AI is pushing software moats from data and interfaces toward actions, inference-cost control, and cross-system orchestration
AI summary card
AI is pushing software moats from data and interfaces toward actions, inference-cost control, and cross-system orchestration
After Goldman Sachs' software and internet conference and Silicon Valley roadshow, it believes AI expands software TAM and makes enterprises more likely to adopt software ecosystems independent of any single large model provider.
- Software companies are shifting from selling seats or features to selling "labor units" or "productivity units," thereby reaching larger enterprise budgets.
- Open-source distilled models, proprietary SLMs, and model routing are weakening the single-point pricing power of frontier model labs, while enterprises also want to reduce dependence on a single model provider.
- Compute and inference capacity have become constraints on AI expansion, shifting value toward runtime architecture, model routing, cost, reliability, and performance commitments.
- AI Native companies are not simply a negative force for SaaS; instead, they are capturing value in the gaps between traditional systems by delivering end-to-end outcomes and cross-system orchestration.
- Goldman Sachs was more positive in its conference feedback on CRM, CRWD, GWRE, IOT, and RBRK, while also watching AI opportunities in identity, workflow, and cybersecurity at OKTA, WDAY, ZS and others.
Report interpretation
Overview
This report summarizes Goldman Sachs' software industry observations after the Private Company Software & Internet Conference, the Annual Silicon Valley Bus Trip, and discussions with VC thought leaders. Covering 30 private companies and 10 public companies, it focuses on how AI is changing software TAM, moats, inference costs, compute constraints, cybersecurity, observability, AI Native, and the competitive landscape for applied AI.
Core views
The core view is that AI is not compressing the software value pool; instead, it may significantly expand software TAM. The value center of enterprise software is moving from record systems, databases, and dashboards to agents, orchestration layers, and action systems that can execute tasks across systems. At the same time, as frontier model capability improves, inference costs rise, pushing software vendors to use open-source distilled models, proprietary small models, and model routing to optimize cost and performance. Goldman Sachs believes enterprises will be more willing to adopt an ecosystem composed of multiple independent software vendors rather than relying entirely on any one model provider.
Analysis framework
Based on conferences and company interviews, the report synthesizes qualitative feedback from public-company management teams, private AI software companies, infrastructure platforms, and VCs, compares the strengths and weaknesses of different software competitive groups in the AI era, and extracts investment implications for TAM, business models, compute, cybersecurity, observability, and applied AI.
Methodology notes
Summarizing trends through interviews with company management teams and industry participants
The report is not driven by a traditional financial model; instead, it distills structural changes in the software industry in the AI era through meetings with multiple companies, a bus roadshow, and VC discussions.
Shifting from data, interfaces, and system records toward actions, orchestration, and inference efficiency
The report argues that the moats of traditional SaaS data and feature modules are being repriced, and that new defenses come from cross-system execution, proprietary models, inference-cost control, workflow implementation speed, and flywheels of real-world deployment data.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- CRMPublic software company; Goldman Sachs was more positive after the meetings
- Strengths
- The Agentforce strategy is clearer, and developer tools will be launched, giving it an opportunity to platformize AI agents and roll them out at enterprise scale.
- Weaknesses
- It still needs to prove that agent work units can achieve scaled adoption and sustainable monetization.
- Comparison
- Compared with vendors that only provide single-point AI applications, Salesforce emphasizes an enterprise platform and ecosystem.
- Risks
- Customer adoption pace, approval cycles, AI costs, and the way value is measured may affect monetization.
- CRWDPublic cybersecurity company; Goldman Sachs was more positive after the meetings
- Strengths
- Demand is improving across multiple modules, and security operations center automation plus AI-driven alert handling could lift demand.
- Weaknesses
- Cybersecurity competition is intense, and budget allocation for new AI capabilities still needs validation.
- Comparison
- In SOC automation, it is in the same benefitting direction as Palo Alto and Huntress.
- Risks
- Rising attacker AI capabilities, customer vendor consolidation, and valuation expectations could create volatility.
- GWREPublic software company; Goldman Sachs was more positive after the meetings
- Strengths
- AI may accelerate cloud migration and expand the insurance software TAM.
- Weaknesses
- Vertical software migration cycles are usually long, and customer system-replacement pace may limit near-term upside.
- Comparison
- Compared with horizontal SaaS, AI in vertical workflows depends more on industry-specific use cases and migration pace.
- Risks
- Cloud migration delays, customer budget constraints, and implementation complexity.
- IOTPublic software company; Goldman Sachs was more positive after the meetings
- Strengths
- Demand is ROI-driven, and it has data and AI differentiation.
- Weaknesses
- It must continue proving that AI can improve customer operating efficiency and support higher monetization.
- Comparison
- Compared with pure software tools, connecting device and operational data may strengthen the data flywheel.
- Risks
- Macro capex, fleet or operating budgets, and the pace of data monetization.
- RBRKPublic cybersecurity and data resilience company; Goldman Sachs was more positive after the meetings
- Strengths
- The Agent Cloud and Identity narrative is expanding, and enterprise demand for resilience is growing.
