The artificial intelligence landscape stands at an inflection point that should profoundly reshape how we think about computational infrastructure. The emergence of agentic AI systems—autonomous agents capable of complex reasoning, coordination, and persistent memory—demands architectural approaches that differ from today's cloud computing paradigms.

This transformation raises critical questions about the $320 billion in projected infrastructure investment by hyperscale providers during 2025 and whether this massive spending targets the right architectural foundations.

Over the first four parts of this Agentic Era series, I’ve established that orchestration creates new competitive moats, that architectural choices between model-centric and workflow-centric approaches determine value capture, that protocols like MCP and A2A form the invisible operating system for coordination, and that outcome-based frameworks must replace traditional SaaS metrics.

For this latest c…

Read more

The marketing department at a mid-sized software company recently made a remarkable decision. Instead of hiring three new specialists to handle content creation, lead nurturing, and campaign analytics, they deployed an agentic AI system for content orchestration and customer interaction management.

The system goes beyond augmenting the human team to increase productivity. It has enabled the creation of “Synthetic Colleagues” that effectively become the marketing team, managing everything from brand-compliant content creation to personalized customer journey orchestration.

Synthetic Colleagues are an emerging pattern that appears likely to reshape software markets over the next decade. In the Agentic Era, we are observing a transition from applications that enhance human productivity to applications that can replace entire organizational functions and assume responsibility for complete workflows.

Traditional AI applications compete for software budgets, typically representing 2-5% of enterprise spending. Agentic applications potentially address much larger budget categories—the salaries, benefits, and operational costs of entire teams. They also facilitate the creation of teams that do not exist because the labor is hard to find. When software can effectively perform the functions of a marketing department or customer service division, the addressable market extends significantly beyond traditional software spending toward the broader costs of organizational cognitive work.

This layer of Agentic AI Applications represents one of the most enticing targets for investors because the economic implications are substantial. A recent report from Market.us forecasts the Agentic AI market to grow from approximately $5.2 billion in 2024 to almost $200 billion by 2034. At its recent Build conference, Microsoft unveiled a series of initiatives to give developers the tools to build Agentic AI applications, a push that places agents at the center of the company’s roadmap.

The result of this phenomenon is a Discontinuity that renders obsolete such classic benchmarks as the Rule of 40. Navigating this turbulent period of transformation requires a new framework. I take the first steps toward addressing the impact of Agentic applications by introducing the Agentic Value Framework. This is my starting point for rethinking valuations and metrics.

The Architecture of Synthetic Colleagues

What makes agentic applications fundamentally different from previous software generations is their architectural sophistication. To understand this distinction, we must first clarify what separates individual AI agents from true agentic systems—and why only the latter qualifies as a "synthetic colleague."

Individual AI Agents: The Specialized Assistant

Traditional AI agents are purpose-built for narrow, well-defined tasks. Think of them as highly specialized assistants: a customer support agent that handles tickets, an email filtering agent that manages your inbox, or a scheduling agent that coordinates meetings. These agents exhibit three core characteristics: autonomy within their specific domain, task-specificity that allows high-performance optimization, and reactivity to environmental inputs.

However, individual agents face fundamental limitations. They operate in isolation, lack persistent memory across interactions, and cannot decompose complex, multi-step objectives that require coordination across different skill sets. A customer support agent can resolve individual tickets efficiently, but it cannot simultaneously analyze support patterns, update knowledge bases, and coordinate with product teams—the kind of integrated workflow that defines actual customer success departments.

Agentic Systems: The Synthetic Colleague

Agentic systems represent a categorical leap from isolated agents to collaborative ecosystems. These are distributed networks of specialized agents that share persistent memory, coordinate through structured communication protocols, and pursue decomposed goals under orchestrated supervision.

Consider a real-world example from recent research: an automated grant proposal system that functions as a synthetic research department. The system consists of four specialized agents—a literature retriever, a compliance alignment agent, a methodology designer, and a formatting specialist—coordinated by a meta-agent orchestrator.

