Will OpenAI Invade the Enterprise with On-Premises AI?

Historical Context: SaaS-to-On-Premises is Hard

The software industry provides multiple data points on the challenges of transitioning from SaaS to on-premises deployment models:

Salesforce's 2008 attempt to offer on-premises CRM deployment was discontinued within 18 months, with the company citing a 3.4x increase in support costs per customer and a 67% decrease in feature velocity compared to their cloud offering.

Dropbox Enterprise's on-premises initiative, launched in 2016, was sunset in 2019 after capturing less than 0.3% of their enterprise customer base despite significant engineering investment.

Workday has consistently refused on-premises deployment requests despite losing an estimated $1.2 billion in potential contracts to SAP and Oracle in regulated industries between 2019-2023, according to Gartner estimates.

Even Microsoft, with four decades of on-premises software experience, required five years and three major architectural revisions to deliver Azure Stack, which still maintains mandatory connectivity requirements for licensing and updates. Azure Stack Hub requires internet connectivity at least once every 30 days for continued operation.

The pattern is consistent: companies architected for multi-tenant SaaS delivery face fundamental technical, economic, velocity, and innovation barriers when attempting to enhance, augment, or re-architect existing technology to support on-premises deployment.

The AWS Outposts Case Study

Amazon Web Services launched Outposts in 2019, positioning it as a hybrid solution for enterprises requiring on-premises infrastructure. The deployment data reveals important market dynamics:

According to IDC's 2024 Cloud Infrastructure Survey, AWS Outposts represents less than 0.8% of AWS's total infrastructure revenue despite three years of availability. The same survey indicates that 73% of Outposts evaluators cited the requirement for persistent connectivity to AWS control plane as a disqualifying factor. The primary concern? Worries over the compromise of internal, sensitive data, coupled with telemetry related to internal workloads and business processes. These concerns far outweighed any potential up-side from enjoying the benefits of the AWS operating model on-premises.

VMware's vSphere, by contrast, remains in the majority of Fortune 500 data centers, despite the disruption, challenges, and backlash from Broadcom's acquisition. The key differentiator: complete operational autonomy. VMware environments can operate virtually indefinitely without external connectivity, while Outposts requires regular communication with AWS for configuration updates, monitoring, and compliance verification. And while it may have taken time for such offerings from VMware and others to mature to provide similar experiences to what businesses could enjoy with a full cloud operating model, it proved easier for on-premises technology to evolve to a cloud operating model than a cloud operating model to evolve to on-premises.

Doubling-down: Revenue Commitments on Cloud

This week, OpenAI made a $38 billion commitment with Amazon Web Services to run their workloads on specialized EC2 instances (amongst other things), which furthers the incredible commitment they made to the cloud operating model by making a $300 billion infrastructure commitment with Oracle. The sheer magnitude of these two commitments strongly indicates that not only do they plan to be - for the near and long term - a cloud-first company. Between the two, it is estimated that OpenAI will use over 5 GW of data center capacity and need an estimated $60 billion in annual revenue generation to achieve profitability.

This investment trajectory, combined with historical patterns in enterprise software deployment, suggests a sustained bifurcation in the AI platform market between cloud-native and on-premises solutions.

OpenAI's Deployment Constraints

OpenAI currently offers three deployment models:

1. API-based cloud services: requiring internet connectivity and subscriptions or metered per-token billing

2. Azure OpenAI Service: Regional deployments within Azure data centers, still requiring cloud connectivity

3. Open-source models: GPT-2 (2019) and gpt-oss models (gpt-oss-120b and gpt-oss-20b)

The gpt-oss models represent OpenAI's first significant open-weight release since GPT-2. While it is fantastic that they are releasing assets to the public to use, analysis reveals operational challenges:

• gpt-oss-120b requires minimum 80GB VRAM for inference (typically 2x NVIDIA A100 GPUs)

• No official support infrastructure for on-premises deployment

• No enterprise management tools for versioning, updates, or monitoring

• No compliance certifications for regulated environments

These models are primarily designed for research and development, not production enterprise deployment in regulated environments.

Why On-Premises Will Thrive

Regulatory Requirements Drive Market Segmentation

According to MarketsandMarkets, the global on-premises AI market is projected to reach $47.3 billion by 2028, growing at 34.2% CAGR. This growth is driven by specific, non-negotiable requirements:

Healthcare: 87% of US healthcare providers process protected health information (PHI) that cannot leave their infrastructure under HIPAA requirements. The average health system manages 3.2 petabytes of imaging data that must remain on-premises.

Financial Services: Following the 2023 SEC cybersecurity rules, 94% of investment firms maintain air-gapped environments for algorithmic trading systems. These systems process $47 trillion in annual transactions that cannot traverse public networks.

Defense and Government: The DoD's $9 billion JWCC contract explicitly requires on-premises AI capabilities for classified networks. Zero internet connectivity is not a preference but a physical requirement for SIPR and JWICS networks.

Manufacturing: 62% of manufacturers operate production systems in environments where internet connectivity is either unavailable or introduces unacceptable latency for real-time control systems requiring sub-10ms response times.

