Meta surged over 11% on September 21st, marking its best single-day gain since April of last year, with market value climbing by more than $190 billion in one session.

The catalyst was the recent launch of Meta's personal AI agent, Muse. Sensor Tower data shows that within six days of release, the app had accumulated over 902,000 downloads on the U.S. iOS store, far surpassing the performance of Meta's previous products; by Monday, Muse had simultaneously claimed the top spot on both the U.S. iOS and Google Play free charts, dethroning ChatGPT.

Markets have responded enthusiastically to Muse's success. Technology analyst Ben Thompson noted in his article "Frontier Overhangs": "Muse Spark 1.3 is not the most advanced model, but that's precisely the warning signal—a model that isn't leading is already sufficient to power a highly sticky personal agent product." Thompson argues that frontier model capabilities have now become "oversupplied," meaning model performance is no longer the most critical moat for AI companies, and the next competitive edge may well be the "application moment."

As model capabilities cross the threshold of sufficiency, the center of gravity in AI competition is shifting. From Meta's Muse on the consumer side to Tencent's WorkBuddy on the enterprise side, a common signal is emerging: what determines product value is increasingly not the model itself, but control over the harness, workflows, and user entry points.

On September 22nd, Tencent mirrored Meta's overnight rally, with Hong Kong-listed shares jumping over 6% in early trading, reaching an intraday high of HKD 463.40.

Meanwhile, research from consulting firms and venture capital is validating the same conclusion from different angles. Gartner warns that by 2030, up to $234 billion in enterprise software spending faces "agent arbitrage" risk; a16z explicitly advises AI application companies to stop pricing by token and shift to outcome-based pricing; ICONIQ's survey of over 300 AI software companies shows that the application layer has become the industry's construction focus, with AI product gross margins expected to rise from 45% in 2025 to an estimated 53% in 2026.

Is the "Application Moment" Approaching?

Thompson introduces a key concept: when frontier model capabilities continue to improve, but actual products and user scenarios cannot keep pace in absorbing those capabilities, a "frontier overhang" emerges—an oversupply of model capacity.

He cites a concrete example: AI narrative game company Fable launched a next-generation model, but market demand was tepid. This wasn't due to poor product quality, but because existing model capabilities were already "good enough" for that scenario—stronger models didn't deliver a user-perceptible experience improvement, which naturally failed to translate into higher willingness to pay or greater market share.

Thompson believes that when performance is insufficient, whoever binds the model tightly to the application layer (harness) wins; once the "good enough" line is crossed, customers begin to care about time-to-market, convenience, customization, and data handling, and modularity gains the advantage.

This detail is indicative. It shows that in certain scenarios, the marginal value of model capabilities is diminishing. When "stronger models" no longer equal "better product experiences," the competitive advantage of the model layer begins to decouple from the commercial value of the application layer.

Thompson further points out that in this context, what truly matters is the "harness"—how model capabilities are effectively organized, orchestrated, and packaged to reliably and consistently complete tasks in specific scenarios. The model is the engine; the harness is the transmission system. Without the latter, no matter how powerful the former, output cannot be effectively delivered. This framework provides a clear coordinate system for understanding the product logic behind Meta's Muse and Tencent's WorkBuddy.

The Consumer Side: Meta Muse's Product Bet

Meta's launch of Muse is a classic "application layer bet." According to Meta's official description, Muse is positioned as a personal AI agent, with its underlying model being Muse Spark 1.3, but the product's core selling point is not model parameter scale or benchmark scores—it's the agent's ability to actually complete tasks: helping users make plans, execute steps, track progress, and genuinely "get things done" in real-life scenarios.

Behind this positioning is Meta's judgment on the AI product competitive landscape: when model capabilities are already relatively abundant, what users truly lack is not "smarter models," but "tools that can help me finish things." Muse aims to become the user's personal execution layer, not merely a Q&A interface.

Looking deeper, Muse's product focus shifts the breakthrough point for consumer-side agents from "is the model smart enough" to "can users confidently delegate tasks, and can the product connect to real-world transactions." The Muse Spark series helps Meta fill in foundational capabilities, but what determines whether an agent enters users' daily lives is not just reasoning and generation abilities—it also includes authorization, trust, and execution chains. To this end, Muse incorporates an isolated runtime environment, credential protection, and approval for critical actions into its product design to establish clear authorization boundaries; it also combines long-term memory with background execution capabilities, allowing the agent to continuously advance toward user goals rather than stopping at one-off Q&A or command responses.

From a business logic perspective, this design has deeper implications. Once a personal agent becomes deeply embedded in a user's daily workflow, it accumulates vast amounts of personal data, preferences, and usage habits, creating hard-to-transfer user stickiness. This stickiness is the true moat—not the capability of any specific model version.

The market's positive reaction to Muse's launch validates this logic to some extent: investor interest in agents that "work" has already surpassed attention on "stronger" models.

The Enterprise Side: Tencent WorkBuddy's Corporate Path

On the enterprise front, Tencent's WorkBuddy exhibits similar product logic. According to Tencent Cloud's official description, WorkBuddy Enterprise Edition's core capability lies not in providing a generic chat interface, but in connecting to various systems and processes within a company, executing work tasks across platforms, and truly embedding AI capabilities into daily business operations.

This represents a critical leap for enterprise AI—from "tool" to "execution layer." In the past, purchasing an AI product was essentially buying a smarter search or Q&A tool; the direction WorkBuddy represents is making AI an agent that can proactively initiate, execute, and complete work tasks, directly participating in business processes.

