AI Models Are Now Writing the Code That Runs AI Chips. Why It Matters

When companies weigh switching AI chips, the hard part is often not the chips. It is the software: the low-level kernels, compilers and libraries that turn a model into fast math on a particular processor. On July 29, at AMD's Advancing AI 2026 event, two of the largest model developers described a new way to attack that problem. OpenAI and Anthropic each said they are using their own AI models to help write and tune the software that runs on AMD's GPUs. For enterprise leaders, the shift points to a future where AI workloads move more easily across hardware, and where the systems they buy are designed around how models and agents actually behave.

How are AI labs using models to build chip software?

OpenAI and AMD described their relationship as moving from product collaboration to full-stack co-design, according to an AMD release published by StorageNewsletter. The work covers the Triton and Gluon programming tools, the LLVM compiler infrastructure and AMD's ROCm software stack. OpenAI's Codex is being used to help engineers write GPU code, and the two companies are researching how AI can automate parts of kernel, compiler and system configuration work.

Anthropic's agreement, announced the same day, follows the same pattern. Alongside a plan to deploy up to 2 gigawatts of AMD Instinct MI450 Series GPUs, Anthropic's teams will use Claude to optimize workloads for AMD hardware and to speed up ROCm development over several years. AMD, in turn, will adopt Claude across its engineering and product development teams.

The logic is simple. Writing and tuning GPU kernels is specialized, slow work, and a platform with a deep library of already-tuned code has a head start. If models can take on a meaningful share of that work, the cost of supporting a second or third hardware platform falls.

What does co-design between model makers and chipmakers look like?

The influence also runs the other way. OpenAI shares early insight into where its models are heading, and AMD uses it to shape silicon, systems and software, according to the release. AMD said that loop informed its Instinct MI400 generation and extends to the MI500 generation that follows. OpenAI has had access to AMD's Helios rack-scale systems for several months and expects to bring them online through deployment partners in the second half of 2026.

Meta described a similar approach. AMD's Helios design is built on Meta's Open Rack Wide specification through the Open Compute Project, and Meta plans to deploy up to 6 gigawatts of AMD GPU-based infrastructure, the companies said in a separate release. Their framing: "The entire infrastructure must be engineered as one integrated AI system."

That release also made a point enterprise architects should note: CPUs are taking on more of the work in agentic AI. Orchestrating agents, running tools and code, and managing data all lean on general-purpose processors, the companies said, and Meta is the lead customer for AMD's next EPYC "Venice" server CPUs.

Why is inference being split across different chips?

A fourth announcement showed where co-design leads. AMD and Cerebras said they will pair Helios systems with Cerebras's wafer-scale processor in a single inference workflow, according to StorageNewsletter. AMD hardware handles prompt processing and large context windows, where throughput matters. Cerebras handles token generation, which is limited by memory bandwidth and where latency matters most. The companies expect the combined system to deliver up to five times higher tokens per second per watt, a vendor estimate, and plan to offer the system first through Cerebras Cloud in the second half of 2026.

The idea of splitting inference into stages and matching each stage to the best hardware is spreading because the stages have different bottlenecks. For enterprises, the practical effect is that "which GPU" becomes a less useful question than "which service, at what latency and cost per task."

What does this mean for enterprise AI teams?

Few enterprises will write GPU kernels. But the trend changes several decisions they do make.

Portability is getting cheaper, so demand it. If model providers can serve the same model across several chip platforms, enterprises have more room to ask for flexible capacity and pricing. AtScale co-founder and CTO Dave Mariani argued in a AI Connect panel that infrastructure choices should stay open because "we don't know where this stuff is going." Our report on AI infrastructure at scale, semantic layers and inference costs makes the case for open formats over single-vendor bets.

Test on your own workloads. Vendor performance claims like a fivefold efficiency gain describe specific setups. Before moving a production workload, run your own evaluation set on the new platform and compare quality, latency and cost. Our coverage of why custom benchmarks beat leaderboards explains how to build one.

Plan for CPUs in agent architectures. If agents spend much of their time calling tools, running code and moving data, CPU capacity and data access speed matter as much as GPU supply. That affects both cloud reservations and on-premises designs, such as the AI factory model NVIDIA and HPE described.

Keep control of what is yours. As the hardware layer becomes more interchangeable, the durable assets are an enterprise's data, evaluations and fine-tuned models. Mistral AI's Aaron Delp, global head of AI technical marketing, told a AI Connect webinar in July that customers should own "the data, the weights and the keys." Our story on why enterprises should own their AI covers what that looks like in practice.

What should leaders watch next?

The first test comes in the second half of 2026, when OpenAI's Helios deployments and the AMD and Cerebras inference service are scheduled to start. Anthropic's first gigawatt on AMD hardware is due in the first half of 2027. The more lasting question is whether AI-written kernels and compilers close the software gap that has kept most AI workloads on a single platform. If they do, enterprise buyers stand to gain the most from a more competitive market for the compute their AI depends on.

Frequently asked questions

How are AI companies using AI to build chip software?

OpenAI said Codex helps engineers write GPU code and the two companies are researching automating kernel and compiler work; Anthropic will use Claude to optimize workloads and ROCm for AMD GPUs.

What is disaggregated inference?

It splits AI inference into stages, such as processing the prompt and generating tokens, and runs each on hardware suited to it. AMD and Cerebras announced such a system on July 29, 2026.

Why do CPUs matter for agentic AI?

AMD and Meta said agents rely on CPUs for orchestration, running tools and code, and managing data, so CPU capacity matters alongside GPUs.

Filed under: Data & AI Infrastructure, Agentic AI