Two compliance ledgers track AI across mainland China, Hong Kong and Taiwan. The developers who reach foreign models through resellers appear in neither.
AsiaPresswire – Somewhere on a mainland tech payroll sits an engineer who did nothing unusual. He pays for a Claude subscription himself, on the two hundred dollar Ultra tier, and he has been using it on company code. By the morning of 4 November 2026 he could read, from three different newsrooms, what his position is worth in the compliance system. TechNode reported that ByteDance has been tightening access to Claude and Codex across the company, and that the restriction reaches staff who pay for the tools out of their own pockets (TechNode, 2026-11-04). The consumer outlet ifanr ran a headline asking whether the sky over the relay stations, the resale shops of the AI world, was starting to fall (ifanr, 2026-11-04). GeekPark’s morning brief added that ByteDance now restricts employees from using personal AI accounts for work (GeekPark, 2026-11-05). None of the three stories printed the name of a single user. The engineer exists in the reporting as a restriction, a headline and a bullet point. This article is about why the people in his position, who number somewhere between a rounding error and a market, cannot be found in either of the two ledgers that claim to count AI in Greater China.
Three newsrooms, one workday
The paper trail of one workday shows how the situation reached him. TechNode’s report of 4 November 2026 stated the rule plainly: “Even if employees pay for Claude or Codex out of their own pockets, they may no longer be able to use these tools freely to work with company code and internal documents” (TechNode, 2026-11-04). The tightening, the report said, had been rolling across the company since the beginning of the period it covered. The verb matters more than the date. Nobody announced a policy with a document number. Employees learned the boundary the way the report described it, through access quietly narrowing, and the reporting itself was assembled from sources inside the company rather than from any published rule.
The same day, ifanr published a piece under a headline that translates as “Anthropic barely flexed, and the sky over the relay stations is already falling?” (ifanr, 2026-11-04). A relay station, zhongzhuanzhan in the trade’s own shorthand, is a reseller that buys or pools accounts on a foreign frontier model, resells access in Renminbi, and shields the end user from the payment and geography checks that the original provider enforces. The ifanr piece carried the regulatory tag, and its premise was a single company action: when Anthropic adjusts enforcement, the resale layer wobbles. The article did not publish a user count. It did not need one for the mood it was describing, which is precisely the problem this article returns to.
GeekPark’s morning brief of 5 November 2026 bundled the ByteDance restriction into a list beside two other items, the mainland test launch of Apple Intelligence and a report that Chinese customers are expected to receive only about 1 percent of NVIDIA’s graphics card orders (GeekPark, 2026-11-05). Read as a list, the brief was three separate stories. Read as a ledger, it was one row in a book where every entry is a company decision and no entry is a user. That is the shape of the whole week’s coverage, and the shape is the finding.
Table 1: What Greater China’s compliance papers counted in the week of 31 October to 5 November 2026
| Date | Document (publisher) | What it counts | What stays outside |
| 2026-10-31 | Claude quota report (QBitAI) | product quota relief by tier | where the users sit |
| 2026-11-04 | ByteDance access tightening (TechNode) | employee tool risk, incl. self-paid use | scale of personal use |
| 2026-11-04 | Relay station market mood (ifanr) | reseller reaction to enforcement | any user number |
| 2026-11-04 | Apple Intelligence China beta (TechNode) | official channel features | the existing relay cohort |
| 2026-11-04 | Bank SinoPac breach (iThome, Taiwan) | incident impact on customers | third-party tool usage |
| 2026-11-04 | MoneySmart breach (Hong Kong Free Press) | affected customer data | affected user totals |
| 2026-11-05 | Digital 2027 Singapore (e27) | measured monthly AI usage, 9 in 10 | mainland user base |
Source: compiled from QBitAI 
, TechNode , ifanr , iThome , Hong Kong Free Press and e27
How the engineer’s own week actually went, whether he moved to a domestic model, a relay plan, or a personal project abandoned at two in the morning, none of the documents say. The honest sentence is short: no relay user has spoken publicly in any of the reports cited here, under a name or a handle attached to a verifiable identity. What can be reconstructed is the position itself, and the position is documented well enough.
How large is the group, honestly
The direct answer is that nobody publishes a count, and the absence is not an accident of secrecy. It follows from how both compliance systems are built. Still, three indirect readings bound the question, and each comes with a stated limit.