The retriever agent scans funding databases and extracts successful proposal patterns. The alignment agent ensures methodology matches specific NSF solicitation requirements. The methodology agent designs research approaches based on extracted patterns and requirements. The formatting agent structures documents per compliance guidelines. Throughout this process, a persistent memory layer stores evolving drafts, funding agency templates, collaborator feedback, and iterative improvements across multiple sessions.

What makes this agentic rather than simply multi-agent is the emergent behavior: the system exhibits decision-making, learning, and adaptation that exceeds any individual component. It doesn't just execute predefined workflows—it develops new strategies based on success patterns, adjusts approaches based on feedback, and maintains institutional memory that improves over time.

While still early, this architectural sophistication is visible across emerging agentic platforms. Within the field of high-growth startups: Writer, Glean, Sierra at horizontal level - Harvey, Hippocratic.ai at vertical level. Amongst incumbents: AgentForce, NowAssist. And the list continues across hundreds of companies racing to develop this new form of software.

The Orchestration Layer: Where Synthesis Happens

The critical innovation enabling synthetic colleagues is the orchestration layer—sophisticated coordination systems that manage task decomposition, dependency resolution, and conflict arbitration across agent networks. Modern orchestrators don't just route requests; they maintain shared context, evaluate agent performance, and dynamically reassign responsibilities based on workload and capability.

The emergence of protocols like Anthropic's Model Context Protocol (MCP) and Google's Agent-to-Agent (A2A) represents the infrastructure breakthrough that makes this orchestration possible. MCP standardizes how agents access external resources and share context, while A2A enables direct agent-to-agent communication with task management and multi-modal coordination. These protocols solve the fundamental coordination challenges that prevented true multi-agent collaboration: standardized communication at machine speed with persistent state management.

This architectural sophistication is what transforms individual agents into synthetic colleagues. Where traditional software requires human coordination between different tools and systems, agentic applications coordinate themselves, maintaining the persistent context, institutional memory, and collaborative decision-making that characterizes effective human teams.

The Evolution Beyond Data Moats

The emergence of applications based on agentic systems coincides with a fundamental shift in competitive dynamics. Traditional software companies built moats through proprietary data accumulation, believing that exclusive datasets created sustainable advantages through better model performance and user experiences.

Agentic platforms operate under different principles. Foundation models trained on broad datasets and accessible via APIs have commoditized basic intelligence capabilities. The competitive advantage no longer stems from possessing unique data, but from orchestrating how synthetic colleagues access, combine, and act upon information in real-time workflows.

This shift is becoming evident in strategic partnerships across the industry. Glean's integration with Snowflake, announced in May 2025, demonstrates the new approach. Rather than building proprietary datasets that would require massive capital investment and years of accumulation, Glean connects directly to existing enterprise data infrastructure. The partnership enables synthetic colleagues to access and reason over data warehouses in real-time without requiring data duplication or custom model training.

This partnership model reveals the new competitive landscape. Data itself becomes infrastructure—valuable but not differentiating. The strategic advantage migrates to the orchestration layer that enables synthetic colleagues to understand context, maintain memory across interactions, and coordinate actions within specific workflow environments.

However, this does not mean all data advantages disappear. Proprietary behavioral data, unique process insights, and customer interaction patterns retain strategic value when they enable better synthetic colleague coordination or workflow optimization. The distinction lies between raw information assets and actionable intelligence that enhances agentic system performance.

The Context Moat: Controlling the Semantic Layer

The context moat emerges from controlling the semantic layer that agents use to understand, remember, and personalize their interactions. Context-rich agents fundamentally outperform generic alternatives because they can make decisions based on accumulated understanding rather than isolated inputs. An agentic system embedded in Notion doesn't just access documents—it understands project relationships, team dynamics, and workflow patterns that inform every interaction.

The strategic advantage belongs to platforms that sit closest to decision-making and real-time workflows. Notion for knowledge work, Microsoft for productivity workflows, Salesforce for customer relationship management—these platforms have natural context moats when they orchestrate correctly.

