SHEYU.AI舍予基业
AI Daily Brief

OpenAI contributed $24.1bn to Microsoft: what the AI ledger means for China's outbound founders

Microsoft's 10-K shows OpenAI revenue of $24.1bn for the year to June 2026, a signal that AI spending is shifting from AI builders to AI users — and to specific, repeatable tasks.

·4 min read
OpenAI contributed $24.1bn to Microsoft: what the AI ledger means for China's outbound founders
AI-generated illustration, not a news photograph

For small and mid-sized companies, AI has moved from "should we try it" to "can we account for it". The hard part is not whether a model is expensive. It is which daily, repeating task you hand over to it.

What Microsoft's filing actually shows

According to Cailianshe, Microsoft disclosed in its 10-K annual regulatory filing with the US Securities and Exchange Commission that revenue from OpenAI reached $24.1bn in the fiscal year ending June 2026. On the growth assumptions Microsoft has previously disclosed for its AI business, one institution estimates that OpenAI accounted for more than half of Microsoft's AI sales in fiscal 2026, possibly around 70%. That is an institutional estimate and the actual figure may differ. Against Microsoft's total revenue, the share is under 10%.

The same report contains a line worth more attention: Microsoft added about $51bn in commercial bookings quarter on quarter, and that increment came entirely from customers other than AI startups.

In plain terms for a business owner: the companies actually paying for AI are shifting from those that build AI to those that use it. Small and mid-sized firms are not spectators.

Why a small outbound company should care

Because the way AI is priced has changed. The old question was which model is cheaper. The question now is which piece of work you hand over, and what that piece of work was worth before.

Revenue sharing, compute procurement and equity ties between the big platforms are remote from most firms. But the 10-K implies a simple structure: AI revenue comes from two kinds of customer — those using AI to build products, and those using AI to run their business. The first competes on technology; the second competes on scenarios. Small and mid-sized companies almost always sit in the second group. That is not a disadvantage; it is a position.

Position determines trade-offs. There is no need to chase every new model. What matters is the one thing in your shop, factory or office that repeats every day and goes wrong every day. Shu Yi, executive director of Sheyu Jiye, puts it directly: when going digital, start with the thing that repeats every day and goes wrong every day.

The common objection: AI is not reliable yet, so wait

That argument is half right.

If your business is low-frequency, non-standard and needs a person to judge each case on site, waiting costs you little. AI cannot make decisions for the boss, and it cannot turn a pile of bad data clean on its own. Shu Yi has also said that AI will not decide for the boss; it puts the numbers the boss needs to see in front of him first.

But if your business is high-frequency, follows a fixed process, and a single mistake means apologising to a customer, waiting has a real cost. Hotels are an example. Of the calls a front desk takes each day, how many are "what time does the restaurant open", "please send up two bottles of water", "can we check out later tomorrow"? None of these is difficult. The difficulty is that when shifts change and staff get busy, requests are missed, answered slowly, or the guest has to explain again.

This is exactly the kind of work AI can now absorb.

A high-end hotel example: how that work gets absorbed

For a high-end hotel with several hundred rooms, the fear is not a lack of guests but not knowing the guest behind the room. Which floor the guest stayed on last time, whether they want a hard or soft pillow, what time they arrive this time — these details sit scattered among several staff.

SHEYU AIHOTEL connects rooms, guests and orders on a single data foundation. Guests submit requests on an in-room tablet or a WeChat mini-program; after a multi-turn AI conversation, the request is written into a summary and turned into a work order, which the front desk takes over and completes from a to-do list without having to interpret the request from scratch. Wake-up calls, deliveries, repairs, late check-out and meal orders can all follow this route.

The boundaries should be clear too: it is built for high-end hotels with several hundred rooms, not for guesthouses with three or five rooms. It standardises the service process, but what that standard should be is still set by people. Shu Yi's view is that service in a high-end hotel cannot depend on individual flair; good service should become a standard that every shift can reproduce.

What outbound teams should look at more closely

One of the most expensive costs in cross-border business is information breaking down between two people. Shu Yi has another line: what guests remember is response speed, and the information held by the front desk and by customer service must be the same set.

That applies beyond hotels. For teams running independent sites, overseas warehouses or cross-border customer service, the problem is often not a lack of tools but tools that each manage their own patch: what a customer says on channel A is unknown to the person on channel B. When choosing software, Shu Yi's standard is worth borrowing: do not look at how long the feature list is; look at whether it lets frontline staff do one step less.

The AI ledger for this round is not about how many salaries are saved. It is about the customers who would have been missed and now are not.

One thing to do now

There is no need to rush into procurement. Take a sheet of paper, list the things your team did more than three times last week and got wrong at least once, and rank them by how much one mistake costs. The one at the top is your entry point.

Which tool to use is the second step. Whether the tool is right depends on whether the first step found the right task.

閱讀繁體中文版 →