VNET’s 1-Gigawatt Quarter Put AI’s Cost of Scale on the Books

AI infrastructure crossed a useful accounting threshold on 18 August. VNET reported more than one gigawatt of wholesale data-centre capacity in service, backed by a near-full order book. The same accounts also showed why a megawatt is not a margin: depreciation from rapid expansion pushed reported gross profit down even as revenue rose.

Two other disclosures made the pattern harder to dismiss as a single-company issue. Pony.ai’s robotaxi revenue grew rapidly from a small base while operating loss still exceeded total revenue. WhiteFiber proposed a new convertible-note issue for data-centre expansion and, unusually plainly, said the proceeds would not remove its need for project finance.

In brief, for the window from 18 August 09:02 to 19 August 09:02 in Tehran:

• VNET reached 1,007 MW of live wholesale capacity and 970 MW of customer commitments at 30 June.
• VNET’s revenue rose 14.2%, but GAAP gross margin fell from 22.5% to 18.2% as depreciation increased.
• Pony.ai’s robotaxi revenue rose to $12.1 million, yet the group recorded a $65.7 million operating loss on $36.2 million of total revenue.
• WhiteFiber proposed $250 million of new unsecured convertible notes, plus a possible $37.5 million option, partly to exchange existing notes and partly to fund expansion.
• The common signal is not weak demand. It is that booked AI demand, delivered capacity, utilization, accounting profit and cash funding are different stages of the same capital cycle.

The evidence ledger

VNET crossed one gigawatt, but depreciation crossed the income statement too

VNET’s 18 August results put 1,007 MW of wholesale capacity in service at 30 June, up 49.4% from a year earlier. Customers utilized 744 MW, up 45.5%, while effective agreements covered 970 MW. The business added 100 MW of live capacity and 57 MW of utilization during the quarter.

That translated into RMB2.78 billion ($409.5 million) of revenue, 14.2% higher year on year. Wholesale revenue grew faster, by 29.3% to RMB1.10 billion, and overtook retail as the company’s main growth engine. Adjusted EBITDA rose 25.4% to RMB918.3 million and its adjusted margin expanded to 33.0%.

The statutory accounts tell the necessary second half of the story. Gross profit fell 7.8% to RMB505.2 million and gross margin narrowed 4.3 percentage points to 18.2%. VNET attributed the decline mainly to higher depreciation from rapid capacity expansion. Adjusted EBITDA removes depreciation and amortization; it is useful for comparing operating cash generation, but it cannot make the assets free.

VNET’s Form 6-K filed on 18 August also shows the financing load around that build. The company held RMB7.21 billion of cash, restricted cash and short-term investments at quarter-end, against RMB4.18 billion of short-term debt and RMB19.24 billion of long-term debt. It raised or refinanced RMB3.77 billion during the quarter and kept its 2026 capital-expenditure plan at RMB10 billion to RMB12 billion.

None of those figures says the expansion is uneconomic. They say that a valid return calculation needs depreciation, financing, commissioning and ramp time beside the order book. Our earlier analysis of Nebius capacity prepayments and deployment gaps made the same distinction from the customer-funding side; VNET’s quarter shows it from an operator’s accounts.

Commitments are not utilization

VNET’s commitment rate was 96.3%, but utilization was 73.9%. That 22.4-point difference is not automatically a demand problem. A customer can sign before halls, power, networking and servers are fully accepted, and new capacity naturally depresses utilization while it ramps. The distinction still matters because depreciation and interest can begin before contracted megawatts produce a full revenue run-rate.

The company classified mature wholesale capacity at 92.5% utilization and ramp-up capacity at 36.6%. Those are more decision-useful figures than a single fleet-wide average. They let an operator ask whether a lower rate is concentrated in newly commissioned assets or spreading into mature sites.

For each campus, the operational bridge should therefore reconcile:

• land and grid rights secured;
• megawatts under construction;
• capacity technically commissioned;
• effective customer commitments;
• customer acceptance and revenue start;
• actual power draw and compute utilization; and
• cash contribution after power, maintenance, rent and finance.

The AI production cost stack applies the same discipline at workload level: capacity becomes valuable only when a useful, reliable service consumes it at a price above its complete operating cost.

CATL moves power deeper into the compute plan

VNET also said it signed a strategic cooperation agreement with battery maker CATL. The parties plan a three-layer “compute-energy” ecosystem spanning gigawatt-scale facilities, distributed networks and what they call a zero-carbon token ecosystem. The accompanying 18 August filing exhibit describes green data centres and direct green-power connections.

The confirmed fact is the agreement. The release does not disclose committed capital, named sites, delivery dates, contracted power, storage duration or customer economics. “Plan to develop” must not be modelled as built capacity.

