In July 2026, three famously low-profile founders in China’s artificial intelligence sector suddenly stepped into the spotlight together, but with dramatically different postures. Zhipu AI (02513.HK) founder Tang Jie announced a “reset,” channeling all IPO-raised funds back into frontier research. MiniMax (00100.HK) founder Yan Junjie, after watching his company’s stock price collapse 77%, pledged to stop drawing a salary and committed 4% of his personal shares to steady the troops. DeepSeek founder Liang Wenfeng, meanwhile, catapulted his company’s valuation from $52 billion to $71 billion within a single month and quietly kicked off preparations for an A-share IPO.

The three events occurred almost simultaneously, pointing to the same underlying logic: in China’s large language model arena, capital is not a multiple-choice question—it is the lifeline. Once a technological advantage stalls, product competitiveness, talent appeal, and the ability to raise the next funding round will all collapse in lockstep. The three helmsmen are now competing for a “ticket” to the era of Artificial General Intelligence (AGI) using entirely different strategies.

Tang Jie’s “Reset”: A Counterintuitive Bet After a Trillion-Yuan Valuation

In January 2026, Zhipu AI listed on the Hong Kong Stock Exchange, becoming the “world’s first publicly traded large model company.” By June, its market capitalization had briefly exceeded HK$810 billion (approximately $103.3 billion), surpassing Xiaomi. Just as the market was expecting a beautiful growth curve, Tang Jie wrote in an internal letter: “Others ring the bell; we reset to zero. This is not a posture. This is conviction.”

He then launched the “Touch High” initiative, demanding that his team de-emphasize commercial monetization pressure over the next two years and concentrate resources on four “technical peaks”: long-horizon tasks, autonomous agents, self-training, and extreme safety governance. Tang’s logic is clear and counterintuitive: the core of AI transformation is not optimizing existing products, but rather a technological revolution that raises the ceiling of human intelligence. An IPO is not the finish line; it is merely the entry ticket to the next, longer game.

Supporting this decision is genuine technical credibility. Zhipu’s GLM-5.2 model already benchmarks against Claude Opus 4.8, and its MaaS platform has reached an annual recurring revenue (ARR) of 1.7 billion yuan (approximately $251.2 million), representing 60-fold year-over-year growth. Tang’s own summary consists of just seven Chinese characters: “Essence, counterintuition, and focus.”

But the “reset” is not without cost. If “Touch High” yields no substantial breakthroughs in two years, market sentiment will flip from “visionary” to “reckless.” Tang is betting on fundamental breakthroughs that benefit the entire application layer, rather than patching up frameworks built by others.

Yan Junjie’s “Last Stand”: A Desperate Fight After a 77% Plunge

Unlike Tang Jie’s composure, MiniMax founder Yan Junjie faces a genuine existential crisis. In early 2026, MiniMax’s Hong Kong IPO surged 109% on its first day, with the stock price peaking at HK$1,330 (approximately $169.67), pushing its market cap to a zenith of HK$410 billion (approximately $52.3 billion). By July 15, however, the stock price had withered to just HK$297 (approximately $37.89), a staggering 77% decline.

On July 9, the company faced its first large-scale lock-up expiration, with approximately 146 million shares—representing 48.9% of total share capital—becoming freely tradable. On the day of the unlock, MiniMax plummeted 17.98% in a single session. In subsequent trading days, the stock price briefly touched an all-time low of HK$209.2 (approximately $26.69).

In the same week as the stock crash, Yan Junjie issued an all-staff letter announcing three decisions: effective immediately, he would no longer receive any salary from the company until it achieves AGI; he would allocate personal shares equivalent to 4% of total share capital to incentivize the core team; and an additional 1% of shares would establish a dedicated fund for the open-source community. Concurrently, MiniMax completed a refinancing round of HK$16 billion (approximately $2.0 billion), with 80% explicitly earmarked for AI infrastructure and model research.

