On October 22, MiniMax (00100.HK) fell below its issue price.

On the same day, Bloomberg, citing sources familiar with the matter, reported that DeepSeek raised approximately 100 billion yuan (about $15 billion) in its latest funding round—one of the largest-ever in China—with participation from CATL and Tencent.

Just days earlier, CITIC Securities updated the progress of MiniMax's A-share listing tutoring (tutorship), preparing for final approval.

It has been just over nine months since MiniMax's Hong Kong debut on January 9.

The timeline is tight: Zhipu (02513.HK), the other "duopoly" player, completed its Sci-Tech Innovation Board IPO tutorship approval in June, aiming to raise 15 billion yuan.

Among peers, Yuezhi Anmian (Kimi) also completed its final pre-IPO round in October, with its valuation soaring to $50 billion. According to rumors, it plans to submit its Hong Kong listing application in Q1 next year.

While some rush toward IPOs, already-listed firms face declining shares.

After initial "scarcity premiums," MiniMax's market cap has plunged 80%, and Zhipu's has lost 1 trillion yuan.

Yet, big models remain cash-hungry, with even OpenAI reportedly in new funding talks, targeting a $1.2 trillion valuation.

Compute costs are exorbitant, R&D burns cash, and profits remain elusive—funding must not stop.

How do the "duopoly" players compare in terms of resources and positioning?

Founded in 2022, MiniMax is among China's earliest "multimodal" AI startups, backed by Alibaba, Tencent, Hillhouse, Sequoia China, and IDG.

MiniMax's growth trajectory is closely tied to its multimodal focus.

Beyond language models, MiniMax monetizes three C-end product lines overseas: AI emotional companion app Talkie (international) and Xingye (domestic), video generator Hailuo AI, along with a suite of models for voice, music, and code.

By 2025, 70% of MiniMax's revenue came from overseas, with total annual revenue reaching approximately 528 million yuan, up 158.9% year-on-year. Financial data shows the company served 212 million individual users and 130,000 enterprise clients globally.

2026 marked a turning point for MiniMax's growth.

First-half revenue hit ~781 million yuan, surpassing 2025's total and growing 283.1% year-on-year. Q2 revenue rose another 81.8% quarter-on-quarter; token consumption in July was 20x that of January. Revenue mix shifted, with open platform and B-end services accounting for 63.4% of total revenue, up 703.1% year-on-year. AI-native product revenue grew 100.9% year-on-year.

Despite near-tripling H1 revenue, MiniMax remains unprofitable: net loss stood at ~2.39 billion yuan, narrowing 11% year-on-year, with cumulative losses reaching ~20.29 billion yuan. Cash reserves stood at 8.83 billion yuan.

At this burn rate, an A-share listing for fundraising is urgent.

Comparing Zhipu offers more insights.

Founded in 2019 as a spin-off from Tsinghua University's lab and labeled "national team," Zhipu's investors include local industrial funds from Beijing, Shanghai, Chengdu, and Zhuhai. Over 80% of its revenue comes from B-end clients, starting with localized deployments for large state-owned enterprises.

In 2025, Zhipu's total revenue reached 724 million yuan, up 131.9% year-on-year, surpassing MiniMax by nearly 200 million yuan.

In H1 2026, Zhipu's revenue hit 954 million yuan, growing 399.7% year-on-year. Open platform and API revenue contributed 825 million yuan (86.5%), a 20x+ increase from 15.2% a year earlier. However, gross margin collapsed from 56.3% in 2024 to 26.38% in H1 2026, as price cuts drove scale.

Like MiniMax, Zhipu remains unprofitable, with H1 net losses at 1.964 billion yuan and cumulative losses around 10.6 billion yuan.

One sustains itself via C-end overseas subscriptions; the other stacks B-end project revenue. This shapes their entirely different capital market narratives.

MiniMax's fundamental risks lie here: emotional companionship and video generation rely on "spend-for-growth" tactics, with high marketing costs and user retention challenges. Fortunately, in H1 2026, API and enterprise services took over as growth drivers, reducing C-end dependency.

So, how do the "duopoly" players' model capabilities stack up?

On the SWE-bench Verified leaderboard—a key engineering benchmark—as of September 2026, MiniMax M2.5 achieved a 75.8% pass rate, outperforming GLM 5 (72.8%), Kimi K2.5 (70.8%), and DeepSeek V3.2 (70.0%). This places it in the global top tier for open-source models, but the benchmark primarily measures "real codebase issue-fixing" ability, not full general intelligence. On abstract reasoning tests like ARC-AGI-2, MiniMax M2.5 and GLM-5 scored ~5%, while DeepSeek V3.2 scored 4%—still far behind leading closed-source models.

