Silicon Valley Coordinates x Datadog Chief Product Officer | HYSTA Chair 李严冰: AI Deployment Enters the ROI Era

Summary

• 李严冰判断，AI deployment has moved from proving whether it can be used to measuring whether it can be governed, trusted, and generate ROI. Enterprises must verify that outputs are correct, safe, and reliable while controlling token costs and total budgets that have far exceeded expectations; she argues that observability capable of explaining what AI is doing, how entities relate to one another, and what value it ultimately creates is becoming critical to scaling deployment.
• 李严冰 sees “AI needs greater observability” as a major opportunity for Datadog to emerge from the SaaS valuation trough. She says the company’s market cap has doubled since the start of the year to an all-time high, and that all of the world’s top 10 AI companies rely on Datadog; the market once worried that AI would replace observability, but the consensus has now flipped to “AI actually needs greater observability” （人工智能实际上需要更高的可观测性）.
• The product boundary is expanding from web dashboards to every workflow involving humans and AI agents. Datadog historically built rich interfaces around human users; going forward, it emphasizes that “wherever we work, our tools should be there too”: engineers should be able to invoke observability from the CLI while coding, and teams should be able to do the same from Slack as they collaborate.
• With frontier models from OpenAI and others potentially completing more tasks autonomously, 李严冰 sees operational complexity, domain expertise, proprietary data, and vertical models as Datadog’s moat. She compares the situation with autonomous driving: a vehicle has only 3 control points—acceleration, braking, and steering—while large-scale software and hardware infrastructure becomes far more complex once AI agents are added; she says Datadog’s specialized models are close in quality to top frontier models, but cost “at least an order of magnitude” less.
• The future will not be dominated by a single frontier model across every use case; post-training and enterprise data will become sources of differentiation. One leading AI lab even asked whether Datadog could label and sell its operational data, underscoring how little specialized-domain accumulation general-purpose models still have; Datadog also acquired reinforcement-learning company AdaptiveML last year to accelerate its internal model investments.
• Coding agents may account for the largest share of Datadog’s internal AI investment, but the real KPI is engineering output, not token consumption. The company links usage to PR velocity, product stability, feature delivery, and security improvements; the more immediate risk comes from employee-built “amateur-grade” agents that could leak business data or API keys, not a doomsday scenario in which “AI kills humans.”

Deep dive

1. China’s AI influence is expanding from research talent into entrepreneurship and commercialization

• 李严冰 cited a Carnegie research report published 3 days ago. The study analyzed 10,000 NeurIPS authors and found that 57% received their undergraduate education in China, including researchers now based in China, the US, and elsewhere.
• She stressed that Chinese influence now extends beyond frontier research to turning AI into businesses: the US has 41 companies valued at more than $1B that were founded or co-founded by Chinese people, several of which participated in this year’s Huayuan Artificial Intelligence Conference.
• As a result, this year’s conference expanded beyond technical frontiers and enterprise applications into scaled deployment and commercial success. The central question is no longer simply whether AI works, but how to deploy it at scale.

2. AI enters the ROI era, with observability becoming critical to scaling

• 李严冰 summarized the enterprise decision chain this way: the initial question was whether AI was truly viable; the next was whether its security and reliability could be trusted; this year, companies discovered that token costs and total budgets had “far exceeded expectations.”
• By observability, she means understanding in detail what AI is actually doing, how entities are connected, and what value it is creating. That gives enterprises a better grip on governance, cost control, and commercial returns, moving them from chasing the trend to deploying AI in a controlled way.
• She said Datadog may be “the first company in SaaS to emerge from the trough,” with its market cap doubling since the start of the year. The case is not based on narrative alone: all of the world’s top 10 AI companies rely on Datadog, and observability has not been displaced by AI—it has become more important because of it.

3. AI agents are redrawing the product boundary, but fully automated large-scale production operations remain far away

• Datadog historically served human users through polished interfaces and dashboards. In the future, both humans and agents will use the product, with interaction moving from the web and mobile devices to wherever work happens, including the CLI and Slack. “Wherever we work, our tools should be there too.”
• 曹卿云’s question gets to the core issue: if OpenAI’s models can already write code, where is Datadog’s irreplaceability? 李严冰’s answer is that coding primarily generates code, while operations involves diverse, mission-critical, interconnected systems whose complexity is not remotely comparable. She said she is not worried about being replaced by AI, particularly because Datadog is also developing its own AI technology.
• She used Aurora, where she previously worked, as a comparison. She described the company as a leading autonomous trucking company and the first in the US—and possibly the only one in the world—to operate commercial autonomous trucks on public highways. Autonomous driving needs to understand only the local environment within its field of view and control acceleration, braking, and steering; large-scale production systems combine software, hardware, infrastructure, and AI agents all at once.

4. The economics of vertical models and proprietary data form Datadog’s moat

• 李严冰 breaks the moat into 3 layers: years of accumulated operational-domain expertise, massive operational data collected at global scale, and sustained investment in research and foundation models.
• General-purpose models offer broad coverage but may be slower and more expensive, and may still fail to answer highly specialized questions. She said Datadog’s specialized models, built on observability data, come very close to top frontier models in quality while costing at least an order of magnitude less.
• She does not believe enterprises should uniformly rely on or abandon frontier models; “it ultimately depends on the specific requirement.” The market will not get every answer from one model or a small number of models. Post-training is already widespread because it can turn a company’s own data into “a true moat.”

5. Enterprise AI spending is quantifiable, but security risk is concentrated in “amateur agents”

• 李严冰 said a leading AI lab and Datadog are customers of each other, and that the lab asked whether Datadog could label and sell its operational data. Her conclusion is that even the most advanced labs lack sufficient accumulation in vertical domains. Datadog acquired AdaptiveML last year to accelerate its investment in reinforcement learning.
• When budgeting for next year’s R&D or AI spending, Datadog found that AI expenditure was growing far faster than intuition would suggest, with coding agents potentially taking the largest share of the budget. The company tracks not only where tokens are used, but also maps the spending to PR velocity, product stability, feature delivery, and security improvements; usage of its customer-facing AI products has grown several-fold.
• Use cases in knowledge work are still in a “hacker” phase of experimentation. Salespeople or HR staff may build their own unoptimized tools or agents for specific needs. 李严冰 believes the real near-term enterprise security problem is not a doomsday scenario in which thousands of agents escape their sandboxes and attack other companies, but the accidental leakage of business data or API keys by these nonprofessional tools. The lack of review and control is the practical obstacle to broader adoption.