The Two AI Certainties Driving a New On-Prem Hardware Supercycle
The Two AI Certainties Driving a New On-Prem Hardware Supercycle
Author: Zion Zhao Real Estate | 88844623 | ็ฎๅฎถ็คพๅฐ่ตต | wa.me/6588844623
Author’s note: This essay is written for education and market literacy, not as financial advice or a solicitation to buy or sell any security. Markets can fall as well as rise, and past performance is not indicative of future results. Educational analysis only. Not financial advice, not a recommendation to buy or sell any security.
TL;DR: Agentic AI Is Here: Why Enterprise Workloads Shift Back On-Prem
Agentic AI is shifting the investment conversation away from legacy software debates and toward a clearer, higher-confidence outcome: enterprises will increasingly replace repetitive human work with always-on AI workflows, and a meaningful share of the compute required will run outside the public cloud.
The first “certainty” is adoption. Agentic systems can plan and execute tasks such as customer support triage, document processing, research, reporting, software maintenance, and operational monitoring. Evidence from real deployments and research already shows material productivity gains in customer service and software development, making continued rollout economically rational. Over time, firms will reallocate labour, reduce handling time, and standardise outputs, especially for high-volume workflows where 80 to 90 percent of requests fall into predictable patterns.
The second “certainty” is where the work runs. Cloud remains essential for training frontier models and for workloads that need elastic capacity. However, agentic workloads inside organisations are often continuous, latency-sensitive, and tightly coupled to proprietary data. Many also face regulatory and contractual obligations around privacy, auditability, and third-party risk. These factors make hybrid architectures the practical end state, with more inference and workflow execution moving closer to the data: on-prem servers, private cloud environments, and even high-performance “AI workstations” for teams.
A further driver is stability. Enterprises typically need consistent behaviour for production tasks such as policy-compliant customer replies, standardised reporting, and repeatable writing styles. Rapidly changing frontier models can introduce output variance and model drift. In contrast, smaller specialised models can be fine-tuned, constrained, version-controlled, and run locally with predictable cost and governance. Cloud economics also matter: always-on agents that query constantly and move data frequently can make fixed-capacity, owned or managed on-prem infrastructure more attractive.
The implication is a broad hardware re-acceleration, not limited to GPUs. Demand should expand across enterprise servers, storage, networking, security, and private-cloud platforms. Companies positioned across this stack include Nvidia, Dell, AMD, Intel, Apple, IBM, Oracle, and others enabling secure, governed deployment. The cloud is not disappearing. The point is that AI expands compute everywhere, and the agentic layer materially increases the share that must live inside the enterprise.
AI is not only reshaping tech stocks. It is reshaping the real economy, corporate hiring, household incomes, and where capital flows next. As agentic AI drives a new wave of on-prem hardware investment, demand will rise for data centres, advanced offices, industrial space, and well-located homes near job clusters. These shifts influence interest rates, rental demand, tenant quality, and pricing power across Singapore property cycles.
If you are buying, we align your purchase with macro timing, unit fundamentals, and long-term liveability. If you are selling, we position your asset with market narrative, pricing strategy, and buyer targeting. If you are renting or investing, we optimise yield, tenant profile, and risk management.
I combine Singapore property expertise with daily macro and cross-asset research, plus strong grounding in Singapore Land Law and transaction structuring. Engage me at 88844623 for a clear, data-driven plan tailored to your goals.
Introduction: The Two High-Confidence Trends in AI: Why On-Prem Hardware Demand Is Set to Re-Accelerate; Wall Street Is Debating the Wrong Battlefield
Public markets have spent the past decade pricing a single dominant enterprise narrative: software eats the world, and the cloud eats the data center. That story drove premium multiples for legacy software vendors and rewarded hyperscalers for building planet-scale infrastructure. Yet the next enterprise wave is not primarily “more SaaS.” It is the agentic layer: systems that can plan, execute, monitor, and improve workflows that used to require humans.
This matters because agents change the economics of compute placement. When work becomes continuous (24 hours a day, 7 days a week), data-intensive, latency-sensitive, and governed by privacy, auditability, and intellectual property constraints, the optimal architecture shifts. Cloud remains essential for frontier model training and elastic workloads, but a meaningful share of inference and workflow execution increasingly gravitates closer to the data: on-device, on-premises, and in private or sovereign environments (Gartner, 2024). (Gartner)
In this essay, I develop two high-confidence propositions and their investment-relevant implications (without making any buy or sell recommendations):
Agentic automation will expand rapidly across repetitive and semi-structured enterprise work.