The first reading comes from the provider’s own financials. Anthropic filed its S-1 prospectus with the US Securities and Exchange Commission on 26 October 2026, listing first half 2026 revenue of US$16.29 billion, against US$1.16 billion in the same months a year earlier, and a first ever quarterly operating profit of US$974 million in the second quarter of 2026 (GeekPark, 2026-10-31). An updated filing put third quarter revenue at US$16.96 billion, up 47 percent quarter on quarter, with an operating loss of US$567 million after the profitable quarter, and roughly US$895.4 billion in compute purchase commitments involving partners such as Amazon and Google (GeekPark, 2026-11-04). Growth at that pace means paid capacity is being absorbed faster than it is being added, and resale access is one of the ways capacity gets absorbed. What the filings do not contain is any geographic or channel split that would separate a relay purchase from a corporate seat. The number proves demand. It cannot prove the cohort.
The second reading comes from product quota. On 31 October 2026, QBitAI reported that Claude had removed the five hour usage cap, with the two hundred dollar Ultra tier users unbundled first, under the headline’s own phrasing that the “$200 Ultra users go first” (QBitAI, 2026-10-31). Quota relief at the top tier is what a provider does when heavy users, the kind who write code all day, are the marginal revenue. Heavy users are exactly the profile of the relay customer. Again the inference points at the group without measuring it.
The third reading is the contrast case, and it shows what a counted user base actually looks like. The Digital 2027 Singapore report, released on 5 November 2026 by We Are Social and Meltwater, found that nine in ten Singapore internet users reach for an AI tool every month, while only about one in three say they are comfortable doing so (e27, 2026-11-05). Singapore measures its users with a survey, a sample and a fieldwork window. The mainland relay cohort has none of those three things. The gap between habit and comfort that Singapore can quantify is the same gap, in the same market region, that nobody quantifies for the relay users.
One more contrast, from the economy next door, makes the blindness sharper. China’s customs authorities announced 39.94 trillion CNY in goods trade for January to September 2026, up 18.8 percent year on year, with a surplus of US$938.1 billion, and full year 2025 goods trade with ASEAN had already crossed US$1 trillion (General Administration of Customs briefing, 2026-10-14). Goods crossing the border are counted per shipment, per tariff line, per trading partner. Services and model access crossing the same border in the same weeks are counted by no one, because the relay layer was designed to make them look domestic. A customs system that resolves trade to the tariff line coexists with an AI market whose customer side has no denominator at all.
Data as of November 2026.
What the statements share
Set the week’s public statements side by side and the common thread is easy to miss because it is a null set. TechNode’s report counts employee risk. ifanr’s headline counts market mood. GeekPark’s brief counts company decisions. QBitAI counts quota. Not one of the four outlets quotes a relay user, a cancelled subscriber, or a developer who moved to a domestic model, and the reason is not hostility. It is visibility. A user inside a relay chain is, by design, invisible to the provider’s account system, to the payment trail, and to any journalist who wants an on the record interview that does not expose the user to the exact enforcement the reporting describes.
The corporate side speaks because it has documents. The regulator side speaks because it has lists. The provider side speaks because it has filings. The relay user has none of the three, only the experience the reports gesture at. In the vocabulary of evidence, every statement about this group is secondhand, and the piece you are reading inherits that limit rather than pretending to clear it.
Where the gap comes from
Two conditions, one on each side of the border, produce this cohort. Either condition alone would not. That is the argument of this section, and it is the load bearing one.
On the supply side, the door never opened
Anthropic does not sell Claude directly inside mainland China, and OpenAI has likewise kept its consumer service outside the market, which is why the relay layer exists at all. The week’s reporting showed the same structure holding at the infrastructure level. Chinese customers were reported to be on track to receive only about 1 percent of NVIDIA’s graphics card orders (GeekPark, 2026-11-05), and Washington’s trade enforcement kept tightening around the region: the Section 301 tariff program that took effect on 24 July 2026 covers 60 economies and applies a 12.5 percent rate to Singapore (Federal Register notice, 2026-07-24). A developer in Shenzhen who wants the current frontier coding agent cannot buy it the way a developer in Singapore can, at list price, on a card, with a support contract. The supply side did not merely forget mainland users. It priced them, sanctioned around them and carded them out.