The Workflow Moat: Embedding into Actual Tasks

Workflow moats arise from deep integration into the specific processes where value gets created. Consider the difference between an AI agent that generates marketing content versus one that drives complete revenue operations—managing lead qualification, nurturing sequences, and deal progression. The content generator provides utility; the revenue operations agent becomes infrastructure.

The firms that control the orchestration layer—where agents are created, connected, and evaluated across workflows—will be valued like infrastructure companies rather than software vendors.

Evolving Beyond the Rule of 40

The rise of synthetic colleagues and Agentic Applications brings us back to the question of how to rethink the metrics that matter when it comes to measuring success and valuation.

The Rule of 40 has served the software industry exceptionally well for over a decade. By establishing that Revenue Growth % + FCF Margin % should exceed 40%, it provided a clear framework for evaluating SaaS company efficiency and balancing growth investments with profitability. This metric guided countless investment decisions and strategic choices that built the foundation of today's software economy.

However, the Rule of 40 was designed for a different economic model, built on three assumptions that agentic platforms challenge:

- Linear Scaling Assumption: More revenue requires proportional increases in costs (sales teams, support staff, infrastructure). Synthetic colleagues scale exponentially. Each additional agent doesn't just add its individual capacity, it enhances the entire network's capability through shared learning and coordination.

- Growth-Profitability Trade-off: Traditional SaaS faces a choice between investing in growth or optimizing for profitability. Agentic systems eliminate this trade-off—network effects simultaneously drive growth and improve margins as coordination reduces operational costs.

- Human-Mediated Value Creation: The Rule of 40 assumes human bottlenecks limit scaling velocity. Synthetic colleagues operate at machine speed with perfect coordination, removing traditional scaling constraints entirely.

Introducing the Agentic Value Framework

Agentic platforms require a completely different approach to revenue calculation and growth measurement. Traditional ARR calculations assume linear scaling where each customer contributes proportional revenue. Agentic platforms generate exponential value through network effects that compound as synthetic colleagues coordinate and learn from each other.

As I argued previously, the advent of GenAI was already making the Rule of 40 obsolete, replacing it with the Rule of 55 that captures the potential productivity gains from AI. The Agentic Era explodes the Rule of 55 because of the potential scope of activities these applications can perform.

The challenge now is establishing a new benchmark formula at a time when the parameters, technology, protocols, and use cases are still in the embryonic stages.

As I am trying to grapple with this moment of Discontinuity, I’ve created the Agentic Value Framework, a starting point that I intend to refine in the coming months as we gather more empirical evidence:

(Network-Enhanced ARR Growth Percentage) + (FCF Margin Percentage) = Agentic Performance Threshold

Let’s look more closely at the critical components of the formula.

Base ARR Calculation for Agentic Platforms:

Base ARR = (Outcome per Synthetic Colleague) × (Platform Value Capture Percentage) × (Synthetic Colleagues per Customer) × (Number of Customers)

This foundation reflects outcome-based economics rather than seat-based pricing. If each synthetic colleague generates $200,000 in annual business outcomes and the platform captures 20% of that value, the revenue per synthetic colleague equals $40,000. A customer deploying one finance synthetic colleague and one marketing synthetic colleague contributes $80,000 to Base ARR. With 1,000 customers, Base ARR reaches $80 million.

Network-Enhanced ARR Growth:

The agentic evolution necessity lies in measuring how network effects accelerate ARR growth beyond linear customer acquisition. Traditional software grows ARR proportionally with customer additions. Agentic platforms experience compounding growth as existing synthetic colleagues become more valuable through coordination and cross-customer learning.

Network-Enhanced ARR Growth = (Traditional ARR Growth Percentage) × (Agent Network Compounding Factor)

The Agent Network Compounding Factor measures how much faster ARR grows due to synthetic colleague interactions versus linear scaling. Early-stage agentic platforms might demonstrate a 1.3x compounding factor, meaning ARR grows 30% faster than customer additions due to coordination benefits. Mature platforms could achieve 2.0x or higher compounding factors as synthetic colleagues develop sophisticated interaction patterns and cross-customer learning accelerates capability improvements.

