AMD is hosting its “Advancing AI” event today in San Francisco, California, with the keynote from AMD’s CEO Dr. Lisa Su expected at 09:30am PT time. While we await the major announcements that company is preparing, we are bringing you a live blog of things that are happening on stage. There are many things to expect from AMD as the company will be showing its best solutions, but some earlier rumors were that we will be seeing the first “Zen 6” designs alongside complete solutions based on that. Updates from AMD’s partners are expected so we can see how AMD hardware actually runs across the world, and there could be a lot of surprises from CEO Dr. Lisa Su. We hope to see a Steve Jobs-style “one more thing” with new Radeon, but since the event is mostly about AI there is little hope. Anyway, stay tuned for more updates as the event starts in about an hour and 20 minutes!
16:25 UTC: We are five minutes away from the start.
16:31 UTC: The event is starting. AMD’s Lisa Su welcomes guests and says an energetic good morning. AMD wants to use computing to solve problems and challenges, the company strives to be a driving force
16:33 UTC: AI is transforming every industry. We are still very early. Compute demand continues to increase 5x per year, models getting better, we see much better capabilities. We are getting specialized models now. This year is the first year when inference is beating training, as 60% of compute is now used for inference, not training.16:37 UTC: Agentic AI is creating a step-change. You task it for a problem, it works until the problem is solved. This is driving demand so much that it is continuing to accelerate, and it is even getting harder to project. By 2030, AI accelerator market will reach $1.4 trillion. This is the size of entire semi market today.
16:39 UTC: AMD is seeing CPU TAM grow so fast that even AMD underestimates the projection. CPUs alone will grow to $220 billion by 2030. Every part of the compute stack will be accelerated by AI.
16:42 UTC: AMD plans three core pillars of corporate strategy: compute leadership, open platforms, and AI everywhere. Lots of hardware is coming!
16:44 UTC: AMD EPYC: Everyone runs on EPYC server CPUs. Now AMD commands 46% of server CPU revenue share, with projections to be even higher. Frontier AI development takes much more than a simple CPU. That requires CPUs, GPUs, networking, all in a single solution that is easy to deploy, easy to service, and very reliable. That is where Helios takes place.
16:47 UTC: AMD integrates GPU, memory, power delivery, coldplates, system management, in a custom module for Instinct MI455X. 320 billion transistors, built on TSMC 3 nm and 2 nm. Uses chiplets and 432 GB of HBM4.
16:49 UTC: Custom networking is key to interconnecting massive compute. Salina DPU and Vulcano AI NIC help with that.
16:50 UTC: AMD claims 15% higher performance compared to NVIDIA “Vera Rubin”, more bandwidth, more inference performance. Helios is officially in full production. Q4 is when shipments start. Leading AI labs are first customers.
16:52 UTC: Lisa welcomes Tom Brown, Anthropic co-founder and chief compute officer. AMD builds big compute clusters, customers put it to good use. Up to 2 GW of Helios will be deployed at Anthropic. “Helios is an amazing machine,” says Tom. Models are performing greatly on AMD platform. Claude AI is doing more of hardware engineering, great opportunity for future collaboration with AMD. Lisa says this is the beginning of multi-year partnership. AMD and Anthropic will work together more to help with scale-up of getting more GPUs online. Security is also very important.
16:58 UTC: AMD compared MI355x and MI455x, massive speedups are now due. High parallelism, more throughput, and up to 18x higher token throughput with Helios and 34x lower token cost. Power is the limiter, Helios brings an average of 10-15% compared to competition. This is up to 30% more tokens delivered per USD spent.
17:01 UTC: OpenAI’s head of infrastructure Sachin Katti joins the stage. He jokingly says attendance at AI event follows scaling laws. Everything is scaling greatly. Models now run long tasks, compute is also scaling. OpenAI always needs more compute, one of the first AI labs to bet on AMD. Helios racks have been tested for GPT-class workloads, and OpenAI sees great capabilities. “We need it earlier,” says Sachin jokingly as every bit of compute is needed. AI is automating significant portion of work of tuning workloads for AMD GPUs. Since AMD is doing open ecosystem, everyone could use optimization, not just OpenAI. This is the tide that lifts all boats, says Lisa. “I need more compute, more quickly” again jokes Sachin. Co-design of hardware is a possible future requirement. MI500 and onward will be co-designed with partners, and OpenAI is here for that.
