In-depth Investigation of API Transit Stations: From Black Gray Products to White Gloves, Where is the Future of Domestic AI?

Every day, millions of API requests are sent from the servers of Chinese developers, entrepreneurs, and even top AI companies. They bypass blockades, pass through hidden third-party transit nodes, and eventually reach the servers of OpenAI, Anthropic, and Google.

These nodes are called "transit stations" by insiders.

They are ubiquitous, yet few know who is behind them. This article will uncover the industry chain of this gray area.

Introduction

If you search for "free codex" on Bilibili, you will find a surprisingly detailed tutorial.

A content creator shared how to use self-built emails to mass-register OpenAI CodeX accounts, combined with receiving code platforms to mass-register payment platform accounts. After the new account is bound to a payment method, it automatically gets a one-month free trial. With this process, he can mass-register a large number of premium accounts, getting a whole month for free.

This initially just looks like a "technical sharing", right?

But he revealed one detail—the GitHub project homepage.

Insiders call this thing a "registration machine", an automated project for mass-registering CodeX. The project left a QQ group number. I searched it up, and it’s already full and expanding into the fifth group. This indicates that the first four groups were filled long ago. Based on 500-2000 people per group, a conservative estimate is at least 2,500 people, up to tens of thousands, are circulating around this project.

Even more "spectacular" is the website he made for the project.

Upon opening the site, advertisements rush to your face. On the left is a promotion for a certain transit station; on the right are clearly priced ready-made ChatGPT premium accounts—these were mass-registered using the method mentioned above.

The advertisement in the middle is the most interesting. The transit station set up a lottery to attract traffic and active community members.

10 Plus accounts. The official price is $20 each, with a total value of 1,400 RMB.

With a cost of 1,400 RMB, they bought thousands of clicks and precise developer users. This is far more cost-effective than running Baidu Ads.

This is the gateway to an industry chain. And the transit station in that advertisement is the first vine we need to trace down.

1. The First Vine - A Personal Site with "All-in-One Bucket" Architecture

The first target is hiding in the Bilibili tutorial's advertisement spot.

Its technical foundation can be described as "crude". The website is hosted on a sub-domain with a .top suffix—a "junk" domain that costs a few bucks a year, representing a complete grassroots approach. Yet, its official site looks decent because it directly uses the template from the open-source project new-api. This project has 30K stars on GitHub, offering a clean interface and comprehensive features.

This genuinely reflects the current ecosystem of transit stations: The technical threshold has been leveled. With open-source wheels, anyone who solves the "supply source" and knows a bit of operations can open their doors to customers. Their real "technical content" has shifted elsewhere—where to mass-register accounts? Where to exploit more free quotas? Where to find cheaper upstream channels? And where to pull in more users? This is their fundamental basis for survival.

Even the site's login method hints at its supply sources: besides the standard GitHub login, there's a portal for a "fleece-hunting station" next to it.

What truly exposes its core operations is its supplier list.

This list is chaotic yet rich. It includes official sources like OpenAI, Google, and Anthropic, cloud vendor channels like Azure and AWS, and even peer transit stations, as well as suppliers dedicated to "reverse engineering". So-called "reverse engineering" means hacking the protocols to forcibly "extract" free or restricted AI web interfaces and package them as APIs for sale.

The sole purpose of this "all-in-one bucket" procurement strategy is: Trading chaos for extremely high availability. If a certain upstream channel gets blocked, the station master can instantly switch to a backup link, leaving the users completely unaware.

Looking again at the "available token grouping" on the left, there hide the code words for supply channels:

Azure Claude / AWS: Represents relatively stable enterprise-class cloud channels.

Anti-gravity model / Ali reverse: This is pure underground tech work, cracking web version interfaces.

CCMAX-Unlimited: Implies some kind of high-concurrency, unthrottled special supply.

I also noticed an unexpected name in the groups: Xiaomi. A while ago, Xiaomi held an event where new registered accounts were gifted a certain quota. It seems someone has already exploited these quotas in bulk and is reselling them here.