Like Rule of 40, this formula does not aim to replace DCF metrics which remain the core foundation but rather measure one of the components of an agentic company’s moat.

However, this framework replaces the Rule of 40's linear assumptions with exponential network effects measurement. The Agentic Performance Threshold should exceed traditional benchmarks—potentially 60% to 100%—reflecting the superior economics available to platforms that successfully coordinate synthetic colleague networks.

Share

Early market signals suggest investors are already recognizing the potential for exponential value creation in agentic platforms. Harvey AI, a legal technology startup, is reportedly in discussions for funding at a $5 billion valuation, according to recent reports. This round, along with similar high valuations across AI-enabled platforms, indicates growing investor recognition that companies deploying coordinated AI capabilities in specific workflow domains may generate value through different mechanisms than conventional software businesses.

Whether these valuations reflect sustainable agentic platform economics or broader market enthusiasm for AI capabilities remains to be demonstrated through operational performance over time.

The Orchestration Platform Imperative

The Agentic Value Framework reveals why orchestration platform control becomes the ultimate strategic imperative. Traditional software companies optimize for user acquisition and retention. Agentic platforms must optimize for coordination dominance.

Orchestration platforms capture value through four compounding mechanisms:

· Direct Outcome Capture: Revenue from synthetic colleagues deployed on the platform

· Network Coordination Premiums: Additional value from coordinated synthetic colleague networks

· Integration Rents: Fees from other platforms routing through the orchestration layer

· Learning Network Effects: Competitive moats from cross-customer synthetic colleague improvement

The winner-take-all dynamics are stronger than traditional software markets because coordination complexity creates natural monopolization. Customers consolidate on dominant orchestration platforms to maximize synthetic colleague coordination efficiency.

This explains why we're likely to see multiple companies achieve $1 trillion valuations in the agentic era—not through incremental software improvements, but through orchestration platform dominance that captures exponential value from network-coordinated synthetic colleagues.

The Strategic Implications

The transition to agentic platforms creates complex competitive dynamics. Incumbent software companies possess significant advantages through existing customer relationships and substantial resources, but face architectural constraints from legacy systems designed for human-centric workflows. Foundation model providers control core reasoning capabilities but often lack workflow integration.

The most compelling long-term position may belong to AI-native startups that can design their entire technology stack from first principles for agentic coordination. These companies can optimize data structures specifically for persistent cross-agent memory, build communication architectures for high-frequency agent interactions, and structure their business models around outcome-based value creation rather than seat-based licensing.

This architectural advantage compounds as system complexity increases, suggesting that ultimate competitive success will flow to companies building agentic systems from their foundational architecture upward.

This transition isn't without risks. As agentic systems scale, they face amplified coordination challenges, error cascades where one agent's mistake propagates through the entire system, and emergent behaviors that weren't explicitly programmed. Unlike traditional software failures, agentic system failures can be subtle, systemic, and difficult to diagnose.

The distributed nature of these systems creates accountability gaps when multiple agents interact to produce outcomes, making it difficult to assign responsibility for errors or unintended consequences. This creates both liability and trust challenges, particularly in high-stakes domains.

However, these risks are architectural problems, not fundamental limitations. Solutions are emerging: retrieval-augmented generation for grounding, tool-based reasoning for accuracy, memory architectures for continuity, and governance frameworks for accountability. The companies that solve coordination at scale will capture disproportionate value.

Software Redefined

The agentic application represents more than an evolutionary step in software development. It's a categorical redefinition. We're moving from tools that amplify human capabilities to systems that replace human functions entirely.

This isn't simply about efficiency gains or cost reduction. It's about fundamentally restructuring how work gets done, who (or what) does it, and how value gets created and captured in digital systems. When applications become co-workers, the entire software industry resets.

The companies that recognize this discontinuity and rebuild their architectures accordingly won't just capture market share. They'll redefine what markets are possible. Those that treat agentic capabilities as incremental features will find themselves competing against synthetic departments that never sleep, never quit, and continuously improve.

The agentic era has begun. The question isn't whether software will replace human cognitive work. It's which companies will orchestrate that replacement most effectively.

