AI Firms Do the Math: The Burn Race Meets the Payback Cycle

In September 2026, AI giants in China and beyond are filing their first-half results, and a subtle shift is underway: on earnings calls, executives are talking less about "how much we invested" and more about "when we pay it back." Tencent's second-quarter capex hit 52.78 billion yuan, up 176% year over year; Alibaba CEO Eddie Wu projected AI-related capex would reach breakeven "within roughly three years"; Zhipu's first post-IPO semi-annual report showed revenue of 954 million yuan, up 399.7%, yet losses still stood at 2.07 billion yuan. The burn race has collided with the payback cycle, and the large-model industry is collectively doing the math.

From "How Much We Spent" to "When We Pay Back": A Shift in Language

For the past two years, AI capex was how tech giants demonstrated commitment — the bigger the spend, the more compelling the story. But in the 2026 H1 earnings season, the tone has clearly changed. Wu gave a concrete timeline on the earnings call: at current average gross margins, AI-related capex could reach breakeven within three years, with the payback period potentially shrinking to around 2.5 years. Tencent president Martin Lau stressed that allocating existing compute resources to Tencent Cloud's leasing services "can generate substantial revenue and a significant return on capital expenditure."

The backdrop is that spending has grown too large to ignore the return question. Tencent's Q2 capex rose 176% year over year and free cash flow turned negative; Baidu's AI business has held above 50% of general business revenue for two straight quarters; Alibaba Cloud's Q2 AI cloud and compute revenue reached 48.44 billion yuan, with adjusted EBITA surging 133% year over year. The investment story is giving way to a return timeline — marking AI investment's shift from "faith-driven" to "math-driven."

Same Story Worldwide: The Seven-Hundred-Billion-Dollar Test

This reckoning is not unique to China. According to Silicon Analysts, the four major U.S. hyperscalers — Microsoft, Alphabet, Amazon, and Meta — spent a combined $433.9 billion on property and equipment in the four quarters through March 2026, against only about $149 billion in depreciation over the same span. Their 2026 guidance totals roughly $700 billion. Amazon's capex has exceeded its trailing operating cash flow (102%), and Meta priced a $30 billion bond in October 2025 — one of the largest investment-grade deals of the year.

The sharper issue is the "depreciation wall": servers depreciate over 5 to 6 years and buildings over 25 to 40 years, so today's income statements carry only a fraction of today's build-out costs. That wave arrives in 2027 to 2028 regardless of AI revenue performance. As LPL Research notes, the market's core question is no longer "how much is being deployed" but "whether today's capital will earn attractive returns over the life of the assets being built" — the essential question of the accounting era.

The Twin Model Makers Pivot: MaaS Becomes the Lifeline

For pure model companies without advertising, gaming, or e-commerce to subsidize AI, the reckoning is more urgent. Zhipu and MiniMax — the "twin stars" of Chinese large models — both completed their pivot to MaaS in H1 2026. Zhipu's open-platform and API revenue reached 825 million yuan, up 2735.7% year over year, with revenue share leaping from 15.2% to 86.5% and API gross margin turning positive from -0.4% to 24.6%. MiniMax's B-side open-platform revenue grew 703.1%, its share rising to 63.4%, while C-side share fell to 36.6%.

Notably, both still spend roughly twice their revenue on R&D, and losses, while narrowing, have not stopped: Zhipu lost 2.07 billion yuan (down 12.1%) and MiniMax lost $358 million (down 11%). Meanwhile, Zhipu's ARR surpassed $1.6 billion in August, with token calls up more than 40x since the start of the year, and MiniMax's July token consumption was 20x January's. Scale is growing, losses are narrowing, margins are turning positive — but "when will we be overall profitable" remains the question hanging over every model company.

Exploring Business Models: E-Commerce Shelves and Overseas Revenue Sharing

The pressure to account has also driven aggressive business-model experiments. On September 3, Tmall launched an AI Space Station (Token Recharge Center), putting subscription products from Alibaba Cloud, Zhipu, Kimi, MiniMax, and others onto e-commerce shelves — Token Plans and Coding Plans sold by subscription period or pay-as-you-go, pushing the industry from project-based revenue toward recurring subscriptions. A bolder experiment is happening overseas: Moonshot AI is in talks with cloud giants including Microsoft, Amazon, and Google about revenue sharing, seeking to launch a new model abroad — not selling tokens from self-built compute, but taking a cut through cloud channels.

Forrester vice president Dai Kun notes that putting Token Plans on e-commerce shelves is essentially productizing, standardizing, and retailing capabilities that were previously enterprise procurement. It cultivates pay-as-you-go habits and signals that the competitive focus is shifting from model capability itself to application enablement, cost control, package design, and customer operations — in other words, model makers must not only build, but also sell.

After the Reckoning: Three Certainties and One Open Question

The collective accounting brings three certainties. First, return on investment becomes the new yardstick for funding and valuation — the "burn money for scale" narrative no longer works indefinitely. Second, MaaS/API has become the consensus revenue pillar for model companies, with business models converging from chaos. Third, cost control and pricing power become the core competitive dimensions of the second half, with domestic chip adaptation and architectural innovation all serving cost reduction.

The open question remains: a 36Kr headline put it bluntly — "revenue up 4x, losses of 2 billion, Zhipu hasn't proven itself yet." The depreciation wall arrives in 2027 to 2028, and whether Moonshot's revenue-sharing model works is still undecided. The reckoning is not the end — it is the industry's coming-of-age ceremony. When "when will it pay back" becomes a required question on every earnings call, it means this industry has finally started speaking the language of business.