By now, the business model of this site is very clear:

1. Sourcing: Exploiting cloud vendor free credits, registering trial accounts, acquiring cheap API Keys, and even reverse-engineering web interfaces.
2. Whitewashing: Accessing the new-api system to uniform all interface formats of different sources into the standard OpenAI format.
3. Distribution: Selling to those developers who need a vast amount of AI calls but don't want to maintain hundreds of accounts themselves.

It solves a real pain point: Chinese developers cannot smoothly use overseas AI services due to networks, payments, account restrictions, and high-price barriers.

But it simultaneously sends a signal: Things here are cheap, but not necessarily stable.

2. The Second Transit Station

In the supplier list of the first transit station, one was marked as the most crucial upstream. Its name lay in the first row.

This is the second vine we need to trace.

Upon opening the site, the style changes dramatically. This is a top-tier .ai domain. Amidst the current AI boom, such domains cost usually in the six figures. The page is no longer a crude open-source template but a meticulously designed "enterprise official website" that highlights enterprise-class services. A company name hangs at the bottom, looking respectable and proper.

This is an entity wanting to run B2B (Business-to-Business) business.

Since there is a company, let's do a background check.

The official website notes a Hong Kong company at the bottom, which indeed can be found; the entity exists. But digging deeper, things go wrong. I unearthed the founder's information—a Chinese entrepreneur, whose company is actually located in Xiamen, categorized as a micro-enterprise.

Checking Qichacha, another Singaporean company pops up.

Thus, I gathered its complete corporate entity list:

Entity Level	Company Name	Registry Location	Core Functions	Legal Liabilities	Actual Status
Top Holding Entity	xxx. LTD.	Singapore	Global holding, overseas collection, App Store/Google Play listing, international compliance	Assumes major international legal responsibilities	Active, current official overseas entity
Domestic Registered Entity	Xiamen xxx Company	Xiamen, China	Domestic domain filing, ICP qualification, domestic corporate collection	Assumes domestic internet service legal responsibilities	Active, pure shell company (0 insured people)
Historical Legacy Entity	xxxx Limited	Hong Kong, China	Early cross-border collections, overseas business signing	Basically holds no legal liability	Dormant, no public business activities

The founder’s background makes this "white glove" positioning even more self-consistent.

He previously ran a cross-border proxy business. Anyone who has done overseas business knows that massive API calls must be combined with proxy IPs, otherwise, the accounts are instantly blocked. Proxies are used to bypass risk control.

Transit stations, fundamentally, are also bypassing risk control—bypassing overseas vendors' account reviews, payment limits, and network blockades against Chinese users. If domestic companies could legitimately buy directly, who would buy from him?

Thus, stepping from cross-border proxy into API transit is not a cross-industry move at all, but a natural upstream and downstream extension of a business chain. Previously selling ladders, now selling water, the users haven't changed, and the demands never wavered.

But an eerie price contradiction arises that I cannot avoid.

Returning to the first transit station's supplier list. It explicitly noted this transit station as the upstream supplier and offered outrageous prices—Opu 4.5, Sonnet 4.5, and Gemini 3 Pro Image Preview were marked down to an 80% discount compared to official prices.

But on this transit station's own site, these three models were blatantly marked at original price.

The exact same supply source: retail original price, wholesale 80% off, a five-fold price difference. What does this mean? Either it possesses some secret channels where costs are close to zero, or what it sells in both channels are fundamentally different things.

Of course, it might be possible that the original webmaster priced it wrongly, and the data got outdated.

However, a boss starting from proxies, entering an industry gray area conceptually built to "bypass risk control"—how official are his so-called official channels exactly? This question could likely only be answered by himself.

The site also links to a company GitHub page, which pretends to be engaging in tech community building.

Clicking through, it’s entirely filled with AI-generated traction projects with meager star counts. Fundamentally, these are SEO and promotional materials, zeroing in contributions to the open-source community. Hanging out a GitHub page is nothing short of rounding out their "enterprise official site" persona better.

Evaluating the scale, this is a typical small-to-medium AI tool company, with 50 to 100 employees, pulling in tens of millions of revenue. It doesn't do fundamental research, rarely trains models, and avoids open source. It only conducts one role: constructing the most hidden pathway possible between overseas AIs and Chinese users.

And this road leads not only to average developers but also to names familiar to us all.