Welcome to Part 3 of the Agentic Era mini-series.

Beneath the GenAI hype cycle, a fundamental transformation is unfolding.

While the market remains focused on prompts, tokens, benchmark scores, and the theoretical path to AGI, a more consequential shift is taking place at the architectural level. This evolution in how AI systems connect and collaborate will likely determine which organizations capture lasting value in the coming decade.

The architectural revolution is redefining:

• How defensible moats are built when everyone has access to the same models

• How SaaS businesses maintain margins when basic AI capabilities become commoditized

• How traditional enterprises can extract asymmetric returns from AI investments

• Which metrics matter when evaluating both AI-native and AI-enhanced companies

Two weeks ago, I kicked off this Agentic Era series to analyze this strategic inflection point and its implications for building durable businesses:

Agentic Era Part 1: A Str…

Read more

Beneath the GenAI hype cycle, a fundamental transformation is unfolding.

We continue obsessing over prompts, tokens, benchmark scores, and the race to AGI. Meanwhile, the truly transformative shift that will create and destroy billions in market value is happening at the architectural level.

AI is evolving from tools we operate to workers that operate autonomously. This isn't just a product evolution; it's a fundamental rewiring of how technology creates value.

This architectural revolution is redefining:

• How defensible moats are built when everyone has access to the same models

• How SaaS businesses maintain margins when basic AI capabilities become commoditized

• How traditional enterprises can extract asymmetric returns from AI investments

• Which metrics matter when evaluating both AI-native and AI-enhanced companies

Last week, I kicked off a series analyzing this strategic inflection point and its implications for building durable businesses in what I'm calling the Agentic E…

Read more

The frontier AI race has shifted from raw model power to orchestration, distribution, and efficiency, creating a Discontinuity moment for investors seeking asymmetric returns.

As benchmark scores converge—GPT-4o, Claude 3.7, and Gemini 2.5 cluster within 5% on MMLU—intelligence is commoditized, and orchestration layers like Anthropic’s MCP and OpenAI’s 1B-user ChatGPT interface define new moats. Seven players—OpenAI, Anthropic, Google, xAI, Meta, Mistral, and DeepSeek—pursue distinct strategies, from interface control to enterprise efficiency, visualized in Figure 1’s radar chart. Investors must prioritize firms building financial fundamentals anchored in these new moats —cash flows, margins, scalability—for DCF valuation, not revenue hype.

This is Part 1 of a short series analyzing this strategic inflection point and its implications for building durable businesses in what I'm calling the Agentic Era.

Subscribe now

The foundation model race has entered a new phase—one defined not by model quality but by the fundamental restructuring of how intelligence translates to market power. This shift presents a classic discontinuity moment: traditional valuation frameworks are failing while new moats emerge at an accelerating pace. The race is still on, and I aim to provide a framework for thinking about where value could ultimately accrue.

The Inflection Point Is Behind Us

While the market fixates on headline-grabbing funding rounds and model launches, a fundamental strategic realignment is underway in AI. At The Information's Financing the AI Revolution summit this week, executives discussed xAI Holdings' $20 billion raise at a $120 billion valuation—the second-largest private funding round in history, surpassed only by OpenAI's $40 billion capital infusion last month. Meanwhile, Alibaba's pointed criticism of DeepSeek ahead of its latest Qwen model release underscores the intensifying competitive dynamics in the Chinese AI ecosystem.

Yet these massive fundraising events mask a critical shift: raw model capabilities no longer determine winners in foundational AI. The competitive frontier has moved decisively toward orchestration quality, creating asymmetric return opportunities for investors who recognize this discontinuity before market valuations fully reflect it.

While capital continues flowing toward model development, the true arbitrage opportunity exists in identifying which players are building defensible orchestration layers that connect model intelligence to distribution endpoints where actual value capture occurs. Those who can identify these orchestration winners stand to capture disproportionate returns as the market gradually recognizes this fundamental shift.

