The AI Infrastructure Reckoning: Who Will Turn Compute, Memory and Power Into Durable Returns?

Author’s Note and Disclaimer:
Zion Zhao Real Estate | 88844623 | 狮家社小赵 | wa.me/6588844623 |  https://linktr.ee/zionzhao

This post is for general information, education, and market literacy only. It does not constitute financial, investment, trading, legal, tax, accounting, or other professional advice, and is not an offer, solicitation, recommendation, or endorsement. Views expressed are personal, general in nature, and subject to change without notice. While reasonable care is taken, no representation or warranty is given as to accuracy, completeness, or reliability. Readers should conduct independent due diligence and seek professional advice. To the fullest extent permitted by law, no liability is accepted for any loss arising from reliance on this material. 

Beyond Nvidia: The New Battle for Memory, Cloud Capacity, Data Sovereignty and AI Profits

The AI Infrastructure Reckoning: Meta’s Cloud Ambition, Micron’s Memory Supercycle and the Return of Rate-Hike Risk

The market is no longer debating whether artificial intelligence is real. It is debating who will capture the economics, who will overbuild, and who will be left financing yesterday’s bottleneck.

Artificial intelligence has entered a more difficult and consequential phase.

The first stage of the AI investment cycle was relatively straightforward. Demand for advanced processors accelerated. Nvidia became the dominant supplier of AI accelerators. Hyperscalers announced larger data-centre budgets. Investors rewarded almost every company connected to computing infrastructure.

The next stage will be more selective.

Micron, Samsung and SK Hynix are demonstrating that the economic centre of the AI boom is shifting beyond processors towards memory, networking, energy and data-centre capacity. Meta is reportedly examining whether to commercialise surplus computing infrastructure, potentially placing one of the world’s largest social-media companies into direct competition with cloud and neocloud providers. Palantir is advancing a vision of enterprise AI built around data sovereignty, open models and controlled deployment. Meanwhile, geopolitical instability and renewed concerns about inflation have brought the possibility of higher interest rates back into the discussion.

The result is not one investment story, but several overlapping contests.

There is a contest for scarce memory.

There is a contest for data-centre capacity.

There is a contest between proprietary frontier models and open alternatives.

There is a contest between infrastructure growth and financial discipline.

There is also a contest between technological optimism and the macroeconomic reality of energy prices, debt costs and central-bank policy.

The AI boom is real. That does not mean every AI investment will succeed.

The decisive question is no longer whether companies are spending. It is whether that spending can produce sustainable utilisation, defensible margins, measurable productivity and returns above the cost of capital.

Micron’s Earnings Confirm That Memory Has Become Strategic Infrastructure

Micron’s fiscal third-quarter 2026 performance was extraordinary. The company reported approximately US$41.46 billion in revenue, compared with US$23.86 billion in the preceding quarter and US$9.30 billion in the corresponding period a year earlier. It also generated approximately US$18.3 billion in adjusted free cash flow while investing US$7.1 billion in capital expenditure (Micron Technology, 2026).

Those figures are not merely evidence of a strong semiconductor quarter. They indicate a structural change in the economics of computing.

For decades, processors received the greatest attention while memory was frequently treated as a more commoditised supporting component. AI has disrupted that hierarchy.

Advanced computing systems do not operate efficiently through processor speed alone. They require enormous quantities of data to be stored, moved and supplied to processors at extremely high speeds. As models grow larger and inference workloads become more complex, memory bandwidth becomes a critical constraint.

High-bandwidth memory, or HBM, is therefore not an optional accessory. It is an essential component of modern AI systems.

A powerful accelerator cannot achieve its theoretical performance if data cannot reach it fast enough. When memory becomes the bottleneck, memory manufacturers can capture an unusually large share of the industry’s incremental profit.

Samsung’s results reinforce this conclusion. The company reported KRW57.2 trillion in preliminary consolidated operating profit for the first quarter of 2026 and subsequently guided to approximately KRW89.4 trillion for the second quarter. These figures must be correctly identified as Korean won, not US dollars, but the underlying message remains remarkable: memory scarcity has created an extraordinary profit cycle (Samsung Electronics, 2026a, 2026b).

SK Hynix’s major US listing further illustrates the market’s appetite for direct exposure to HBM. Investors increasingly recognise that the AI supply chain cannot be understood through Nvidia alone.

The processor may remain the architectural centre of the system, but economic value often moves towards the scarcest indispensable input. During an oil shortage, the producer of the constrained resource can temporarily earn more incremental profit than downstream businesses. The same principle now applies to AI memory.