On February 24, 2026, Anthropic published an industry-shocking report, explicitly naming three top Chinese AI labs: DeepSeek, Moonshot (Kimi), and MiniMax. The report provided definitive data—these labs utilized around 24,000 fraudulent accounts generating over 16 million interactions.

What were they doing? In industry slang, this is called a "distillation attack"—crazy API calls to Claude, extracting out its logical reasoning, thought chains, and Agent capacities bit by bit, and feeding them to their proprietary models.

And these prominent AI enterprise factions would never use their official IPs to perform these tasks. If an official account gets banned, it's a severe compliance accident sufficient to trigger international lawsuits. But if an agent's account gets banned, they can shrug: "I just bought the services, I don't know how they handled it."

This embodies the "professional value" of transit stations. For tech giants, transit stations provide more than just sock puppets; they provide complete "adversarial engineering capabilities": when accounts are banned, new ones pop up instantly. Massive attack traffic is layered inside normal user requests, much like swarming ants, making the platform’s AI audit system fail to distinguish them.

Because the matter exploded, the US government pushed the PAIP Act in April 2026. The controls aren’t merely about "banning accounts" anymore; they are now focused on targeted sanctions against proxy service providers furnishing infrastructures for distillation attacks.

Looking backward, those convoluted offshore architectures, the seemingly excessive compliance designs, were likely not just guarding against OpenAI, but against these truly devastating legal consequences.

3. The Third Transit Station

This is also a supplier of the first transit station; though marginally lower in status than the second, it remains crucial.

This one uses generic Chinese pinyin allied with .ai for domains—a tier-two Top-Level Domain that costs around 100,000 RMB. Compared to the first station, it's fairly decent; compared to the second, it still lacks quite a bit of substance.

Its after-sales pipeline involves the standard gray-production trio: QQ groups, Telegram, and emails.

But what substantially shocked me was that it imitated vast model enterprises, providing Resource Packs and Coding Plans. This is not merely selling APIs; it's practically marketing "Developer Meals". Functionally, its product layout has increasingly drawn towards the formalized subscription models from certified manufacturers, trying to ascertain a steadier user relationship.

The platform embodies strict positioning: tilting towards developers, sometimes enterprises. Lacking the whole string of offshore company architectures and grand website layouts akin to the second, it understands superiorly than the first station what developers demand—beyond merely lowering prices, a bundled, predictive stability matters more. It occupies a role hovering midway between the grassroots forces and official corps.

4. Horizontal Comparison - A Chart Exposing the Segregated Realms

Three Transit Stations: Grassroots, White Gloves, Strikers, individually exemplifying three different ecological sectors stretching traversing this industrial chain. Aligning them reveals astonishing dissonances.

Dimension	Transit Station A	Transit Station B	Transit Station C
Domain Strategy	Secondary + .top	Top-Level + .ai	Secondary TLD + .ai
Domain Cost	Below 10 RMB	Above 200k	Above 100k
Website Operation Period	Under 6 Months	Over 2 Years	Over 10 Months
Tech Base	Open Source new-api	Self-developed	Self-developed
Supplier Structure	Bulk signup + Extracted + Peers	Official + Unknowns	Official + Unknowns
Ratio to Official GPT5.4	36x Cheaper	Original (Debatable)	5% Cheaper
Model Count	115	200+	200+
Promotional Tracks	QQ Groups + GitHub Ads	Self-Media + Rebate invites	Self-Media + Rebate invites
After-sales	QQ Groups	Email + Live chat	Email + Live chat
Target User Base	C-End freeloaders	B-End Local Enterprises	C-End Developers
Company Backdrop	Individual	HK + Singapore	Unknown

Browsing through the chart, the disconnects across these pivotal parameters appear piercingly glaring:

Domain translates to Class. One operates a sub-ten-bucks sub-domain, one drops two-hundred grands on a Top-Level .ai, and the final balances with a one-hundred-thousand-range tag. The domains inherently signal the ambition scopes and commitment inputs of their handlers respectively. The first chases brisk profits, the second nurtures long-term stakes, and the third strives upward to elevate itself.

