25 years at the forefront of enterprise software investing.
Investing across the software lifecycle with tailored equity and credit solutions.
Vista’s latest news and insights from the forefront of enterprise software and AI investing.
A differentiated investment partner for advisors and their clients.
AI systems that execute autonomous, multistep, goal-oriented actions with minimal human intervention. Agentic AI moves beyond traditional AI models that respond to human prompts, acting as 'workers' that execute tasks and workflows. To learn more, read our whitepaper on Agentic AI and the Future of Enterprise Software.
The interface where AI can deliver business value, from productivity tools to vertical software and AI applications including SAP, Atlassian, Salesforce, ServiceNow and Vista Equity Partners' portfolio companies.1
The direct cost of delivering an AI-powered product, primarily inference costs paid to model providers. Unlike traditional cloud hosting costs, which scale modestly with users, AI COGS scale directly with usage intensity. To learn more, read our whitepaper on Understanding Inference and the Economics of Enterprise AI.
The frameworks, policies and technical controls that help ensure AI systems operate reliably, transparently and in compliance with regulatory requirements and organizational or industry standards. AI governance encompasses audit trails, explainability, reproducibility and accountability mechanisms.
The physical and software systems required to build, train and run AI models at scale. This includes data centers, AI chips, networking, power, cooling and cloud platforms.
The coordination and sequencing of multiple AI agents, models or automated steps to complete a complex, multistage workflow. Orchestration dictates how tasks are assigned, how agents work together and how unexpected errors are flagged for human review, enabling AI systems to manage entire business processes rather than isolated tasks.
For example, consider what happens after a car accident or house fire — you call your insurer. That call starts First Notice of Loss, a high-volume, time-sensitive workflow: verifying information, gathering documents, assessing risk and fraud and deciding next steps. Vista's Agentic Factory worked with Duck Creek, a Flagship portfolio company, to apply Agentic AI here. A claim orchestration agent oversees six specialized sub-agents, each trained for one of the tasks mentioned above and trained to bring in a human for any claims involving injury or exceeding a set value threshold. This is AI orchestration.
A thermal management method used in data centers and computing hardware in which fans and airflow systems reduce heat generated by processors, GPUs and other components. Air cooling is the baseline approach for standard compute deployments given it is simpler, cheaper and faster to deploy than alternatives like liquid cooling.2 Vector Core Compute (VC2), Vista's enterprise AI inference cloud, utilizes air cooling to deliver sustainable, low-cost inference.
In our view, AI is a connected ecosystem built across layers of technology and waves of value creation. Layers represent the market segments that help make AI possible, including: hardware and semiconductors, digital infrastructure, energy, foundation models, cloud and delivery platforms, data and development platform, AI applications and software.
A physical server in a dedicated environment where a company's data and systems are not shared with others and the customer gets exclusive access to the hardware. Think of it as renting an entire building instead of a single office suite in a shared tower — no other tenant's activity can slow down your systems or create a security risk. Bare-metal deployments are preferred for high-performance workloads where direct hardware access reduces latency and maximizes GPU utilization.3
An existing data center or infrastructure site that is retrofitted or upgraded to support new or expanded AI workloads. Brownfield deployments are typically faster and cheaper to bring online in comparison to greenfield sites.4 Reusing existing structures and utility infrastructure, brownfield retrofits reduce carbon emissions compared to new construction and can reduce capital costs by 30 to 50 percent relative to a brand-new greenfield build.5 Vector Core Compute (VC2), Vista's enterprise AI inference cloud, converted a brownfield for its downtown Los Angeles facility to deliver sustainable, low-cost inference.6
A building where all the major telecom and internet companies (e.g. AT&T, Verizon, Comcast) physically connect their cables and equipment. Rather than each network laying its own cables across a city, all plug into one central location, enabling direct interconnection between networks. Carrier hotels provide near-instant access to the broader internet without needing to lay new fiber-optic cable, and are therefore ideal brownfield sites for inference data centers. Vector Core Compute's (VC2) first data center in downtown LA is located next to a carrier hotel with advantaged connectivity.
