Recently, SenseTime—an AI company that has carried the banner of the "Four Little Dragons of AI" for eight years—released a positive profit alert: It expects to record a profit of RMB 500 million to 700 million for the first half of 2026, compared to a loss of RMB 1.489 billion in the first half of 2025. This marks SenseTime's first instance of merged profitability since its listing.

One announcement ignited a stock rally, but behind it lies a series of nearly concurrent events: Zhipu unveiled its new-generation foundation model GLM-5.3, boasting a 50% improvement in programming capabilities over its predecessor; Anthropic reported Q2 revenue exceeding $11.5 billion, up over 13x year-on-year; OpenAI's annualized revenue surpassed $40 billion, doubling from last year; DeepSeek's API prices fully increased at midnight on August 17. When these signals align, they point to the same conclusion: The AI industry is approaching its "moment of truth."

01. A Symbolic 'First'

What makes SenseTime's profit alert noteworthy is its significance.

This is not a quarterly fluctuation for a minor player but one of China's most iconic AI unicorns achieving merged profitability for the first time. The announcement attributes the profit improvement to two key factors: First, reduced losses in its core business segments in H1 2026 compared to the previous year; second, gains from fair value changes in its strategic AI ecosystem investments.

It is important to note objectively that SenseTime's "turnaround" is not purely from operational cash generation—its adjusted net loss (excluding non-IFRS items like investment gains/losses) is expected to narrow by 60-70% year-on-year, indicating the core business remains unprofitable, albeit with significantly reduced losses. However, the directional shift is real: In 2025, SenseTime's revenue hit RMB 5.015 billion, up 32.9% YoY (a record high and the fastest growth in three years), while net losses narrowed by 58.6%. EBITDA turned positive for the first time post-listing in H2 2025. The trajectory from "chronic losses" to "positive EBITDA" to "merged profitability" describes a classic industry transition from investment phase to harvest phase.

SenseTime CEO Xu Li stated bluntly at the annual results briefing: "2026 marks the true scaling (mass-scale) outbreak (breakout) phase for AI." His roadmap emphasizes native multimodality, agent-native interfaces, ultimate (ultimate) cost reduction per unit of intelligence, and scaling (mass-scale) visual AI monetization—in short, AI must evolve from an auxiliary tool into genuine productivity.

02. The Technical Race: The 'Post-Training' Revolution Behind GLM-5.3

Two days before SenseTime's profit alert, on August 14, Zhipu officially launched its new-generation foundation model GLM-5.3. A subtle yet telling detail: The foundation model itself remained unchanged from GLM-5.2—what evolved was "post-training."

Post-training refers to the phase after pre-training where model capabilities are further refined through techniques like instruction tuning and reinforcement learning. If pre-training is like constructing a building, post-training is the interior Decoration (renovation)—the same foundation can yield vastly different results based on execution. Through ultimate (extreme) post-training scaling, Zhipu significantly raised the model's intelligence ceiling: Programming capabilities improved by 50% over GLM-5.2, achieving top open-source rankings on benchmarks like TerminalBench 3.0 and Agents' Last Exam, with some high-difficulty tasks even surpassing Anthropic's Claude Opus 4.8. The company announced it would release model weights two weeks after launch.

The significance of this technical approach lies in efficiency. While overseas giants continue burning cash on larger-scale pre-training, domestic models prove that "renovation" can yield stronger capabilities from the same foundation. This epitomizes the industry's shift from "parameter racing" to "engineering excellence"—in the commercialization era, intelligence density per unit cost matters more than total parameter count.

Concurrently, Alibaba open-sourced its Qwen3.8 model series on August 14, bringing its total open-source models to over 460 for the global community. Qwen's global downloads surpassed 3 billion, with over 300,000 derivative models. The thriving open-source ecosystem is exporting China's large model capabilities to global developers—a trend corroborated by Hugging Face's report that Chinese open-source models account for 41% of global downloads.

03. Overseas Parallel: AI Giants Growing Revenue 'Multiples'

Zooming out, this inflection is global.

Anthropic reported Q2 revenue exceeding $11.5 billion, up over 13x YoY—though partly driven by base effects, the growth trajectory is staggering. OpenAI's annualized revenue surpassed $40 billion, doubling from ~$20 billion at the end of 2025, with July monthly revenue growth exceeding 20% across channels like ChatGPT subscriptions, AI coding tool Codex, and enterprise ChatGPT Work. Both companies have filed confidential IPO documents, with Anthropic potentially going public as early as this fall at a rumored $2 trillion valuation.

