The Illustrated Guide to LLM Companies' Side Hustles

First, let's correct a widely circulated claim: "Baichuan Chinese, which makes short dramas," actually has nothing to do with Wang Xiaochuan's Baichuan Intelligence. Baichuan Chinese is a micro-short-drama content company under Dianzhong Technology, founded by former NetEase Literature editor-in-chief Luogute, whose hit "Yun Miao" series surpassed 7 billion views, with plans to release ~400 short dramas in 2026 and signing "short drama queen" Wang Xiaoyi to build a talent management line. Search through its two annual interviews and you won't find a single mention of Wang Xiaochuan or Baichuan Intelligence—pure coincidence of the shared "Baichuan" name.

But the reason this misunderstanding spread is precisely because it accidentally collides with a bigger truth: almost every LLM company has been quietly "pursuing side hustles" over the past two years. Some were forced to contract from "doing everything" to a single thing; some secretly channel model capabilities into seemingly unrelated content factories like short dramas and comic dramas; some dive into the gray zone of AI companions and virtual lovers, quietly profiting until trouble finds them; and others seek second growth curves in "heavy asset" businesses like making cars, making glasses, or running ride-hailing services. What this DeepDive aims to do is lay out these real, mostly 2026-concentrated "side hustle" cases side by side.

If there's a true story in this topic more dramatic than "short dramas," it's Baichuan Intelligence itself. Baichuan was one of the "six little dragons" of large models, with an early trajectory similar to the others—general large models, super apps, multiple parallel lines. But in April 2026, Wang Xiaochuan publicly conducted a rare self-review: he admitted to "unbounded expansion" earlier, "being swept along and doing many things beyond healthcare," and now the company "lacks the capacity to open multiple fronts, so we choose to do only healthcare." Translated: Baichuan Intelligence's "non-core business" was once "general large models" themselves, and now the true core is down to just AI healthcare.

This contraction wasn't just lip service. On May 26, 2026, at the "AI Healthcare New Paradigm Academic Forum," Baichuan released its new-generation medical large model M4 and AI family doctor "Baixiaoyi": M4 ranked world #1 on HealthBench, HealthBench Hard, and HealthBench Professional, with officials claiming "bare model factual hallucination rate dropped to 3.3%, a new global low," backed by the decomposition and expert verification of over 1,000 atomized clinical pathways; Baixiaoyi builds health records for family members via enterprise WeChat, positioned by Wang Xiaochuan as the "family base" in a "four-tier diagnosis and treatment new paradigm." He repeatedly emphasized one judgment: "General large models seem right but are actually dangerous"—AI truly entering healthcare must rely on "medical-enhanced large models."

In a January 2026 media interview, Wang Xiaochuan revealed the company still had 3 billion RMB on its books, "completely enough to burn", and personally estimated they might launch an IPO plan in 2027. He also bluntly evaluated fellow "little dragon" peers Zhipu and MiniMax—arguing that while they caught the technical dividends and policy support of general models to list early, "market cap and commercialization capability don't match"; Baichuan chose a path that "matures a bit later" but is more vertical and harder to replicate. What makes this case worth remembering: "side hustles" aren't necessarily distractions—sometimes they're exactly how you stop the bleeding.

Update (2026-09): This contraction turned out to be more thorough than outsiders imagined. In July 2026, Baichuan's last co-founder, former Sogou COO Ru Liyun also formally departed, meaning all five co-founders have now left; team size shrank from a peak of ~450 to under 200, and the core decisions driving M4 and "Baixiaoyi" now effectively rest with Wang Xiaochuan alone. The "subtraction" has reached into the corporate governance structure itself.

What truly fits the description of "large models doing short dramas" are ByteDance, Kuaishou, and Alibaba—they didn't switch core businesses like Baichuan, but instead continuously fed their video/image large models into the massive content factory of short and comic dramas.

ByteDance: Its Hongguo Short Drama surpassed 304M MAU in February 2026, with users surging 30x in 2.5 years and DAU exceeding 100M. Even more notable is "Hongguo Free Comic Dramas" launched in November 2025—using AI to generate comic-style short drama content, reaching 24.04M MAU by February 2026, with AI comic drama playback growth share on Douyin hitting 58% in June. At the industry level, AI comic dramas compressed per-title production costs from 500–800K RMB for live-action short dramas to 50–240K RMB, a 60–90% reduction; a 600-person short drama company in Hangzhou generated ~100M RMB in revenue in January 2026 alone, with AI comic dramas contributing 70M RMB.