Orchestration Is the New Moat

The recent convergence in benchmark scores between leading models (GPT-4o, Claude 3.7, Gemini 2.5) reveals that basic intelligence is becoming commoditized.

Top models—GPT-4o, Claude 3.7 Sonnet, and Gemini 2.5 Pro—achieve near-identical scores (within ~5%) on MMLU and other benchmarks, reflecting commoditization of general intelligence. Additional players (Meta’s Llama, Mistral’s Large, DeepSeek’s R1, xAI’s Grok) show competitive but slightly lower scores, only reinforcing the trend.

What isn't commoditized—and should not be for the foreseeable future—is the orchestration layer: the complex systems enabling models to seamlessly access data, perform multi-step reasoning, and integrate into existing workflows.

Orchestration creates moats by embedding intelligence into user-specific applications, which are harder to replicate than model improvements. Anthropic’s Model Context Protocol (MCP), adopted by OpenAI and Google in Q1 2025, standardizes data-model integration, locking in enterprise users. OpenAI’s ChatGPT drives 500 million weekly users (or perhaps even 1 billion!) through intuitive interfaces.

This orchestration advantage manifests through several critical capabilities:

✅ Connection protocols establishing standardized interfaces between models and external systems, like MCP, adopted by OpenAI and Google in Q1 2025

✅ Agent frameworks guiding models through complex multi-stage reasoning processes

✅ Tool integration systems extending model capabilities through specialized services

✅ Interface design patterns translating raw intelligence into intuitive user experiences

The players who master these dimensions will capture disproportionate value regardless of whose model performs marginally better on the next benchmark.

The Orchestration-Distribution Nexus

Orchestration forms the moat, but distribution is the vector for its value. Orchestration ensures intelligence is delivered effectively—through seamless, user-specific workflows—creating defensibility that competitors struggle to replicate. Distribution, meanwhile, is the reach of that intelligence, leveraging existing channels to scale adoption. The most successful players combine orchestration advantages with established distribution vectors.

OpenAI's interface orchestration, delivered through Microsoft's distribution network, is more powerful than either component alone. Similarly, xAI’s orchestration embeds AI into Tesla’s vehicle fleet and X’s 500 million+ users, creating always-on availability. Anthropic’s MCP, adopted by OpenAI and Google, standardizes agentic workflows, enhancing orchestration across platforms. As Google DeepMind CEO Demis Hassabis noted, MCP is “rapidly becoming an open standard for the AI agentic era”. This orchestration-distribution nexus drives strategic divergence, rewarding specialized approaches over generic capabilities.

Strategic Divergence: Seven Distinct Paths

The frontier AI landscape has fractured into seven distinct strategies, each pursuing unique defensibility.

Driven by model commoditization, companies are building moats through orchestration, distribution, or efficiency, not raw intelligence.

We focus on seven prominent players—OpenAI, Anthropic, Google, xAI, Meta, Mistral, and DeepSeek—excluding others like Alibaba (Qwen) or Cohere for clarity. Unlike traditional markets where competitors converge, this rapid divergence creates asymmetric return potential as winners emerge.

OpenAI: Market Leader with Increasing Friction

OpenAI maintains a dominant market position through ChatGPT's 1B weekly active users and Microsoft's extensive distribution network. Yet vulnerabilities are emerging through interface stagnation and the retirement of their plugin ecosystem. Their dependency on Microsoft's infrastructure also limits strategic flexibility despite recent efforts to diversify cloud partnerships (e.g., Oracle).

The fundamental question isn't whether ChatGPT will remain popular, but whether OpenAI can maintain its interface advantage as competitors develop more flexible orchestration systems.

Anthropic: The Trust Premium

Claude has established a clear differentiation through its safety-first approach, creating particular resonance with enterprise customers seeking reliability over raw capabilities. Claude 3.7 Sonnet introduces a hybrid reasoning model that delivers both fast responses and careful consideration—a technical innovation with significant market implications.

With $14 billion in funding and strategic partnerships with Amazon and Google, Anthropic has secured distribution while maintaining independence. Also, its Model Context Protocol (MCP) is a key differentiator, thanks to its adoption by OpenAI and Google, establishing Anthropic as an orchestration leader.