Wall Street is now valuing AI companies using a new framework: Not "burning cash for users" but a combination of 2028 revenue projections and revenue multiples. Anthropic projects 2028 revenue to reach $190-200 billion—a figure itself representing the capital market's most aggressive bet on AI commercialization.

04. From 'Burning Cash' to 'Making Money': The Return of Pricing Power

Domestic AI industry changes are equally dramatic, with pricing as the most visible signal.

At midnight on August 17, DeepSeek's new API pricing took effect: Its flagship V4-Pro model saw output prices rise from RMB 6 to RMB 27 per million tokens during peak hours (a 350% increase), while introducing a peak-valley pricing mechanism akin to electricity systems. Just three weeks prior, "rock-bottom prices" dominated domestic large model competition; now, the former "price disruptor" is leading the charge. Previously, Zhipu had raised API prices three times, with Moonshot AI's Kimi flagship output price reaching RMB 100 per million tokens, and Tencent Cloud's Hunyuan seeing some interface prices surge over 460%.

The confidence to raise prices stems from demand: Daily domestic large model token calls have reached 140 trillion, up over 1,000x from early 2024, with agent applications consuming 100x more compute per task than ordinary dialogues. As call volumes explode 1,000-fold, compute costs become a rigid constraint, making low-price subsidies unsustainable. Price hikes are essentially a market-driven rebalancing of supply and demand, marking AI's shift from "traffic logic" to "value logic."

SenseTime's turnaround, Zhipu's price hikes, and Anthropic's revenue surge may seem like independent events but are actually three facets of the same industrial transformation: Technology is strong enough, demand is vast enough, and charging has become viable.

05. The Flip Side of Commercialization: Profit Quality and Path Divergence

However, the "moment of truth" also reveals diverging profit quality and path selection.

SenseTime's merged profit includes gains from fair value changes in AI ecosystem investments—a boost from rising valuations of portfolio companies. The true test of business model viability lies in operational performance after excluding non-recurring items: Adjusted net losses persist, albeit narrowed by 60-70%. For investors, distinguishing "accounting profit" from "operating profit" is crucial—the former signals an inflection point, while the latter confirms value.

Path divergence within the industry is also intensifying. Small-to-medium AI tools without proprietary large models, relying solely on API repackaging, face soaring costs after head (leading) vendors' collective price hikes, accelerating market consolidation. Players able to optimize costs through off-peak scheduling, cache hit rate improvements, and model tiering gain competitive edges. The AI application market is evolving from "wild growth" to "survival of the fittest"—a hallmark of industry maturity.

A deeper divergence exists: Foundation model vendors monetizing through "model royalties," application vendors monetizing through "scenario landing," and cloud vendors monetizing through "compute infrastructure" have vastly different profit timelines. Anthropic and OpenAI earn model premiums, SenseTime profits from ecosystem investments and visual AI scale, while domestic cloud vendors capitalize on compute demand expansion—together forming distinct quadrants of the AI commercialization landscape. This inflection will not benefit all equally.

06. The Inflection Year: The Story's Turning Point Has Just Begun

Viewed through a longer lens, 2026 for AI resembles 2016 for mobile internet—technology maturation, demand explosion, and initial commercialization model validation, with the industry shifting from "storytelling" to "accountability."

A set of data underscores this turning point's magnitude: Anthropic's Q2 revenue grew 13x YoY, OpenAI's annualized revenue doubled in a year, SenseTime achieved merged profitability for the first time post-listing, DeepSeek transitioned from "rock-bottom" to "rational" pricing, and domestic open-source models accounted for 41% of global downloads. Each milestone was once deemed "impossible," yet their concurrent emergence can only mean one thing: AI's commercialization inflection point has truly arrived.

Of course, risks must be acknowledged. The quality of profitability (accounting vs. operating), demand elasticity post-price hikes, the realizability of overseas IPO valuations, and whether Chinese AI firms can sustain iteration amid high-end compute restrictions will determine if this inflection represents a solid turning point or a temporary rebound.

From burning cash to recouping investments, from subsidies to pricing power, from concepts to revenue. The AI industry has taken eight years to reach its "moment of truth"—and the story's turning point has only just begun.

- End -