Zhipu’s H1 Revenue Surges 400% as API Pivot Cuts Gross Margin in Half
Zhipu AI Technology (02513.HK) delivered a revenue explosion in the first half of 2026—but the same pivot that drove a 27-fold surge in API income also cut gross margins in half, exposing the costly arithmetic of transitioning from high-margin enterprise software deployments to a volume-driven, cloud-native business model.
The Beijing-based AI developer reported H1 2026 revenue of RMB 954 million (US$132.5 million), a 399.7% year-on-year increase that already surpasses its full-year 2025 revenue of RMB 724 million (US$100.6 million) by 32%. Shares closed up 9.63% at HK$1,195 on August 31—the same day the company was formally added to the MSCI China Index—giving it a market capitalization of HK$556.4 billion (US$71.3 billion). Yet the stock remains nearly 60% below its all-time high of HK$2,980 touched on June 22, a gap that encapsulates the market's unresolved debate over when growth translates into sustainable earnings.
API Volumes Explode, Reshaping Revenue Mix Overnight
The single most consequential data point in Zhipu's interim results is not top-line growth but the velocity of its revenue-mix rotation. Open-platform and API revenue surged 2,735.7%—roughly 27-fold—from RMB 29.1 million in H1 2025 to RMB 825 million (US$114.6 million), lifting its share of total revenue from 15.2% to 86.5% in just twelve months.
The inverse is equally striking: on-premise localized deployment revenue—historically Zhipu's cash engine—contracted 20.5% to RMB 129 million (US$17.9 million), while private enterprise large-model deployments collapsed 54.6% to RMB 67 million (US$9.3 million). The share of on-premise business in total revenue fell from 84.8% to 13.5%.
Board Secretary Xiao Lei framed the shift as structural rather than cyclical during the earnings call: "The revenue-composition swap that occurred in H1 2026 is an inevitable consequence of model capability crossing a generational threshold." Management articulated the commercial logic as a four-stage progression—selling models, selling API calls, selling subscriptions, selling end-to-end task outcomes—arguing that once GLM (General Language Model) crossed the threshold of autonomously executing complete engineering projects, recurring API and coding-plan subscriptions became the dominant commercial form.
Operational metrics corroborate the narrative. As of August 31, Zhipu's MaaS (Model-as-a-Service) platform annualized revenue run rate (ARR) stood at US$1.6 billion on a monthly basis and US$2.0 billion on a weekly basis. Enterprise and developer users exceeded 7.4 million; token call volumes grew more than 40-fold from the start of 2026; paying daily active users rose 603%; and average API selling prices doubled, up approximately 101% since January.
Margin Compression Signals the True Cost of Scaling Cloud Infrastructure
The profitability picture is more complicated. Gross profit reached RMB 252 million (US$35 million), up 163.7% year-on-year, but overall gross margin fell from 50.0% to 26.4%—a 23.6-percentage-point compression driven entirely by the mix shift rather than deterioration in any individual segment.
Critically, the API business itself turned gross-margin positive, improving from -0.4% to 24.6%. But because it now constitutes 86.5% of revenue while the high-margin on-premise segment has shrunk to 13.5%, the blend pulls the consolidated figure sharply lower. Enterprise AI agent gross margin declined from 64.6% to 35.6%; enterprise general large-model gross margin fell from 58.5% to 42.3%. The fastest-growing segment remains the least profitable one.
Cost of sales surged 635.4% to RMB 702 million (US$97.5 million)—a growth rate that outpaced even the 399.7% revenue increase—driven by rising compute service fees as Zhipu scales inference capacity. Management acknowledged that domestic compute supply remains constrained by heterogeneous chip architectures, with large-scale domestic chip availability still in early stages. Xiao Lei projected that within three to six months, a wave of advanced domestic chip suppliers would enter volume production, which could meaningfully reduce unit inference costs.
Operating Losses Widen Once Accounting Noise Is Stripped Away
The reported net loss of RMB 2.072 billion (US$287.8 million) narrowed 12.1% year-on-year, but this improvement is largely an accounting artifact. The prior-year period included RMB 429 million in fair-value losses on financial instruments issued to pre-IPO investors; post-listing, that line item collapsed to RMB 22.1 million in H1 2026. Strip out that effect alongside share-based compensation and listing expenses, and adjusted net loss widened 12.1% to RMB 1.964 billion (US$272.8 million). On a pure operating basis, the loss expanded 13% from RMB 1.899 billion to RMB 2.147 billion (US$298.2 million).