Anthropic’s vulnerability could lie in scale: without a direct consumer presence, it risks becoming dependent on integration partners for distribution and market feedback.

Google: The Awakened Giant

Despite pioneering much of the fundamental research enabling the generative AI revolution, Google initially struggled to translate technical advantages into product leadership.

Gemini 2.5 Pro, integrated across Search, Android (2B+ users), and Workspace, leads in multimodal tasks like image reasoning and coding, per Alibaba’s Qwen3 tests (April 2025). In a recent post, Alberto Romero explains how Google is “winning on every AI front.”

Google's infrastructure capabilities—built on custom Tensor Processing Units (TPUs)—provide meaningful cost and performance advantages, while its data access creates natural orchestration opportunities through personalization.

Its primary vulnerability remains organizational coordination challenges between research excellence and product execution. Organizational misalignment between DeepMind and product teams is a known issue, delaying Gemini’s market impact.

xAI: The Hardware Integration Play

Elon Musk’s xAI embeds AI into Tesla vehicles and X’s 500 million+ users, creating ambient intelligence.

While Grok's model performance may not lead benchmarks, its ubiquitous presence creates a different form of defensibility—one based on passive integration rather than active interface superiority. The question isn't whether Grok is the most intelligent model, but whether its integration advantages create sufficient value to overcome technical limitations.

Meta: Ecosystem Over Ownership

Meta’s Llama 4 bets on open-source scale, with 1B+ downloads and 170K+ GitHub stars. Llama 4 Scout’s 10-million token context window and Maverick’s multimodal capabilities showcase technical excellence, but monetization via WhatsApp/Instagram integration is unclear.

The bet is that value will accrue through integration with Meta's core platforms, but the path to capturing that value remains unclear.

Mistral: Technical Excellence Seeking Application

French startup Mistral has emerged as a technical leader, delivering high-performance models like Codestral and LeChat that excel in high-value domains such as coding and enterprise efficiency. By focusing on precision over breadth, Mistral has carved out a premium niche, earning credibility with clients like BNP Paribas, Axa, and the French Ministry of Defense in Q1 2025. These partnerships signal early enterprise traction, positioning Mistral as a trusted partner in Europe’s AI ecosystem.

While Mistral’s orchestration capabilities and distribution channels lag behind giants like OpenAI or Google, its capital-efficient approach and open-source models, such as Mixtral, are fostering a growing developer community. To capture greater value, Mistral must strengthen its user-facing systems or secure high-profile integrations. Mistral represents today a technically superior player with the potential to scale into a leading enterprise AI provider if and only if it leverages its European foothold effectively.

DeepSeek: The Efficiency Disruptor

DeepSeek has fundamentally reset economic expectations through innovative engineering approaches, delivering GPT-4 quality at a fraction of the computational cost. Its "mixture of experts" architecture and reinforcement learning techniques represent genuine architectural innovation rather than mere scale advantages.

By releasing R1 under MIT licenses, DeepSeek is following Meta's ecosystem play but with a critical efficiency advantage. While export controls on advanced semiconductors may limit global scaling, this approach could prove decisive in markets with computational constraints, potentially creating an asymmetric advantage in emerging economies.

Share

The Fundamental Metrics Shift

Traditional benchmarks like MMLU do not tell the full story, capturing intelligence but not market power or financial value. With GPT-4o, Claude 3.7, and Gemini 2.5 converging, operational and financial metrics—tied to orchestration—define defensibility:

• Interface retention: User stickiness within orchestration environments (OpenAI's 1B weekly users)

• API value density: Revenue per computation rather than raw usage volume

• Embedded distribution: Integration depth within existing platforms and workflows

• Developer ecosystem: Community size and contribution velocity (Meta's 170K+ GitHub stars)

• Task automation rate: Success percentage on complex multi-step processes

• Efficiency ratio: Performance relative to computational investment

These metrics reflect orchestration’s role in practical defensibility. Interface retention (e.g., OpenAI’s users) depends on user-friendly orchestration layers. Revenue per AI task measures how orchestration drives profitable API usage, as seen in Anthropic’s Claude. Complex task success rate captures orchestration’s ability to handle multi-step workflows, critical for enterprise adoption.