A larger-than-consensus portion of that automation will run outside the public cloud, driving a renewed cycle in enterprise and edge hardware.
Trend 1: Agentic Automation Will Keep Expanding, Because the Economics Are Too Compelling
From “assistants” to “systems of work”
Early generative AI adoption looked like productivity augmentation: copilots for writing, coding, and summarization. The agentic layer is the next step: a software system that can decompose a goal into tasks, call tools and APIs, retrieve data, draft outputs, request approvals, and continuously monitor results. Recent surveys document fast-moving progress in large language model (LLM) agents and tool-using architectures, alongside persistent limitations in reliability and evaluation (e.g., Wang et al., 2024; Xi et al., 2023). (arXiv)
Even with limitations, the direction is clear: firms are already deploying AI to reduce handling time, increase resolution rates, and reallocate human labor toward higher-complexity cases. In contact centers, peer-reviewed evidence shows measurable productivity and quality improvements when AI assistance is introduced (Brynjolfsson et al., 2025). (OUP Academic)
In software development, controlled experiments show substantial speed gains with AI coding assistance (Peng et al., 2023). (arXiv)
These are not theoretical gains. They represent a shift in unit economics: firms can replace portions of labor cost with compute and software. Institutions and policy bodies also increasingly track the exposure of tasks and occupations to generative AI and automation, suggesting broad-based impact even if the magnitude varies by sector and regulation (OECD, 2025; ILO, 2025; World Economic Forum, 2025).
Why “automation pressure” is structural, not cyclical
The agentic layer scales because it targets the most common enterprise pattern: repeatable workflows with stable input schemas and defined outputs. Examples include:
Customer support triage, deflection, and knowledge retrieval
Document drafting, contract review, and clause extraction
IT service desk operations and internal knowledge search
Finance operations such as reconciliations, reporting support, and variance explanations
Sales enablement tasks like account research, call summarization, and follow-up drafting
Bank of America, for example, has publicly stated that its virtual assistant usage reduces call volumes and that its internal virtual assistant has reduced IT service desk calls by more than 50 percent (Bank of America, 2025a, 2025b). (Bank of America)
The key point is not that every job is “replaced.” It is that a large fraction of repetitive work becomes compressible, and the organization’s next constraint becomes throughput, governance, and deployment architecture. That leads directly to Trend 2.
Trend 2: The Agentic Era Favors Hybrid and On-Prem Compute More Than Markets Assume
Cloud is not “forbidden” in regulated industries, but compliance changes the architecture
A common oversimplification is that healthcare, finance, and legal workflows “cannot use the cloud.” In reality, many regulators allow cloud usage under specific controls, contracts, and safeguards. For example, U.S. HIPAA does not categorically ban cloud computing; it requires appropriate administrative, physical, and technical safeguards, and typically a compliant arrangement with the provider (HHS, 2022). (HHS.gov)
Similarly, U.S. financial regulators emphasize risk management, governance, oversight, and third-party controls for cloud usage (FFIEC, 2020). (HHS.gov)
In the European context, controller-processor obligations and security requirements under GDPR drive careful contractual and technical design, regardless of whether workloads are on-premises or hosted (EDPB, 2020).
So the real issue is not “cloud or no cloud.” The issue is that agentic workloads often combine:
Sensitive internal data (customer records, trade secrets, regulated content)
Continuous querying and data movement
Requirements for auditability, reproducibility, and stable outputs
Tight latency and reliability expectations for operational workflows
These constraints increase the value of keeping certain models and inference close to the data, or within infrastructure the organization directly governs.
Why stable outputs matter for enterprise agents
Enterprises do not only want intelligence; they want consistency. A customer support policy bot must answer in a compliant way every time. A newsroom automation workflow must follow style guidelines. A regulatory reporting assistant must be auditable.
Frontier hosted models evolve quickly, and model behavior can drift across versions. Research has documented substantial behavior changes over relatively short periods in widely used LLM services (Chen et al., 2024). (Harvard Data Science Review)
Regulators and central banks increasingly discuss model risk topics like drift, governance, and validation as core operational issues (Monetary Authority of Singapore, 2024). (Monetary Authority of Singapore)
This creates a pragmatic enterprise pattern:
Use frontier models (often cloud-hosted) for the extraordinary: prototyping, complex reasoning, or tasks where variation is acceptable.