The platforms through which businesses access and deploy AI applications — Amazon, Microsoft and Google are generally viewed as leading providers in this layer.7
The general-purpose processor in a computer responsible for executing a broad range of tasks. CPUs manage coordination, data preparation and non-AI workloads, but are less suited than GPUs or other purpose-built AI chips for the intensive compute required to train and run AI models.8 Intel introduced the world's first commercially available CPU in 1971.9
The environments where AI is trained, refined and integrated, from companies like Snowflake and Databricks.10
A facility housing the servers, networking equipment, power systems and cooling infrastructure required to store, process and deliver compute capacity at scale. In 2025, OpenAI signed more than $1.4 trillion of infrastructure deals to build out the data centers it needs to meet the compute demand required to train and run its large language models.11 By 2030, supporting Gen AI and Agentic AI could require up to $5.2 trillion in AI-specific data center investment globally.12
The concept of maintaining control and visibility over where data resides, who can access it, and on what terms it's used to train or inform AI models. Organizations in regulated industries often require data sovereignty or single-tenant infrastructure — a dedicated environment where a company's data and systems are not shared with other customers — to keep proprietary data and IP under their control.
Facilities that store data, host models and deliver the compute capacity required to train and deploy AI at scale. This includes companies like Digital Realty, Equinix and Ciena.13
Model training and inference require reliable, low-cost power through renewable generation, grid modernization, battery storage and more. Companies providing these solutions include Siemens Energy, NextEra Energy and Brookfield Renewable.14
The deployment of AI models, software and infrastructure within large organizations to automate processes, support decision-making and enhance operational efficiencies across business functions.
A security and compliance certification used to evaluate the security and reliability of technology infrastructure and services. FedRAMP is the U.S. government's standardized process for vetting and approving cloud software as secure enough to use for federal agencies. It involves extensive security reviews that help ensure a product meets specific data protection, encryption and monitoring standards deemed "compliant" for purposes of federal agencies handling highly sensitive data.
A general-purpose AI model trained on broad datasets that can be used for a wide range of tasks through prompting. Foundation models serve as the starting point for most modern AI products rather than being trained from scratch for each use case. This designation is relative and evolves as new models are released. Examples of foundation models today include Anthropic's Sonnet.
The most advanced, highest-capability version of AI models available at a given point in time. Frontier models are appropriate for tasks requiring advanced reasoning but are not the primary choice for most enterprise workflows.15 This designation is relative and evolves as new models are released. Examples of frontier models today include Anthropic's Opus or Fable.
A class of AI models that produce content such as text, images, audio, video or code.
A processor designed for rendering graphics. GPU has become a dominant hardware for AI training and inference. GPUs can perform large numbers of mathematical operations simultaneously, making them well suited for the intensive compute AI workloads may require. NVIDIA popularized the term "GPU" and later enabled developers to use GPUs for general-purpose computing tasks like AI and data processing. However, the underlying hardware and technology dates to the 1980s.16
A brand-new data center build or infrastructure site built from the ground up designed to meet specific power, cooling and density requirements. Greenfield sites can offer greater flexibility to accommodate high-density AI compute infrastructure but require longer lead times and higher upfront capital than existing brownfield retrofits. Construction timelines can range from 12 to 36 months, with a 3- to 5-year power queue.17
Units of electrical power used to measure the capacity of data centers and the infrastructure that supports them. 1 GW = 1,000 MW.
The physical foundation of AI computing, including advanced chips and semiconductors from companies such as Nvidia, AMD, Intel and Broadcom.18
A category of the largest computing infrastructure providers that operate data centers at massive scale, delivering compute, storage and networking services globally (e.g. AWS, Google Cloud, Microsoft Azure).
The variable cost incurred every time a model is used. Unlike model training, inference costs are paid by whoever is using the model, whether that be an individual or a business. Model providers charge for inference based on usage, which is measured in units called tokens. To learn more about inference and the economics of enterprise AI, read our whitepaper.
The time elapsed between submitting a prompt to an AI model and receiving a response — for the person using the AI, it's how long you wait for a reply, similar to how long a webpage used to take to load in the early internet era. Latency is a primary performance metric alongside throughput for inference. Low latency is critical for real-time applications such as conversational AI, coding assistants and agentic workflows, while higher latency may be acceptable for larger-scale processing tasks. We believe that improvements in latency, driven by hardware advances, model optimization and inference architecture, can be relevant to the competitiveness of AI infrastructure and model providers.
