Microsoft Maia 300 First to Adopt TSMC 2nm, Shipments May Fall Below One Million
nashnova research
Microsoft's Maia 300 is set to become the first AI chip built on TSMC's 2nm process, with mass production planned for 2027 — yet supply-chain estimates put annual shipments at under one million units, exposing the gap between process leadership and volume competitiveness across the custom-silicon camp.
Microsoft gets 2nm first — so why are volumes so low?
Maia 300 is expected to enter mass production in 2027, ahead of comparable chips from Nvidia, AMD, and Google on the same TSMC node.
Yet supply-chain estimates peg annual shipments at under one million units; the previous Maia 200 shipped only about 300,000 units in total.
This means → Microsoft earns the "first mover" technology label, but its actual share of the AI accelerator market remains tiny. Process leadership ≠ order-book leadership.
Why is TSMC handing a $2 billion packaging deal to GlobalFoundries?
TSMC and GlobalFoundries signed a five-year, ~$2 billion agreement under which GlobalFoundries will produce silicon interposers — connector boards that link multiple chips inside a package — and embedded deep-trench capacitors (eDTC, tiny energy-storage components built into silicon) needed for advanced packaging.
TSMC previously made these components in-house or outsourced them to VIS, a company it holds a stake in. Shifting the work to a U.S.-based foundry marks a clear strategic pivot.
In plain terms = TSMC's big U.S. investments target leading-edge process and packaging lines, not mature-node wafer fabs — but packaging technologies like CoWoS still need mature-node components. Rather than build another mature fab in the U.S., TSMC taps GlobalFoundries' existing local capacity; if Washington imposes tariffs on Asian-made semiconductors, domestic production also hedges that risk.
Why has Google's next-gen TPU slipped?
Humufish, Google's next-generation TPU co-developed with MediaTek, was originally slated for late-2027 mass production. Current estimates push that to Q1 2028; more conservative supply-chain forecasts place initial low-volume production in Q2 2028, with ramp-up only in the second half.
The delay stems not from wafer fabrication or packaging issues but from Google's repeated revisions to the compute-chip design.
This reflects a shift in Google's chip-development model. Under the earlier Broadcom partnership, Broadcom handled most of the design, physical implementation, and foundry coordination; Google mainly set specifications. With MediaTek, Google has taken on more design responsibility — gaining autonomy but also absorbing the schedule risk that comes with frequent design changes.
Will MediaTek suffer from Google's delay?
Google has already increased orders for the previous-generation TPU V8t and extended its shipping lifecycle to compensate MediaTek for R&D costs and the Humufish slip.
This means → MediaTek's 2027 TPU-related shipments to Google may actually rise — Google is effectively paying for the delay by buying more of the older product.
EMIB-T yields sit below 40% — can this packaging route survive?
Humufish will use Intel's EMIB-T advanced packaging technology — embedding silicon bridges directly into the IC substrate rather than relying on a full silicon interposer — with both compute and I/O chips fabricated by TSMC.
The leading substrate supplier, Japan's Ibiden, currently achieves EMIB-T yields below 40%; Taiwanese suppliers may be even lower.
In plain terms = Humufish is EMIB-T's first large-scale commercial deployment. Whether yields can climb to an economically viable level before mass production begins is the make-or-break test for this entire packaging approach.
市场有风险，内容仅供研究参考，不构成投资建议。