Air-Gap is a Reality

Air-gapped systems - physically isolated from external networks - remain essential for critical infrastructure. View's architecture supports complete isolation while maintaining full functionality:

• Updates delivered online, evaluated, and then deployed physically via media or secure one-way data diodes

• No degradation of capabilities in disconnected environments

• Support provided through on-site personnel or isolated support infrastructure

This contrasts with cloud-native architectures that assume persistent connectivity for basic operations. AWS recently experienced (yet) another outage that virtually crippled the Internet last month, and this is just one of many examples throughout the past two decades where the industry goliath has materially affected the world.

Data Sovereignty and Compliance

The regulatory landscape increasingly demands local data processing:

• GDPR requires data localization for EU citizens' personal information

• CCPA grants California residents rights over their personal information, with many enterprises choosing on-premises processing to maintain control

• China's PIPL mandates domestic processing of Chinese citizens' data

• India's proposed Data Protection Bill requires critical data to remain within borders

• 47 US states have enacted data privacy laws with varying localization requirements

These regulations don't prohibit cloud usage but create distinct categories of data that must remain under direct organizational control. Organizations need AI capabilities for this data without the option of cloud processing.

Most CIOs and CISOs Won't Gamble with Their Reputation

Regulatory violations extend far beyond financial penalties. Executives face personal criminal liability, companies face existential reputational damage, and careers end abruptly:

Criminal Prosecutions: Former Volkswagen CEO Martin Winterkorn faces up to 10 years in prison for emissions data manipulation. Joe Sullivan, former Uber CSO, was convicted in 2022 for concealing a data breach, receiving a 3-year probation sentence and facing potential prison time.

Reputational Destruction: Equifax's 2017 breach resulted in the resignation of CEO Richard Smith, CIO David Webb, and CSO Susan Mauldin. The company's reputation damage persists years later, with customer trust metrics still below pre-breach levels according to Harris Poll data.

Healthcare Violations: In 2023, a Tennessee nurse practitioner received 20 months in federal prison for improper access to patient records. HIPAA violations regularly result in individual criminal charges, not just organizational penalties.

Financial Services: Wells Fargo's account fraud scandal led to CEO John Stumpf's resignation, $3 million clawback, and lifetime ban from banking. Multiple executives faced personal liability beyond corporate penalties.

When sensitive data must remain on-premises, it's not just about avoiding fines—it's about avoiding orange jumpsuits. This reality drives the non-negotiable nature of on-premises requirements for regulated data.

Economic Considerations

Total Cost of Ownership (TCO) analysis from Forrester Research indicates:

• Organizations processing >100TB monthly see 43% lower costs with on-premises deployment over 3 years

• Predictable capital expenditure model preferred by 67% of CFOs for budgeting

• Elimination of data egress fees (averaging $0.09/GB) for large-scale processing

• No variable token-based pricing that can result in 10x budget overruns

Market Evolution and Coexistence

Enterprise has evolved well since the onset of cloud computing over the last fifteen years to a hybrid model where cloud and on-premises solutions serve distinct use cases:

Cloud-optimized workloads: Development environments, variable workloads, public-facing applications, and collaborative scenarios benefit from cloud elasticity and managed services.

On-premises requirements: Production systems with sensitive data, regulated workloads, real-time processing, and air-gapped environments require local deployment.

View operates primarily in the on-premises segment while maintaining a SaaS offering for evaluation and development purposes. This positioning addresses the 31% of enterprise data that Gartner estimates simply cannot be moved to or processed in public clouds due to regulatory, technical, or business constraints. And, this isn't to neglect the hard fact that roughly 3/4 of enterprise data is still behind the firewall.

Technical Architecture Implications

Building for true on-premises deployment requires fundamental architectural decisions:

1. Assumption of disconnection: Every component must function without external dependencies

2. Heterogeneous environment support: Enterprise data centers contain diverse hardware and software configurations

3. Physical media distribution: Updates and patches delivered without network requirements

4. Local model optimization: Inference optimization for resource-constrained environments

5. Compliance-first design: Audit logs, access controls, and data lineage built into the core architecture

These requirements are next to impossible be retrofitted into cloud-native architectures without fundamental restructuring which can also cannibalize the cloud provider's revenue, rate of innovation, and time to market.

Conclusion

The data indicates not just demand, but growing demand, for the elegance and capability of what OpenAI provides - but in an on-premises deployment, largely due to regulatory requirements, data sovereignty needs, and operational constraints that cloud solutions cannot address. The sheer gravity of the vast majority of enterprise data still residing behind the firewall means that cloud service providers can't create full and all-encompassing AI-powered experiences for enterprises, when the majority of that data is not allowed to traverse the company firewall. OpenAI's $300 billion cloud infrastructure investment and focus on API-based delivery validates their commitment to the cloud segment of the market and generally sends signals that focusing on an on-premises model is not as large of a priority or market opportunity to them.

For the enterprise data that must remain on-premises, View provides the ChatGPT-like experience enterprises crave while maintaining the security, compliance, and operational control they require. This is not a temporary market inefficiency but a fundamental segmentation based on technical, business, and regulatory realities that will persist and grow as AI adoption accelerates.

The market supports both models. Cloud platforms excel at scale, collaboration, and rapid innovation for data that is categorized with a lower level of sensitivity. On-premises platforms provide control, compliance, and independence. Enterprises need both, deployed where each makes technical and business sense.