MAS Capital Markets Services Licence no. CMS101936 · Fund Management
Who owns the front door
The AI race has moved from the smartest model to the assistant that gives everyone else their orders. Three weeks in: where each contender stands, how our first calls have fared, and what would prove us wrong.
Where we stand. For three years the AI race was about the smartest model. That race is ending in a draw: models are converging, and they can be rented. The contest has moved up a floor, to the assistant you actually talk to, the one that books, buys and files on your behalf and hires everything underneath it to do so. Meta, Apple, Microsoft, OpenAI, Google and Amazon are now fighting for that front door. Nobody owns it yet. This note maps the fight, marks our earlier calls to market, and sets out what would prove us wrong.
From the smartest model to the front door
The model is no longer the moat. Meta, OpenAI, Google, Anthropic and SpaceXAI all ship comparable frontier models, and most of them will rent theirs to anyone. What sets a product apart now is the layer above: the app you open, the habit you form, the permissions you grant, and whether the thing can actually do the job rather than describe it.
The clearest evidence is who is chasing whom. Meta launched Muse, a personal agent that shops, books and fills in forms, on 8 September. It reached No. 1 among free US iPhone apps by 18 September, ahead of ChatGPT, on roughly three million downloads in under three weeks (third-party estimates; Meta has not disclosed usage). OpenAI, with by far the largest existing user base, is reported by The Information to be building a task agent and weighing a Muse-style assistant in response. Leading on the model did not secure the front door.
And the agent itself is becoming a commodity. Muse is, in Meta’s own words, “heavily inspired” by OpenClaw, a free open-source agent. Microsoft’s new always-on Autopilot is built on a “hardened” version of the same code, as described in Alex Heath’s interview with Satya Nadella (and reported earlier by GeekWire). OpenAI hired OpenClaw’s creator in February. Three of the largest companies in the field are cooking from one public recipe. When everyone has the recipe, the edge is the kitchen: distribution, trust, the checkout, and who can afford the computing bill.
Three ways to get paid at the door
Advertising does not vanish when an assistant does the shopping. It changes shape, into one of three business models.
- Meta: Muse is free for most users. Mark Zuckerberg said Meta will profit by “taking a small fee from transactions” (as reported from Connect).
- A toll lets Meta earn from the purchase itself, without having to turn Muse conversations into ad targeting.
- Microsoft: the Copilot seat stays; the agents that do real work are billed in credits at US$0.01 each (Microsoft’s rate card).
- A single agent task can run to several dollars (a partner’s reading of Microsoft’s planning tables), so a heavy user could cost more than the seat itself. The meter is how Microsoft grows without selling more seats.
- OpenAI (ChatGPT plans) and SpaceXAI (Grok Bot, from US$120 a seat a month for teams).
- Simplest to understand; hardest to defend if a rival gives the same thing away and charges a toll instead.
The money at stake is large and, so far, still growing. In the April–June quarter, Google Search earned US$63.3bn (+17%), Meta’s advertising US$59.4bn (+27% at constant currency), Amazon’s advertising US$19.8bn (+26%) and YouTube US$11.1bn (+13%): about US$154bn in one quarter (63.3 + 59.4 + 19.8 + 11.1). All of it predates Muse; July–September holds barely three weeks of it.
Three camps: let it in, keep it out, own the phone
- Walmart, Best Buy, Gap, Sephora, Wayfair, Instacart and Expedia have joined Muse; Shopify’s merchants via Shop Pay; payments through Stripe, PayPal and Shop Pay.
- Each has its own reason: for Walmart, reaching shoppers who start outside Amazon; for Expedia and Shopify’s merchants, one more channel alongside paying search engines.
- Amazon blocked Muse on 21 September.
- It is arming its own agents instead: its Seller Assistant is “built on Amazon Bedrock with Anthropic’s Claude models”, and Alexa+ serves its own shoppers.
- Apple launched Siri AI in beta on 14 September, the week Muse topped its App Store. It acts across iPhone apps, including WhatsApp and Outlook, through Apple’s own developer hooks.
- Google makes Android and builds Gemini into it, and supplies the models behind Apple’s Siri.