But rankings don't tell the full story.

MiniMax's differentiators have always been cost-effectiveness and multimodality.

In pricing announced around March, M2.1's input cost was just $0.3 per million tokens—about 17% of OpenAI's equivalent model and 55% of Zhipu's. Its voice model, Speech-2.6, ranked among global leaders in ELO scores, while video model Hailuo ranked 4th on SuperCLUE's text-to-video leaderboard—the only startup product in the top six. The October-released M3.1 Flash upgraded "next-stage" capabilities like agentic tasks, computer use, and cybersecurity, with official claims of nearly 3x faster online inference and 70% lower costs.

After analyzing fundamentals, MiniMax's position becomes clear.

While not the strongest in overall model capability, it offers a rare domestic combo of "full modality + cost-effectiveness + C-end overseas reach."

Thus, MiniMax's A-share narrative will hinge on three keywords: cheap, scalable, and user-rich.

Who's paying for big models?

To raise capital, firms must clarify how they'll make money.

At September's Bund Summit, Jiang Xiaojuan from the University of Chinese Academy of Social Sciences outlined five commercialization paths explored by the industry:

1. Private custom deployments for governments and enterprises; 2. MaaS/API charging based on usage; 3. "Cloud-model integration" (free models but paid cloud compute); 4. Token-based pricing for agent interfaces; 5. C-end subscriptions.

Maturity levels vary, but API currently drives incremental revenue.

Previous surveys by CNR showed that the primary payers are AI-native startups and agent developers—those building applications atop large models—not traditional IT budgets in enterprises.

In Zhipu's H1 report, open platform API revenue accounted for 86.5%, mostly from developers.

MiniMax follows the same path, with open platform and enterprise services contributing 63.4%.

This explains why big model firms slash prices and token costs: they're vying for developers' wallets, who are inherently price-sensitive.

C-end subscriptions also matter. MiniMax's Talkie/Xingye and Hailuo AI monetize via membership fees, ranging from tens to thousands of yuan monthly. Many Hailuo paying members are creators seeking high-quality dance or combat animations. However, this ceiling is clear—AI-native apps' user willingness to pay hasn't yet proven capable of sustaining a $100 billion company's ambitions.

Overseas cases offer reference points.

Anthropic generated $16.8 billion in Q3 2026 revenue, up 45% quarter-on-quarter (Q1: $4.7B, Q2: $11.6B). OpenAI's annualized revenue nears $50 billion, backed by 1.2 billion weekly active users and 2.5 million enterprise clients.

Like their Chinese counterparts, overseas AI giants face the same dilemma: impressive revenue growth but no closed "profit loop."

To date, Anthropic's cumulative funding exceeds $125 billion, with operating losses still at $8 billion. In October, it borrowed $9 billion in subordinated debt from Blackstone at over 10% interest, while IPO materials revealed plans to spend up to $750 billion more on compute. Higher financing costs reflect weaker investment appetite or difficulty raising capital.

OpenAI, valued at $1.4 trillion, is still raising $30 billion in fresh funding.

Viewed together, the conclusion is clear: the API business model is gaining traction, and revenue growth is real, but a "profit inflection point"—where revenue growth outpaces cost growth, and gross margins cover R&D and compute—remains elusive.

Globally, no big model startup has yet delivered a "self-sustaining" annual report.

Profitability claims mostly signal emerging scale effects, insufficient to offset impending $100 billion+ capital expenditures.

[Chaoyang Capital Theory] dubs this stage the "paid validation phase": money is coming in, but not enough to outpace burn rates.

Burn rates are endless; funding is a never-ending ticket

Since profits remain elusive, firms must explain their perpetual cash needs.

Ultimately, big models burn cash on three fronts: compute, R&D, and "next-stage" arms races.

Compute is the biggest expense.

Anthropic's roadshow projected $750 billion in future compute spending—an extreme but industry-representative figure.

Tencent's books reveal compute's toll: 2025 capex hit 79.2 billion yuan, with UBS forecasting 170 billion yuan in 2026. In Q1 alone, new AI products dragged Tencent's operating profit down by ~8.8 billion yuan.

Giants can offset AI costs with profits from gaming, advertising, and cloud services. Independent big model firms lack this lifeline and must fund every compute dollar themselves.

R&D is the second cost. In H1 2026, Zhipu spent 2.131 billion yuan on R&D—2.2x its revenue. MiniMax spent ~1.982 billion yuan, ~2.5x its revenue.