The Contrasts behind Pricing. GPT-5.4 portrays models that sell 36 times cheaper, normal-priced, or 5% cheaper. Yet "Original price" masks the underlying crevices we discovered prior: If they wholesale identical inventories at 80% discounted metrics downriver, does "original" remain pristine, or are they exclusively displaying this frontally for you? Anything 36 times cheaper obviously houses tainted issues—they directly brandish their discounts transparently. The one tagging with identical metrics, however, harbors deeper abysses beneath it.

Fragmenting the Consumer Strata. C-End freeloaders, B-End regional enterprises, and C-End domestic developers. The tiers differ across three spending budgets and their comprehension of "stability." The first pool shifts when accounts vaporize instantly upon bans. The second's loss evokes commercial penalties and project debacles—necessitating corporate frontals, legal frames to masquerade into pure-hearted SaaS platforms.

An Unavoidable Void. Station C writes "unknown" towards official firm backings. A hub running upward of ten months clutching over two hundred models built atop proprietary schemas seldom boils down to mere indie ventures. Neglecting naked operations akin to A, discarding multi-shore framings analogous to B, it picks complete silence. Often, such deliberate taciturnity signals another painstakingly concealed shell game.

Three Transit Stations epitomize triad segments atop this commercial sequence: Grassroots Plunderers, Enigmatic Enterprise Gloves, and Aspirational Aggressors towards Official ranks. They aren’t combatants amongst themselves—A serves downriver of B—co-existing on shared trails absorbing disparate profit streams up along the path.

5. Beneath the Surface — E-commerce Platforms

If the aforementioned hubs dealt fundamentally with "Developer Business," the token exchanges on Taobao define the very bottom of this hierarchy.

Searching "CodeX" on Taobao uncovers overwhelming search results boasting meticulous operational phases. Keyword unifications are profound—"Domestic direct links", "Stabilized applications", "Issuance of Invoices", "Corporate transaction acceptance"—they target squarely upon local demographic weaknesses.

Their starting tags range dirt-cheaply: Ten bucks for two hundred calls spanning twenty-four-hour windows. Visually unmatchable, yet gazing deeper exposes calculating algorithms disjointed from authentic platform measures. Legitimate realms measure across 'tokens', not 'calls' nor 'queries.' Exploiting technological ignorance, these vendors specifically pamper layman mentalities.

Worsening matters, their currency is labeled as "USD". Obviously, this fails reality checks—a basic conversion of yuan into real-dollar quotas spells huge losses for them otherwise. These Taobao corners function on similar transit site backends pushed over to e-commerce shells. You are paying physical currencies into proprietary metrics, manipulated natively on customized scaling values handled implicitly behind doors! Sellers flex immense dictation adjusting rules universally over margins and refunds without clear-cut validations!

Going by the sellers' claims, 26 RMB can buy CodeX's "$50 USD" quota. This conversion ratio looks tempting, but keeping the previous transit stations in mind, it is essentially taking cheap inventory exploited upstream, cloaking it in a custom pricing module, and reselling it to laymen who know nothing about the industry. Multipliers can be adjusted, pricing units can be custom-defined, and refund policies are utterly up to the seller. Even those transit stations with their own websites never promised unbreakable prices. Once it reaches the Taobao level, transparency plummets to zero.

Browsing the comment section, the demands are highly concentrated. Someone asks if an invoice can be issued—yes, and it supports corporate payments. Someone asks if it can be used without a VPN—yes, via a domestic direct connection. These are gray demands that official channels simply cannot fulfill but exist massively in reality. Invoices and direct connections have become the core competitive advantages for Taobao token sellers.

But the deadliest issue within this entire chain is Trust.

You fundamentally have no idea what model is being called behind the screen. How can an ordinary person verify this? There are indeed people in the market who have created "model verification tools" to help users check whether the returned results genuinely come from the purported models. However, this verification faces a dead end: it’s impossible to stare at it 24/7. Sellers might use the real model during the day and switch to an inferior one late at night—no one would know. A user who bought "CodeX quotas" for 26 RMB might find the results generated at 2 A.M. completely switched to a cheap model trying to cut corners.

With no supervision and zero guarantees, everything relies solely on the seller's conscience. And unfortunately, in an industry chain starting at 10 bucks, "conscience" is probably the most worthless commodity.

