Conference Season Field Dispatch — June 2026

Three Conferences Covered In Depth

This field dispatch synthesizes on-the-ground intelligence and management commentary from three major technology and capital markets conferences that converged during the first week of June 2026. Together they represent the most concentrated signal window of the mid-year semiconductor and AI infrastructure calendar.

Computex 2026 — Taipei, June 2–5

Computex remained the flagship venue for silicon product launches, with NVIDIA's Vera Rubin platform generating the most sustained investor attention. Jensen Huang's keynote confirmed a 2H 2026 ramp timeline for Vera Rubin, with Blackwell supply normalization expected by mid-year. QUALCOMM reinforced the Oryon compute roadmap for both PC and edge AI applications, while INTEL used the event to highlight 18A foundry client traction and the expanding ecosystem of external design wins. System-level themes centered on co-packaged optics, 800G+ connectivity for rack-scale AI clusters, and the transition from datacenter-centric inference to edge-distributed architectures.

Evercore Global TMT Conference — San Francisco, June 2–4

Evercore's investor-facing format provided the most direct read-through on institutional sentiment. Portfolio managers and sell-side analysts converged on a consistent view: AI infrastructure spend remains durable through 2026 and into 2027, but the composition is shifting from GPU procurement toward the surrounding memory, connectivity, and custom silicon stack. MARVELL's management reinforced the custom silicon thesis, citing expanding engagements with hyperscaler customers for networking ASICs and connectivity fabric. Memory dynamics were a recurring discussion point, with WESTERN DIGITAL providing context on enterprise SSD demand and early indicators of DRAM tightness in DDR5 configurations.

BofA Global Technology Conference — June 2–3

BofA's event attracted C-suite participation across the semiconductor supply chain. INTEL's management articulated the most detailed public commentary on 18A foundry positioning, framing the node as critical to domestic semiconductor sovereignty objectives. The Token Economy theme — the economics of inference at scale, including cost-per-token trajectories and the implications for memory bandwidth — was prominently featured across multiple presenting companies. Agentic AI infrastructure requirements were discussed in the context of continuous inference workloads that place structurally higher demands on HBM bandwidth than batch-based training tasks.

Five Structural Themes Across All Three Events

Despite different formats and audiences, the same five structural themes emerged consistently across Computex, Evercore Global TMT, and BofA Global Technology. Each represents a durable investment consideration that extends well beyond the conference cycle.

Company Spotlights

Five companies generated the most sustained management and investor discussion across the three conferences. The table below summarizes the key signal and investment read-through from each.

Investment Framework: What the Conferences Confirmed

Across 72 hours and three distinct venues, the consistent message was that the AI infrastructure investment cycle is broadening and deepening simultaneously. The GPU-centric phase of hyperscaler capex is giving way to a more distributed architecture — one where memory bandwidth, custom silicon, connectivity fabric, and edge inference share the capital allocation equally with compute.

For investors, the signal hierarchy that emerged from Conference Season June 2026 is as follows: NVDA remains the structural anchor of the AI compute stack, but the adjacent opportunity set — HBM supply (SK Hynix, Micron, WDC), custom silicon (MRVL, Broadcom), foundry sovereignty (INTC), and edge compute (QCOM) — is now generating institutional interest that was previously concentrated solely on the GPU layer.

The Token Economy and Agentic AI themes represent a forward demand signal that is just beginning to shape capital allocation decisions. When inference workloads become continuous rather than batch-driven — as agentic architectures require — the memory bandwidth and total system power budgets change materially. This creates a second-order demand pull for HBM, DDR5, and ultra-bandwidth connectivity that is not yet fully priced into the sector.