Structural Growth Does Not Abolish the Semiconductor Cycle

The bullish case for memory is powerful, but it must not become careless.

Artificial intelligence may lengthen the memory cycle and increase its structural growth rate. Hyperscale data centres, sovereign AI programmes, enterprise adoption, robotics, autonomous systems and continuous inference could support stronger demand than the personal-computer and smartphone cycles of the past.

Micron has also emphasised multiyear strategic agreements intended to improve revenue visibility. Long-term customer commitments could make demand more durable and reduce some of the volatility that historically defined the DRAM industry (Micron Technology, 2026).

Nevertheless, structural demand and cyclicality can coexist.

Railways, telecommunications networks and internet infrastructure all experienced genuine long-term demand. Each also suffered periods when capital was deployed faster than commercially viable utilisation.

Record semiconductor margins inevitably attract new investment. Micron, Samsung and SK Hynix are expanding production capabilities. Those investments take time, particularly for technically demanding HBM products, but new capacity will eventually reach the market.

Investors should therefore ask four questions.

How much additional supply will become available between 2027 and 2030?

Will AI demand grow quickly enough to absorb it?

Will hardware architectures become more memory-efficient?

Can current pricing power persist after scarcity begins to ease?

The correct conclusion is not that memory profits must immediately collapse. Nor is it that extraordinary margins are permanently guaranteed.

The investment opportunity lies in distinguishing a durable structural upgrade from a temporary scarcity premium.

Why Great Earnings Can Still Produce Falling Share Prices

Investors often assume that outstanding earnings should automatically produce a rising stock price.

Markets do not work that way.

A share price reflects expectations before the earnings announcement. Once expectations become extreme, even historic results may fail to surprise investors sufficiently.

This helps explain why semiconductor shares can decline after Micron or Samsung reports record profits. The market may not be disputing current demand. It may be questioning whether profitability has reached a cyclical peak, whether capacity expansion will eventually weaken prices, or whether future growth can continue at the rate already implied by valuations.

The critical issue is the second derivative.

Revenue can remain high while its growth rate slows.

Margins can remain exceptional while declining from an unsustainable peak.

A company can remain fundamentally strong while its shares fall because the market had priced in an even stronger outcome.

This is not necessarily evidence that the AI thesis has failed. It is evidence that valuation and expectation discipline still matter.

The Hyperscaler Capital Expenditure Boom Is Real

The strongest argument against the claim that AI infrastructure demand is imaginary is that some of the world’s most financially sophisticated companies are committing enormous amounts of capital to it.

Meta has guided to between US$125 billion and US$145 billion of capital expenditure in 2026. Microsoft has indicated that demand for Azure and AI services continues to exceed available capacity, while its calendar-year 2026 investment could approach US$190 billion. Amazon, Alphabet, Oracle and specialised infrastructure providers are also expanding aggressively (Meta Platforms, 2026; Microsoft Corporation, 2026).

The contractual evidence is equally significant.

CoreWeave reported approximately US$99.4 billion in revenue backlog as of March 31, 2026. Nebius announced a five-year agreement to provide Meta with US$12 billion of dedicated capacity, alongside a potential additional commitment of up to US$15 billion under specified conditions. Oracle reported approximately US$638 billion in remaining performance obligations (CoreWeave, 2026; Nebius Group, 2026; Oracle Corporation, 2026).

This is not the behaviour of companies conducting a limited experiment.

AI infrastructure is becoming one of the largest capital-allocation programmes in corporate history.

However, the existence of spending does not settle the debate. It changes the debate.

The relevant issue is no longer whether demand exists. The issue is whether the infrastructure can be delivered on time, utilised efficiently and monetised at attractive returns.

Backlog Is Important, but It Is Not Cash

Large remaining performance obligations and contractual backlogs provide important evidence of future demand. They should not be confused with recognised revenue, operating income or cash flow.

AI infrastructure contracts often depend on facilities being completed, power connections becoming available, accelerators being delivered and service-level requirements being satisfied.

A signed customer cannot produce revenue until the infrastructure is operational.

This is particularly important for specialised neoclouds. They must secure land, electricity, cooling systems, transformers, networking equipment, processors, memory, financing and technical personnel before they can fulfil their obligations.

Backlog therefore proves that customers want capacity. It does not eliminate execution risk.

