Meta's Muse keeps working after users close the app, pointing to a potential growth opportunity for makers of server processors
AI agents like Meta's Muse unlock a new computing dynamic.
Meta Platforms says its new Muse agent can keep working after a user closes the app. For investors in Intel, AMD and Arm, that capability deserves attention.
An assistant that continues researching, using websites and completing tasks creates computing work long after a person stops typing, and every hour of that work runs on hardware that some company has to buy.
The way Meta (META) built Muse makes the point concrete. Muse runs in its own virtual computer in the cloud, a software-defined machine that behaves like a stand-alone PC, with a second AI agent checking its internet activity. Many virtual computers can share a physical server, but each still needs processing power to run software and carry out tasks.
I see that additional work as a credible growth opportunity for makers of server central-processing units, the general-purpose chips that run operating systems, applications and most everyday software in a data center. If agents become routine tools for businesses and consumers, demand for the processors supporting them could give a further lift to shares of Intel (INTC), Advanced Micro Devices (AMD) and Arm Holdings (ARM). These stocks have already benefited from CPU optimism, each roughly tripling so far this year. I see wider adoption of agents as a potential source of sustained growth in sales and earnings.
To see why agents lean on CPUs, consider an AI agent asked to fix a software problem. Producing a suggested repair is one step. The agent may also search files, run the revised program, inspect the results and then try again. Those activities require conventional computing alongside the specialized processors that run large AI models.
Arm, which licenses processor designs to other chip makers, described the CPU's role in an agent system as preparing requests, retrieving information, running software tools and checking results. Graphics processing units, or GPUs, remain essential for demanding model calculations. As agents undertake longer jobs, however, the surrounding work can become a larger part of the system's workload. That could give data-center operators another reason to buy or upgrade server processors alongside the GPUs they are already racing to install.
Cheaper or more capable models could encourage people to delegate more work to agents. Even if generating an answer becomes less expensive, executing the resulting tasks can create additional demand for CPUs. Whether that translates into more hardware purchases depends on whether usage grows faster than efficiency improves.
Imagine a small business that compares supplier prices once a month because someone has to gather the information. An agent could check daily, investigate alternatives and prepare an order for approval. The company would be running more searches and software operations with the same staff. Repeated across thousands of companies and applications, that is a plausible route to a larger CPU workload.
The math can also cut the other way. Some automation will simply replace computing work people already do. Providers can also consolidate tasks on existing machines or optimize their software so each server handles more agents. That is why I would watch growth in actual usage and hardware purchases together. The bullish case requires enough additional activity to outrun those savings.
If the additional activity leads to more hardware purchases, Intel and AMD are positioned to benefit. That spending could bring them additional orders from cloud providers and businesses already using their processors. Their latest results show the scale of their data-center businesses.
Intel reported second-quarter data center and AI revenue of $6.3 billion, up 59% from a year earlier. AMD reported data-center revenue of $6.7 billion, up 107%, with growth driven by both EPYC server processors and Instinct AI GPUs.
Both companies are positioning their server chips to work alongside AI hardware. Intel's quarterly report describes work with partners on infrastructure combining its Xeon server processors with specialized AI chips. AMD has a similar opportunity through EPYC, alongside its accelerator business, which sells chips built to speed up AI calculations. In July, AMD said Meta was testing sixth-generation EPYC server platforms in its labs.
Arm wants a share of the same business Intel and AMD are pursuing. In March, it introduced its own Arm AGI CPU, expanding beyond licensing processor technology into selling finished chips. Meta is the lead partner and co-developer. Arm therefore has a direct commercial route into the infrastructure of the company behind Muse.
That relationship does not establish which processors run Muse, nor does it set the value of future orders. But it does challenge the assumption that success for a large agent platform automatically means more business for Intel and AMD. A customer the size of Meta can choose among competing processor platforms, and Arm now participates through both licensed technology and its own products.
Nvidia (NVDA) is competing for this work, too, starting from its dominant position in AI GPUs. Its Vera CPU is designed for code execution, software tools and the isolated environments where agents operate. Nvidia offers it as a host for accelerated systems, where it manages the GPUs beside it, and for stand-alone CPU workloads. That gives Intel and AMD a rival able to sell the processor with the accelerators customers are already buying.
Don't miss: America's power grid is running out of juice - and these stocks stand to gain
With four serious contenders, price alone will not settle the contest. For customers, the purchasing decision comes down to the cost of completing useful work across the whole system. A processor that handles more work without drawing more power may justify a higher price. A cheaper chip can disappoint if it leaves expensive accelerators waiting for data. Software compatibility and the effort required to deploy a new platform also affect that calculation, which favors chips that existing data-center software already supports.
I would look for CPU makers whose chips save customers enough money across an AI system to justify a premium. As agents spread, the ability to deliver those savings could give suppliers more room to defend their prices and profit margins. That would make CPUs a more valuable part of the AI business.
-Jurica Dujmovic
Comments