Snowflake's AI Agent Economy: The 65-Customer Concentration Behind the 22% Surge

When Snowflake reported earnings that sent its stock soaring 22%, the market narrative coalesced around a single phrase: the AI agent economy. The headline numbers were undeniably impressive—revenue of $1.55 billion against $1.48 billion expected, adjusted EPS of $0.62 versus $0.45 anticipated, and product revenue growing 37% year-over-year. Yet, as I studied the earnings transcript and the accompanying analysis, I found myself less interested in the beat itself and more disturbed by what the celebratory coverage glossed over. This isn't a story about AI magic. It's a story about concentration, consumption metering, and whether Snowflake is building a moat or a pricing trap. And the deeper I dug, the clearer the picture became: Snowflake has engineered a brilliant mechanism for monetizing the AI wave, but it has done so by creating a dependency structure that may prove fragile precisely when it needs to scale most. The context here is essential. Snowflake is not an AI lab. It is not training frontier models. Its strategy, as articulated through the launch of its CoCo coding agent and CoWork analytics agent, is fundamentally an engineering play—an architectural bet that the value in enterprise AI lies not in the model itself but in the orchestration layer that sits atop trusted, governed data. This is a profound philosophical divergence from the path taken by companies like Databricks, which has invested heavily in open-source model development and acquisitions like MosaicML. Snowflake's wager is that in the enterprise, data gravity trumps model intelligence. The AI agent is not the product; the AI agent is the amplifier for the data cloud. Every task an agent performs, every SQL query it generates, every pipeline it optimizes, consumes underlying compute and storage resources. The agent is, in the most literal sense, a consumption engine. This design philosophy has yielded impressive early metrics. CoCo has reached 9,100 accounts, adding over 2,000 in a single quarter. CoWork sits at 5,800 accounts. The company attributes about 50% of its product revenue growth to AI-specific offerings. Remaining performance obligations (RPO) have climbed to $9 billion, up 30% year-over-year, providing strong near-term visibility. The net revenue retention rate of 126% suggests that existing customers are expanding their consumption. Indeed, these are indicators of product-market fit that would make any SaaS founder envious. But here is where my audit training kicks in. Account count is not active usage. Net revenue retention is an aggregate measure that masks underlying distribution. And the most critical number in the entire report is not the $1.55 billion in revenue or the 37% product growth—it is the disclosure that 65 customers each contribute over $10 million in annual revenue. That is 0.45% of the total customer base driving the lion's share of the growth narrative. Let me be clear about what this means, because I think the market is missing the structural implication. Snowflake's AI agent strategy is not yet a broad-based platform phenomenon. It is a whale-driven phenomenon. The customer references cited in the earnings materials—Indeed deploying CoWork and CoCo for their data teams, Sayari migrating 12 billion records with CoCo—are exactly the kind of large, data-intensive enterprises that have the engineering resources to integrate AI agents into complex workflows. They are also the customers most likely to have negotiated favorable pricing. The 126% net revenue retention, which the market reads as a signal of healthy expansion, could equally be read as a sign that Snowflake's largest customers—the ones with the leverage to demand custom terms—are the ones driving the expansion. The long tail of 14,500 customers may be consuming AI agents at a much lower rate. This concentration risk is not merely an investor concern; it is a product strategy concern. If the AI agent economy is to be the foundation of Snowflake's future valuation—and a $60 billion market cap suggests it is—then the agents need to be a commodity that any enterprise can deploy, not a bespoke solution for the Fortune 100. And here we encounter the crux of the matter. Snowflake's consumption-based pricing model is brilliant in its alignment with the AI agent's function. Every agent invocation generates revenue. The more autonomous the agent, the more compute it consumes, the more storage it fills with logs and intermediate results, the more money Snowflake makes. This is a masterful alignment of value capture with value creation. But it carries within it the seed of a potential backlash. When customers deploy an AI agent, they are effectively signing a blank check for compute resources. A cost-conscious CFO will begin to ask: what is the ROI of this agent? How many queries did it actually run? How much of my data platform bill is attributable to the agent versus my human analysts? If the agent is saving time but costing more in infrastructure than the time saved is worth, the math collapses. The contrarian angle here is not that Snowflake is a bad company—it is clearly executing well. The contrarian angle is that the AI agent economy, as Snowflake is currently building it, may be a high-end luxury product