Apple is the camp we under-weighted. Siri AI runs on “Apple Foundation Models, custom-built in collaboration with Google and its Gemini models”, on the device and on Apple’s own servers, where Apple says personal data is “not stored nor made accessible to Apple or anyone else”. But it is slow and narrow: a waitlisted beta, English only, not in the EU on the iPhone or in mainland China, with no “agent mode” (MacStories) and no shopping. Apple and Google hold the strongest structural cards, the phone itself. Neither has yet turned that into the habit Muse built in three weeks.
Which middlemen are most exposed? Our working rule: exposure = how interchangeable the product is × how much the business depends on someone else’s front door. A hotel room sold on three sites at the same price is a comparison an agent does well. A company that owns its door, as Amazon does, can defend. One that depends on search pages and now on assistants is more exposed.
| The market’s first verdict | Past month | What is in the price, and what is not |
|---|---|---|
| Meta | +30% | Paying for the door-owner, before a single Muse revenue figure exists. |
| Apple | +9% | Near its 12-month high on iPhone demand. Siri is barely in the conversation. |
| Microsoft | +4% | Most of it on the Copilot relaunch day (+3.7% on 25 September). |
| Alphabet | +1% | Owns Android and supplies Siri’s models; its own agent still a side tab. |
| Amazon | −4% | The wall is up; whether it protects the ad business is untested. |
| Expedia · Booking | −21% · −22% | Sold as the middlemen an assistant can go around. Not every point of it is about assistants. |
| Schwab · LPL · Allstate | −9% · −13% · −12% | Part of the selling was linked to agents making it easier to switch provider (the “consumer inertia” idea, Bloomberg citing Goldman). We treat the link as unproven: rates moved sharply too. |
One-month share-price change to the close of Friday 25 September 2026 (FMP); describes market reactions, not a view on any security.
Two kinds of front door
At home, the door is a habit. At work, it is a permission. Consumers use whatever is already open on their phone, which favours Meta with WhatsApp and Apple with the iPhone. Companies use whatever IT can control, which favours Microsoft’s bet on identity, compliance and an audit trail for every agent.
Trust is already being tested. On 22 September a researcher showed that malware on a Mac could hijack Muse; Meta fixed it within a day. On 25 September Reuters, citing The Information’s review of an internal Meta incident report, said a separate SEV-2 flaw could expose a user’s cloud computer holding “emails and files”; Meta added a clearer safety warning. Reuters did not verify it; Meta had not commented. For banks, insurers and wealth managers, this is the whole question: an assistant that can act on your behalf is only as useful as it is safe.
China already ran the experiment
The recipe Muse borrowed went through a full cycle in China this spring. In March about 1,000 people queued at Tencent’s Shenzhen headquarters for help installing OpenClaw, local governments offered grants, and every big cloud firm launched a copy. Within days Beijing restricted it at government agencies, state firms and banks, citing security. By September the hype had gone. Three things killed it: cost, setup and security. In our read, the gains stayed with whoever owned the chat app: Tencent through WeChat, Alibaba and ByteDance through DingTalk and Lark. Muse is built for the second act: hosted, with vetted add-ons, no real passwords or card numbers seen by the agent, computing paid by Meta, and WhatsApp as its WeChat. China’s regulators act in days, the West’s in months.
1 · The door may have no moat either. If users switch at no cost, owning today’s favourite guarantees little (Paul Kedrosky’s “moat-less AI”; Gemini lagging despite Google’s reach supports it).
2 · The numbers disagree so far. No company has yet disclosed a fall in clicks attributable to assistants.
3 · Downloads are not habits. Meta has disclosed no daily-user figure, and repeat use decides this.
The value in AI is moving to the front door: the assistant that holds the customer and hires everything below it. The models become suppliers, even to the most self-sufficient company in technology. The shops become plumbing. Advertising turns into a toll, a meter or a subscription paid to whoever owns the door.
But nobody owns it yet, and it may not stay owned. Meta has the habit and the fastest start. Apple and Google own the phones; Apple has the trust but is the slowest, and Google supplies Apple but has not yet made its own agent a habit. Microsoft has the office and the permission. Amazon is defending its own. Our base case is a contested door for at least the next year, not a winner.