Research and development spending of RMB 2.131 billion (US$296 million) rose 33.6% year-on-year and equated to 2.23 times H1 revenue—meaning Zhipu spent RMB 2.23 in R&D for every RMB 1.00 earned. Sales and marketing expenses fell 14.8% to RMB 178 million, and general and administrative costs dropped 44.2% to RMB 103 million, demonstrating deliberate overhead discipline. But R&D's structural growth continues to overwhelm those savings.
Cash and equivalents stood at RMB 3.994 billion (US$554.7 million) at period-end. However, IPO proceeds are nearly exhausted: net proceeds of approximately HK$45.88 billion (RMB 39.7 billion, or US$5.5 billion) have been deployed, representing over 93% of total net IPO fundraising. A HK$31.4 billion (US$4.0 billion) placement completed in July now serves as the primary liquidity buffer for the next phase of investment.
GLM Roadmap Bets on Self-Training Architecture to Differentiate
Goldman Sachs maintained a Neutral rating, noting that GLM-5.3—released during the reporting period—is an iterative update sharing the same base architecture, total parameter count, and activated parameters as GLM-5.2. The bank cited the still-pending flagship GLM-5.5 and narrowing performance gaps versus peers as factors compressing Zhipu's valuation premium. Morgan Stanley, by contrast, raised its price target to HK$1,800, arguing that improved compute access and the July placement provide a stronger growth runway. CMB International kept a Buy rating, pointing to GLM-5.3's advances in coding, agentic tasks, and cybersecurity as near-term positive catalysts.
Founder and Chief Scientist Tang Jie addressed skepticism over Zhipu's parameter-constrained approach directly. He argued that model scale must be assessed across three axes simultaneously—parameter count, training data volume, and compute allocation—and noted that domestic training datasets currently range from 30 to 50 trillion tokens. Under compute constraints, marginal returns from parameter expansion alone are limited. GLM-5.3's end-to-end coding completion rate improved more than 50% over GLM-5.2 through one month of extended long-horizon task environment training and reinforcement learning, without any architectural change.
Tang disclosed that the next-generation GLM-6.0 will follow a Full Self-Training paradigm—a model capable of autonomous self-purification across pre-training, mid-training, and post-training stages, including self-directed training termination and error correction. He described the core challenge as not scale but self-judgment: the model's ability to autonomously determine when training should stop and how errors should be corrected. Ethical and social governance dimensions will also be incorporated into subsequent model research, Tang added.
On the commercial application side, Zhipu is extending coding capabilities into what it terms "Co-work" scenarios—professional-grade complex workflow automation rather than general office productivity. Cybersecurity represents the most validated deployment: since GLM-5.2, collaboration with domestic security teams has identified 2,436 expert-verified, deduplicated vulnerabilities across 269 real-world codebases, including more than 1,000 classified as high-severity. Legal, financial analysis, and data analytics verticals are in earlier-stage commercialization, with varying timelines tied to reliability thresholds and verification mechanism maturity.
Impact Assessment: The Margin Trough Is a Deliberate Transition, Not a Structural Failure—But the Clock Is Running
Zhipu's H1 2026 results present a coherent strategic logic: sacrifice near-term margin to capture API volume at scale, improve unit economics as compute costs fall, and ascend the capability ladder toward higher-value autonomous task completion. The trajectory of API gross margin—from negative to 24.6%—and the 101% increase in average API selling price suggest the unit economics are moving in the right direction.
The risk, however, is timing. With over 93% of IPO proceeds deployed, R&D spending running at 2.23x revenue, and operating losses expanding on an adjusted basis, Zhipu's ability to sustain this transition depends heavily on the July placement capital, the pace of domestic chip supply normalization, and whether GLM-5.5 and GLM-6.0 can widen the performance gap before competitors close it. Management's own formulation—"when will this substitution actually translate into improved profitability?"—remains the central question that neither the H1 results nor the current analyst consensus has answered.
Related Coverage:
Zhipu Undercuts DeepSeek With GLM-5.3 Flash on 100,000+ Domestic Chips