Operationally, the partnership still points to a real shift. Energy procurement, storage, grid connection and workload scheduling are becoming part of one infrastructure design rather than separate vendor conversations. A buyer should require metered evidence for renewable matching, storage losses, backup duration and curtailment—not rely on a “zero-carbon” label whose boundary is unspecified.

Pony.ai showed revenue scale before profit scale

Pony.ai’s Q2 release at 05:00 EDT on 18 August reported $36.2 million of total revenue, up 68.8% year on year. Robotaxi service revenue rose 691.2% to $12.1 million, helped by a seventh-generation fleet and joint deployments. The global robotaxi fleet reached 1,975 vehicles at 30 June, and capital expenditure increased to $32.2 million from $9.6 million.

Large growth rates need their denominator. Robotaxi revenue began at $1.5 million a year earlier. Group gross profit was $6.4 million, while operating expenses were $72.1 million and operating loss was $65.7 million. The operating-loss margin narrowed sharply, from 285.6% to 181.5%, but remained larger than total quarterly revenue.

Pony.ai also recognized a $25 million impairment on prepayments for long-term investments that it deemed unrecoverable after strategic changes. Its non-GAAP measure excludes that impairment. The full [financial-results exhibit](https://www.sec.gov/Archives/edgar/data/1969302/000110465926098113/tm2623382d1_ex99-2.htm) provides the reconciliation; readers should preserve the GAAP result when judging capital allocation even if the adjustment helps compare recurring operations.

The useful deployment metric is not fleet count alone. Watch paid rides per active vehicle, revenue per vehicle-hour, remote-assistance load, insurance and maintenance, geographic utilization, partner capital contributions and cash payback on each fleet generation. A world model may reduce engineering work across cities, as Pony.ai says, but that benefit needs to appear in those operating ratios.

WhiteFiber proposed more convertible capital—and said it would not be enough

At 16:01 EDT, WhiteFiber announced an intended $250 million private placement of senior unsecured convertible notes due 2032, with an option for another $37.5 million. Pricing, interest, conversion terms and completion were still subject to market conditions at the cutoff.

Part of the proceeds would fund cash consideration for exchanges of existing 4.5% notes due 2031. The remainder would support property, construction, energy-service agreements, GPU servers, partnerships and working capital. WhiteFiber explicitly said additional project financing, such as construction loans, would be required to complete those initiatives.

This would be the company’s second large convertible financing in 2026. Its March-quarter Form 10-Q says January’s $230 million issue produced about $102.5 million of net proceeds after a zero-strike call premium, discounts and expenses, and added roughly $10.4 million of annual cash interest.

The new announcement is therefore a financing proposal, not $250 million of new construction cash. Final usable proceeds will depend on pricing, exchange cash, fees and any hedging transactions. Conversion can reduce cash repayment pressure but may dilute shareholders; unsecured status does not remove claims on future cash flow. The right operating question is how much fully funded, power-secured, customer-backed capacity the transaction can bring to service.

The market’s message was a higher proof burden

The day’s share-price tape is consistent with that accounting focus, but it does not prove a collapse in AI demand. The Associated Press’s 18 August US close report recorded a 1.3% fall in the Nasdaq, including declines of 7.0% for Micron, 2.3% for NVIDIA and 3.2% for Broadcom. AP connected the pressure to high yields and renewed concern about whether AI investment will earn enough.

High long-term rates increase discount rates and the cost of debt just as operators need more capital. Yesterday’s NVIDIA–OpenAI guarantee analysis showed suppliers absorbing residual-value risk to unlock projects. WhiteFiber’s proposal shows a smaller operator reaching again for convertible capital. VNET shows that even strongly committed capacity creates depreciation before every megawatt is utilized.

These are linked financing mechanisms, not interchangeable credit risks. The lesson for operators is to report the bridge from contract to cash with enough granularity that lenders, customers and boards do not have to infer it from headline gigawatts.

Limits and what to watch next

The quarterlies are unaudited, company-selected snapshots. VNET’s CATL agreement contains few binding economics; Pony.ai’s overseas vehicle figures include agreements under negotiation; WhiteFiber had not priced or closed its notes by the cutoff. Currency translations are conveniences, and non-GAAP measures differ across issuers. One market session is context, not a forecast or investment recommendation.

The next evidence should be concrete:

• WhiteFiber’s final note size, coupon, conversion premium, exchange amount and net expansion proceeds;
• VNET’s ramp-up utilization, gross margin after depreciation and cash conversion against its RMB10–12 billion capex plan;
• named, financed projects under the CATL agreement rather than ecosystem language;
• Pony.ai’s paid utilization and unit economics as the fleet approaches its year-end target; and
• whether high yields change customer contract terms, supplier guarantees or commissioning schedules.

AI demand is visible in orders, rides and financing. The 18 August disclosures made the missing test equally visible: whether each new unit of physical scale can move from commitment to utilization to durable cash return before the cost of capital catches it.