However, in this ocean, it’s not only the small fish swimming around.

6. Entry of the Sharks: Fu Sheng and Justin Sun Eye the Same Cake

While grassroots webmasters were quietly raking in cash via the gray market, genuine sharks eventually caught the scent of blood.

This time, petty players stayed home. One is the CEO of a listed company, and the other is the controversial king of the crypto world. Almost concurrently, they reached into the same waters.

Cheetah Mobile: Fu Sheng “We Are Not a Transit Station”

Fu Sheng, Chairman and CEO of Cheetah Mobile, helm of a listed enterprise with over a million followers on Douyin/TikTok—the exact guy who drunkenly yelled at Zhou Hongyi in a group chat not long ago.

Here he comes, presenting a transit station attempting to mask itself otherwise.

He refuses to admit they are transit stations. But let's establish a strict reality check: the definition of a transit station is not delineated by whether they dabble in dark-gray markets or not. When there is something wedged between you and the primary AI model enterprise, dictating requests away from direct official connectivity—you function strictly as an intermediary station, period. The term "Transit Station" naturally stinks within common reputation, making evasion understandable, yet undeniable size and facades don't whitewash definitions.

Since it’s driven by a listed entity, reviewing UI visuals or framework structures becomes trivial. Let’s target directly at models and pricings.

ChatGPT-5.4 reflects an 85% markdown tag equivalently—15% cheaper than official paths. The model range runs limited, fielding roughly 37 options—clinging tightly to restrained layouts compared to the rampant 200+ selections littering general proxy sites, acting instead like a legit SaaS structural selection. Another tagging reveals the core strategy: DeepSeek-V4 sits at merely a quarter against official metrics! Naturally, this slashing defies sustainable realities; they are violently hemorrhaging investment pockets aggressively securing user mindsets to claim the early traffic domain.

Fu Sheng swoops in throwing cards embedded in compliance standards and branding. Ironically, the platform fails to manifest a direct corporate owner attached; nobody knows accurately which entity signs off, bills, or invoices the revenues drawn! How does a corporate-run pipeline maintain blurring this key feature? It beckons potential enterprise users into cautious second thoughts!

Justin Sun: A One-Letter Top-Tier Declaration

Justin Sun joined the game too.

He omitted preliminaries entirely and outright grabbed a prime top-tier AI domain—featuring solely a single letter! Such scales cost effortlessly within millions. Domain mirrors declarations: one letter broadcasts boldly, stating "I have arrived, I am serious, and I rule at the top-tier."

On the domain's landing, yet again—nobody can directly tell which company holds ground. Their pricing game twists remarkably distinctive: GPT-5.4 maintains absolute 1-to-1 official equivalence!

Not exactly cheap.

Coupled with an ultra-expensive domain alongside strict full-price structuring, Justin Sun plainly steers away from the price wars altogether! His trajectory signals an entirely different road: wielding an ultra-tier domain generating immense branding leverage, consolidating legitimized settlement channels targeting legitimately wealthy corporate clientele exclusively! Scarcity and raw trust represent his true weaponizations over pricing numbers!

Sharks have arrived, will the waters run clearer?

Fu Sheng slices 15% off official marks, Justin Sun peddles raw official metrics. One blasts venture capital aggressively securing tractions, the other constructs branding off colossal domains. These respectful sharks might visually contest with indie proxy nodes—in truth, they lock horns ferociously among themselves atop entirely separate evolutionary chains!

Shark invasions hardly guarantee the obliteration of gray markets. Sharks engulf mid-to-high level shoals consisting of enterprises insisting on invoices, contracts, and stable frameworks. The low-level small fry demographic remains continually lingering around the £10 Taobao proxy realms feeding on cheap, 200-call-rate scraps.

The division of this industry's layout becomes terrifyingly transparent.

The Final Chapter: The Legal Red Line and the Endgame Forecast

Coming this far into the investigation, a fundamental question must be addressed: Why is it called a "transit station"?

The antonym for a transit station is a direct connection. Direct connection means dynamically buying services straight from giant model vendors directly sending device pings into international servers—while additionally functioning strictly within legit internet connectivity pipelines. However, this trajectory stays notoriously severed for immense swaths of domestic denizens.

