The selloff in global tech stocks on October 8 can be traced back to a single disclosure standard.

According to a report by the Financial Times, OpenAI told investors its annualized revenue reached roughly US$50 billion as of the end of September, about US$20 billion less than the roughly US$70 billion figure that had circulated widely in the media a month earlier.

After the news broke, the Nasdaq fell 1.25%, the Philadelphia Semiconductor Index dropped 3.39%, Nvidia (NVDA.US) slid 2.94% and lost more than US$168.4 billion in market value overnight, Oracle (ORCL.US) fell 5.48%, Micron, SK Hynix and SanDisk each dropped more than 4%, and in the optical communications sector Applied Optoelectronics plunged 13.58% while Coherent lost 9.63%.

But the most important half-sentence in that report comes later: the US$20 billion gap mainly stems from a difference in how revenue is counted, and does not mean OpenAI's actual revenue suddenly declined.

The difference lies in the fact that companies like Anthropic include sales revenue from cloud partners such as AWS and Google Cloud in their annualized revenue, while OpenAI does not adopt this method; looking at the company's own growth rate, OpenAI's overall third-quarter revenue run rate grew 77%, and its enterprise business run rate grew 107%.

The same fact, measured with different rulers, can produce a US$20 billion gap and a big red candle. This is precisely where this supply chain is most easily misread right now — the market still has no unified standard for metrics such as "annualized revenue," "backlog orders" and "computing capacity."

Take the most direct example. Three overseas pure-play computing rental providers use three different sets of terms: CoreWeave (CRWV.US) talks about revenue backlog, equal to remaining performance obligations plus estimated revenue to be recognized, and voluntarily discloses how much of that has already begun delivery; Nebius (NBIS.US) talks about RPO (remaining performance obligations, a statutory disclosure metric under accounting standards ASC606/IFRS15) and contracted backlog, and discloses the amount of each major order, the yield per megawatt, and the proportion of capital expenditure covered by customer prepayments; IREN (IREN.US) has the clearest tiering — recognized revenue, annualized operational capacity (operating ARR), and signed but not yet operational (contract ARR). At the same company, these three tiers of figures can differ by more than a hundredfold.

Across all three frameworks, "signed" and "received" are kept clearly separate. In China, by contrast, most companies use the single term "orders in hand" to cover every status.

GBA AI COMP (01396) took a relatively restrained approach in its 2026 interim results: intended AI computing orders in hand exceeded 37 billion yuan, of which more than 20 billion yuan had been delivered and was generating stable billing; more than 95% of new orders this year carry customer prepayment terms, and more than 95% are long-term contracts of three to five years. Only when both figures are given together can readers know how much is reserve and how much is realization.

Range Intelligent Computing Technology Group Company Limited (300442.SZ) disclosed a physical metric — AIDC revenue of 1.995 billion yuan, up 126.24% year on year; VNET (VNET.US) used "total operational capacity of base-type IDC of 1,007 megawatts, signed rate of 96.3%, installed capacity of 744 megawatts, and installed rate of 73.9%"; GDS (09698) described things in terms of signed capacity and a domestic investment plan of 30 billion to 50 billion yuan over the next three years.

Companies such as Sharetronic Data Technology Co.,Ltd. (300857.SZ) and Jiangsu Lettall Electronic Co.,Ltd. (603629.SH), which expanded from manufacturing or electronics into this field, tell yet another story — the share of computing business revenue in total revenue.

On October 8, the parts of the A-share and Hong Kong markets that fell were precisely those being dragged along by this confusing mix of standards.

That day, the Shanghai Composite fell 0.79%, the ChiNext Index dropped 3.15%, the STAR 50 fell 4.82%, and turnover expanded to 1.68 trillion yuan; Hong Kong's Hang Seng Tech Index fell 2.89%, hitting a new low since September 2024; CPO, semiconductor and optical communications stocks led the decline, with Yuanjie Semiconductor Technology Co.,Ltd. (688498.SH) and others hitting the daily limit down.

Yet in that same week, this company disclosed that its first-half data center revenue had grown 640% year on year.

Share prices are driven by standards and expectations, while revenue is driven by delivery. The two were never synchronized in pace to begin with.

So before comparing valuations across such names, do one thing first: translate each company's "orders" into the same ruler — how much is signed, how much is delivered, and how much has actually been paid for.

After translating, then compare. You will usually find that the multiple relationships look completely different from what you saw at first glance.

This is also why, from the same interim report data, some people read "better than expected" while others read "pie in the sky." The disagreement is not in the data itself, but in which layer of the standard you are reading.