The Next AI Inflection Is Now
The next inflection point in AI is not coming. It is here, and it happened on 8 September when Meta shipped Muse.
Muse is the first mass-market agentic AI that an ordinary person can actually use. Not a chat window that answers questions — an agent that connects to your email, your calendar and your payment methods and completes multi-step work with minimal supervision. Free tier, then $20 and $100 a month. iOS, Android, web, and the glasses next.
The part I think most people are underrating is the architecture. Each user gets a dedicated virtual machine in Meta’s cloud. That is not a detail, it is the whole unlock. It is what makes it safe enough to hand an agent your inbox and your credit card, and it is why this feels usable in a way the demos never did. Isolation plus ease of use is the combination that gets a normal person over the line.
Competitors will ship the same thing. Google, OpenAI, Apple — all of them, within a year. But Meta is first with distribution that no one else has, and first mover with billions of users already logged in is a real position. This may be the moment Zuckerberg finally leads a platform instead of chasing one. He missed mobile, he bought his way into social, he spent a decade and a fortune on a metaverse nobody wanted. This one he is ahead on.
Disclosure: I opened a position in Meta last week and added to it as recently as yesterday. I may add more. That is a change from earlier this month, and it is the most conviction I have had in Meta in years.
An aside for whoever at Meta is listening: build the earbud. The glasses are impressive and a large share of people will never wear a camera on their face in public. An AI earbud is the same agent with none of the social cost — no lens, nothing pointed at anyone, just a voice in your ear. It is the non-creepy version of the same product and it would sell to people the glasses will never reach.
The investable question is not whether Muse is good. It is who gets paid when a few hundred million people each run a persistent agent in a private VM. So I put that to my own tooling and asked for the beneficiaries, and separately for the businesses that get hurt. Both lists below are its output, not my portfolio. I own exactly one name on either one, disclosed above.
Who benefits — from consumer-scale agentic AI generally, and Meta’s per-user VM architecture specifically:
- Memory — MU, and the Korean makers. A dedicated VM per user is the most RAM-hungry way anyone has chosen to deploy AI. Shared inference amortises memory; isolation does not. If this architecture wins, memory demand goes up more than compute demand.
- Compute — NVDA, AMD, TSM, AVGO. Persistent per-user agents mean inference running continuously rather than in bursts. Different load shape, much larger in aggregate.
- Power — VST, CEG, TLN, GEV, ETN. Hundreds of millions of always-on VMs is a generation problem before it is a chip problem.
- Thermal and data-centre plumbing — VRT, MOD. Same logic, less glamorous, fewer people looking.
- Optical interconnect — LITE, COHR, ANET. Isolated VMs talking to shared model weights is an east-west traffic problem inside the building.
- Identity and security — OKTA, CRWD, PANW, ZS. An agent holding your credentials and payment methods is a new attack surface, and every enterprise is about to ask who is allowed to act on whose behalf.
- Payment rails — V, MA. Agent-initiated transactions still settle somewhere, and authentication of a non-human buyer is a service someone gets paid for.
Who gets hurt — the premise being that a great deal of corporate profit is collected from consumer friction, and an agent that never gets tired is friction’s natural enemy:
- Personal-lines insurance — PGR, ALL, TRV. Renewal profit depends on people not re-shopping. An agent re-shops every policy every year, automatically. The cheapest carrier gains; everyone else loses a margin pool they have relied on for decades.
- Broadband and wireless — CMCSA, CHTR, VZ, T. Promotional-rate roll-offs and switching hassle are a deliberate revenue line. Agents cancel, negotiate and port numbers without getting bored or intimidated.
- Subscription businesses that bank on forgetting. Any model where a meaningful slice of revenue is people who meant to cancel. Agents audit recurring charges monthly.
- Consumer banking fee income — regional banks especially. Deposit inertia, overdraft, a savings rate two hundred basis points below the best available. All of that survives on nobody checking.
- Travel and ticketing intermediaries — EXPE, BKNG, LYV. Opaque pricing and fee stacking are less durable when the comparison is instant and exhaustive.
- Anyone whose pricing power is confusion. Complex fee schedules, tiered plans designed not to be compared, loyalty programmes that are arithmetic in a trench coat.
And the honest objection, which is aimed at the position I just opened: if agents do the shopping, what happens to advertising? Meta’s entire business is persuading humans. An agent is not persuadable. That risk is real, it is not near-term, and it is the thing I will be watching in Meta’s own numbers more closely than subscription revenue.
Positioning: long Meta and adding, still long the rest of the AI names, still holding a large cash position. Nothing here is a recommendation — the lists above are an AI’s answer to a question I asked, published because the question seems more useful than most of what I read this week.