MetaX reported RMB612 million of H1 profit. Excluding nonrecurring items, it lost RMB49 million

MetaX cofounder, CTO, and chief hardware architect Peng Li presents the Xiyun C600 at WAIC in Shanghai on July 27, 2025. The photograph is from MetaX's official event account.[4]

In the photograph, MetaX cofounder and chief hardware architect Peng Li holds up the Xiyun C600 GPU under the violet lights of the 2025 World Artificial Intelligence Conference. The chip entered mass production in May 2026. It is the kind of tangible milestone that makes a young semiconductor company's progress easy to narrate: launch the next accelerator, move it into production, and let higher revenue turn engineering into earnings.[1][4]

MetaX's first-half accounts do not move in that simple sequence. They show RMB612.45 million of net profit attributable to shareholders, but also RMB661.31 million of net nonrecurring gains and losses. On the regulator-defined line that removes those items, MetaX still recorded an RMB48.86 million loss.[1]

Those figures are not competing estimates. All three appear in the same company filing. As of September 3, 2026, the useful dossier is therefore not “profitable or unprofitable?” It is a reconstruction of which clock just turned, which one only appeared to turn, and what would make the distinction durable.

The bridge is complete inside the filing

China's securities regulator defines nonrecurring gains and losses as transactions that are unrelated to normal business, or that are connected to it but sufficiently special or occasional to interfere with a normal reading of operating performance. Public companies must make the classification with reference to economic substance, industry characteristics, business model, and materiality.[3]

MetaX did not invent an unofficial adjustment after the result. Its half-year report supplies a formal reconciliation. The largest starting item was RMB926.28 million from holding and disposing of financial assets, before the related tax and other offsets. Inside that amount, RMB887.07 million was a fair-value change on trading financial assets. MetaX states that this one line equaled 105.75% of total pretax profit and marks it as not sustainable.[1]

The rest of the bridge is smaller but still important. The filing includes RMB9.34 million of qualifying government subsidies and RMB0.60 million of asset-disposal gains; it also records negative RMB14.07 million of other non-operating items, negative RMB39.01 million of other qualifying items, and deducts RMB221.83 million for the tax effect. The resulting net nonrecurring amount is RMB661.31 million.[1]

The arithmetic then closes without a residual:

RMB612.45 million reported attributable profit minus RMB661.31 million of net nonrecurring items equals a RMB48.86 million attributable loss excluding those items.

That does not make the reported profit fictitious. Fair-value changes belong in the accounts under the applicable rules, and an asset that rises in value improves the holder's balance sheet. It does make the economic source different from selling more GPUs at a margin. MetaX itself draws that boundary by calling the RMB887.07 million fair-value gain unsustainable. The headline result measures the whole reporting entity; the deduction line is a narrower earnings-quality test. It is not operating profit—it still includes financing, tax, impairment, and share-based-compensation effects—but it removes the items MetaX classified under the regulator's nonrecurring rule.[1][3]

A second-quarter turn is hidden inside the half

The half-year total obscures the strongest underlying earnings evidence in the filing. MetaX's first-quarter report recorded RMB561.92 million of revenue, a RMB98.84 million attributable loss, and a RMB102.93 million attributable loss excluding nonrecurring items.[2] Subtract those exact first-quarter figures from the cumulative first-half figures, and the implied second quarter looks different:

revenue of RMB761.68 million;
reported attributable profit of RMB711.29 million; and
attributable profit excluding nonrecurring items of RMB54.07 million.

The last figure is the consequential one. It suggests that MetaX crossed the narrower profit line during the second quarter even though the six-month total remained negative on that basis. It also prevents an overly bleak reading of the half-year result: the earnings direction improved by more than the headline “still a loss” admits.[1][2]

But RMB54.07 million is a calculation from two cumulative reports, not a separately labeled quarterly measure published by the company. It represents one quarter, and both reports were unaudited. A single positive interval can reflect shipment timing, product mix, expense timing, impairments, or tax effects that do not repeat. The number earns a place in the dossier because the subtraction is reproducible; it does not yet earn the status of a settled run rate.

Revenue carries more operating weight than the profit headline

MetaX's first-half revenue reached RMB1.324 billion, up 44.67% from a year earlier. Cost of revenue rose 40.92% to RMB566.30 million. Calculated directly from the filing, that puts gross margin near 57.2%, compared with about 56.1% in the prior-year half. Meanwhile, the loss excluding nonrecurring items narrowed 75.83% from RMB202.15 million.[1]

These figures suggest improving operating leverage without isolating it: sales grew faster than their direct cost, gross margin edged higher, and the regulator-defined loss contracted sharply. They also have to carry a large research organization. MetaX spent RMB524.81 million on R&D in the half, up 15.28% and equal to 39.65% of revenue. Sales expense grew faster than revenue, rising 73.23% to RMB126.91 million as the company expanded marketing, ecosystem work, and share-based compensation for sales staff.[1]