What this is not: a call that advertising collapses, or a verdict on any company. It is a map of where value is trying to go, and of the evidence that will show whether it gets there.
Marked to market: the six calls we made on 25 September
| The call | Today | Evidence so far |
|---|---|---|
| Another security headline within six months | HIT, EARLY | Three days later, on a report (Reuters citing The Information), and via the cloud computer rather than the email inbox we expected. |
| OpenAI ships a Muse-style consumer assistant | LIVE TEST | OpenAI’s DevDay is Tuesday 29 September. Leaks of an always-on assistant are unconfirmed. |
| The download spike fades by Christmas | NOT YET | Estimated downloads are still rising. No usage figure disclosed. |
| Muse moves into WhatsApp shopping | NOT YET | No announcement since Connect. |
| Meta’s results put a first number on Muse; ad growth above 20%; spending outlook rises | LATE OCT | Meta’s July–September results, expected around 28 October. |
| A retailer now blocking Muse lets it in on negotiated terms within 12 months | NOT YET | Amazon’s block stands. |
And one miss of our own: our 24–25 September map left out Apple’s 14 September Siri AI launch. This note corrects it.
Questions to carry forward, and what each would tell an investor
Each test maps to one of the positions set out under “What it means for portfolios” below: the door owner, the renter, the middleman, the brain supplier and the owner of computing.
| The question | What would answer it | What it would mean for the investment case |
|---|---|---|
| Is OpenAI a follower or a fast follower? 29 Sep (DevDay) | A consumer assistant that acts, shipped rather than previewed. | A strong launch makes the door more crowded and costlier to hold. It favours the brain suppliers and computing over any single door owner. |
| Is ad money actually moving? ~28–29 Oct, then Q4 | Google’s paid-click growth; Amazon’s ad growth; any Muse fee line in Meta’s results. | Clicks slowing while a fee line appears would strengthen the door owner case and weaken middlemen reliant on search. If not, the shift is slower and the travel sell-off may have run ahead of the evidence. |
| Does Microsoft’s meter show up? Late Oct | Any disclosure of Copilot credit revenue, and whether it is booked with Office or with Azure. | Visible usage revenue would show the work door can earn more without more seats. |
| Do the phone owners turn position into habit? Oct → June 2027 | Siri AI leaving the waitlist and reaching the EU and China; an agent mode or checkout; Gemini’s agent promoted from a side tab. | Yes: the phone owners start earning at the door and Muse’s lead narrows. No: the habit stays with the chat apps. |
| Does “inertia” income erode? Oct results | Cash-sweep balances at brokers and renewal rates at insurers. | Erosion would mark businesses that rely on customers not shopping around as exposed. Stable numbers would say September’s selling was about rates. |
| Is Muse a habit or a hit? Q4 | A daily-user figure; the US ranking through December; the Muse Charm device (due December). | A disclosed habit would support the premium the market already pays the door owner. A fade would say the door has no moat. |
| Does the wall hold? 12 months | A second big retailer joining Amazon’s block, or Amazon negotiating terms. | If it holds, owning your own traffic is a defence. If it cracks, every store eventually pays the toll. |
What would change our mind
- If one assistant clearly wins both home and work by year-end, the “contested door” base case is wrong and the winner deserves a far higher value than we allow.
- If assistants stay free with no toll, meter or subscription that sticks, the door is a cost centre, not a toll gate, and the value stays with the owners of computing.
What it means for portfolios
The useful question is not which assistant is best, but where each business sits relative to the door:
- Owns a door (the chat app, the phone, the office suite): the strongest position, but it carries the computing bill until the toll, meter or subscription pays for it.
- Rents a door: gets reach without the cost of building it, but lives by the landlord’s rules.
- Sits behind someone else’s door, selling interchangeable goods: the most exposed; this is the middleman an assistant can go around.
- Supplies the brain to several doors: gains volume but gives up pricing power, because the door owner can swap suppliers or hide the supplier’s name.
- Owns the computing: paid whoever wins, the least contested part of the stack. But not immune: paid either way is not paid at any price, and our separate work suggests chip orders are running ahead of the power available to plug them in.