Scores range from 1 (weak) to 10 (exceptional), based on qualitative and quantitative evidence from April 2025 data. Note that the scores presented here are based on our proprietary analysis and reflect only the author’s view.

Discounted Cash-Flow (DCF) valuation demands predictable cash flows, high margins, and low churn, unlike AI’s compute-heavy costs. Google’s 2004 IPO ($106m profit on $962m revenue) showed scalable margins, a benchmark that AI firms will have to meet.

As an example, OpenAI’s $12.7B 2025 revenue projection (from $3.7B in 2024) ignores negative margins ($5B losses) and a modest share of paid users (20m per recent information, i.e., 4% on the basis of 500m weekly users). The Information recently reported on revenue projections of $125B by 2029 with an estimated gross margin of 70%, but we were not provided with unit economics behind those figures.

Yes, frontier AI companies, like any business, will ultimately be valued on their ability to generate free cash flow, not just revenue. Investors must develop conviction now that these firms can build defensible moats through the metrics above, ensuring they thrive in a DCF-driven future.

When Raw Power Still Matters

While orchestration and distribution have become primary differentiators, model capabilities remain relevant in specialized domains where precision is paramount. Breakthrough innovations in reasoning (like OpenAI's O-series) or multimodal understanding (like Gemini 2.5) could still create meaningful advantages in sectors like healthcare, finance, and specialized enterprise applications.

The critical insight isn't that model quality doesn't matter, but that it matters primarily when translated through effective orchestration into distinct user value. Raw intelligence without orchestrated application increasingly represents unmonetized potential rather than market advantage.

Beyond model quality, regulatory shifts (e.g., EU AI Act enforcement) or competition for AI talent could reshape the landscape, particularly for smaller players like Mistral or DeepSeek.

The Investment Implications

We're witnessing a classic discontinuity moment—where traditional valuation frameworks fail but where the greatest investment alpha is generated. For investors navigating this landscape, four key questions determine potential returns:

✅ Value Translation: How effectively does the company convert model intelligence into user-specific applications?

✅ Distribution Leverage: Does the strategy leverage existing channels or require building new ones?

✅ Efficiency Balance: Does the approach optimize both performance and computational economics?

✅ Strategic Resilience: Can the model withstand competitive responses and regulatory evolution?

Companies that convincingly address these questions—balancing user value, scalable channels, cost discipline, and regulatory agility—will build the financial fundamentals required for DCF-driven valuations, capturing outsized returns as the market matures.

Investors must act swiftly to identify these moat-builders before valuations fully reflect their defensibility. The AI revolution may be new, but the financial domain remains timeless: only those firms with robust fundamentals will thrive in the long run, delivering the cash flows that justify today’s bold bets.

The Path Forward

The frontier AI market is entering a phase of strategic crystallization where winners and losers will be determined not by benchmark superiority but by moat construction. Orchestration quality—not raw model intelligence—will increasingly determine market leadership.

By 2027, expect consolidation as orchestration leaders acquire or outpace smaller players, while hardware integration and efficiency redefine the competitive landscape.

For institutional investors, the arbitrage opportunity exists in identifying which companies are building orchestration advantages that markets haven't yet fully priced. The convergence in model capabilities creates a false perception of commoditization that obscures the emerging moats in orchestration, distribution, and efficiency.

The companies that turn intelligence into distribution will capture the majority of value in this next phase of frontier AI development. Everything else is just noise.

About Decoding Discontinuity

Generative AI represents a Discontinuity. In this environment, static defenses fail. What's needed is a dynamic approach that assesses not just current advantages, but resilience and adaptation potential in the face of technological discontinuities. Enter the Durable Growth Moat.

My work bridges the gap between cutting-edge AI research and real-world business impact, enabling higher returns through actionable insights. Want to talk about GenAI and Discontinuity? Just reach out.

Subscribe now

Read the archives