- Weaknesses
- It still needs to prove that identity and agent security expansion can translate into sustained revenue growth.
- Comparison
- It has a differentiated narrative at the intersection of data resilience and identity security.
- Risks
- Competition, changes in customer security-budget priorities, and uncertainty around new product adoption.
- OKTAPublic identity security company; covered as a key focus in the report
- Strengths
- Agent identity management was cited in the report as one of the next security product cycles, and Okta may benefit from demand for non-human identity and agent identity management.
- Weaknesses
- The identity security market is competitive, and the monetization pace for AI-agent identity is still unclear.
- Comparison
- Competes with large platforms such as Microsoft at the identity layer.
- Risks
- Platform competition, customer consolidation, and uncertainty around the evolution of agent identity standards.
- WDAYPublic application software company; covered as a key focus in the report
- Strengths
- Workday maps flex credits to work units, aligning with the pricing direction from seats to productivity units.
- Weaknesses
- It needs to prove that work-unit pricing can be understood by customers and drive incremental budget.
- Comparison
- Compared with traditional SaaS seat pricing, work units are closer to value-based billing.
- Risks
- Customer acceptance, AI gross margin, and the pace of workflow-automation deployment.
- ZSPublic cybersecurity company; covered as a key focus in the report
- Strengths
- It is already seeing more agentic traffic on the network and may benefit from agent-traffic security and bandwidth-related demand.
- Weaknesses
- The report still leaves open the question of whether cybersecurity vendors can capture this upside through hybrid user-and-bandwidth models.
- Comparison
- Compared with other security vendors, Zscaler is more directly exposed to the network-traffic layer.
- Risks
- Unclear monetization model, customer budgets, and intensifying competition.
Key data
- Conference coverage30 private companies and 10 public companiesFrom the Private Company Software & Internet Conference, the Annual Silicon Valley Bus Trip, and VC discussions.
- Public-company meeting listADSK, CRM, CRWD, GWRE, IOT, OKTA, RBRK, WDAY, ZS; plus pre-quiet-period discussions with PANW and SNOWThe report explicitly lists the relevant public companies.
- More constructive public companiesCRM, CRWD, GWRE, IOT, RBRKGoldman Sachs said it was more positive on these companies after the meetings.
- Distilled model costAbout 2% of the cost of training the base modelThe report says industry discussions pointed to distilled-model costs being significantly lower than training the base model.
- Superhuman model routing97% of API calls go through proprietary LLMsUsed to illustrate how proprietary models and routing capabilities can reduce dependence on frontier model providers.
- Intercom cost reductionCosts fell 80% after rebuilding a proprietary rerankerShows the potential inference-cost advantage of proprietary models.
- Baseten GPU clusterApproximately 40K-80K GPUsMostly Hopper and Blackwell, used to illustrate that inference capacity and financing commitments are becoming bottlenecks.
- Incremental GPU commitment costAbout 5K additional GPUs correspond to roughly $100-125 million in annualized commitmentsBaseten management used this example to show the sharply rising capital intensity of additional inference capacity.
- Locus Robotics deployment scaleApproximately 17K robotsUsed to illustrate the importance of real-world deployment data flywheels in applied AI.
Impact & implications
The investment implication is that the central question for software has shifted from "Will AI compress SaaS?" to "Which software companies can turn AI into billable productivity, controllable inference costs, and auditable cross-system execution?" The more advantaged assets are companies with clear ROI, the ability to embed AI into core workflows, proprietary model or runtime capabilities, and exposure to high-priority budget areas such as security and identity; the more pressured ones are traditional software vendors that rely on seats, dashboards, or a single functional module and struggle to prove value or execute across systems.
Risks
- If frontier model providers stop subsidizing tokens or significantly change pricing, the gross margins of software companies dependent on external models could be hit.
- Imbalances between inference supply and demand, GPU delivery lead times, and multi-year prepayment commitments may constrain AI application expansion.
- Internal enterprise approval, governance, audit, and determinism requirements may slow AI automation deployment relative to technical feasibility.
- AI Native companies may compress the value of traditional SaaS seats, interfaces, and modular products.
- Although multi-cloud and neocloud inference deployments improve availability, they also increase cost, operational complexity, and vendor-management difficulty.
- In security, improved attacker capabilities may change the offense-defense balance within 12-18 months.
What to watch
- Whether enterprises move from seat-based pricing to labor units, work units, or outcome-based pricing.
- Whether proprietary SLMs, open-source distilled models, and model routing continue to reduce inference costs.
- Whether AI agents create new budgets for network traffic, identity management, SOC automation, and patch governance.
- Whether product releases and customer adoption at CRM, CRWD, GWRE, IOT, and RBRK validate Goldman Sachs' more positive conference feedback.
- Inference GPU supply, Hopper/Blackwell delivery cycles, neocloud capacity, and multi-year compute-contract pricing.
- Whether AI Native companies can continue to erode the whitespace of traditional SaaS with shorter deployment cycles and higher PoC conversion rates.
- Whether real deployment-data flywheels form a repeatable moat in robotics, defense, and edge autonomy for applied AI.