Human-written opinions, AI deep dives, single-ticker research, and factor notes — all the analysis output in one place.
The same week server CPU demand was validated again, a new challenger launched specifically to attack the unit economics of the same workload — surfacing the bifurcation point of the inference economy
Korean exports printed +154% YoY in May and an Nvidia HBM-cut headline hit the tape; in Taipei a record NT$177.4B foreign sell coexisted with Nan Ya PCB +29%; in Tokyo capital rotated into rare-earth names. The week's cross-market puzzle is that the binding constraint moved below the silicon.
Power, land, and rack design — not chips — have become the new bottleneck for AI capex
Behind the Meta/NVIDIA headlines, ON, OpenAI, and IBM simultaneously put capital behind a single proposition: inference moves off the data center.
x86's two-horse race is over — two Arm-based server CPUs landed launch customers in the same 72 hours, and only one company collects royalties on both
Qualcomm's Dragonfly C1000, OpenAI–Broadcom Jalapeño, and Argentum's $4.1B GB300 deal show new silicon now ships only after a single hyperscaler signs the entry check.
While automotive DRAM allocations get cut and DDR5 16Gb spot holds at $46.667, the memory oligopoly is expanding capital structure first, capacity second.