Use smaller, specialized models for production workflows where consistency, cost predictability, privacy, and auditability dominate. These can be deployed on-premises or in controlled environments, often fine-tuned and constrained to a narrower domain.
Techniques like parameter-efficient fine-tuning (for example LoRA) reduce the cost of specializing models to a firm’s data and style requirements (Hu et al., 2021). (arXiv)
The “always-on” cost curve: why continuous agents can pressure cloud economics
Public cloud is excellent for elasticity and speed. But always-on, high-frequency workloads can produce bill volatility and unfavorable unit economics, particularly when data egress, storage access patterns, and 24 by 7 inference dominate. Cloud cost management remains a top reported challenge in industry surveys, and hybrid approaches are widely expected to persist (Flexera, 2025; Gartner, 2024). (Flexera)
This does not mean “the cloud is dead.” It means “the marginal workload mix is changing.” Agentic workflows look less like periodic batch jobs and more like continuous internal operations. That profile often benefits from:
Fixed-cost capacity amortization (owned or leased hardware)
Locality to proprietary data stores
Predictable performance and latency
Governance and security controls aligned to internal policies
Fact-Checking Key Examples Often Used to Illustrate “Cloud Exit”
Dropbox: not anti-cloud, but proof that scale changes the math
Dropbox is frequently cited because it moved a large portion of its storage off AWS into its custom infrastructure, Magic Pocket. Dropbox publicly stated it was storing and serving over 90 percent of user data on Magic Pocket (Dropbox, 2016a, 2016b). (Dropbox Tech)
The lesson is not that every company should replicate Dropbox. The lesson is that at sufficient scale and workload stability, owning infrastructure can improve cost efficiency and control.
Netflix: cloud compute plus proprietary delivery network
Netflix is sometimes mischaracterized as “leaving the cloud.” In reality, Netflix completed a major migration of systems to AWS, while also building its own Open Connect content delivery network for video distribution (Netflix, 2016; Netflix, 2019). (Amazon Web Services, Inc.)
Again, the lesson is hybrid optimization: keep what is core, continuous, and latency-sensitive in specialized infrastructure, while leveraging cloud where it is most efficient.
Washington Post layoffs: be careful with attribution
It is tempting to attribute layoffs directly to agents replacing workers. But specific claims require verification. For example, reports in early 2026 described significant Washington Post staff reductions on the order of roughly one-third, not “50 percent,” and public reporting attributes these decisions to a mix of financial and strategic factors rather than a single AI-driven causal mechanism (The Guardian, 2026; Pew Research Center, 2026). (abZ Global)
The broader point remains: automation pressure is rising. But responsible analysis avoids asserting one-to-one causality when evidence is incomplete.
The Hardware Renaissance: What Actually Expands When Agents Go On-Prem
If a growing portion of inference and workflow execution shifts closer to the data, the demand expansion is not limited to GPUs. It is a stack-wide buildout:
1) “AI workstations” and departmental inference servers
A new category is forming between consumer PCs and data center clusters: compact systems capable of running useful models locally for teams and small enterprises. Nvidia’s DGX Spark (previously announced as Project DIGITS) is a concrete example of a desktop AI system positioned for local AI development and inference (NVIDIA, 2025). (HHS.gov)
Apple’s silicon roadmap, high-bandwidth unified memory designs, and ecosystem integration also support on-device inference use cases, especially where privacy and local responsiveness matter. This is not about replacing cloud training; it is about distributing inference across endpoints.
2) Enterprise servers, storage, and networking
Agentic workflows do not only “think.” They retrieve, write, index, monitor, and log. That stresses:
Storage throughput and IOPS (especially for retrieval augmented generation and enterprise search)
Networking inside the enterprise perimeter (east-west traffic)
Security and key management (including confidential computing patterns)
Monitoring and governance systems for audit trails
3) Private cloud and “cloud-in-your-data-center” offerings
A major adoption path will be managed private infrastructure that looks like cloud operationally but runs in a customer-controlled environment. Vendors like Oracle and IBM explicitly position offerings around hybrid and on-prem deployment models, reflecting customer demand for control and compliance. (Oracle)
Company Positioning: Who Benefits and Why (Conceptually, Not as a Recommendation)
In this essay, I will touch on Apple, Nvidia, Dell, Intel, AMD, IBM, Oracle, and Qualcomm. Their potential benefit pathways differ:
Nvidia: Extends from data center training dominance into enterprise inference appliances and “AI factory” reference architectures, widening its total addressable market beyond hyperscalers. (HHS.gov)
Dell: Positioned as an enterprise integrator and server vendor as firms stand up on-prem AI stacks, often combining compute, storage, and services. (Dell)
AMD and Intel: Benefit if AI becomes ubiquitous across endpoints and enterprise servers, with demand spanning CPUs, accelerators, and platform-level optimizations. (Intel CDRD)
Apple: Benefits if “personal and small business agents” become a mainstream on-device workflow category, especially where privacy and portability are valued.
