Phil Hughes

Honored to meet Senior Secretary @JungWooHa2 today in Seoul to discuss South Korea’s ambitious vision for sovereign AI. @AMD is committed to partnering to grow and expand the AI ecosystem in support of Korea’s AI G3 vision.

This $AMD-Samsung MOU is more significant than the headline suggests. It’s not just a memory supply agreement. It’s a supply chain alignment signal at the CEO-to-Chairman level, and it has a foundry kicker nobody’s talking about. Here’s how I read it: 1.Samsung is already AMD’s primary HBM partner, supplying HBM3E for MI350X and MI355X. This MOU extends that to HBM4 for the MI455X. Memory bandwidth is a rack-scale differentiator, and Samsung’s HBM4, the industry’s first to reach mass production on its 1c DRAM process with a 4nm logic base die delivering up to 3.3 TB/s, gives AMD a supply partner that is moving fast. AMD is not diversifying supply here. AMD is doubling down on Samsung as primary supplier for its most important AI accelerator. 2.The MOU covers DDR5 optimized for Venice, AMD’s 6th Gen EPYC with up to 256 cores, the CPU behind the Helios rack-scale platform. Having one supplier tightly integrated across HBM4 for the GPU and DDR5 for the CPU creates system-level optimization that matters when you’re building rack-scale AI infrastructure. This is how you close the gap with NVIDIA’s co-design playbook. 3.The part most people will miss: the MOU includes exploratory discussion of Samsung foundry services for next-gen AMD products. AMD has been loyal to TSMC for good reason. If AMD is even opening a foundry dialogue with Samsung, that tells me AMD is proactively building optionality as wafer demand continues to outstrip supply. This is strategic maturity, not desperation. 4. The signing with Lisa Su, held at Samsung’s Pyeongtaek complex, followed by dinner with Chairman Lee Jae-yong That’s not a procurement exercise. That’s a strategic partnership being elevated to the highest corporate levels. Context matters: Samsung announced HBM4 mass production for NVIDIA’s Vera Rubin platform earlier this week at GTC. Samsung now has HBM4 production commitments for both major AI GPU platforms, a notable positioning shift for a company that was struggling with NVIDIA HBM qualification a year ago. For AMD, this MOU, combined with the Meta 6GW deal, the OpenAI commitment, and the Oracle 50K MI450 deployment, signals that the infrastructure pieces around MI455X and Helios are solidifying ahead of H2 2026 shipments. Execution is everything from here. Unrelated, I met with Paul Cho today at GTC, President of Samsung Semi. Six Five video coming shortly.

Inference performance isn't a single number, it’s a universe of variables. On an equal footing, #AMD Instinct MI355X delivers a materially lower cost-per-token for high-interactivity workloads. You can read more in our latest blog. @roaner #Inference amd.com/en/developer/r…

AMD and @Samsung are deepening our strategic collaboration to support next-generation AI infrastructure. From HBM4 to DDR5, we’re optimizing memory solutions for AMD Instinct GPUs, EPYC processors, and the AMD Helios platform. By advancing memory and system-level innovation together, we’re enabling performance and efficiency at scale. Learn more: bit.ly/40EFg5G

@BNightengale Future mayor of Henderson, NV Bryce Harper

Sheer class

A lot of bold performance claims being made about next-gen AI CPUs lately. But where are the benchmarks?

@ryanshrout I'm sure you are mr. club 30

I'm a sucker for the @NVIDIA and @Disney Imagineering tie ins.

@BenBajarin @satyanadella no reliance on connectors is HUGE

I like this, since I firmly believe having agents work together across different models will be a thing. This is also similar to Perplexity Computer, which I haven't tried but heard good things about. Good strategy for Copilot. Very smart move @satyanadella!

Proud to partner with @AkashSystems and @MiTACcomputing to unveil the world’s first diamond-cooled AI servers powered by @AMD #Instinct MI350X GPUs. 💎 akashsystems.com/post/akash-sys…

Thank you President Chiang and the entire @LifeAtPurdue community for a great visit. Loved meeting some brilliant students and spending time talking about the future of computing and AI.

@patrick_oshag Can you DM me?

I’d love to profile / interview Lisa Su There’s way too little out there about her given her incredible track record and importance Can anyone connect me to her or her team?

Big day for @AMD and @Meta as we expand our strategic partnership to deploy 6GW of Instinct GPUs. Excited to partner across silicon, systems and software to help Meta push the boundaries of AI at massive scale!

A lot to like about the $AMD and $Meta deal. A few things most interesting besides the tens of billions, likely over 4-5 years (that's how long it will take to fill 6GW). Co-optomized Instinct for Meta: By being willing to make specific optomizations to the GPU for Meta workloads, it acts much closer to a custom ASIC than a merchant one. AMD being willing to do this, and telling the market they are willing to do this, makes Instinct more attractive IMO than in some cases their efforts in custom ASICs. Multi-generational deal: The time it will take to deploy 6 GW likely takes until 2029-2030, so this is a long term product commitment to AMD. Benefit the open ecosystem for AMD: Meta should also help accellerate the open ecosystem in favor of AMD hardware given their contributions to the community (like PyTorch). A single customer at this scale, is a net positive for the Instinct ecosystem thus adding even more credibilty to AMD competiting in this space.

Speed is the moat. MI455X is both fast, and is the fastest I have seen in execution for bringup of a complex GPU platform. MI455X is right on target for shipments in 2H2026 - Irrespective of what @SemiAnalysis_ says - rev your engines because speed is coming.

Never have truer words been tweeted

The Texans would be approximately 8000 times cooler if they were the Oilers. Still pisses me off.

Believe it or not this is a full Windows PC. Very interesting and innovative idea from @HP. Running @AMD Ryzen. hp.com/us-en/desktops…

Great to kick off #CES2026 as we work with our entire ecosystem of partners to bring AI everywhere, for everyone. We brought Helios, our rack-scale AI solution to Vegas, showcased our expanded portfolio with Instinct MI455X, MI440X for enterprise, and MI430X for supercomputing, and launched the new Ryzen AI 400 series and Ryzen AI Halo bringing powerful AI to personal devices and the edge. Thanks to our partners @gdb, @drfeifei, @Mkratsios47, @ramin_m_h, @LumaLabsAI, @blueorigin, @abscibio, @AstraZeneca, @illumina, @G_Bionics for kicking off #CES2026 with us.

OK OpenAI’s Greg Brockman @gdb on the $AMD CES stage. Brockman talks progress and the future of personalized assistants. Brockman says he needs more compute for agents that operate for hours, days for every person on the planet perpetually. ‘Needs billions of GPUs’. “No one has a plan to build that scale.” Why OpenAI is using MI455 per GB. He created that attached slide. His words, not AMDs. Bandwidth, performance, and HBM memory footprint for inference optimization.

@AMD The path to yotta scale infrastructure if a combination of compute, open rack arch, open fabrics, rack scale solutions, and that it all needs to be turnkey. This is the goal of Helios. "The worlds best AI rack"