ASSETCASTING
CompanyEvent 2026.10.12

Verifying the Maia 300 HBM spec: the basis for 288GB and open questions

This case checks the October (month 10), 2026 Maia 300 reveal — 288GB HBM4 per chip, 96 per rack, 2027 deployment — against official and press materials, with the Maia 200 record of sole SK hynix HBM3E supply and the current undisclosed-supplier status.

Updated Oct 13, 2026

What does Maia 300 add to HBM demand?
Verified facts

The announcement: 288GB per chip, 96 per rack, 2027 deployment

Microsoft's official blog says multi-gigawatt Maia 300 deployments start in 2027, making Maia the primary inference system for its products and services. Reports describe the 3rd-gen chip on TSMC N2P with 8 HBM4 stacks per chip for 288GB and 18TB/s within 950W, 96 per rack, claiming up to 7x the rack-scale token throughput of Maia 200 and 40% more tokens per megawatt than leading GPU systems. Reuters reported more than 100 ten-thousand chips deployed next year, with compute up to 13.8 FP4 and 6.9 FP8 peta-ops versus the Maia 200 (216GB HBM3E, 7TB/s, 750W).

Sources [1][2][3]
Verified facts

Versus the predecessor: 216GB to 288GB, supplied by SK hynix

Maia 200 paired HBM3E 216GB with 7TB/s, while Maia 300 moves to HBM4 288GB with 18TB/s — about 33% more capacity and 2.5x bandwidth. Maia 200 was unveiled on January 26 (month 1), 2026 with SK hynix sole-sourcing 6 12-layer HBM3E stacks. No checked official or press source names the Maia 300 HBM supplier.

Sources [2][4]
Interpretation

What the new customer's arrival means, and open questions

96 per rack times 288GB per chip is about 27,648GB (about 27.6TB), matching the video's figure, so per-rack HBM totals grow even as compute shifts off Nvidia. But the 7x and 40% are vendor claims without independent verification, as reports note, and the 100 ten-thousand deployment is a Reuters-reported plan figure. With the HBM supplier undisclosed, neither SK hynix retention nor Micron or Samsung entry can be concluded.

Sources [2][3]
Limitations

Limits of this case

Spec figures rest on the official blog and press quotations; no independent benchmark was checked. CXMT-related claims belong to the video and a separate case and are not re-verified here, and supplier and shipment figures remain plans and outlooks.

Sources

  1. Microsoft to expand Azure Maia system to multi-gigawatt deploymentsAccessed Oct 13, 2026
  2. Microsoft Maia 300 AI chip arrives with monstrous 288 GB HBM4 memory and 18 TB/s bandwidthAccessed Oct 13, 2026
  3. Microsoft Unveils Maia 300 AI ChipAccessed Oct 13, 2026
  4. HBM4 production nears as Samsung, SK hynix vie for AI chip dominanceAccessed Oct 13, 2026

Related learning notes

As China's CXMT Shakes the Market, MS Packs HBM Into Its Next-Gen ChipIt checks the Maia 300 specs, deployment plan, and prior supply-chain record behind the video's HBM demand message against official and press materials — including the open question of the undisclosed supplier.Microsoft unveils its Maia 300 AI chip to cut its reliance on Nvidia | The 8:30The custom-chip discussion connects to the existing Maia specification review. Separate vendor efficiency claims, deployment plans and the undisclosed supplier. This does not verify the video's later AI safety reporting.[Hong Jang-won's Bull and Bear] Semiconductor Stocks Need New DopamineThe existing Maia case makes the video's custom-chip and HBM discussion concrete. Distinguish conventional DRAM competition from HBM demand, while recognising that an announced design does not guarantee a particular supplier's revenue.