IBM and Oracle: Benefit as governance-heavy industries choose hybrid stacks and managed private deployments that emphasize compliance, auditability, and integration. (Red Hat)
Qualcomm: Positioned if AI PCs, phones, and edge endpoints broaden inference demand and specialized NPUs become a standard feature across devices.
A useful framing is that agentic deployment expands the hardware TAM in two directions at once:
Upward: more data center training and specialized inference clusters.
Outward: more distributed inference at the edge, in offices, and within private environments.
The Decision Framework Enterprises Will Actually Use
Rather than “cloud versus on-prem,” enterprises tend to segment workloads:
Best fit for public cloud
Frontier model training and large-scale experimentation
Bursty workloads with uncertain demand
Rapid prototyping and cross-region deployment
Workloads where vendor-managed security and reliability reduce internal burden
Best fit for on-prem or private environments
Highly sensitive data and strict governance requirements (HHS, 2022; FFIEC, 2020) (HHS.gov)
Stable, always-on agents with predictable capacity needs
Latency-critical workflows and offline resilience
Applications requiring stable outputs and version control for auditability (Chen et al., 2024; MAS, 2024) (Harvard Data Science Review)
This is why hybrid is not a transitional phase; it is likely the steady state (Gartner, 2024). (Gartner)
Risks, Misconceptions, and the “Hard Parts” of On-Prem Agents
A hardware renaissance is plausible, but not frictionless.
On-prem is not automatically safer. Security depends on patching, identity management, logging, and operational maturity. NIST guidance emphasizes security and privacy controls in cloud and hybrid contexts, highlighting shared responsibility and governance needs (Jansen & Grance, 2011; NIST, 2023). (NIST Publications)
Model governance becomes operationally central. Enterprises must manage drift, evaluation, and audit trails, especially in regulated domains (MAS, 2024). (Monetary Authority of Singapore)
Not all tasks need frontier intelligence. Overpaying for capability can be as wasteful as underinvesting in governance. The winning architectures will be right-sized.
Energy and cooling constraints remain binding. Distributed inference increases power density in offices and private data centers, requiring facilities upgrades and procurement discipline.
Conclusion: Two Trends Worth Anchoring On
The agentic layer is not a speculative novelty; it is a natural continuation of enterprise automation with new capabilities. Empirical research and real deployments already show productivity improvements in coding and contact-center settings (Peng et al., 2023; Brynjolfsson et al., 2025). (arXiv)
As agents become always-on co-workers inside firms, compute placement becomes a first-order strategic decision. Regulatory constraints rarely imply “no cloud,” but they do impose governance requirements that often favor private or hybrid deployment (HHS, 2022; FFIEC, 2020). (HHS.gov)
And because stable, auditable production workflows often prefer controlled environments, on-prem and private infrastructure demand can re-accelerate, pulling through a broad hardware stack rather than only a single chip category.
The market may continue debating whether legacy SaaS moats erode or adapt. That debate is real, but it is also hard to handicap. The higher-confidence observation is simpler: agentic workflows increase total compute, and they expand the number of places compute must live. That is the setup for a multi-year hardware buildout across endpoints, enterprises, and sovereign environments.
If you are an international buyer, a China Chinese family office, a South East Asia investor, or a Singapore-based upgrader, the question is not only which property to buy.
It is when, why, and how it fits into your broader wealth plan, residency or education pathway, and risk management across cycles.
That is exactly how I advise.
I am a Singapore real estate professional with a foundation in economics, global affairs, and portfolio construction, as well as deep familiarity with Singapore Land Law, Business Law, and regulatory considerations. I also maintain the discipline of an Officer Commanding in the Singapore Armed Forces: structured planning, rigorous due diligence, and execution with accountability. I do not rely on marketing headlines. I dedicate hours daily to studying macroeconomics, markets, geopolitics, policy shifts, and cross-asset signals, then distill them into clear, practical insights like the essay you just read.