Both "outpace revenue with R&D" to secure next-gen model capabilities.

In 2025, MiniMax's R&D spending was 1.689 billion yuan versus 580 million yuan in annual revenue—losses stem largely from R&D, a fate shared by model firms.

The third cost is subtler but explains persistent cash needs.

As competition enters the "next stage," standalone text capabilities are insufficient. Multimodality, agents, computer use, and self-improvement raise engineering stakes.

MiniMax's M3.1 Flash tripled inference speed and cut costs by 70%, but its base model iterated from M2 to M2.7 in 1–2 months, while video model Hailuo upgraded from 1.0 to 2.3 in just four months.

Cheaper, better models drive higher usage; higher usage inflates compute bills; bigger bills demand more funding. This self-reinforcing cycle keeps big model firms perpetually hungry.

Primary markets keep pouring in cash, but attitudes are diverging.

DeepSeek's ~100 billion yuan round valued it at 500 billion yuan pre-money (~148x price-to-sales), with CATL and Tencent as major investors and founder Liang Wenfeng contributing ~20 billion yuan. Yuezhi Anmian raised $3.5 billion in June at a $35 billion valuation, soaring to $50 billion four months later.

Secondary markets reacted tellingly: on October 22, MiniMax broke its issue price, while Zhipu's valuation fell 1 trillion yuan from its peak.

Primary markets assign valuations; secondary markets slash them—perhaps because the market believes AI's long-term story but now scrutinizes "burn-for-growth" efficiency.

For MiniMax, an A-share listing is a race against time: capitalize on primary market enthusiasm before it fades entirely and secure a larger A-share pool.

A-share gambit: Odds of success? Valuation discounts?

MiniMax's A-share journey began in late May, with the second tutorship progress report now out, stating preparations for final approval.

Zhipu moved faster, completing June tutorship approval and targeting 15 billion yuan in fundraising (12 billion for general-purpose base models, 2 billion for MaaS platforms).

Success hinges on smooth tutorship approval, regulatory clearance, and market acceptance of valuation during the launch window.

Historical precedents exist. Cambricon (688256.SH), which also listed unprofitable on the Sci-Tech Innovation Board, now has a ~576.3 billion yuan market cap and trades at over 170x TTM P/E, having once surpassed 1 trillion yuan in valuation.

This shows that the Sci-Tech Innovation Board's pricing tolerates "high R&D, unprofitable, hard-tech" firms—a narrative regulators and markets understand.

But valuations are trickier.

Ignoring broader Hang Seng declines, Hong Kong valuations for the duopoly are at historic lows.

On October 22, MiniMax closed at HK$160.7—a new low—valuing it at ~HK$57 billion (352 million shares). In March, it peaked at HK$1,330, nearing HK$400 billion in market cap.

Zhipu closed at HK$453 on October 22, valuing it at ~HK$220.9 billion—down over 80% from its June 22 peak of HK$2,980 (~HK$1.07 trillion market cap).

Even after crashes, MiniMax trades at ~40x price-to-sales (annualizing H1 revenue), while Zhipu trades at ~150x (using price-to-sales for loss-making firms).

For A-share context, consider Cambricon's 170x P/E, reflecting a "domestic compute scarcity premium."

To earn a similar premium, big model firms must convince A-share investors of two things: (1) API and enterprise service revenue can sustain triple-digit growth, and (2) gross margins can rise to cover R&D.

MiniMax's H1 gross margin was 17.9%—far from that threshold, but trending upward.

If it lists successfully, valuation will likely anchor between two points: Hong Kong's current prices won't demand deep discounts, but Sci-Tech Innovation Board's AI scarcity premium won't be fully granted either.

Zhipu's 15 billion yuan fundraising target offers a regulatory and market-accepted "absorption capacity" reference.

The real significance of the A+H dual-platform for MiniMax lies in adding insurance to its financing: the Hong Kong stock market provides overseas liquidity and access to international investors, while the A-share market offers refinancing capabilities and valuation re-rating favored by mainland investors. If sentiment cools on one side, the other can still sustain operations.

The trade-off, however, is the dilution pressure from the dual-platform structure and the scrutiny of the same narrative in two markets—the price comparison effect serves as the most direct constraint on management.

The capital story in the large language model industry is, when stripped down, not overly complex: burn cash, raise funds, and repeat.

Going public simply shifts the financing channel from the primary to the secondary market, while returning to the A-share market further expands the financing radius.

Relying solely on financing is not a sustainable strategy in the long run; the most critical card at the table remains the ability to 'generate self-sustaining cash flow.'

Source: Chaoyang Capital Theory