Market analysts note that Yan is playing the “sincerity” card. When a stock is in freefall, no amount of CEO PowerPoint slides carries weight. He put his own interests on the table: forgoing a salary means his economic interests are synchronized with company value; committing personal shares means his personal wealth rises and falls with the team’s options.

But promises can only solve so much. Product competitiveness does not improve simply because of a founder’s sacrifice, nor will the market’s collective skepticism about the “large model business model” dissipate due to a personal gesture. JPMorgan downgraded MiniMax’s target price twice in one week, from HK$300 to HK$240 (approximately $30.62), citing concerns over equity dilution from the financing.

A deeper issue lies in the fragility of MiniMax’s business model. Over 70% of the company’s revenue comes from overseas consumer-facing products, including the virtual companion app Talkie and Hailuo Video, yet the gross margin for its consumer business is a mere 4.7%. In 2025, MiniMax reported revenue of $79 million against R&D expenses as high as $253 million, with annual losses exceeding 12.8 billion yuan (approximately $1.9 billion). The flagship M3 model released in June triggered a massive trust crisis among developers due to billing model changes, with actual costs for equivalent tasks surging by up to 257%, alongside accusations of “double standards” in API pricing between Chinese and international markets.

As of July 14, MiniMax’s total market cap stood at approximately HK$72 billion (approximately $9.2 billion), while Zhipu AI, which listed in the same period, maintained a market cap of roughly HK$745.6 billion (approximately $95.1 billion). The gap between the two has widened to over tenfold.

Liang Wenfeng’s “Blitz”: From Rejecting Capital to a $71 Billion Valuation

If Tang Jie is playing defense and Yan Junjie is stopping the bleeding, then DeepSeek founder Liang Wenfeng is waging a capital blitz.

According to multiple media disclosures, DeepSeek has fully initiated preparations for an A-share IPO, planning to submit application materials within the year and complete the listing by 2027. The company just closed its first major funding round exceeding 50 billion yuan (approximately $7.4 billion), with a post-money valuation of about $52 billion. Just weeks later, a new funding round has already begun, targeting a pre-money valuation of $71 billion (approximately 480 billion yuan). The valuation has multiplied sevenfold within three months.

The investor roster is nothing short of stellar: Tencent contributed 10 billion yuan (approximately $1.5 billion), CATL put in 5 billion yuan (approximately $738.8 million), while JD.com, NetEase, and IDG Capital each invested 3 billion yuan (approximately $443.3 million). The National AI Industry Fund also joined as a co-investor. Founder Liang Wenfeng personally contributed approximately 20 billion yuan (approximately $3.0 billion), making him the single largest investor.

Even more noteworthy is the financing structure. Apart from the National AI Industry Fund, the other industrial investors are entitled only to dividend distribution rights, with no board seats or operational voting rights. All external shares are subject to a mandatory five-year lock-up period with no transfers permitted. Through his capital contribution and structural design, Liang controls nearly 78% of voting power. Analyst Michael Taiwo commented on this: “Scale is what people applaud. Ownership is what you eat.”

Liang’s confidence stems from three things: a technical reputation that the global market buys into, the personal wealth from High-Flyer Quant anchoring the bulk of the financing, and a sufficiently long history of “rejecting capital.” In 2025, High-Flyer Quant achieved a return rate of 56.55%, with annual revenue of approximately 8.6 billion yuan (approximately $1.3 billion). Liang holds an 85% stake, yielding billions in dividends annually, providing a steady stream of self-owned capital to fuel large model R&D.

According to the Bloomberg Billionaires Index, Liang Wenfeng’s personal net worth has risen from approximately $16.7 billion to $36 billion, making him the eighth-richest person in China, surpassing Anthropic co-founder and CEO Dario Amodei and OpenAI co-founder and President Greg Brockman.

The Dual Test of Technological Catch-Up and Commercialization

Behind the capital frenzy, the technological and commercialization challenges facing DeepSeek are equally severe.