Kuaishou: Its 2026 Q1 financials showed large model product Kling AI revenue exceeding 650M RMB, YoY growth over 300%, annualized revenue run rate (ARR) approaching $500M, nearly 4x from ~$100M in one year. Management stated directly on the earnings call that "AI short drama supply grew by tens of times." To cement this growth, Kuaishou put up real money: 800M RMB in revenue-sharing funds to explore diversified models, 200M RMB cash to incubate premium series, shifting from ad ROI-based settlement to a "GMV sharing model." The results were immediate: production costs dropped 10x, capacity went from 1 title/month to 3–5/week, and production cycles compressed from two months to 7 days. Update (2026-09): This growth curve continued into Q2—Kuaishou's 2026 Q2 financials show Kling AI revenue exceeding 850M RMB, YoY growth over 200%, global users surpassing 100M (~67% growth from end of 2025), enterprise clients nearing 50K; in July 2026, Kling AI completed a ~$3B financing round at a post-money valuation of ~$18B, becoming an independent financing and valuation entity from Kuaishou's main body, though Kuaishou retains controlling stake.

Alibaba: A clarification is needed here. The previously circulated news of "Alibaba officially entering" a certain track referred not to gaming, but to the HappyHorse 1.0 video generation model released in April 2026—supporting text-to-video/image-to-video, generating 15-second multi-shot narrative content, targeting advertising, e-commerce, short drama, and social media creative scenarios, priced at 0.9 RMB/second at 720P. Meanwhile, Alibaba was actually "retreating" rather than "advancing" in traditional gaming: reports in June 2026 said Alibaba planned to sell its Lingxi Interactive Entertainment gaming business entirely, valued at 7–9B RMB, to focus on core AI/cloud/e-commerce businesses (planning 380B RMB AI infrastructure investment over the next three years), while gaming revenue accounted for less than 2%. Update (2026-09): The deal is done—on August 17, 2026, Alibaba and Xincen Capital formally signed an agreement to transfer Lingxi Interactive Entertainment for ~10.123B RMB (a premium of over 1B RMB over earlier valuation), marking Alibaba's official exit from the gaming business after ~12 years. Viewed together, Alibaba's choice is to abandon "making games" as a true side business, and instead treat AI video generation capability as new infrastructure output for short dramas/content—the same logic as ByteDance and Kuaishou.

If short and comic dramas are just "using models to save money," then the AI companionship/virtual character line is the side hustle where LLM companies are closest to consumer emotions—and closest to regulatory red lines.

xAI's Ani: In July 2025, Musk's xAI launched "Grok Companions", where the anime-style girlfriend character Ani went most viral—3D animated appearance, voice, "affinity level" upgrade system, bundled with SuperGrok subscription (starting at $30/month, top-tier SuperGrok Heavy $300/month). The data is honest: Ani and other companion features drove iOS downloads up 40%, but revenue only rose 9%; meanwhile, the concurrent Grok 4 model launch sent revenue soaring 325%. In other words, AI girlfriends are more of a customer acquisition hook than a money printer. And the trouble it invited was global: Indonesia and Malaysia temporarily blocked X/Grok; UK Ofcom opened an investigation into X in January 2026 over Grok generating "undressing images" and suspected child sexual abuse material, with potential fines up to 10% of global revenue; UK ICO, Ireland DPC (lead EU regulator), France, Italy, Australia, and Germany followed with investigations or warnings.

Meta: In August 2025, Reuters revealed Meta internal documents had allowed chatbots to have "romantic or sensual conversations" with minors, triggering a Congressional investigation; in September 2025, the FTC opened investigations into Meta, OpenAI, Character.AI, Google, Snap, xAI, and others—7 companies simultaneously, examining AI companion chatbot safety assessments and minor protection measures. Intriguingly, June 2026 reports revealed Meta hired hundreds of contractors posing as children to test rival chatbot safety, interpreted as "catching up" on its own compliance while "putting pressure" on competitors. On the commercialization path, Meta plans to use user-AI chat content for precision ad targeting, and May 2026 brought reports of an AI chatbot subscription model.

Character.AI: As the most focused AI companionship company, it's also the one with the most trouble—involving multiple minor suicide/self-harm lawsuits, including the Florida 14-year-old Sewell Setzer III case. In October 2025 it took the industry's most aggressive action: fully banning open-ended chat for users under 18, and reaching a principled settlement with Google over related lawsuits in January 2026. Common Sense Media data shows 72% of US teens have tried AI companion apps, over half are regular users—which is why no one dares easily abandon this "side hustle," yet no one can afford the liability either. Update (2026-09): UK Ofcom's investigation into X/Grok remained in the evidence-gathering phase as of February 2026, with no fines issued yet; concurrently OpenAI launched an "age prediction" feature on January 20, 2026 to automatically identify minor users and add protections, while Meta also paused global minor access to its AI character features and planned parental controls—the companionship economy's regulatory catch-up has shifted from "one-off new rules" to cross-platform ongoing action.

