AMD @ BOFA Global Tech Conference 2026. Executive Summary & Full Transcript
Speakers: Jean Hu (CFO) and Matt Ramsey (Head of Investor Relations), Advanced Micro Devices.
Hosted by Vivek Arya, BofA Securities semiconductor research.
Source: AMD fireside chat transcript, BofA Securities 2026 Global Technology Conference
Date: June 2, 2026
Executive Summary
The session was dominated by one theme: the rise of agentic AI and what it means for CPU demand. Hu framed the inflection bluntly, arguing the workload has shifted from “answering questions” to “orchestration,” database access, and heavy tool execution, all of which lean hard on CPU performance.
AMD’s data-center CPU business grew more than 50% year over year in Q1, and management guided Q2 CPU revenue up more than 70% year over year, with roughly two-thirds of that growth coming from unit expansion rather than ASP.
On total addressable market, Hu walked through AMD’s escalating CPU TAM math: the company first sized the 2030 data-center CPU opportunity at $60 billion (18% CAGR) at its November analyst day, then raised it to more than $120 billion on the Q1 earnings call as agentic adoption accelerated.
AMD splits the opportunity into three buckets:
Traditional general-purpose compute (a roughly $25-30 billion 2025 market, steady but slower growth),
GPU “head node” CPUs (growing fast as GPU-to-head-node ratios climb),
Standalone agentic AI racks, which Hu argued will be the majority (over 50%) of the larger TAM.
Ramsey reinforced the structural story: as AI capex shifts from training toward inference, and as chatbot inference becomes agentic inference, the CPU-intensive work between inference steps (post-processing, data retrieval across cloud/ERP/CRM systems) drives demand for current Turin parts and the upcoming 256-core, 2nm Venice CPU launching in a couple of months.
On x86 versus ARM, management declined to frame it as an architecture war, emphasizing TCO, platform breadth (5 generations from 8-core to 256-core), the entrenched x86 code base, and a decade of RAS and security validation with hyperscalers, which they argue gets AMD into “essentially every RFQ.”
Value share reached roughly 46% in Q1, and AMD credits its TSMC partnership for the ramp. Supply remains tight (notably 3nm), but Ramsey suggested AMD secured more allocation from TSMC earlier than investors appreciate, while warning the “starting point” of investor expectations was too low.
On data-center GPU, AMD reiterated the MI450 / Helios rack-scale ramp is on track. Per Ramsey’s comments at the session, the MI450 has sampled and a number of customers have full Helios racks in their own data centers running production workloads in test.
Launch begins Q3, scaling into Q4 (”billions of dollars of scale”) and stepping up again in Q1, supported by the ZT Systems acquisition for system-level design. Per management’s commentary at the session, AMD expects more multi-gigawatt customers beyond OpenAI and Meta, and said both anchor customers are now forecasting above AMD’s original 2027 plan, with warrant structures aligning upside.
On memory, Hu acknowledged unprecedented DRAM cost inflation, stressing long-term supplier agreements and a strategy to share cost increases with customers rather than exploit shortages for pricing. Ramsey noted most server/AI DRAM is procured by OEMs, ODMs, and hyperscalers, so it does not run through AMD’s P&L, though some consumer and gaming input costs are reflected in guidance.
Bottom line: Management’s message was confident and supply-constrained-bullish - agentic AI as a durable CPU demand driver, an on-track MI450 ramp, and demand visibility extending into 2027 and beyond.
Full transcript below the fold. Let’s dig in…