The hardware chronology offers a plausible commercial backdrop, not a product-level explanation. MetaX says its N100 inference line entered mass production in April 2023, the C500 training-and-inference line in February 2024, and the C600 in May 2026. The company attributes the half's revenue growth to higher GPU shipments, but does not disclose units, average selling prices, revenue by accelerator generation, or gross margin by product.[1]

That missing split matters. C600 was in mass production for only the final two months of the reporting period. The accounts cannot tell an outside reader how much of the second-quarter improvement came from C600, mature C500 and N100 products, customer mix, or expense timing. The launch photograph proves that the product existed onstage; the May production date moves it one step further. Neither proves that C600 caused the quarterly turn on the profit-excluding-nonrecurring measure.[1][4]

Cash kept a slower clock

Operating cash flow was negative RMB1.297 billion, widening from negative RMB882.90 million a year earlier.[1] That outflow was almost as large as the half's reported revenue and more than twice the headline profit. An automated financial-statement screen published by Sina Finance also flagged the divergence between earnings and operating cash, along with the scale of receivables relative to half-year revenue.[6]

The balance sheet shows where some of the pressure accumulated. Trade receivables were RMB969.23 million at June 30, up 33.55% from year-end. Prepayments climbed 163.24% to RMB2.196 billion. Inventory remained substantial at RMB1.432 billion, although it was below the RMB1.496 billion year-end balance. MetaX attributes the cash outflow to higher purchasing and, in its risk discussion, to the continuing loss excluding nonrecurring items, collection timing, and procurement intended to protect supply.[1]

This is not an immediate-liquidity argument. MetaX also held RMB4.352 billion of cash and RMB4.351 billion of trading financial assets at period end. Rather, it is an earnings-quality boundary. The RMB887.07 million fair-value remeasurement is not cash collected from GPU customers, while receivables and prepayments absorb funds before they become—or without necessarily becoming—operating cash.[1]

For the next reporting period, cash conversion is at least as informative as another profit headline. Receivables growing alongside revenue can be ordinary for a scaling supplier. It becomes a problem if collection days lengthen, credit losses rise, or purchasing commitments continue to outrun customer cash.

Day-0 releases are a supporting ledger, not a profit bridge

MetaX's software cadence supplies evidence that the operating organization kept building while hardware volumes grew. In the half-year filing, the company says its MXMACA stack supports more than 6,000 CUDA applications and has native adaptations for more than 1,000 models. It also reports more than 110 million network API calls to its open community between the community's February 2025 launch and July 28, 2026.[1]

On August 28, MetaX announced release-day support for GLM-5.3 and said it had completed 35 Day-0 model adaptations since December 2025. That is a vendor-reported release ledger: it shows that the team can respond repeatedly to upstream model changes and publish deployment paths on its own accelerator stack.[5]

It does not measure inference speed, accuracy, stability, utilization, or customer revenue on the resulting workloads. An adaptation announcement can range from a runnable image to a deeply optimized production deployment. MetaX's aggregate API-call count likewise lacks unique-user, workload, and paid-usage denominators. These disclosures are evidence of engineering cadence, not a basis for assigning the RMB54.07 million calculated second-quarter result to software compatibility.

The distinction protects both sides of the analysis. Ignoring software would miss a necessary part of selling non-Nvidia GPUs. Treating every supported model as commercial proof would repeat the accounting error in a new form: mixing a useful but differently defined measure into the result one wants to explain.

What would make the turn repeatable

Three tests can resolve the ambiguity without waiting for another spectacular fair-value move.

First, MetaX needs to report more than one quarter of positive attributable profit excluding nonrecurring items. The calculated second-quarter crossing is the best current signal; consecutive positive quarters would show that it was not merely shipment or expense timing.

Second, the cash ledger must begin to converge. Faster customer collections, slower growth in prepayments, and operating cash outflow narrowing relative to revenue would turn accounting progress into funded operations. A renewed loss excluding nonrecurring items combined with another large cash outflow would falsify the earnings-turn thesis even if reported profit stayed positive.

Third, product disclosure needs a denominator. C600 shipment volume, revenue mix, repeat orders, or a stable product-level margin would connect the May production milestone to the accounts. Customer workload evidence could do the same for MXMACA: pinned versions, hardware configuration, sustained utilization, and repeat paid use are more informative than another cumulative adaptation count.

MetaX's first half contains a real improvement and a real optical effect. The optical effect is the easy one: a financial-asset gain pushed reported profit well above the result generated after nonrecurring items were removed. The improvement is narrower but more valuable: subtracting the first quarter indicates that the latter measure turned positive in the second.

The next filing does not need a larger headline. It needs the slower clocks—profit excluding nonrecurring items and operating cash—to keep moving in the same direction.

cronfeed.work