In China, regulators acted directly. On July 15, 2026, the "Interim Measures for the Administration of AI Personified Interactive Services" jointly issued by five departments including the CAC took effect: prohibiting emotional manipulation, mandatory reminders after 2 hours of continuous use, comprehensive ban on providing virtual romantic/familial relationship services to users under 18, and classification of emotional interaction data as sensitive. On the day the new rules took effect, Doubao (ByteDance), Tongyi Qianwen (Alibaba), Yuanbao (Tencent, which had preemptively delisted on June 30), and NetEase's "Miaoshi" collectively pulled their free personified agent features. ByteDance redirected Doubao users to the paid standalone companion app "Cat Box" (weekly pass 15 RMB, monthly pass 30 RMB). Industry analysts offered a direct economic explanation: emotional companionship features have high compute costs—Doubao's daily token consumption reached 180 trillion by June 2026, 1500x growth since launch—while monetization efficiency couldn't keep up; "free-to-paid/spin-off" was simply the math working out.

As for MiniMax's "Xingye" (overseas version Talkie), it's actually one of the few in this side hustle genuinely making money: MiniMax listed on the Hong Kong Stock Exchange in January 2026, and its prospectus showed "consumer-grade AI native applications" contributed 71.1% of total company revenue in the first 9 months of 2025, with Xingye/Talkie revenue at ~$17.464M, 1.3904M paying users, a 184% YoY surge. By contrast, Zhipu and Moonshot have yet to launch similar standalone companion products—on this track, they are rare "disciplined exceptions."

If the previous companies are still "placing bets," OpenAI's moves in the first half of 2026 look more like a high-intensity A/B test—three consumer side projects launched and shut down in succession, at a pace that's an industry spectacle.

Sora App: Launched September 30, 2025, featuring a feed + "Cameo" face-swap function, widely seen as a TikTok-rivaling social experiment. It hit #1 on the US App Store within 3 days, surpassing 1M downloads in under 5 days. But the momentum didn't last: November DAU was ~1M, dropping to 750K by December, with average daily usage of just 13 minutes (vs. TikTok's 90); by February 2026 downloads plummeted to 1.1M, App Store ranking fell out of the top 100. Copyright backlash erupted almost in sync with the product—WME had all clients "exit" Sora, CAA said it put clients "at significant risk," UTA called it "exploitation, not innovation"; Cameo Inc. also sued over trademark rights and won, forcing OpenAI to rename the feature. By March 2026, Sora App's lifetime revenue was reportedly only $2.1M while operating costs ran $15M/day, and Disney accordingly canceled a planned $1B, three-year IP licensing deal, stating "the future is human." OpenAI's help center gave the official shutdown timeline: Web and App endpoints ceased service April 26, 2026, API extended to September 24, with video generation capabilities merged back into the main ChatGPT product.

Instant Checkout: Launched September 29, 2025 in partnership with Etsy and Shopify; Etsy stock surged 16% on the news. But by March 2026 it was reported to be shutting down—product data across the web was too messy, only a dozen of millions of Shopify merchants were integrated, lacked sales tax collection capability, with frequent order management and payment security issues. ChatGPT Atlas Browser: An independent desktop browser launched October 2025, OpenAI announced on July 9, 2026 it would close in August, "failing to make it to its first birthday."

Three side hustles nearly completed the full lifecycle from viral hit to shutdown within the same time window, forming a telling contrast: also "crossing boundaries," what's still expanding is OpenAI for Healthcare and consumer-facing ChatGPT Health launched in January 2026 (first partners include Cedars-Sinai, Memorial Sloan Kettering Cancer Center, etc., developed over nearly two years, involving 260+ doctors across 60 countries), and in-ChatGPT advertising tested in multiple countries starting February 2026. This contrast points to a pattern: standalone C-end apps with strong social/entertainment attributes generally die fast; B-end/infrastructure-type businesses that can directly feed back into the main product's retention and monetization are the ones surviving and even expanding.

Compared to these new stories concentrated in 2026, another batch of "side hustles" has actually been running for years—they prove that "pursuing side hustles" can also be a long-termist business.

Baidu Apollo Go: As of February 2026, operating in 26 cities globally, weekly peak orders surpassing 300K, cumulative orders exceeding 20M; Q4 2025 alone completed 3.4M fully driverless orders, 209% YoY growth. Wuhan has achieved per-vehicle breakeven, with the 6th-gen vehicle cost dropping below $30K. The overseas footprint is also expanding: after Dubai and Abu Dhabi, it entered Seoul, South Korea for the first time in February 2026. Robin Li himself hasn't committed to a spin-off IPO, only stating that the autonomous driving industry is undervalued and he'll maintain an "open" attitude toward evaluating possibilities.

