If Everyone Gets a Personal AI Agent, Which Stocks Win?

The internet will no longer be used by us but by our agents — that is the line Mark Zuckerberg put on stage at Connect, and Meta’s Muse is its product form. Positioned as a personal superintelligence, Muse is not a developer tool that lives in a terminal but a personable helper that takes over daily jobs. Bora Özkent reads the shift through his early Grokbot use; in his framing Muse and Grokbot are two shades of the same family, one built for mass distribution and one for tinkerers. Distribution, more than raw technique, becomes the decisive advantage.

From Chat to Delegation

We used to ask AI for answers; now we hand it goals and permissions. A single instruction — “plan a trip within my budget” — turns the agent from a chat box into an operator: it interprets the request, fires tools, gathers options and moves step by step inside the apps we authorize. Özkent’s Lisbon–Porto line makes the loop tangible. The agent spins up its own virtual computer, opens and closes travel and hotel apps, collects restaurant and ticket options and proposes the most balanced combination. On the first run he withheld card details and stayed in the planning-and-tracking layer; on the next run he leans toward handing over payment. That small delay signals that trust precedes technique.

A daily log accompanies the thesis: YouTube spam and fraudulent crypto pitches are scanned, removed and blocked on a routine basis; emails are triaged, noise archived and sales or marketing opportunities surfaced. For a researcher who already tracks markets and runs the latest large models in his own terminal, the threshold for consumer agents is different: friction is lower and setup is hidden. That is why Muse’s reach through Facebook, WhatsApp and Instagram to billions looks like a distribution multiplier rather than a single launch, with a U.S. App Store top rank as an early read-out. Grokbot remains the tool for the technically curious, while Claude and ChatGPT hold agent-like twins whose configuration and interface barrier stays higher — a gap that could shift share in the near term.

The future imagined is not one agent but many agents negotiating. My agent wants the best holiday within budget while the hotel’s or merchant’s agent counters from the other side. The idea of skipping a large hotel platform to let my agent bargain directly with the hotel’s own agent captures how intermediary margin thins. Özkent frames it without hype: we used to hop between apps; in the new regime we may not even know the app names, the agent picks the source by itself. Seen from above that accelerates price discovery and transparency; seen from below it squeezes storefronts that live on commission and ads.

Why Amazon Closed the Door

The most visible clash is Amazon blocking Muse from shopping its storefront under an ‘unauthorized agent access’ warning. Publicly framed as security, Özkent reads it as economics: Amazon now earns more from ads than from retail itself, and impulse discovery inside the basket depends on human browsing. Agents do not watch ads; they compare. An agent that roams across Shopify, local shops or a producer’s own site picks the cheapest and fastest delivery and may even bargain. That thins Amazon’s ad pool and brokerage power. As an investor in Amazon, Özkent judges the move unwise because the direction of travel — agents gaining leverage — looks hard to reverse by decree.

The pressure extends beyond shopping to how software is packaged. Instead of subscribing to an entire suite, an agent can call the right module at the moment of need; subscription clutter with forgotten memberships gets found and cleaned, recurring invoices disappear. Inside enterprises the same turn appears: a per-seat license model built for humans meets a world of fewer humans and more agents that use tools more efficiently. That does not mean every software company loses, but revenue will drift from seats to outcomes, and seat-based pricing becomes contestable. Just as Booking’s storefront power is debatable when an agent can bypass the aggregator, general SaaS pricing faces the same test.

What does the extra delegation cost physically? Every handoff creates inference compute behind the scenes. Deloitte’s projection for 2026 puts two-thirds of inference compute in the hands of agents, with growth accelerating. The sharp rally in Intel, AMD and ARM — all critical to inference — on the days Muse expanded its reach is therefore not surprising. Özkent simplifies the loop: the agent parses the instruction, fires tools, retrieves options, re-runs the large model to compare, validates and returns to the user. Both model and tool layers run together. When those layers mature into agents that bargain with each other, demand for data-center power, cooling and interconnect rises in step.

How Money Flows and Where the Map Points

Two monetary tracks run through the story. First is subscription: Muse is free today but unlikely to stay free, Grok already commands a premium. Deloitte sketches 35 to 45 successful consumer apps generating a $35–45 billion subscription pool, including niches most listeners have not heard of. The second and larger track is commerce itself: McKinsey estimates agentic shopping of $3–5 trillion by 2030, a scale made intuitive by comparing it to roughly five times Turkey’s ~$1 trillion scale. In that shadow, storefronts face an estimated $234 billion revenue-at-risk pool. Slow, medium and fast adoption paths will bend the curves differently at different price points, but the direction holds: intermediated shopping and seat-based software weaken, orchestration and infrastructure strengthen.

Among large caps the picture is mixed. For Meta, Muse offers subscription and per-transaction take-rate on top of a billion-plus user base; operating the models is expensive and whether people hand over identity and payment credentials remains uncertain. At Microsoft, Azure can feed on agent workloads while the license side feels pressure from seat contraction. At Google, search advertising is the most fragile seam because an agent skips the ad to select the product. Nvidia sits across training and inference, though not dominant in inference alone; rising demand alone does not guarantee translation into profit, pricing and margin decide. ‘Chip demand is up’ is not a thesis until it ends with ‘and it converts to cash’.

Smaller, infrastructure-close names make a sharper bet. Okta (OKTA) sits as a control plane for agent identity — which agent is real, which is spoofed, who holds which permission. JFrog (FROG) anchors the software supply chain, storing, scanning and shipping the artifacts that agents will multiply as they generate code; demand for an artifact vault grows with agent-authored software. Rambus (RMBS) widens the processor-memory path with interface chips and licenses, a lever when context retention drives memory bandwidth; memory makers broadly get the same tailwind. Vertiv (VRT), referenced as Envent, owns power distribution and cooling for high-density racks — a direct play on heat. DigitalOcean (DOCN) gives small teams a cloud base to run agents without building servers. Akamai (AKAM) lands with its 11-billion-scale cloud-edge tie-up discussed with Anthropic, owning the layer that carries and protects agent traffic. Common risk across them is that server architectures and platform transitions are not yet settled.

The headwind list is cleaner. Seat-based SaaS such as Salesforce can see its license pool shrink as fewer humans and more agents do the work. Travel storefronts like Booking Holdings can see price discovery and loyalty erode when an agent bypasses the aggregator for the property itself. Any ad-centric model feels the same squeeze when an agent ignores the ad. Özkent frames three adoption paths — slow with a 50 million paid-user threshold, medium and fast — each bending subscription and commerce revenues differently, but the move from ‘user running across screens’ to ‘user commanding a single assistant’ stays fixed. His reminder, as a Meta investor, that the stock ran fast while infra spend stays heavy, is a call to discipline: excitement without conversion into free cash flow does not sustain price. The video therefore ends not with a close but with an invite to continue deep dives in his School community and a pointer to boraozkent.net, framed as research focus rather than investment advice.