rather than a mass-market infrastructure play. Consider the pricing dynamics. Snowflake's gross margins hover around 75%, which is excellent for a data platform. But AI agent operations introduce a new cost layer: the inference cost of the underlying LLM. Whether Snowflake is using Anthropic's Claude, Meta's Llama, or a fine-tuned open-source model—and the company has not disclosed its model sourcing—each agent invocation carries a variable cost that is fundamentally different from a traditional SQL query. The marginal cost of a data warehouse query is low; the marginal cost of an LLM inference is significantly higher, especially for complex multi-step reasoning. If Snowflake cannot control these inference costs, or if it passes them directly to customers in a way that triggers bill shock, the 126% net revenue retention could reverse as customers rationalize their AI spend. I have spent the past several months studying the intersection of AI agents and institutional data governance, and I believe there is a deeper issue that the market is not pricing. The security and compliance surface area of autonomous agents is still poorly understood. An AI agent that writes and executes code on a data platform is effectively a new class of database user—one that operates at machine speed and with machine scale. Traditional role-based access control (RBAC) was designed for human users with predictable patterns. An AI agent, particularly one that is designed to be autonomous, subverts those patterns. It may access data in ways that are not anticipated by the human administrator. It may, as we have seen in academic research on prompt injection, be manipulated into executing operations that violate data residency or compliance requirements. Snowflake has robust security infrastructure—SOC 2, ISO 27001, granular RBAC, audit logs. But these were designed for a world where humans issue queries. The world of autonomous AI agents requires a new paradigm of governance, one that includes model-level audit trails, explainable decision logs, and real-time guardrails against malicious prompts. The enterprise customers Snowflake is courting—financial institutions, healthcare providers, government agencies—are precisely the ones that will demand these capabilities before they allow agents to run unchecked on their most sensitive data. The competitive landscape intensifies these concerns. Databricks has been building its AI capabilities for years, with a particular focus on open-source tools like MLflow and Delta Lake that have strong developer mindshare. More importantly, Databricks has been aggressive in positioning itself as the platform for both data engineering and AI at scale. The cloud providers—AWS with Bedrock and Redshift, Azure with OpenAI integration and Synapse, GCP with Vertex AI and BigQuery—are all moving to bundle data and AI services in ways that could squeeze Snowflake's intermediate layer. A customer already deeply integrated into AWS is a natural candidate for AWS's native data plus AI stack, especially if the pricing is more predictable than consumption-based metering. Snowflake's moat is its data ecosystem—the network effects of having data stored in its platform, the ease of data sharing across organizations, the cross-cloud deployment flexibility. But if AI agents become the primary interface for data access, the value shifts from the storage layer to the orchestration layer, and that is a layer where competition is fierce and the barriers to entry are lower. From an infrastructure perspective, Snowflake's AI agent strategy will undoubtedly increase demand for compute and, specifically, for GPU resources. The company's multi-cloud strategy provides flexibility, but it also means Snowflake has limited control over GPU supply. In a world where NVIDIA's H100 and H200 GPUs remain constrained, and where the largest cloud providers prioritize their own AI workloads, Snowflake could find itself in a position where its growth is gated by its cloud partners' allocation decisions. This is not a theoretical concern; it is a structural risk embedded in the multi-cloud architecture. The inference cost of AI agents is a direct function of GPU availability and pricing. If GPU costs remain high, Snowflake's margins will compress, or it will pass the costs to customers, which could dampen adoption. The market's 22% surge in response to the earnings report suggests that investors are pricing in a future where Snowflake becomes the operating system for the AI agent economy. At a price-to-sales multiple of approximately 10 times forward product revenue, the market has already assigned Snowflake a significant AI premium. This premium is justified if you believe that AI agents will fundamentally transform enterprise data operations, and that Snowflake will be the primary infrastructure layer for this transformation. But it is not justified if you believe, as I do, that the current growth is disproportionately driven by a small cohort of large customers who may be deploying agents as an experimental luxury rather than a proven cost-saving tool. The philosophical shift inherent in Snowflake's strategy is worth articulating. The company is moving from a paradigm of "data as an asset" to "data as a fuel." In the old paradigm, enterprises stored data and retrieved