17:10 UTC: CPU time now. AMD reflects on EPYC since launch. Today, 5th gen EPYC “Turin” is the best server CPU. However, compute never stands still. Agentic AI drives CPU development. Hence, AMD made EPYC “Venice” CPUs for helping agentic workloads do work very fast. Run 1000s of agents fast. 302 billion transistors built on 2 nm. Up to 512 thread per socket, 256 cores. Biggest leap in EPYC CPU generations. Venice is not one chip, it is an entire family of chips. Venice-HF 96-core, up to 5 GHz, ships inside Helios. Venice Dense brings 256 C / 512 T per CPU. AMD “Verano” will be more power efficient low power memory, faster fabric. Venice-X will also be available with 3D V-Cache.17:17 UTC: AMD Venice is a performance monster, delivers more performance per watt and more performance density. Compared to x86 the gap is big, but compared to leading Arm processors, x86 gap is even larger. Venice leads across categories. 20% higher per core performance compared to competition. Paired with core density, up to 2.2x increase. AMD claims that x86 still has an advantage, besides Venice having the best performance. Customers have a choice to get what they need, with core counts and performance targets that they want. Venice is in full production now, with Q4 seeing rollout.17:22 UTC: Meta’s head of infrastructure Santosh Janardhan is here. Demand for content is exponential, once Santosh met with Lisa to ask for CPU deployment is now turning into multi-year partnership. Systems now need to be co-designed and co-created. Meta is one of AMD’s biggest partners with million of CPUs deployed. Meta’s workloads are evolving fast, all of that is revolving around the CPU, despite AI running on GPUs. Think about CPUs and GPUs now as co-joined teams. Meta started with MI300X GPUs and now is integrating MI455X for everything. From recommendation systems to the latest Muse Spark models, everything will be running on MI455X. AMD is also co-designing with Meta. Looking out in the future, AMD-Meta collaboration will need to work on systems that are not in isolation, but early co-design. Sit down in a room today and design something for 2028, two years into the future.
17:31 UTC: Inference is segmenting into more areas. Getting ultra-low latency tokens is what everyone wants. Only disaggregated compute will be capable of delivering that. Cerebras is partnering with AMD to achieve that and Andrew Feldman joins the stage. Cerebras has wafer-scale chips for fast inference. AMD is partnering with Cerebras to get models inferencing fast. AMD and Cerebras are building a disaggregated solution with Helios and Cerebras wafer-scale engine. Later this year the solution will arrive through Cerebras cloud, afterward for customer deployments.17:38 UTC: Vamsi Boppana is joining the stage for ROCm update. AMD has been moving fast to get developers used to ROCm, software releases go out every 6 weeks for fast iteration. Open source and abstraction is what drives ROCm development. Everything now runs on AMD platforms from day zero. Software libraries and domain-specific languages have been made to assist open-source AI development. AI is generating kernels for fast model deployment, shockingly good, as Vamsi describes it. AI generated code is so good that is making it straight to production. ROCm.AI platform has been announced to give developers AI-driven platform for agents to understand AMD platforms. ROCm.AI uses AI skills, optimizations, libraries, configurations, and everything in between to get a model tuned to every AMD GPU/CPU configuration. For example, MiniMax M3 getting optimized for MI355X accelerator, this is simple as the system is doing it by itself. From start to deployment, an agent does the work. ROCm.AI delivered 38% more token output by optimizing by itself. Up to 3.3x more performance can be delivered by ROCm.AI than vanilla deployment. Even training is improved.17:49 UTC: Philippe Tillet from OpenAI joins the stage. He built Triton, which builds the base of AI GPU programming now. High-performance GPU kernel development is now much easier. Now AI is helping program AMD GPUs, GPU kernel engineers utilize it. In the past 6 months, agents became excellent at doing optimizations.
17:55 UTC: ROCm is also about HPC and science, which also runs on the same system. One open stack for every workload. Instinct MI430X leads in HPC workloads with FP64, delivering 288 TeraFLOPS of FP64, 432 GB of HBM4. It ships in first half of 2027. EU and US national labs will be deploying these for science. Basically the only HPC accelerator now in production. NVIDIA is not competing in HPC much, ie FP64 data.