Why does this matter for you?
Because Singapore property decisions are increasingly influenced by forces outside the project brochure: global liquidity cycles, interest-rate expectations, currency dynamics, capital flow trends, regulation, and geopolitical risk. A real estate agent who understands only property may help you view units. A real estate advisor who can connect property with macro and multi-asset positioning can help you make the right decision with the right timeline, structure, and downside protection.
Real estate also has a unique role in a serious portfolio: it is typically less volatile than equities and crypto, can provide stable rental income that resembles dividend cashflow, and has historically offered long-term capital appreciation potential in a supply-constrained, well-governed hub like Singapore. The objective is not to “bet” on property. It is to allocateintelligently, diversify your risk, and build resilient wealth.
If you want a partner who can advise you across the full stack, market timing, project selection, legal and regulatory considerations, negotiation strategy, and portfolio fit, I am available for a private consultation.
Message with:
Your profile (investor, family office, occupier, student family, institutional),
Your timeframe (0 to 3 months, 3 to 12 months, 12+ months, etc),
Your goals (capital preservation, yield, appreciation, residency planning, schooling proximity).
I will revert with a clear, structured plan and curated options aligned to your risk profile and objectives.
Disclosure
This essay is for education and discussion only. It does not constitute financial advice, investment recommendations, or guarantees of outcomes. Company examples are illustrative of business models and deployment pathways, not endorsements.
References (APA 7th Edition)
Bank of America. (2025a, April 8). AI adoption by BofA’s global workforce improves productivity, client service, and employee support [Press release]. (Bank of America)
Bank of America. (2025b, August 20). A decade of AI innovation: BofA’s virtual assistant Erica surpasses… [Press release]. (Bank of America)
Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2), 889–942. (OUP Academic)
Chen, L., Zaharia, M., & Zou, J. (2024). How is ChatGPT’s behavior changing over time? Harvard Data Science Review. (Harvard Data Science Review)
Dropbox. (2016a, March 14). Scaling to exabytes and beyond (Magic Pocket infrastructure). (Dropbox Tech)
Dropbox. (2016b, July 6). Moving 500 petabytes of user data into our Magic Pocket. (Dropbox Blog)
European Data Protection Board. (2020). Guidelines 07/2020 on the concepts of controller and processor in the GDPR.
Federal Financial Institutions Examination Council. (2020). Joint statement on risk management for cloud computing. (HHS.gov)
Gartner. (2024, November 19). Gartner forecasts worldwide public cloud end-user spending to total $723 billion in 2025[Press release]. (Gartner)
Hu, E. J., Shen, Y., Wallis, P., Allen-Zhu, Z., Li, Y., Wang, S., Wang, L., & Chen, W. (2021). LoRA: Low-rank adaptation of large language models. arXiv. (arXiv)
International Labour Organization. (2025). Generative AI and jobs: A refined global index of occupational exposure.
Jansen, W., & Grance, T. (2011). Guidelines on security and privacy in public cloud computing (NIST Special Publication 800-144). National Institute of Standards and Technology. (NIST Publications)
Monetary Authority of Singapore. (2024). Information paper on AI model risk management. (Monetary Authority of Singapore)
National Institute of Standards and Technology. (2023). AI risk management framework (AI RMF 1.0). (NIST Publications)
Netflix. (2016, January 5). Completing the Netflix cloud migration. (Amazon Web Services, Inc.)
Netflix. (2019). Open Connect overview [PDF]. (Open Connect)
NVIDIA. (2025). DGX Spark (Product information). (HHS.gov)
Organisation for Economic Co-operation and Development. (2025). OECD work on AI, jobs, and skills: Generative AI and the labour market.
Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). The impact of AI on developer productivity: Evidence from GitHub Copilot. arXiv. (arXiv)
Pew Research Center. (2026). Washington Post announces staff reductions amid financial pressure (Reporting and context). (Pew Research Center)
The Guardian. (2026, February 11). Washington Post to cut jobs… (abZ Global)
U.S. Department of Health and Human Services, Office for Civil Rights. (2022, December 23). Cloud computing (HIPAA guidance). (HHS.gov)
World Economic Forum. (2025). The future of jobs report 2025.
Comments
Post a Comment