2: Why Now: Three Economic Forces Are Converging

“What happens to Big Tech when advanced intelligence no longer has to be rented from the cloud for every task?”

Apple and Meta have crossed one of the most consequential competitive boundaries in technology. On October 15, Meta introduced Muse Spark Open, an 816-billion-parameter multimodal model, and with Apple demonstrated that this class of model can run locally across two M5 Ultra Mac Studios using Apple's MLX framework.

For the first time, the emerging contest is not simply over who builds the most powerful intelligence. It is over who controls where that intelligence runs, who owns the interface, who retains the data and who captures the economics of inference.

The cloud will remain indispensable for training and demanding workloads. But its assumed dominance over advanced inference is now being challenged. The Great Split is moving from the ownership of AI models toward control of AI deployment.

Apple's unified-memory design allows up to 512 GB per Mac Studio; two connected systems meet the model's requirements, while its mixture-of-experts architecture activates only part of its parameter count during inference.

The weights are available under a license with commercial restrictions — open-weight does not mean unrestricted open-source. The demonstrated setup is expensive and specialized, aimed at developers and demanding professionals. It does not yet prove an equivalent model can run economically on an iPhone or laptop. But the direction is unmistakable: the economic frontier of AI is moving from remote capacity toward locally accessible capability.

First, AI infrastructure financing is becoming more expensive, exposing capital-intensive firms whose future revenues must support heavy borrowing. Second, advanced models are becoming cheaper to deploy — mixture-of-experts, compression and quantization let sophisticated models run on less active compute. Third, agentic AI is changing the economics of usage: an autonomous agent can make hundreds of model calls per workflow.

The relevant cost is no longer the price of a single response — it is the total cost of intelligence consumed across a workflow. Reuters reported on September 22 that Apple was already pitching its new Macs against rented data-centre capacity. The mechanism is straightforward: more agents → more inference → greater sensitivity to token costs → demand for local capacity.

The computer is evolving from a device that executes user commands into a system that executes user intentions. In agentic computing, the user delegates objectives to software that selects applications, coordinates services and completes tasks.

That changes the strategic role of the operating system. The next platform controller may be the company determining which agents receive permissions, which models execute tasks, how data is accessed and where computation occurs — a new category of competition: AI-native pervasive computing, with workloads allocated by cost, latency, privacy and performance. The likely architecture is hybrid, not exclusively local.

Apple controls the hardware environment: silicon, unified memory, operating systems, developer tools and distribution. Meta controls a model ecosystem: frontier research, open-weight distribution and social platforms. Apple demonstrates that sophisticated external models run on its hardware; Meta reaches developers and enterprise workloads beyond its own apps.

But Apple wants to preserve control of the user experience, while Meta wants its intelligence broadly available everywhere. This is competitive complementarity, not permanent alignment — cooperation in deployment, continued rivalry over the future interface to personal computing.

The shift is from centralized intelligence delivery toward economically optimized deployment, producing four effects: the inference cost curve becomes a competitive weapon; memory architecture gains strategic value; enterprise sovereignty acquires commercial value; and the interface becomes the control point. Lower cost per task can increase total usage — a larger AI market with a very different distribution of profits.

Five Takeaways For Investors

the investor hierarchy

• Follow inference economics, not parameter counts: the metric is cost per completed task, adjusted for latency and reliability.
• Revalue memory and system architecture — processors, memory and interconnects working together.
• Watch operating-system control: Apple, Microsoft and Google may capture value as models become interchangeable.
• Distinguish cloud displacement from market expansion: local inference can cut some cloud spend while lifting total usage.
• Invest in economic capture — businesses converting lower intelligence costs into recurring revenue and free cash flow.

The first AI revolution made intelligence powerful. The next will make intelligence pervasive — and the greatest economic prize will belong to whoever controls its deployment.