Dipanwita Mallick argues that cost and privacy concerns are keeping many enterprise AI prototypes from reaching production.
2
Enterprises are moving toward hybrid AI infrastructure, combining private environments with cloud flexibility and keeping compute closer to data to reduce latency.
3
HP's proposed private AI approach uses managed infrastructure, zero-trust networking, a flat monthly fee, and Nvidia GPUs instead of asking each company to build everything itself.
Summary
Dipanwita Mallick describes the infrastructure problems enterprises face when moving from AI prototypes to production. Multi-agent workflows can use millions of tokens in a session, which makes a fully cloud-based stack expensive. Building private infrastructure also brings hardware, security, networking, and maintenance costs. Privacy adds another barrier because sensitive enterprise data moves between APIs, agents, memory stores, and chat interfaces. Mallick says customers want more control over data while keeping the speed and tools offered by cloud platforms. Her answer is hybrid AI infrastructure, with workflows moving between private and cloud environments and compute placed closer to the data. She presents HP Edge as a managed private AI option connected to an enterprise network through zero-trust connections. HP provides hardware, storage, networking, and data services, charges a flat monthly fee, and uses Nvidia GPUs. In the question session, she describes the next step as workstation-as-a-service with optional platform integration.
Mallick says enterprises are launching AI projects and prototypes at record scale, but few deliver business value after the prototype stage. She cites an IDC study saying a company might build 33 AI prototypes and move only four into production, which she describes as an 88% failure rate for scaling AI. She also cites a BCG study that found 11% of companies were able to unlock business value from AI at scale. Those companies reportedly saw a 30% higher EBIT than enterprises that remained in pilot or prototype phases.
Multi-agent workloads make cloud costs harder to control
Mallick says cloud platforms are attractive because they provide tools and let teams start quickly. The cost becomes difficult when multi-agent workflows consume millions of tokens per session. Each API call adds to the bill, so running the entire AI stack in the cloud can become expensive. Moving everything to a do-it-yourself on-premises setup does not solve the problem automatically. That approach requires capital spending on infrastructure and creates continuing operating costs. She says both options can become impractical over the long term.
Sensitive data needs more control as agents exchange information
Mallick says agentic AI depends on data, and that data often includes an enterprise's sensitive information. Data moves between APIs, agents, memory stores, and chat interfaces. Enterprises want visibility into how data is handled, where it resides, and who can use the system. She identifies privacy and security as a major reason companies hesitate to move from prototypes into production. Her argument is that infrastructure decisions must address data movement and control alongside performance and cost.
Mallick says agents need to communicate with other agents in real time, which makes latency important. Enterprises want to avoid moving data repeatedly between their own systems, the cloud, and other locations. The direction she describes is for AI to move closer to where the data resides. This leads customers toward private AI infrastructure, while they still want cloud flexibility for experimentation, development, testing, and deployment. She presents hybrid infrastructure as the emerging preference because it combines control with access to cloud tools.
Do-it-yourself private infrastructure carries its own burden
Building a private AI environment in an office, colocation facility, or another preferred location may sound appealing, but Mallick says it creates many operational demands. An organization has to procure and maintain hardware, define security protocols, and operate the environment. The result can require both capital expenditure and operating expenditure. She says HP does not recommend that companies handle the whole private infrastructure build alone. The alternative she presents is a managed private environment that can connect with cloud workflows.
HP Edge extends enterprise infrastructure through managed services
Mallick explains HP Edge as an extension of an organization's on-premises infrastructure. The infrastructure may sit in an HP-managed data center, but a zero-trust connection links it securely to the enterprise network. HP provides managed security, networking, storage, and data services, with the option for co-management. She says the approach is intended to give enterprises control over their private AI environment without requiring them to operate every part of it themselves. The infrastructure can support workflows that move between private and cloud environments.
HP's commercial model uses a flat fee and Nvidia GPUs
Mallick names several parts of HP's offering. The connection between HP and the enterprise network is designed as a secure pipeline for data to move in and out. Customers are charged a monthly flat fee rather than a usage-based price, which she says makes budgeting easier. She also says the infrastructure includes Nvidia GPUs for demanding AI workloads and multi-agent workflows. The offering is positioned as a way to provide private AI infrastructure while preserving the ability to use cloud services.
The next step is workstation-as-a-service with integration support
In the question session, Mallick says her team is offering high-end HP workstations as a service. The systems include high-performance GPUs, including Blackwell GPUs, and other hardware for AI workloads. Her team works with customers to understand what they need and can integrate an orchestration platform when required. She says the goal is to avoid giving enterprises a fragmented set of components. The service is intended to provide a more complete environment for running AI workloads.
"We don't want to offer fragmented solutions to their enterprises, right? Things are already so complicated."Dipanwita Mallick17:59
Who should watch
You are deciding whether a production AI system should run fully in the cloud, privately, or across both environments.
Your multi-agent workflows move sensitive data between agents and services, and you need to think through control, location, latency, and operating cost.
You are evaluating managed GPU infrastructure or workstation-as-a-service instead of building and operating a private AI environment yourself.
