MiniMax Triples Alibaba Cloud Spending Ceiling to US$1.2 Billion as Compute Hunger Bites

Three days after reporting interim results that stunned the market, Shanghai-based MiniMax has filed another set of numbers that may matter more for understanding where China’s AI industry is heading. According to Hong Kong stock exchange filings made on Wednesday, the company has raised the three-year purchase ceiling on its cloud computing deals with Alibaba Group Holding by 220 per cent, to US$1.2 billion.

The backdrop is simple and stark: MiniMax burned through roughly two-thirds of its annual Alibaba Cloud budget by the end of June, forcing a dramatic upward revision of the deal. Under the updated agreement, which now runs through 2028, MiniMax’s annual spending limits jump from US$125 million to US$400 million for 2027, and from US$135 million to US$500 million for 2028. This year alone, the company now plans to spend up to US$300 million on Alibaba Cloud services — nearly triple its original US$115 million cap.

For a company that IPO’d in Hong Kong in January and reported first-half revenue of US$116.6 million, a US$1.2 billion three-year cloud commitment is a statement of scale. It says MiniMax expects its compute consumption to keep compounding at a pace that its original agreement simply could not contain.

Why the Budget Blew Out

The revisions reflect MiniMax’s expanding compute needs for both model training and live inference — and the distinction matters. Training runs for new frontier models (the M-series large language models, the H3 video-generation model) are episodic, budgetable line items. Inference — the ongoing cost of serving every API call, every Hailuo AI video render, every agent workflow running on MiniMax models — scales with adoption, and adoption is what exploded.

The company’s own H1 2026 numbers explain the pressure. Revenue grew 283 per cent year over year to US$116.6 million, powered by a 700 per cent jump in enterprise business. By July, token consumption on MiniMax platforms had grown to 20 times its January level. When your inference traffic multiplies by an order of magnitude in six months, no cloud budget drafted the previous year survives contact with reality.

There is also a quieter line item in the filings worth flagging. In a separate revision, MiniMax expanded its API service budget with Alibaba, raising the 2026 cap from US$650,000 to US$7.5 million — an over-tenfold increase that brings its three-year API spending ceiling to US$62.5 million, a nearly 20-fold jump from previous limits. That Alibaba itself is selling API services back to one of China’s most prominent model developers is a neat illustration of how entangled China’s AI stack has become: the same hyperscaler that is a MiniMax investor and shareholder is also its landlord, utility company, and upstream supplier.

The Connected-Transaction Pattern

Because Alibaba is a related party — it led MiniMax’s 2024 financing round and participated as a cornerstone investor in the January IPO, which raised US$618 million — these cloud purchases are classified as connected transactions under HKEX rules. That is why the spending caps exist in the first place: Hong Kong’s listing framework requires caps and independent shareholder scrutiny on deals between a listed company and its investors.

The cap raise itself is routine governance; what it reveals is not. Annual caps are typically set with generous headroom. When a company burns two-thirds of a full-year cap in six months and triples the ceiling, it is effectively telling the market that its demand forecast broke upward faster than expected — and that Alibaba Cloud is capturing that demand almost wholesale.

For Alibaba, this is the revenue side of its own AI capex story. The company has committed at least RMB 380 billion (US$53 billion) over three years to AI and cloud infrastructure, and landed China’s AI challengers as anchor tenants. MiniMax’s US$1.2 billion commitment makes it one of the larger single-customer pipelines visible in that buildout.

What It Says About China’s AI Economics

The MiniMax filing is a useful datapoint in a debate playing out on both sides of the Pacific: does inference demand actually pay? MiniMax’s interim results suggest it can — gross margin improved from 12.1 per cent to 17.9 per cent as token volumes scaled, and management repeatedly frames inference efficiency as the gating factor for its “Intelligence with Everyone” strategy. But the spending filing shows the other side of the ledger: revenue growing 283 per cent against a cloud bill that is being reset three times higher.

The bet embedded in the new US$1.2 billion ceiling is that revenue growth keeps outrunning compute cost. MiniMax’s enterprise segment is growing at 700 per cent, token consumption at 20x in six months. If those curves hold, US$400-500 million a year in cloud spend is cheap rent for the position. If they flatten — and analysts noted the company is behind the pace needed to hit full-year forecasts — the same ceiling becomes a fixed cost pressing on a company still posting wide losses.

There is also a sovereignty angle that is easy to miss. Every dollar of this US$1.2 billion flows to a domestic hyperscaler. While US labs divide their spend among Nvidia-equipped American clouds, China’s model developers are consolidating onto Alibaba Cloud, Tencent Cloud and Huawei’s stack — a consequence of export controls that make Nvidia frontier GPUs scarce and expensive. MiniMax’s recent work on inference cost reduction, including its open-weight model releases, is aimed squarely at making each of those cloud dollars buy more tokens.

The Takeaway

Strip away the filing language and the story is this: one of China’s fastest-growing AI companies, seven months public, has just told the Hong Kong exchange that it needs three times the compute it thought it did. The cap raise is governance housekeeping; the demand curve behind it is the real news. Inference is no longer a rounding error in China’s AI P&L — it is the primary line item, and the clouds that own the GPUs (or their domestic equivalents) are the structural winners of this phase of the race.

Whether MiniMax’s revenue engine can keep pace with a US$500-million-a-year cloud habit by 2028 is now one of the more interesting questions in Chinese AI finance. The company itself, in the person of CEO Yan Junjie, has compressed the bet into a single line: “Intelligence can scale almost without limit; energy and compute cannot.”