Apple’s arrangement shows what official entry costs. To ship Apple Intelligence in the mainland, Apple partnered with Alibaba and routed the service through Qwen, the Alibaba model, opening a developer beta on 4 November 2026 alongside the iOS 27.3 test build, with a public release planned for December 2026 (TechNode, 2026-11-04). Official entry required a domestic partner, a domestic model and a feature set cut to fit. The door opens, when it opens, on the terms the compliance system sets.
At home, substitutes arrived with conditions
The second condition is the demand gap that domestic models have not yet closed. The same Apple Intelligence beta that landed on 4 November carries a first batch of features that includes live translation, writing tools, visual intelligence and photo cleanup, while the next generation Siri announced at WWDC26 stays outside this update (GeekPark, 2026-11-04). Translation and writing are table stakes. The relay market, on the week’s own evidence, is concentrated somewhere else, in agentic coding tools, Claude Code and Codex, the products the ByteDance restriction names specifically. Policy is pushing the domestic stack to close the gap: the 15th Five Year Plan sets a target of 7 percent average annual growth in research and development spending for 2026 to 2030 (Chinese government work report commentary, 2026-03-05), and analysts project 4.5 to 5.0 percent GDP growth across the plan period (Galaxy Securities chief economist Zhang Jun, via China.com.cn, 2025-10-24). Both figures are targets and forecasts, not accomplished facts, and both say the substitution is funded and scheduled rather than finished.
Now run the two conditions against each other. If the supply side were open, the relay user would be an ordinary customer with a receipt, and the provider’s ledger would count him. If the domestic substitutes were complete, the relay user would have no reason to exist, and the domestic regulator’s ledger would count him where he actually buys. The cohort persists because neither is true: the foreign door is closed by the provider’s compliance, and the domestic shelf lacks the specific tool, the coding agent, that the workload requires. Single cause explanations fail in both directions, which is why the group has lasted through successive enforcement rounds.
Why corporate compliance hardened in the same week
The employer’s rule, the one that reached the engineer who pays his own way, did not come from nowhere. It landed in a region that spent the same week absorbing data breach disclosures, and the chronology is worth keeping because it makes the corporate logic legible rather than paranoid.
In Hong Kong, the finance portal MoneySmart told customers on 4 November 2026 that it had become aware of a breach the previous Thursday, after hackers compromised internal IT systems, and that “These systems hold personal data that customers share with us when they compare or apply for financial products through MoneySmart” (Hong Kong Free Press, 2026-11-04). In Taiwan, Bank SinoPac held an emergency press conference on the evening of 4 November 2026 and confirmed that abnormal transfers traced to hackers exploiting a security setting in a vendor’s system, with initial reporting putting the impact at about 400 customers and roughly NT$100,000 per transferred amount before the bank corrected the mechanism (iThome, 2026-11-04). In the same market, the e-commerce ERP vendor Mallbic suffered a security incident that pushed the shop platform Shopline to rotate API keys, exposing data merchants had imported (iThome, 2026-11-04). Hong Kong policy responses moved in the same direction: a lawmaker proposed at a Legislative Council meeting that estate owners’ votes be verified through a government managed digital identity platform to fight bid rigging (Hong Kong Free Press, 2026-11-05).
Each incident is a different system with a different owner. The pattern they share is the one a corporate security office reads: third party tools, vendor settings and personal accounts are where data leaves. An employee who pipes company code into an account bought through a reseller is, in that frame, not a productivity story. He is the vendor setting. The ByteDance tightening that TechNode reported is what that frame produces as policy, and it explains why the rule reaches self paid subscriptions. The company cannot tell, from its side of the network, which personal subscription is clean, so it treats the category as the risk. The engineer’s own money bought him no exception, because the ledger his employer keeps does not have a column for payment source.
What the official channel changed that week
The strongest argument against this piece runs through the same documents, and it deserves a full statement rather than a footnote. Official channels are widening. Apple Intelligence reaches mainland iPhones in December 2026 with Qwen inside Siri (TechNode, 2026-11-04). Claude lifted its five hour cap at the end of October (QBitAI, 2026-10-31). Domestic models absorb more of the workload every quarter, backed by the R&D growth target of 7 percent a year. If the official shelf keeps filling, the relay cohort shrinks by attrition, and the blind spot closes itself. On that reading, this article describes a transitional generation of users, not a permanent class.