it when needed; the value was in the accumulation. In the new paradigm, data is consumed by autonomous agents that continuously process, analyze, and act upon it. This is a fundamental redefinition of the data platform's role in the enterprise. The platform becomes not just a repository but a substrate for autonomous action. Snowflake's bet is that this paradigm shift will favor the platform with the most data, the most robust governance, and the most seamless integration between stored data and executing agents. The bet may be correct. But the transition is not without friction. I was reminded, in analyzing this, of the ICO boom of 2017. I spent three months auditing the whitepapers of 42 failed ICOs, and 85% lacked a sustainable value proposition beyond speculation. The parallel is not exact—Snowflake has real revenue and real customers. But the dynamic is similar. In a bull market for AI, every company wants to tell an AI story. Snowflake's story is more credible than most because it is grounded in a consumption model that directly ties AI adoption to revenue growth. Yet the same mechanism that makes the story credible—consumption-based pricing—also creates the vulnerability. When customers run AI agents, they are not just buying a software license; they are increasing their consumption of the underlying data platform. This is a beautiful business model because it aligns the vendor's incentive with the customer's usage. But it can also be a trap. A customer who deploys an AI agent that fails to deliver meaningful productivity gains will not just cancel a subscription; they will reduce their entire data platform consumption. The downside is amplified just as much as the upside. The question of who bears the responsibility when an AI agent makes a mistake is one that Snowflake, and the industry as a whole, has not adequately addressed. If an agent, operating on a Snowflake platform, executes a erroneous data transformation that leads to a compliance violation, who is liable? The customer who configured the agent? Snowflake, which provided the platform and possibly the underlying model? Or the model provider, which may be an undisclosed third party? This ambiguity is a significant barrier to adoption in regulated industries. I have spoken with data governance officers and compliance leads at financial institutions, and the refrain is consistent: they are not prepared to let an autonomous agent run without a clear audit trail and a clear delineation of responsibility. Snowflake's enterprise-grade security infrastructure provides a foundation, but the agent layer introduces new risks that the traditional framework does not fully cover. Let me also address the question of what is not in the report. The analysis does not disclose the specific revenue contribution of the AI agent products, nor the gross margin of those products. It does not disclose the model sourcing strategy—whether Snowflake is using a single third-party model, a multi-model approach, or a fine-tuned open-source model. It does not disclose the task success rate of the agents—how often they complete tasks without human intervention, how often they require correction. These are not minor details. The profitability of the AI agent business, the long-term cost structure, and the reliability of the agents in production are the key variables that will determine whether Snowflake's transformation is a genuine evolution or a temporary narrative-driven re-rating. The comparison to Databricks is instructive. Databricks, with its acquisition of MosaicML, has bet on open-source models as the foundation for enterprise AI. This gives Databricks greater control over the model layer and potentially lower inference costs in the long run. Snowflake's choice to remain model-agnostic—and, notably, to keep its model sourcing undisclosed—provides flexibility but also opacity. Over time, as inference costs become a larger share of the cost structure, this opacity could become a competitive disadvantage. If Snowflake is paying premium API rates for a third-party model, and Databricks is running optimized open-source models, Databricks could undercut Snowflake on price for AI workloads. This is a slow-burning threat, but it is real. There is also the question of the long tail. Snowflake's AI agent adoption is currently concentrated among its largest customers. The 9,100 CoCo accounts and 5,800 CoWork accounts are impressive in absolute terms, but they represent a fraction of the overall customer base. If Snowflake is to justify its valuation, it needs to bring AI agents to the mid-market and the SMB segment. This requires a different product approach—more templated workflows, more vertical-specific solutions, and a pricing model that is predictable enough for smaller organizations to budget. The current consumption-based model, which is essentially a blank check for compute, is not necessarily attractive to a mid-sized company with a $50,000 annual IT budget. Snowflake may need to introduce fixed-price agent packages or usage caps to capture this segment. The fact that it has not yet done so is a telling sign that the current strategy is optimized for the enterprise segment. The industry impact of Snowflake's AI agent strategy extends beyond the company itself. The shift from manual querying to