超威半导体（AMD）：能否与Meta达成合作？

Advanced Micro Devices (AMD): May I Meta you? This morning AMD and META announced a sizeable infrastructure partnership,intending to deploy 6GW of AI GPUs as well as a deepened collaboration on CPUs (withMeta as a lead customer on 6thGen EPYC Venice and Verano parts). Some of the GPUs,while based on AMD’s MI450 platform, will be customized to Meta’s workloads. Stacy A. Rasgon, Ph.D.+1 213 559 5917 Mark Shmulik+1 917 344 8508 The deal appears almost identical to the OpenAI deal announced last Octoberdownto the magnitude (6GW), timing (first shipments beginning 2H26 and “multi-year” horizon),and (unfortunately) equity giveaway (with a warrant for up to 160M shares issued to Metadepending on purchases and AMD stock price thresholds, ending at $600). Alrick Shaw+1 917 344 8454 Similar to OAI deal, the Meta deal should be accretive to EPS all else being equal,though we note that consensus already models in a sizeable AI ramp for thecompany.Our (very rough) math at the time of the OAI deal suggested EPS accretion inthe ballpark of ~$3.50/GW or so (Exhibit 1), and at this point we have no reason to believethe Meta dea would be hugely different. that being said, consensus already models closeto $30B in AI GPU revenues for the company in FY27 and over $40B in FY28 (Exhibit 2),already likely larger than OAI can deliver and therefore probably already incorporating theprospect of new deals like this. Arpad von Nemes+1 917 344 8461 Wenhuan Chang+1 917 344 8546 Deeksha Pandey+1 917 344 8447 And while we understand the rationale, we would really love to see large deals donewithout the need to “incentivize” customers with equity.At this point for AMD gainingscale for their parts and their ecosystem is critical, hence we understand the need to attractlarge customers in any manner possible. However, we absolutely do believe that one has toview their competitiveness through that lens, and would love to see deals done without theneed to give away large chunks of the company in return. In that light perhaps the OAI deallaying out this framework was not a great precedent (if you are a large company looking tosign a deal, why would you want something worse than what OAI got?). We wonder what thewarrant count will be if AMD gets to their 2030 targets... Does this mean anything for NVDA and AVGO?The “customization” point will probablyraise some eyebrows especailly for AVGO. That being said, this is not an ASIC (it’s just avarient, and doesn’t require a new tapeout) and we note that Meta has been a long termMIXXX partner. More broadly we believe much of what we have been seeing across thespace is more a function of compute scarcity than anything else; overall we don’t anticipatechange to the AVGO META partnership at this point (and frankly AVGO has plenty of othercustomers to keep them busy as their broader TPU architecture grows in dominance). Andof course NVDA just announced their own large META deal as well. From Meta’s standpoint,while this may continue to prompt questions about the timing ofpeak capital intensity, its emphasis on “efficient inference compute” and “deliver personalsuperintelligence” further underscores itslong‑termAI ambitions, reinforcing our viewthat following its model launch, Meta will move quickly to launch and scale productsaggressively. We make no changes to our AMD model at this time.We rate AMD MP, $235 PT. Werate META OP, $900 PT. BERNSTEIN TICKER TABLE INVESTMENT IMPLICATIONS AMD (Market-Perform, $235.00):AI expectations remain high, but a new deal with OpenAI has the prospect to drive further(possibly substantial) growth NVDA (Outperform, $275.00):The datacenter opportunity is enormous, and still early, with material upside still possible. AVGO (Outperform, $475.00):A strong 2025 AI trajectory seems set to accelerate into 2026, bolstered by software, cashdeployment, and superb margins & FCF. Meta (Outperform, $900):Confidence on sustainable growth from core business with further upside from AI productlaunches. DETAILS EXHIBIT 1:Similar to AMD’s deal with OAI, the Meta agreement could potentially result in roughly ~$3.50+ in EPSaccretion per GW deployed assuming no incremental opex (probably a bit less in practice). EPS Accretion per GW Calculation assumes no incremental opex, with direct fallthrough of incremental gross profit. Assumes 13% tax rateSource: Company reports, Bernstein estimates and analysis APPENDIX - FINANCIAL FORECASTSU.S. SEMICONDUCTORS AND INTERNET