The commercial implications of this shift are equally profound. When AI can complete tasks across multiple software systems, the "functional value" of traditional enterprise software faces direct challenges. Gartner clearly stated in a July report that the rise of agentic AI puts up to $234 billion in enterprise application software spending at risk—because agents can bypass or replace workflows that previously required multiple independent software products, potentially eroding the "functional moats" of traditional software.

For enterprise software companies, this is a structural pressure that demands serious attention.

Repricing Commercial Value: From Tokens to Outcomes

The "good enough" phenomenon in model capabilities is driving a fundamental transformation in AI business models.

Venture capital firm a16z directly highlights this trend in its report: the real value of AI applications comes from the data, tools, workflows, and quantifiable results they integrate and deliver—not from which model was called or how many tokens were consumed. In other words, the pricing logic of "selling model capabilities" is yielding to the pricing logic of "selling outcomes."

a16z outlines three pricing models: selling model access, priced per token; turning models into useful work, priced by value units customers recognize, typically credits; and delivering clearly attributable business outcomes, priced by results. Among 50 enterprise AI technology buyers surveyed, 27 preferred "credits tied to identifiable work," while 14 preferred per-token pricing.

This has direct implications for the value distribution across the AI industry chain. If models themselves gradually become a foundational capability—similar to computing power in the early cloud era—then the companies truly capturing commercial value will be those application-layer players that build concrete products, vertical workflows, and verifiable ROI on top of models.

McKinsey provides corroborating evidence from the enterprise perspective in its report: as token costs continue to decline, the core question enterprises ask when evaluating AI investments has shifted from "how strong is this model" to "how much quantifiable efficiency gain and economic return can this agent deliver." The verifiability of ROI is becoming the central criterion in enterprise AI procurement decisions.

This means AI application companies must answer a different question—not "what model did we use," but "how much money can we save our customers and how much efficiency can we add."

Industry Structure: Application-Layer Competition in a Multi-Model Era

ICONIQ's survey of more than 300 AI software companies reveals a clear industry trend: nearly two-thirds of surveyed companies are developing horizontal or vertical AI applications, and multi-model usage has become the norm.

The fact that "multi-model usage has become the norm" is itself a powerful testament to the commoditization of models. When enterprises can flexibly switch between different models based on task requirements, model providers' bargaining power is suppressed, while application-layer platforms that effectively integrate multiple models and build stable workflows gain stronger negotiating positions.

ICONIQ's report also tracks changes in AI revenue and gross margins. The importance of this data dimension lies in its ability to help the market determine whether the application layer's commercial value capture is actually happening or remains mere narrative. For investors, this is the key metric distinguishing "AI application stories" from "AI application businesses."

Looking at the industry structure, the emerging landscape is: the model layer (OpenAI, Anthropic, Google, Meta, etc.) provides foundational capabilities, the application layer (various vertical agents and workflow platforms) transforms those capabilities into deliverable results, and traditional enterprise software companies face the pressure of being bypassed or replaced. The value distribution among these three layers will be the most dynamic phenomenon worth observing in the AI industry over the coming years.

The Seating-License Logic of Traditional SaaS Is Being Pierced Through

If agents can complete tasks across systems, employees no longer need to frequently enter every traditional software interface. The software may still run, but the interface recedes into the background.

Gartner projects that by 2030, up to $234 billion in enterprise application software spending will face risk from agentic AI, representing roughly 20% of enterprise application SaaS spending.

Gartner managing partner George Brocklehurst says: "Agentic AI changes the economic logic of software. Agent systems deliver outcomes directly, bypassing interface-dependent applications and making software invisible." The traditional link between seat count and software revenue growth may therefore be weakened.

The pressure falls first on vendors that rely on dashboards, feature modules, and per-seat licensing fees. Enterprise buyers won't increase budgets just because there's an AI button added; they'll ask whether agents reduce manual operations, shorten processes, improve conversion rates, or lower unit costs.

But this is not the end of traditional SaaS. Vendors that possess industry data, customer relationships, system connections, and business process control can still reinvent themselves as agent execution layers. The risk lies in clinging to interface-based pricing; the opportunity lies in occupying the position of task execution and outcome delivery.

Does Model Leadership Still Translate Directly into Revenue?

It should be clarified that the above analysis does not conclude "models don't matter." Frontier model capabilities remain the foundation of the entire AI industry—without sufficiently strong models, the application layer cannot be built. The question genuinely worth discussing is: can model leadership still directly translate into product advantages and commercial value?

Based on current evidence, this translation path is becoming longer and more complex. The case of Fable's new model facing weak demand shows that in certain scenarios, stronger models can no longer deliver perceptible user value improvements; the product logic of Meta Muse and Tencent WorkBuddy shows that major tech companies are already building differentiation through "harness + workflow + user entry points" rather than relying solely on model capabilities; the research from a16z and McKinsey shows that enterprise customers' willingness to pay is shifting from "model capabilities" to "quantifiable outcomes."

This means AI industry competition is entering a new phase: model capability is a necessary condition, but no longer a sufficient one. Whoever can transform general-purpose models into concrete product experiences, stable workflows, and verifiable commercial returns will truly secure a favorable position in the next stage of competition.

This shift also points to a valuation framework worth re-examining: in AI, model parameter scale and benchmark rankings certainly matter, but the depth of application-layer workflows, the stickiness of user entry points, and the verifiability of ROI may be becoming more predictive indicators of commercial value. When models become "good enough," the battlefield for the next war has already quietly moved.