agent-driven workflows will transform the labor market for data professionals. The traditional data analyst, who spends hours writing SQL queries and building dashboards, will increasingly become a supervisor of agents—defining objectives, monitoring outputs, and intervening when the agent goes astray. This is a profound change in job function. It raises the bar for entry-level data roles, which may accelerate the polarization of the data job market: high-skill roles that manage agents will command premiums, while routine data tasks will be automated. Snowflake, by embedding agents into its platform, is accelerating this transition. The company is not just building a product; it is shaping the future of data work. Whether this is a positive or negative development depends on one's perspective. For the enterprise, it promises significant efficiency gains. For the individual data analyst, it is a threat to the existing skill set. The human cost of the AI agent economy is not in the marketing materials. The regulatory environment is another variable that could shift the trajectory. As AI agents become more prevalent in enterprise data operations, regulators will inevitably take notice. The European Union's AI Act, which is already being implemented, will impose requirements on high-risk AI systems, and autonomous agents that operate on enterprise data may fall into this category. Snowflake, as the platform provider, will need to ensure that its infrastructure supports compliance with emerging regulations. This is both a cost and an opportunity. A platform that can certify its agents as compliant with the EU AI Act, or with sector-specific regulations in finance and healthcare, will have a significant competitive advantage. Snowflake has the resources and the expertise to build this capability, but it is not a trivial undertaking. As I reflect on the totality of the evidence, I come back to a core tension. Snowflake has demonstrated that AI agents can be a powerful engine for growth. The financial results are real, the customer references are credible, and the architectural bet is coherent. But the concentration of growth in 65 customers is a warning sign that the narrative is ahead of the reality. The AI agent economy, as Snowflake is currently monetizing it, is a whale-driven phenomenon. The question is whether it can become a mass-market phenomenon. This will require a transition from a product that is integrated into the data platform to a product that is accessible to any enterprise, regardless of size. It will require pricing models that are predictable and that do not penalize customers for agent failures. It will require transparency about the underlying models and the cost structure. And it will require a governance framework that gives customers confidence that agents will not become liabilities. There is a future in which Snowflake successfully navigates this transition and becomes the foundational layer for the AI agent economy. In that future, the company's value is not just as a data warehouse but as the operating system for autonomous enterprise operations. The data that has accumulated in Snowflake's platform becomes the fuel for a new class of applications that continuously optimize supply chains, monitor financial risks, and generate insights that no human analyst could produce at scale. This is a compelling vision, and it is one that I am sympathetic to as someone who has long argued that the true value of blockchain and Web3 lies in trustless coordination and transparent data flows. The principles that make AI agents valuable—automation, efficiency, scalability—are the same principles that drive the decentralization movement. But the implementation is everything. A centralized platform that controls the data, the compute, and the agents may be more efficient, but it also concentrates power in ways that run counter to the ethos of decentralization. Snowflake is building a centralized AI agent economy, and it will be very profitable if it succeeds. But the concentration risk—in the customer base, in the compute infrastructure, in the model supply chain—is the mirror image of the decentralization that I have spent my career advocating for. The takeaway, then, is not that Snowflake's AI agent strategy is flawed—it is not. The takeaway is that the market is pricing in a smooth, linear progression toward AI agent ubiquity, while the reality is more contested. The path from 65 large customers to 14,000 customers is not automatic. It requires product simplification, pricing innovation, and a robust governance framework. It also requires Snowflake to navigate an increasingly competitive landscape where the cloud providers and Databricks are not standing still. The 22% surge is a vote of confidence in the strategy, but it is also a high bar. Snowflake will need to deliver not just one strong quarter, but several consecutive quarters of broad-based AI agent adoption across its entire customer base, not just the whales. I will be watching the next few quarters with particular attention to the disclosure quality around AI product margins, model sourcing, and the expansion of CoCo and CoWork accounts beyond the enterprise segment. The story of Snowflake's AI agent economy is still being written, and the next chapter will be the most decisive.