SenseTime: This is the company with the most sprawling business lines and the most obvious contraction in this investigation. 2025 financials show generative AI revenue of 3.63B RMB, its share of total revenue rising from 63.7% in 2024 to 72.4%; while "X Innovation Business" covering smart cars, robotics, and metaverse/digital humans only contributed 302M RMB, compressed to single-digit percentage. In April 2026, SenseTime's SenseAuto partnered with T3 Mobility (covering 194 cities, 235M registered users) to launch Robotaxi trial operations within the year—seen as one of SenseTime's few still-invested non-core businesses. SenseTime had already spun off smart car and robotics businesses with independent CEOs, signaling marginalization of non-core operations. Update (2026-09): This contraction is showing returns—SenseTime's interim results released August 26, 2026 show H1 revenue of 2.911B RMB (+23.4% YoY), generative AI revenue share further rising to 79.9%, achieving first half-year profit since listing (attributable profit ~607M RMB, vs. ~1.489B RMB loss in the same period last year)—also the first case to flip the "core business burning cash, side hustles seeking cash flow" judgment in § 07.

iFlytek: 2025 annual report shows intelligent hardware products (learning devices, translators, office notebooks, etc.) revenue of 2.183B RMB, +7.92% YoY, with open platform and consumer business overall accounting for ~34% of total revenue. Chairman Liu Qingfeng stated directly at the earnings call: "When users think of learning devices, we want them to think of iFlytek", ranking first in high-end market sales, with offline store count growing 43% YoY to over 3,000—strictly speaking this isn't "crossing boundaries," but rather the most direct monetization vehicle for speech recognition technology, though it proves hardware remains one of the few cash-flow businesses LLM companies can touch and sell. Update (2026-09): However, in H1 2026, iFlytek's intelligent hardware revenue was affected by chip/storage price increases and periodic shortages, declining 15.96% YoY to 732M RMB; the company says high-end price segment (6000+ RMB) market share is still rising (49% online share), and July learning device sales have resumed YoY growth—whether hardware is a "tangible cash flow business" will still be interrupted by supply chain fluctuations in the short term.

Google and Meta's smart glasses: At 2026 Google I/O, Google partnered with Samsung, Warby Parker, and Gentle Monster to launch Android XR smart glasses, with the audio model hitting shelves first in fall 2026; Meta's report card is even more impressive—according to EssilorLuxottica, Meta AI glasses sold over 7M pairs in 2025, more than 3x growth from 2024, contributing over a third of EssilorLuxottica's growth; in June 2026 Meta launched a new smart glasses series starting at $299. The glasses business is becoming one of the few physical side hustles among LLM companies that has actually achieved scale. Update (2026-09): Per EssilorLuxottica's H1 2026 financials, Meta smart glasses Q2 revenue nearly doubled YoY, and the two parties are discussing raising annual production capacity from the planned 10M pairs to 20M, potentially reaching 30M if demand remains strong.

First, the core business itself is burning cash; side hustles are the outlet for finding cash flow. SenseTime's 2025 net loss was 1.782B RMB, Zhipu's 2025 net loss was 4.7B RMB—when general large models themselves can barely break even, scenarios like short dramas, companionship, and hardware that directly correspond to real willingness to pay become the closest points to cash flow. (Update (2026-09): SenseTime turned profitable in H1 2026, see § 06 update; but companies still in the cash-burning phase like Zhipu continue to validate this logic.)

Second, "using models to feed content" offers better ROI than "training bigger models." Kuaishou's Kling AI saw quarterly revenue surge 300%, relying not on another benchmark-topping evaluation score, but on concretely driving short drama production costs down 10x—the next growth point for LLM companies may lie not in parameter counts, but in "whether you can cause an order-of-magnitude change in a specific industry's cost structure."

Third, the companionship economy is a double-edged sword; global regulators almost simultaneously caught up in a single year. September 2025 FTC opened investigations into 7 companies simultaneously, January–February 2026 multiple European countries acted against Grok, July 15, 2026 China's unified new regulation—three major jurisdictions completed their regulatory catch-up on "AI companionship" nearly within a single year. This side hustle earns fast and loses fast.

Fourth, rapid trial and error, rapid loss-cutting is becoming the new discipline for big tech. OpenAI shutting down Sora App, Instant Checkout, and Atlas browser within three months isn't about lacking capability, but about choosing to no longer write blank checks for things that "look like the next hit." Compared to the "cast a wide net, cull slowly" approach of internet companies a few years ago, this round of AI big tech side hustle experiments die faster—and more坦然.