Tweet 1 – Hook: BofA raised its price target on AMD from $550 to $620 today. The announcement was clean, polished, and devoid of on-chain proof. But this is not coverage of a token launch or a DeFi protocol—it is a financial institution projecting value on a chipmaker. Yet the logic chain is identical. Assumption is the adversary of verification.
Tweet 2 – Context: AMD is the second-largest GPU supplier for the AI compute market. Its MI300X family powers inference workloads in AWS, Azure, and Google Cloud. The narrative is simple: AMD is the only viable alternative to NVIDIA in a market where hyperscalers desperately need to reduce dependency on a single vendor.
Tweet 3 – Core Insight (Part 1): The BofA price adjustment is a direct bet on AMD's ability to capture AI compute share. But the model is built on three assumptions. First, that CoWoS advanced packaging supply from TSMC will scale without bottleneck. Second, that ROCm software ecosystem will close the gap with CUDA. Third, that hyperscaler orders will materialize at the projected volume.
Tweet 4 – Core Insight (Part 2): Let us start with CoWoS. TSMC's capacity for this packaging technology is finite. AMD, NVIDIA, and Broadcom are all competing for the same wafers. Any delay in capacity expansion—whether due to equipment shortage, labor constraints, or geopolitical friction—directly caps AMD's AI chip output. The price target assumes no supply side failure. That is a leap.
Tweet 5 – Core Insight (Part 3): ROCm is AMD's answer to CUDA. But adoption remains niche. Major AI frameworks like PyTorch and TensorFlow are optimized first for NVIDIA. AMD's compatibility layer improves, but the gap is not measured in months—it is years. Hyperscalers allocate resources to both, but the cost of migrating production workloads from CUDA to ROCm is non-trivial. The bull case assumes frictionless migration.
Tweet 6 – Core Insight (Part 4): Then there is the supply chain dependency. AMD is a fabless company. It relies entirely on TSMC for advanced process nodes and on Samsung/Hynix for HBM3e memory. If TSMC's 3nm yield disappoints or HBM supply tightens due to SK Hynix's own capacity constraints, AMD's product roadmap slips. The price target does not account for these externalities.
Tweet 7 – Contrarian Angle: Here is what the bulls got right. AMD's chiplet architecture is a structural advantage. By tiling smaller dies, AMD achieves higher yield per wafer and lower cost per compute unit compared to NVIDIA's monolithic approach. This is a genuine technical edge. In a market where total cost of ownership matters—especially for inference workloads—AMD has a pricing lever that NVIDIA cannot easily match.
Tweet 8 – Contrarian Angle (Part 2): The sovereign AI narrative is real. Governments in India, Japan, and Europe are building state-funded compute clusters. They prioritize supplier diversity. AMD is the default choice for these projects because it offers an open-source stack and is not perceived as a monopoly. This creates a demand base insulated from hyperscaler procurement cycles.
Tweet 9 – Contrarian Angle (Part 3): The PC inventory cycle is turning. AMD's Ryzen AI chips with integrated NPUs are positioned for the AI PC wave. Windows 12's AI features will drive replacement demand. This is a tailwind that the market underestimates because it focuses entirely on data center revenue.
Tweet 10 – Takeaway: I do not dispute AMD's trajectory. But a price target of $620 requires that three conditions hold simultaneously: CoWoS capacity scales perfectly, ROCm adoption accelerates, and hyperscaler orders are fulfilled without material delays. Each assumption carries a non-trivial risk premium. The market is pricing AMD as if these risks do not exist. I prefer to check the hash on the supply chain before accepting the narrative at face value.
Tweet 11 – Final Word: BofA's upgrade is a bet on narrative. AMD's engineering is real. But the margin of error is thin. If CoWoS capacity misses by 10% or ROCm adoption stalls, the PE compression will be swift. Follow the liquidity. Follow the tapes. The ledger remembers everything.