For AMD, from 🤫 hussh
Open compute, owned by the people who run it.
AMD believes compute should be open. So do we. EPYC and Instinct can power a grid of AI factories in garages and warehouses. Ryzen AI can run a private agent on the machine a person owns. ROCm keeps every layer open and inspectable. You build the open hardware. We would love to put 🤫 Private Agent One on it, owner by owner, consent-first.
What AMD and 🤫 make possible.
ROCm is your open answer to closed stacks, and Ryzen AI runs large local models on 128GB of unified memory.
We will build Agent One's inference on ROCm and run it on Ryzen AI: owner-operated, open, no lock-in.
The private agent that shows ROCm and Ryzen AI at their best, on the person's own desk.
Data Center is now more than half your revenue, with EPYC and Instinct as the growth engine.
Our garage and warehouse AI-factory grid is a distributed buyer of EPYC and Instinct, running on ROCm.
A second supply path for AI factories, with committed volume spread across thousands of owners.
You align incentives with your biggest customers. The OpenAI warrant is the public template.
A milestone-based hardware commitment and shared marketing for the grid rollout.
A proven structure to grow the open-compute grid together.
EPYC, Instinct, Ryzen AI and an open stack.
AMD reports four segments: Data Center, Client, Gaming and Embedded. The story is the data-center AI ramp, EPYC and Instinct, now more than half the company.
Up 34% year over year, with about $6.8B of non-GAAP net income and about a 52% non-GAAP gross margin.
Up 57% year over year. The first quarter Data Center passed half of AMD's revenue.
A record for EPYC, the clear share-taker in the data center.
The XDNA 2 NPU. Ryzen AI Max pairs it with up to 128GB of unified memory to run large models locally.
AMD's open-source, LLVM-based software stack, native to the major frameworks, across Windows and Linux.
OpenAI will deploy 6 GW of AMD GPUs, with a warrant that aligns incentives on both sides.
Serious data-center silicon and an open-software spirit. Everything a private, owner-run agent needs.
EPYC server CPUs
A record share of server-CPU revenue. The workhorse of the modern data center, and the CPU tier of any compute grid.
Instinct AI GPUs
MI350 shipping now, with MI450 and the Helios rack coming in the second half of 2026, anchored by OpenAI.
Ryzen AI on the device
50 TOPS XDNA 2 NPUs and Ryzen AI Max with up to 128GB of unified memory. Large models, on the machine the person owns.
ROCm, the open stack
Open source, LLVM-based and native to PyTorch, vLLM and JAX, with a bridge for porting CUDA code. The opposite of lock-in.
Chiplet leadership
AMD pioneered chiplet packaging with Infinity Fabric, the cost and scaling edge behind EPYC and the next-generation Venice.
Xilinx and Embedded
FPGAs and adaptive SoCs: flexible edge and embedded compute for specialized nodes.
Where AMD is going.
Every one of these priorities has room for a private agent the person owns.
Scale data-center AI
Lisa Su calls Data Center "now the primary driver," with growth guided above 60% a year over three to five years.
Anchor customers at hyperscale
The 6 GW OpenAI agreement, with a warrant, puts Instinct at the largest scale in AI.
Open software
Close the gap with open, framework-native ROCm on Windows and Linux. Openness and cost as the strategy.
On-device AI
"AI everywhere, for everyone": Copilot+ class NPUs across the client line, and Ryzen AI built for developers.
Rack-scale systems
Helios moves AMD from selling chips to selling integrated racks, competing at the AI-factory level.
Process leadership
EPYC Venice on TSMC 2nm with up to 256 cores, the next data-center CPU.
Instinct MI350, Helios and Ryzen AI Max.
What AMD shipped and announced most recently, and where Agent One plugs in.
Instinct MI350 series
Shipping since mid-2025: CDNA 4 and 288GB of HBM3E, driving Data Center growth in Q1 2026. MI450 is sampling for the second half of 2026.
Helios rack
Announced for the second half of 2026: 72 MI455X accelerators, EPYC Venice CPUs and Pensando NICs, unified by ROCm. Supercomputing you own and operate.
Ryzen AI Max
16 Zen 5 cores, a 50 TOPS NPU and up to 128GB of unified memory. Runs quantized models of about 100B parameters locally, in a Ryzen AI Halo desktop priced below rival developer boxes.
ROCm 7 and 7.2
Native MI350 support, FP4 and FP8, day-zero PyTorch and vLLM, and wider Windows and Linux support.
Buy together. Build together. Sell together. Take risk together.
AMD silicon fits both halves of our north star: the distributed AI-factory grid and the agent on the person's own machine. ROCm's open spirit fits how we build.
Build: the AI-factory bill of materials
EPYC Turin and Venice CPUs, Instinct MI350 and MI450, and Helios racks, unified by ROCm.
We will standardize the 🤫 Warehouse and Factory One bill of materials on this stack, a second supply path when GPUs are sold out.
AMD gains a distributed, owner-run buyer. The owner runs an open AI factory that is theirs.
Build: the agent on Ryzen AI
Ryzen AI Max, with 128GB of unified memory and a 50 TOPS NPU, runs large models locally.
Ryzen AI as reference silicon for 🤫 Agent One on the machine the person owns, built on ROCm.
AMD shows on-device AI at its best. The person owns their AI, their data and their compute.
Sell: ROCm-native reference systems
AMD leads with total cost of ownership and an open stack.
Puppy One nodes as ROCm-native AMD reference systems, sold together into edge and telecom communities that want the lowest cost per watt.
AMD gains edge volume and an open-stack showcase. We gain a spine we can market together.
Take risk: aligned incentives
AMD co-invests with customers through structures like the OpenAI warrant.
A milestone-based hardware commitment and shared marketing for the grid rollout.
A proven way to grow the open-compute grid together.
The people building AMD.
Dr. Lisa Su
Chair and CEO
Architect of the EPYC and Instinct turnaround, scaling data-center AI on an open stack.
Jean Hu
EVP, CFO and Treasurer
Leads financial strategy and the capital allocation behind the AI ramp.
Mark Papermaster
EVP and CTO
Leads the Zen and CDNA roadmap: Venice, MI450, chiplets and rack scale.
Forrest Norrod
EVP, Data Center Solutions
Leads EPYC and Instinct in the data center, AMD's growth engine.
Open compute, owned by the communities that run it.
AMD's open ROCm and open racks match how we build: consent-first, owner-operated, open. It is the technical spine for "Own your AI. Own your data. Own your compute." And 🤫 Agent One is free for life to every federal and state government employee and every past or present service member who helps build the grid.