits the AI hardware industry into two camps: chipmakers chasing maximum bandwidth with premium HBM, and chipmakers optimizing for cost with commodity memory. Microsoft's bet on HBM4 at this scale - reportedly alongside talks with TSMC for over 300,000 Maia 300 units in 2027, a number that could eventually exceed one million chips deployed ^[10]- represents a demand signal that memory suppliers will be pricing into their roadmaps well before the first production rack goes live.