NVDA, TSM, MU: Nvidia custom HBM hits 30% more bandwidth while UBS flags 91% demand surge and 10% supply shortfall + Nvidia is working with Samsung and TSMC to design the base die for NVHBM, its custom HBM, per Seoul Economic Daily. Shifting the memory controller off the GPU onto the base die and shrinking the PHY 60% lifts bandwidth up to 30% and cuts power 15% versus standard HBM4E. + Samsung's own 4nm base die lines sit only at Hwaseong Line 17 and part of Pyeongtaek P3, too thin to absorb custom volumes. Industry estimates put TSMC 3nm yield above 90% against roughly 60% for Samsung Foundry 2nm SF2, so Nvidia is consolidating base die production at TSMC. + UBS projects HBM demand next year at 61.5 billion gigabits, up 91%, against shipments of only 54.6 billion gigabits, leaving a 10% shortfall. TSMC is answering with two new 3nm lines at Fab 18 in Tainan plus 3nm at its second Arizona plant and second Kumamoto plant. Samsung is hedging by validating the base die on its own 2nm process. The demand math keeps compounding. JPMorgan puts the HBM market at $64B in 2026, $160B in 2027 and $280B in 2028, with memory shortfalls near 20% in each of the next two years. Every point of that shortfall is pricing power for whoever ships qualified HBM first, which is why base die work is consolidating at the highest yield foundry.

It seems SA was too bullish on the new GW added per year in the next few years?

Latest portfolio updates and my comments on: - Elon’s latest comments on China’s DRAM supply and what they mean for my memory thesis. - Why AI demand could continue to surprise the market. - The biggest risk I see in the AI infrastructure buildout. - My latest portfolio changes, including a new position. - Why my AI GW tracker tells a different story from market consensus. New portfolio tracking, valuation, and AI analysis features on the new platfrom. Read the full article published on the paid member research platform. Thank you all!

Lisa Su just put numbers on the agentic wave: 14x growth in agentic traffic this year alone. Inference is now 60% of the AI market. And she's calling a $1.5T data center accelerator TAM by 2030. The buildout isn't slowing.

NVDA, META, GOOG — 100K+ GPUs: Meta bets on Nvidia's open networking for Vera Rubin factories + Meta is adopting Nvidia Spectrum-6 102.4Tb/s Ethernet switches plus Spectrum-X Multiplane for agentic AI factories of more than 100,000 Vera Rubin GPUs, the biggest hyperscaler commitment to Nvidia's open networking stack yet. Meta is also rolling out its own N-IceCube switch running FBOSS software. + Nvidia opened the Vera Rubin rack blueprints through the MGX 1.2 spec: racks up to 250kW, rack scale blind mate liquid cooling, cable free compute trays and a liquid cooled 5,000A busbar. An Open Data Center Spec v0.7 with AMD, Google and Meta, and a grid resilience spec that restores near full load in under 2 seconds, complete the package. + Google added Falcon Transport support to future ConnectX-10 SuperNICs, giving Nvidia two Tier-1 hyperscaler anchors for the open factory spec at once.