MiniMax and Z.ai: positive API margins, substantial company losses

MiniMax and Z.ai’s first-half 2026 results answer one part of the commercialization question: APIs and subscriptions generate revenue, and service efficiency can improve gross profit. They leave another unresolved—how those profits cover training, research and the broader operating base.

MiniMax reported $116.6 million of revenue, up 283.1%, at a 17.9% gross margin. Z.ai reported RMB953.9 million, up 399.7%, at 26.4% overall and 24.6% API gross margins. Different currencies and classifications are retained. The useful comparison is the business model and funding requirement, rather than an artificial converted ranking. ^[1][2][4]

Recurring usage changes the revenue mix

MiniMax’s open platform and other enterprise services generated $73.93 million, or 63.4% of revenue versus 30.3% previously. API use, paying accounts and Token Plan contributed. AI-native product revenue still rose 100.9% to $42.64 million; 60.8% of revenue came from outside mainland China. These are distinct routes to monetization, not a single token-resale business.

Z.ai's switch was larger. Open Platform and API moved from 15.2% of revenue a year earlier to 86.5%, while enterprise general-purpose large-model revenue fell 54.6% to RMB67.0 million. On-premise deployment fell from 73.7% of the prior full-year mix to 13.5%. Revenue is moving from project acceptance and one-time licenses toward metered calls and subscriptions, tying pricing, retention, concurrency and service reliability together.

Z.ai says MaaS token volume was more than 40 times its level at the start of 2026 by the announcement date, Coding Plan volume was more than 23 times higher and average API selling price rose about 101%. Volume and price rose together, but these remain company-defined statistics without independent cohort retention, net-price or promotion data.

August ARR cannot be added to first-half earnings

Yicai attributed an August ARR above $800 million, roughly 80% enterprise, to MiniMax management. Caixin attributed $1.6 billion for Z.ai to annualizing August revenue. Neither appears in the interim filing, and full refunds, discounting, retention and duration bridges are unavailable. The media attribution remains material. ^[5][7]

ARR annualizes a current month or contract state. It is not first-half recognized revenue, cash collection or guaranteed revenue for the next twelve months. The figures help measure August velocity; they should not be added to reported revenue or compared across the companies without aligned definitions.

The next filing must test three conversions: ARR into quarterly revenue, API margin durability and contract liabilities plus operating cash flow. If ARR rises without a matching movement in recognized revenue, margin or cash, promotions, customer concentration or the annualization method may explain part of the signal.

How lower inference cost reaches API margins

Z.ai’s interim announcement describes domestic chips as its main inference resource, a cluster above 100,000 chips and an 80% decline in unit token inference cost from year-start. Its API gross margin reached 24.6%. Pricing, mix and scale also affect margin, so the entire improvement cannot be attributed to hardware substitution. ^[4]

Call reporting adds that owned clusters, leased compute and purchased services cover both training and inference. More than 100,000 chips therefore does not mean 100,000 company-owned assets or provide a direct capex estimate. Suppliers, effective online chips, power draw, throughput, SLA and the cost mix across sourcing modes remain undisclosed.

MiniMax discloses only that infrastructure efficiency lifted total gross margin to 17.9%. Yicai reported management saying M3 and H3 are being adapted to domestic chips, that a large domestic-compute cluster is expected to come online and that M3.1 targets inference cost at roughly one third of M3's launch level. Those are plans and targets, not yet an operating cluster-cost-margin bridge comparable with Z.ai's filing.

1. 01
   Tokens enter billing

   API calls and subscriptions turn model use into recurring revenue.

2. 02
   Clusters absorb concurrency

   Models, engines, networks and scheduling determine whether installed compute creates effective tokens.

3. 03
   Unit cost falls

   Scale and engineering must cover cloud services, chips, networking and operations.

4. 04
   Margin tests the economics

   Price and cost converge in segment gross margin; cash flow remains the next gate.

Training still requires outside capital

MiniMax’s $296.9 million research expense rose 138.8%, mainly through training cloud costs. Company-defined cash balance was $1,322.8 million and bank borrowing $133.6 million. July net equity placing proceeds of about HK$9.444 billion and convertible proceeds of about HK$6.433 billion added funding. The broad cash measure includes financial assets, restricted cash and time deposits; it is not all unrestricted cash equivalents. ^[1][2]

Z.ai spent RMB2.131 billion on R&D, up 33.6%. It ended June with RMB3.994 billion of cash and cash equivalents and RMB2.225 billion of bank loans, while adjusted net loss was RMB1.964 billion. Positive API gross profit has arrived, but listing proceeds and debt still support a larger R&D and operating base.