The response is about the ledgers, not the window. Attrition reduces the size of the group. It does not create a count, because the reason the group is uncounted is structural: the provider’s books record accounts, the regulator’s books record licensed services, and the corporate books record sanctioned tools. A user can leave the relay market entirely and still never appear in any of the three, as long as the market he enters is measured only through its suppliers. Singapore’s nine in ten, counted monthly, is what closing the blind spot looks like when someone decides to do it. Nothing in the mainland’s current architecture produces that survey, for this cohort, at any size.
There is also a fairness point inside the counterargument. The transition, however fast, runs on people who took compliance risk with imperfect information, paid real money and are now being squeezed from both sides at once, by the provider’s enforcement above and the employer’s rule below. A market that closes a window should be able to say how many people were in it. Greater China’s AI compliance regime, in the first week of November 2026, cannot.
The counting method
Definition first. A relay user, for this piece, is a person physically or corporately inside mainland China who reaches a foreign frontier model through a third party reseller account rather than a direct provider subscription. The definition is behavioral, not legal, because no legal category names this group.
The denominator is unknown, and the piece does not estimate one. What Table 2 lists are the indicators that were checked and what each can and cannot carry. The time frame is the week of 31 October to 5 November 2026 for events, the third quarter of 2026 for financials, and January to September 2026 for trade statistics. Where a figure is a target or a forecast, the text keeps the target grammar. No sampling was possible, so no sample size is claimed; the estimated reading of demand comes from provider revenue growth, which is a proxy for absorption, not a count of persons. Nothing in the method adjusts for currency, seasonality or churn, and none of the proxies should be summed.
Table 2: Indicators checked for sizing the relay-user group, and the limit of each
| Indicator (unit, period) | Value | What it can proxy | What it cannot show |
| Anthropic revenue (US$ billion, Q3 2026) | 16.96 | paid capacity being absorbed | geographic or channel split of buyers |
| Revenue growth (percent, Q3 vs Q2 2026) | 47 | acceleration of demand | who the marginal buyer is |
| China goods trade (trillion CNY, Jan-Sep 2026) | 39.94 | the measured economy beside the unmeasured one | service or model-access flows |
| R&D spending growth (percent per year, 2026-2030) | 7 (target) | pace of domestic substitution | whether the coding-agent gap closes |
| Apple Intelligence China release (date) | Dec 2026 (planned) | official channel widening | size of the existing relay cohort |
Source: compiled from GeekPark on the updated S-1 figures, GeekPark on the original filing, and the General Administration of Customs trade briefing; the R&D figure is a plan target, and the December release is a stated plan, not a shipped fact.
The sourcing rule follows from the method. Every fact in this piece comes from published reporting, a regulatory filing as covered by the press, or official statistics, and each carries an inline attribution with date. The situation of the unnamed engineer is a composite reconstructed strictly from those documents: no individual was interviewed, no pseudonymous character accuses any named institution, and no quotation was invented. Direct quotes appear only where a source published them, including the TechNode sentence on self paid access, the ifanr headline, the MoneySmart breach statement and the QBitAI quota report. Where a group would naturally supply testimony and none exists in the record, the text says so rather than filling the silence. A count of this group would require one of three things: a provider disclosing channel geography, a survey with a sample frame, or a regulator publishing enforcement totals by user type. As of the first week of November 2026, none of the three exists. The group is real, its position is documented, and its number is not.
Limitations of the survey base
- The relay-user group has no published denominator; every size statement here is a proxy inference, not a measurement.
- Anthropic’s financials are covered through secondary reporting of the S-1 filings, one step from the primary document.
- All user-side evidence is secondhand; one named interview could change the texture of the conclusion.
Source notes
- TechNode , ByteDance reportedly tightens access to Claude and Codex, 4 November 2026
- ifanr , relay station market piece, 4 November 2026
- GeekPark , morning brief with ByteDance personal AI restriction, 5 November 2026
- GeekPark , Anthropic S-1 filing coverage, 31 October 2026
- QBitAI , Claude removes the five hour cap, 31 October 2026
- Hong Kong Free Press , MoneySmart breach disclosure, 4 November 2026
- iThome , Bank SinoPac incident press conference, 4 November 2026
- e27 , Digital 2027 Singapore findings, 5 November 2026
- General Administration of Customs , January to September 2026 trade data, 14 October 2026
This article was prepared by the expert team at Global Infinity Labs, drawing on insights and data provided through its MCP data services.
About Global Infinity Labs
Global Infinity Labs is an AI innovation company specializing in MCP data services.
Email: support@globalinfinitylabs.com