According to Artificial Analysis test data from April, DeepSeek’s previous-generation flagship model V3.2 ranked 16th globally in overall performance, 17th in coding capability, and 15th in agent capability. Domestic rivals such as Alibaba’s Qwen, Moonshot AI’s Kimi, Zhipu GLM, and MiniMax have all pulled ahead. While a V4 Preview version has been released, the official version remains delayed. DeepSeek’s own technical report also acknowledges that V4-Pro-Max “slightly lags behind GPT-5.4 and Gemini 3.1 Pro” in reasoning, placing it roughly three to six months behind the frontier.

On the commercialization front, DeepSeek’s current annualized revenue is approximately $400-500 million, with core income derived from paid API calls for the V4 series. The revenue structure is singular and lacks major enterprise clients. Liang himself has candidly told investors that priority will be given to frontier technology R&D, and short-term commercialization is not the focus.

To tackle the challenge of computing power self-sufficiency, DeepSeek plans to channel the massive financing into building gigawatt-scale proprietary intelligent computing centers and developing in-house AI inference chips, pursuing a dual-track strategy of “overseas high-end GPU procurement + domestic chip adaptation.” The V4 model has already completed adaptation for Huawei chips.

Regarding the listing venue, DeepSeek has bypassed the Hong Kong channel in favor of mainland China’s STAR Market, primarily relying on the fifth set of listing standards tailored for AI large models, introduced by the Shanghai Stock Exchange on June 17. This framework does not require companies to meet traditional profitability metrics, only demanding core technological advantages, scaled product deployment, and complete compliance filings.

The Talent War Fuels the Capital Race

The real reason Liang Wenfeng broke his years-long “no fundraising” iron rule may be hidden in talent attrition data. According to multiple industry media reports, Luo Fuli, a key contributor to the V3 model, left to join Xiaomi’s AI division; researcher Guo Daya moved to ByteDance; and another core member, Wang Bingxuan, chose to join Tencent. Three core researchers were poached by tech giants within the same month.

Statistics show that in China’s 2026 campus recruitment market, the median monthly salary for algorithm positions has surpassed 24,000 yuan, with top talent easily commanding over 50,000 yuan per month. For core researchers who have already made a name in the industry, the compensation packages offered by big tech firms are far more attractive than those from a startup without a public valuation. More importantly, there are stock options—frontier AI labs retain core teams by offering options priced against a formal valuation. DeepSeek had never raised funds before and had no external valuation, making it impossible to price options at a level employees would find credible.

Just ten days after the first funding round closed, DeepSeek swiftly posted 33 urgent job openings, subsequently launching a global expansion hiring plan covering all departments, with multiple core teams set to double in size. This money, in essence, is a talent retention fee.

Three Philosophies, One Gamble

The three founders represent three starkly different survival strategies in China’s current AI landscape: Tang Jie opts for “isolated” deep scientific cultivation, buying the upper limit of the future with a reset; Liang Wenfeng chooses “extreme expansion” through a capital sprint, buying decision-making autonomy with sovereignty; Yan Junjie selects “empathetic” incentive symbiosis, buying time to survive through shared burden.

Stripped to the core, all three are doing the same thing: maintaining an absolute conviction in technological breakthroughs amid the tides of capital.

For Chinese large model companies, the path of “make money first, then invest in technology” simply does not exist. Once a technological advantage stalls, product competitiveness, pricing power, talent appeal, and the ability to raise the next round will all collapse in tandem. None of the three philosophies is absolutely correct, but at a given moment, in a given situation, each represents the most honest choice a founder can make.

As global AI competition enters a white-hot phase over the next two years, the market will test not only these companies’ model capabilities but also how these helmsmen, under extreme pressure, lead their respective teams across the valley of death from “tools” to “superintelligence.” The history of China’s AI is being slowly written by these three men who have chosen to “bet their lives,” in every single moment of decision.