Both companies reported no significant or outstanding period-end capital commitments. That does not imply low future compute demand. Cloud services, leases and purchased compute can be recognized as operating expense, short-duration contracts or other arrangements. Cash, debt, cloud-service expense, lease liabilities and later financing need to be read together.

Measure the gap using gross-profit dollars

MiniMax disclosed roughly $20.8 million of gross profit against $296.9 million of research expense, about 7.0% coverage. Z.ai reported RMB251.611 million of gross profit and RMB2,131.171 million of research expense, or about 11.8%. Neither comparison covers all selling, administrative, financing and other items; neither is a net-profit model. ^[2][4]

Holding first-half research spending and margins fixed, gross profit sufficient only to cover research would require about $1.659 billion of MiniMax revenue at 17.9%, or RMB8.073 billion for Z.ai at 26.4%. These static sensitivities are not forecasts or cash break-even targets. Margins, spending and non-cash items can all change.

Mix can also produce apparently conflicting signals. Z.ai’s API margin improved from negative to 24.6%, while group margin fell from 50.0% to 26.4% as cloud revenue became dominant. The group percentage does not establish deterioration in API economics; segment gross-profit amounts matter.

MiniMax’s net loss narrowed 11.0% to $358.0 million while adjusted loss widened 111.2% to $293.0 million. Z.ai’s net loss narrowed 12.1% to RMB2.072 billion while adjusted loss widened 12.1% to RMB1.9641 billion. Different directions warn against treating narrower statutory losses as lower recurring spending. Neither loss measure equals operating cash burn.

Compare incremental gross profit with incremental research

Model research is not a factory that can be completed and left unchanged. New models, post-training and product competition require repeated investment. Positive gross margin does not establish steady-state earnings; incremental customer contribution must increasingly cover the cost of remaining competitive.

MiniMax can reuse model development across consumer products and its platform, but monetization differs. Consumer offerings require distribution and retention; APIs require developer integration, concurrency and reliability. Shared research can improve reuse while preventing simple product-level profit allocation from group accounts.

Z.ai’s shift from local projects to cloud APIs changes cash timing. Project revenue may follow milestones, while usage follows consumption and settlement. More recurring activity does not necessarily mean advance payment. Credit, refunds, stored balances and settlement terms determine collections.

MiniMax says July token consumption was 20 times January’s. Comparing that monthly pace with first-half revenue cannot establish a realized token price. Z.ai’s greater-than-40-times activity at announcement likewise differs from the half-year revenue period. Consistent time windows and paid-versus-free classification are required. ^[2][4]

In an illustration, doubling volume while cutting price 20% increases revenue 60%. If unit direct cost also falls 20%, margin can remain unchanged; a cost decline of only 10% pressures it. At larger volumes, small unit differences create substantial profit or loss, making token growth alone inadequate.

Trial accounts, stable production customers and very large buyers have different retention and support costs. More accounts can diversify demand while raising operating complexity. Repeatable paid production work is stronger support for long-term capacity than free traffic peaks.

Cash balances should not be divided mechanically by adjusted losses to estimate runway. Restricted and investment assets, non-cash items, accruals and subsequent financing differ. Use available liquidity and scheduled payments for training, inference, interest and capital investment.

The favorable trajectory is larger incremental gross profit, falling research intensity and improving operating cash. Rapid sales and better service margins can coexist with even faster training expenditure. Commercial progress and incomplete capital returns can both be true.

Revenue supports demand, not a direct capacity conversion

Financial statements give inference demand a revenue basis. Z.ai describes operating domestic-chip capacity, while MiniMax serves international customers. Overseas sales do not establish an owned facility in every customer region: owned clusters, colocation, leases and cloud purchasing send demand to different asset owners.

Public disclosures omit compute per token, average sequence length, cache hit rate, batching efficiency, chip models and PUE. Forty-fold and twenty-fold token growth cannot be translated into forty-fold and twenty-fold GPUs or megawatts. Lower unit cost, model architecture and utilization can absorb a substantial share of the increase.

Over the next 12 to 36 months, the discriminating signals are sustained API-margin expansion, adjusted losses shrinking with revenue, effective online domestic-cluster scale and cost per completed task, and ARR converting into recognized revenue and operating cash flow. Improvement across all four would support financeable, durable data-center capacity.