17:59 UTC: AMD’s Dan McNamara joins the stage. Claims AMD is the only company spanning every part of the compute stack. Customers are co-designing hardware with AMD. Some customers want per-core performance, some customers want bandwidth. No single SKU satisfies every scenario. Hence, why AMD has many SKUs for each customer. AMD claims that it is beating competition at a minimum of 2.5x improvement for the most important tasks. Every type of clients needs to be addressed, and everyone will get their powerful compute product. MI350P is finally coming for everyone today. Up to 260 billion parameter models on a single GPU, so everyone can run local models. AMD is one of the first customers itself for autonomous threat detection. Enterprise AI will be deployed locally, for the enterprise itself to control.
18:10 UTC: Jeremy Legg, CTO of AT&T joins the stage. AT&T is burning a trillion tokens per month, that number is moving at double-digit percentages. The company handles 300,000 calls per day for support scenarios, transcribes them, and acts on it. He notes that world-class teams are needed, but AI is helping them accelerate in every scenario. The company believes in data sovereignty, AMD is helping them run open-source models and actually post-train models so they run these models on their own data centers, with no data leaking to outside labs. AT&T is the first telecom company to train a model on AMD hardware. Closed-source models might be good, but open-source models are what AT&T needs as well. OTel 2.0 model launches today, which is a custom open source model from the company.
18:16 UTC: Jack Huynhh joins the stage. He talks about personal AI that are the next-generation productivity multipliers. Your PC is one of the most powerful devices, but potential remains untapped most of the time. Local AI is the future as data remains personal. Potential is just surfacing. Local models are evolving so fast that it took only 7 months for a 9 billion model to outperform a 120 billion model. Iteration is happening fast, and powerful models will be running on local machines like Ryzen AI 400 and Ryzen AI MAX. For example, Ryzen AI Halo is AMD’s most powerful local device for developers to test, build, and create applications on their desk, with data remaining private. All of that will be on open-source software. AMD partners with Hugging Face to help optimize for open source AI models on Ryzen AI Halo. Every Ryzen AI Halo box will include a yearly subscription to Hugging Face Pro, for free.
18:23 UTC: Gorgon Halo arrives to push up to 192 GB of memory, models with up to 300 billion parameters on a single system. Powerful system gets more powerful.
18:25 UTC: Cisco’s Jeetu Patel joins the stage. Human token consumption is nothing compared to agents, hence infrastructure needs to adapt. PCs will be running agents on their own and doing work on their own. AMD and Cisco are working together to make sure that these systems remain secure, agents remain isolated enough, and that the entire infrastructure gets more layers of control for agent behavior for safety and security. These are where runtime guardrails are coming, so nothing gets past pre-defined security. This happens through the full-scale software and security policy enforcement. Cisco already provides management of the entire estate, monitoring tokenomics, and everything else. Every single agent can be monitored and get their costs contained. Cisco provides this management apparatus for select customers now, but early Fall will be the date for the US customers.18:32 UTC: Physical AI enters the stage. AMD will be powering this area as well, but now with a foundation with a specialized platform, a robotics brain. AMD Kria AI system on module. The system is about 3x faster than NVIDIA Jetson Thor, built on AMD hardware and software. The Kria AI robotics platform arrives with AMD ROCm and RAS. This will be the brain of every robot that interacts with physical world. It also joins open robotics ecosystem.18:38 UTC: Dr. Lisa Su comes back to round up everything. AMD previews EPYC “Florence” and next-gen Instinct. MI500 is previewed with massive scale up design. MI500 will deliver the largest generational leap ever, bigger than the current. Over 2,000x performance increase in just four years. Co-design is happening with customers now. New Helios systems will be delivered every year. Lisa is wrapping the event up, thanking partners. Leaving the final thought, she says that the biggest change is not what might be possible, but actual impact that we are seeing today with AI. AMD will be at the center of it with new hardware, software, and infrastructure. 30,000 AMD engineers are making it happen.18:47 UTC: That concludes today’s Advancing AI live blog!
Full TechPowerUp Coverage of AMD Advancing AI 2026
Partner and Ecosystem Announcements