A thermal management method that uses water or other coolant fluids to reduce heat from processors and GPUs, offering significantly greater efficiency than air cooling for high-density compute environments. Liquid cooling is required for high-density GPU racks. Compared to air cooling, liquid cooling is more complex and costly to deploy.19
A sophisticated AI system trained on vast text data to understand and generate human language. LLMs learn statistical patterns in language — grammar, facts, reasoning styles and context — by processing huge volumes of text. They then use those patterns to predict what text should come next when given an input, known as a prompt.
A subset of AI in which systems learn patterns from data and improve performance over time without being explicitly programmed for each task. Machine learning is the technical foundation underlying most AI products, including large language models (LLMs) and recommendation engines.
A system that classifies each incoming AI prompt or request and directs it to the least expensive model capable of handling it effectively. Not every task requires a frontier model (Anthropic's Opus or Fable); a well-designed router matches workload complexity to the appropriate model tier, reducing inference costs without sacrificing accuracy. Vector Core Compute (VC2) routes AI workloads utilizing Intel Xeon CPUs.
The process of teaching an AI model to produce an accurate output by adjusting its internal settings based on patterns it learns from large datasets. Training is compute-intensive, typically performed once or periodically. Frontier AI training is rapidly growing and getting increasingly expensive, with compute use rising roughly five times a year and the largest training runs expected to top one billion dollars by 2027.20
AI models capable of processing and generating multiple types of data (text, images, audio and video) within a single system. Frontier models can be multimodal, which may enable richer enterprise applications than their text-only predecessors.
An AI model whose "weights" — the core parameters that define how the model works — are publicly released. This allows an individual or company to run the AI model on their own infrastructure (i.e. chips, data centers and cloud computing) rather than paying a commercial provider to deliver the model and support it. Meta's Llama is an Open-Source Model. An example from the internet era is the Linux Operating System versus Microsoft or MacOS, where developers can access a free blueprint but have to build and maintain it themselves. Open-source models can perform at or near competing commercial models for many enterprise tasks, often at a significantly lower cost, though they typically require more internal technical resources to deploy safely.21
The practice of designing and refining prompts given to an AI model to reliably produce accurate, relevant and well-formatted outputs. Effective prompt engineering is a core skill for teams deploying AI in enterprise workflows.22 Roles requiring prompt engineering skills, regardless of title, increased threefold between 2024 and 2026.23 As AI becomes embedded in everyday work, the ability to communicate effectively with AI systems is becoming a fundamental workplace skill.
A purpose-built inference chip developed by SambaNova Systems that processes data in a fundamentally different way than traditional CPUs and GPUs.2425 RDUs are optimized for the data movement patterns common in LLM inference, offering potential advantages in speed and energy efficiency for specific AI workloads.
Infrastructure that dedicates all compute and storage resources to one customer with no sharing across other users. Organizations in regulated industries often require single-tenant infrastructure to ensure their data is kept separate from other customers for security and confidentiality purposes.
Security and compliance certification used to evaluate the security and reliability of technology infrastructure and services. SOC 2 is an enterprise data-handling standard that assesses how a service provider manages data security, availability and confidentiality. SOC 2 is commonly required by enterprise customers.
The full cost of acquiring, operating and maintaining a technology asset over its lifetime, encompassing capital expenditures such as hardware procurement, facility build-out and ongoing operating expenses.
The basic unit of text that AI language models read and generate — roughly equivalent to three-quarters of a word or 7 characters. When a model is used, it processes input tokens (such as a question and any context) and generates output tokens (the answer). The model provider charges for both based on how many tokens are being consumed.
Tokens-per-second measures inference speed.
Tokenomics refers to the cost and revenue economics of token generation, encompassing the price charged per token, the compute cost to produce it and the margin in between.
AI systems designed to perform specific, predefined tasks using rule-based logic, statistical methods or narrow machine learning techniques. Unlike generative AI, traditional AI does not produce novel content; it classifies, detects or predicts within a fixed problem scope (e.g., fraud detection, image recognition, recommendation engines) and requires human input at every stage. See the evolution from Traditional AI to Generative AI and now Agentic AI here.
