AINEWS
Code interpreters and rule based tools put CPUs back into serving AI models. The GPU era didn't end; it stopped being the whole stack.
When DeepMind's AI solved IMO problems at silver-medal level in July 2024, the winning system did not run on GPUs alone. It leaned on a symbolic code interpreter that executed generated Python, ran the math, and fed the result back into the neural model, a hybrid loop that, by definition, needed CPUs alongside the accelerator farm. The same pattern shows up in xAI's Grok 4 release, where the model ships with built-in code execution, and in OpenAI's o3 family, which the company describes as a chain that invokes tools the user never sees.
They are the new shape of the frontier.
For most of the past decade, the answer to "what runs a large language model" has been simple: GPUs. Pure neural networks are dominated by parallel matrix multiplication, the workload GPUs are built for. CPUs, the general-purpose workhorse of classical computing, were treated as a host processor that mostly moved bytes around. That framing defined the post-2012 deep-learning era. Model size, training compute, and GPU count all moved together, and chipmakers, cloud providers, and AI buyers all built plans around that single axis.
A Substack post by Gary Marcus (CPUs and the rise of neurosymbolic AI) argues the era ended around mid-2023, when frontier companies "quietly started incorporating things like symbolic code interpreters." The claim is contested. NVIDIA and parts of the AI press read tool use as ordinary scaffolding rather than a paradigm shift. The underlying observation is harder to argue with: the systems that just won math olympiads and shipped to paying customers are not pure neural networks.
Neurosymbolic AI is the academic name for stacks that combine neural networks with explicit symbolic reasoning, often through code, formal logic, or external solvers. A 2024 arXiv systematic review treats it as a distinct workload class, not a marketing label. A separate workload paper at ISPASS24 profiled the resource mix those systems actually need, and the mix is not GPU-only. Every code interpreter call, every database lookup, every search-tool invocation is a CPU job the accelerator sits idle through.
That composition changes three things a chip strategy team can plan around.
Inference bills are no longer dominated by accelerator minutes alone. The DeepMind IMO run spent a meaningful share of its compute on the symbolic search loop, not the neural rollout. Enterprise buyers running tool-using agents will see the same shape on their invoices: GPU time plus CPU time plus orchestrator time, billed separately or not, but present either way.
A symbolic interpreter is a new attack surface and a new dependency. Code execution means model behavior is shaped by whatever runtime executes the tool, which means a different vendor, the interpreter author, the sandbox provider, or the cloud, sits in the critical path. Claude's tool-use design) and Grok 4's code-execution mode both place the runtime in front of the user, even when the model name is the headline.
Marcus's follow-on argument (How o3 and Grok 4 accidentally vindicated neurosymbolic AI) and a parallel post on his New Savanna blog (What we've got in frontier models is now neurosymbolic) make the capability case: the systems that solve IMO problems, run long-horizon agents, and ground their answers in retrieved evidence are doing things pure transformers cannot. That is not a forecast. The IMO silver medal is a dated result from a named system.
The GPU era is not over. Hyperscaler capex is still going into accelerators at a record clip. What has changed is the unit of work: the question is no longer "how many GPUs" but "what is the CPU, symbolic-runtime, and accelerator mix for a given inference call." Anyone buying, building, or pricing frontier AI on a GPU-only spreadsheet is missing half the bill and most of the control surface.
The next data point worth watching is not another model launch. It is the first hyperscaler capex line that breaks out CPU and accelerator spend separately, the way cloud bills now break out storage and egress. When that line item appears, the GPU-only story is officially over for accounting purposes, too.