Two Chinese frontier labs just reported revenue up nearly 300-400% with sales spending down. That inversion is not a cost story. It is a go-to-market reset for the entire AI industry.
# When Revenue Grows 400% While Sales Budgets Shrink
In its H1 2026 filings, Zhipu (Z.ai) posted revenue of 954 million RMB, up 399.7% year over year — while selling and marketing expenses fell 14.8% to 178 million RMB. MiniMax did the same in the same quarter: revenue of 786 million RMB, up 283.1%, with selling and distribution costs down 17.9% to 181 million RMB.
There is no precedent for this curve in traditional software. SaaS companies spend 20-40% of revenue on sales precisely because customers need to be persuaded that they need the product. AI companies are discovering something different: when the product is intelligence itself, the product does the selling.
**The salesforce is being replaced by the model. That is the real story in these filings.**
## From Deals to Self-Serve
Look at where the money actually comes from.
A year ago, Zhipu's revenue was dominated by on-premise deployments — shipping the model into enterprise data centers at a few million RMB per contract, with revenue recognized on delivery acceptance and repurchases dependent on relationship maintenance. That business requires an army of account executives.
In H1 2026, Zhipu's open platform and API revenue grew from 29.1 million RMB to 825 million RMB — now 86.5% of total revenue. Customers no longer wait for a sales visit. They register, top up a balance, and pay per token. In the company's own words, revenue has shifted from one-time recognition to ongoing, and the company now has "foreseeable recurring revenue" for the first time. It also put GLM Coding Plan subscriptions on Tmall, sold like any consumer product: Lite at 118 RMB per month, Pro at 538, Max at 1,078. Distribution has converged on e-commerce economics.
MiniMax took a different route to the same destination. Known for consumer products like Hailuo, it saw open platform and enterprise services revenue grow 703% to 497 million RMB — overtaking its consumer apps to become the largest revenue source at 63.4%. Developer and enterprise accounts passed 2 million, a 10x increase since late 2025; July token consumption was 20x January; annualized revenue run-rate crossed $800 million in August, with roughly 80% from business customers. Nobody cold-called 2 million developers. They came because the model worked.
## The Sales Function Doesn't Disappear — It Moves
The instinct here is to read falling sales costs as belt-tightening. That would be a mistake, because the sales function hasn't vanished. It has migrated along the customer lifecycle.
API adoption is nearly frictionless. Switching from one foundation model to another is a few lines of code for shallow use cases — which means those customers have essentially zero loyalty. Price slightly lower elsewhere, and they are gone.
Deep integration is the opposite. Once a model sits inside core production workflows — proprietary APIs, fine-tuned weights, embedded tooling, production data flows — switching becomes painful. Zapier's 2026 survey found that among enterprises that tried to switch AI vendors, only 42% said the process went smoothly. The other 58% failed or paid far more than expected. Zhipu's own numbers show the deep-binding trend: token volume up more than 40x since the start of the year, paying daily active users up 603%, top-ten customers' daily usage up 98x.
So the commercial motion splits. Nothing to sell at the front of the funnel; a great deal to do at the back. What deep customers need — API stability commitments, fine-tuning support, industry solutions, prompt escalation — used to be brokered by account executives. Now it must be delivered by customer success engineering, documentation, and developer community at scale.
That is the actual risk in these numbers. If the cut sales budget was redeployed into solutions architects and post-sales support, this is a precise reallocation. If it was a blanket reduction while the company waits for organic growth, renewal risk accumulates silently — especially as B2B revenue share climbs and enterprise customers demand SLAs, security audits, and dedicated response that no API endpoint provides on its own.
## The Intelligence Cost Curve
Here is a way to frame what is happening across the industry.
**The Persuasion-to-Product curve:** in enterprise software, the cost of acquiring a customer historically lives in people — sales and marketing spend that scales roughly with revenue. AI foundation models are the first software category where the acquisition cost is carried by the product itself. A model good enough makes the adoption decision for the customer: not adopting it is the competitive risk. The decision cost approaches zero, so the persuasion spend approaches zero.
The filings confirm it. Zhipu spends 12x more on R&D (2.13 billion RMB) than on sales; MiniMax spends 11x more (2.0 billion RMB). These are not software companies with unusually small sales teams. They are a different category of company — closer to a utility selling intelligence per token than a vendor selling licenses.
The curve has a corollary: the basis of competition moves downstream. When adoption is self-serve and switching shallow usage is trivial, revenue retention depends on how deeply customers embed the model, and how well the vendor supports that embedding. Sales' OKRs shift from new logos to net revenue retention. The role that remains looks less like a salesperson and more like a hybrid of engineer and consultant.
## What To Do With This
If you run an AI or SaaS company: audit which of your acquisition costs exist to persuade customers of a need they already feel. In AI-adjacent categories, that spend is being competed away by products whose capability sells itself. Redirect it toward retention and expansion infrastructure — docs, SDKs, customer success engineering — before a competitor does.
If you sell into enterprises as a vendor or reseller: your value is no longer brokering access to a model. It is the deep-integration work that creates switching costs — workflow embedding, data pipelines, compliance, SLAs. The shallow-integration middleman is being disintermediated by the API itself.
If you are an investor: watch the ratio of R&D to sales expense at AI companies, but watch the customer success function even more closely. A lab cutting sales while keeping light post-sales support is a company with a fragile revenue base at any ARR figure. The filings show who cut; only the next renewal cycle will show who cut correctly.
If you are a developer or buyer of AI: your switching cost is your leverage, and it is decaying by the quarter. Decide deliberately how deep to integrate, because every workflow you wire into one vendor's API is a future negotiation position — and every shallow usage you leave multi-homed is free optionality.
The model is the salesforce now. The people who used to sell the model are being hired back to keep it sold.
---
*Sources: Zhipu (Z.ai) and MiniMax H1 2026 financial disclosures as reported by 36Kr (Sept 15, 2026); Zapier 2026 enterprise AI vendor-switching survey.*