Therefore, transit points execute one distinct directive: setting up proxy server points domestically. A user ping reaches this domestic point first to be routed externally toward foreign hubs afterwards!

What does this indicate?

This falls fundamentally under illegal circumvention tapping into global internets! Disregarding how attractively disguised they are, any foreign AI proxy presenting local interfaces crosses into statutory breach domains right from inception point architectures!

Descending strictly beyond this boundary line, legal frictions crash across three cascading planes:

First Threshold: Illegal Business Operations. Proxies fundamentally transact telecommunication value-added features illegitimately. Marketing information transit capacities to locals devoid of licensing credentials falls flat beneath telecom authorizations entirely.

Second Threshold: Infringement by Providing Tools Invading/Illegally Accessing Computer System Configurations. Services waving flags of "Interface reverse engineering" lean exclusively upon cracking network communication paths tapping into authorized architectures illegally! Facilitating such hacks and charging entry defines crystalline violations procedurally!

Third Threshold: Fraud alongside Data Security Violations. Large-scale exploiting via fraudulent subscriptions scraping trial cash breaches fundamental swindling! Exporting civilian data through contraband internet pipes illegally across international servers inherently ignites strict data-safety charges respectively! Triple thresholds represent independent criminal liabilities collectively.

However, courtlines hardly comprise the exclusive threats proxy operators suffer from. Defensive retaliation originating upstream simultaneously constricts heavily!

OpenAI and Anthropic witness rapid enhancements into their scanning mechanisms heavily weaponizing precise behavioral biometric 'fingerprinting' structures. Eras permitting trivial evasions via swapping IPs plus spamming newly scripted signups are closing swiftly; mouse winning-odds diminish exponentially! Simultaneously, streams of supply vanish continuously—Google, Microsoft, among cloud giants tighten limits enforcing stronger registration walls cutting gifting allocations. Ultimately, these historically cheap limitless inventory veins will drain completely.

Justice lines restrict externally whilst source avenues suffocate at their roots; wedged heavily between these two crushing parameters paints an unequivocally crystalized endgame path going forward!

Over eighty percent of indie runners alongside small group fractions will melt away silently amongst these collapsing pressures. Chat rooms disassemble, proxy links yield infinite error pages—erasing them as completely unproven occurrences altogether. Without proper legal backbones shielding them, dealing defensively is genuinely non-existent.

A rare fractional cut of robust syndicate frames won't die easily; instead forcing them backward into far deeper obscure crypt markets shedding away open marketing entirely transacting strictly internally beneath dense security parameters exclusively restricting entry altogether.

The evacuated consumer voids finally invite assimilation handled dynamically via Cheeta or Justin Sun-styled semi-legit platforms equipped adequately with vast corporate bankings covering immense legal buffers surviving offshore! These titans hardly enter intending toward exterminating shadows; instead charging resolutely simply capturing the abandoned user spoils left completely vulnerable beforehand!

Standing statically observing this whole theater—one peculiar question overshadows concerns over proxy-bustings vastly: Why exactly does this market flourish robustly anyway?

Thriving proxy hubs directly manifest twisted necessities stemming between the local delay traversing domestic AI breakthroughs matching alongside ravenously demanding consumer pools respectively! Millions of devs, businesses, coupled with startup nodes never truly wish upon violating laws voluntarily—they just crucially thirst for code strings functioning seamlessly generating operational solutions guaranteeing they avoid dropping dead entirely behind progressing era parameters!

Blocking proxy nodes stands as standard statutory jobs; yet strictly enforcing proper, mighty domestic AI parameters yielding capable power thresholds comparable across globally competitive fronts—enabling developers the sheer grace completely absolving "stealing fired promethean elements" exclusively guarantees this obscure abyss dissipating permanently off radar charts finally!

Before that very horizon crosses lines, millions upon millions of these exact API paths shall endlessly tunnel onward through identical shadows.

Thanks for tuning in my friends! Should this content serve properly interesting, be kind enough directly pressing like, forward, and save interactions without hesitations!

Fancy catching my latest releases instantly onwards, be certain slamming a star highlight ⭐ avoiding missing me next time around!

That hits the stopping point for today folks.