Zhipu AI shares surge 35% as annual revenue more than doubles

Zhipu AI (HK:2513) reported 2025 revenue of HK$724.3m, up 131.9% YoY, and shares rallied as much as 35% to a record HK$938.0. Net loss widened 60% to 4.7 billion yuan and gross margins contracted, even as on‑premise deployments made up 74% of revenue and cloud revenue jumped nearly 300% (but at lower margins). The company cited strong demand and GLM adoption—its platform surpassed 4 million registered users by March 2026 and nine of China’s top 10 internet firms are deeply integrated with GLM—supporting the sector rally despite ongoing profitability risks.

Analysis

The market is rewarding execution on AI model iteration, but the durable winners will be those that convert model adoption into sustainable unit economics rather than just headline traction. A shift in product mix toward hosted/cloud delivery compresses gross margins because hosting, networking and customer support create recurring variable costs — that dynamic favors capital-light model licensors only if they retain pricing power or can monetize higher-tier services.

Second-order beneficiaries are predictable: inference-optimized hardware and software (GPU/accelerator vendors, NICs, inference runtimes) and specialized systems integrators that eliminate friction for on-prem deployments. Conversely, generalist cloud resellers and high-touch professional services face margin pressure as open-source model stacks lower switching costs and enable self-service deployment by savvy enterprise IT teams.

Key catalysts and risks are asymmetric across horizons. Over days–weeks, flows and sentiment swings (positioning around headline AI winners) will drive price moves; over months–years, two structural forces matter most: (1) commoditization of base models (downward pricing pressure) and (2) geopolitically driven constraints on advanced semiconductors or software exports, which can either choke supply or protect domestic incumbents. A reversal could be triggered by a meaningful improvement in monetization (tiered enterprise pricing, long-term contracts) or by regulatory/controls tightening that narrows the competitive set.

Contrarian angle: current prices imply flawless scale-out without sustained margin erosion. That is a high bar — if customers shift to lower-cost hosted alternatives or if model costs per token fall materially, multiples will re-rate quickly. We prefer to express conviction through structured, time-boxed trades that capture further adoption while limiting exposure to a rapid margin re-pricing.

Key Decisions for Investors

• Long 2513.HK via a 3-month call spread (buy 20% OTM call, sell 50% OTM call) sized 2–3% NAV on a <=10% pullback; rationale: captures upside if adoption/contracting continues while capping premium paid. Risk: limited to paid premium; reward: asymmetric (target 2x–4x premium if sentiment/sticky contracts follow).
• Buy NVDA (NVDA) for 6–12 months (allocate 3–5% NAV) to play durable on‑prem inference demand; hedge with 1–2% NAV of 3–6 month puts to limit 20% drawdown. Risk/Reward: downside protected, upside participation if compute cycle sustains (30–100% potential depending on next earnings/cycle).
• Pair trade (3–9 months): long pure-play model vendor exposure (equal notional to 2513.HK position) / short a large China cloud operator (e.g., BABA) to capture margin-compression divergence. Risk: cloud capex consolidation or hosting monopolies that win scale; reward: captures re-rating if software/edge vendors preserve higher unit economics.
• If already long winners, buy 2–4 week out-of-the-money protective puts (or collars) to guard against headline-driven reversals; keep delta small to avoid drag but large enough to limit 25–35% downside over the next earnings/positioning window.