Customer concentration is another concern. A provider may report an enormous contracted pipeline while depending heavily on one or two counterparties. If those customers delay expansion, renegotiate requirements or experience financial pressure, the provider may face material consequences.

The strongest infrastructure businesses will be those that convert contracted demand into diversified, profitable and recurring revenue without taking excessive balance-sheet risk.

Free Cash Flow Cannot Be Dismissed

One of the most important disagreements in the AI investment debate concerns free cash flow.

Bullish investors correctly argue that capital expenditure can temporarily suppress free cash flow even when a company is making intelligent long-term investments. A data centre must be financed and constructed before it can generate revenue.

Operating cash flow and free cash flow therefore answer different questions.

Operating cash flow indicates whether the underlying business is producing cash.

Free cash flow indicates how much remains after capital investment.

A company can have a healthy core business and weak free cash flow during an expansion cycle.

However, free cash flow still matters.

Persistent negative free cash flow can increase debt, cause shareholder dilution and reduce strategic flexibility. If interest rates remain elevated, financing costs can materially weaken project economics.

Oracle illustrates both sides of this issue. The company reported approximately US$32 billion in operating cash flow and US$638 billion in remaining performance obligations, but heavy infrastructure spending contributed to approximately negative US$23.7 billion in free cash flow during the fiscal year (Oracle Corporation, 2026).

The bullish interpretation is that Oracle is investing ahead of extraordinary contracted demand.

The cautious interpretation is that it is assuming substantial financing and execution risk to serve rapidly evolving customers with infrastructure that may become technologically obsolete faster than expected.

Both statements can be true.

The correct question is not whether capital expenditure is large. It is whether the expected return on that expenditure exceeds the company’s financing cost and compensates shareholders for the risk.

Meta’s Reported Cloud Expansion Is Strategically Logical

Meta has reportedly been developing plans to commercialise excess AI computing capacity through some form of external cloud offering.

This should not yet be described as a fully launched neocloud business. The structure, scale and timing remain uncertain. Meta may provide model access, managed AI services, raw computing capacity or a narrower combination of products.

Nevertheless, the strategic logic is clear.

Meta is already building enormous computing infrastructure for recommendation systems, advertising optimisation, content generation, business messaging, assistants and advanced model development.

Selling unused or temporarily available capacity could improve asset utilisation, offset depreciation and energy costs, diversify revenue and strengthen the ecosystem surrounding Meta’s models.

Cloud computing fundamentally involves pooling infrastructure and allocating it flexibly across users. This can reduce the economic cost of idle capacity and improve utilisation (Armbrust et al., 2010).

Meta could also use external demand to justify even larger infrastructure investment. If the company can serve its internal AI ambitions while earning revenue from third parties, its data centres become both a strategic capability and a monetisable commercial asset.

That is a more persuasive investment narrative than simply describing every dollar of capital expenditure as an internal research expense.

Meta’s Entry Does Not Automatically Destroy Neoclouds

The immediate reaction to Meta’s possible expansion was that specialised providers such as CoreWeave and Nebius would face a powerful new competitor.

Competition would certainly intensify.

Meta has immense financial resources, technical talent, purchasing power and experience operating large-scale data centres. It could also price external capacity aggressively because its advertising business absorbs part of the infrastructure’s fixed cost.

Yet Meta’s entry may validate the neocloud market rather than invalidate it.

A company does not normally enter a market it believes has no customers.

If Meta believes enterprises and model developers will pay for advanced computing capacity, that supports the argument that AI infrastructure is becoming a substantial commercial category.

The effect on current neoclouds will depend on differentiation.

A company that merely rents interchangeable processors may experience price pressure.

A provider that offers superior deployment speed, networking, orchestration, inference optimisation, developer tools, customer support or access to scarce hardware may remain valuable.

Meta is also a significant customer of specialised providers. Its decision to build an external service does not automatically cancel its contractual commitments.

The AI market is increasingly defined by customer-supplier-competitor relationships. Companies may buy capacity from one another, invest in one another and compete with one another at the same time.

Meta may purchase external infrastructure because a neocloud can deliver a particular region or hardware generation faster than Meta can build internally. Meta can simultaneously operate its own systems and sell temporary surplus capacity.

The market is becoming a network of interdependence rather than a simple supply chain.

Meta Must Build a Platform, Not Merely Own Processors

Owning large quantities of accelerators is not sufficient to create a successful cloud business.

Enterprise customers expect predictable uptime, cybersecurity, identity controls, storage, billing, technical support, regulatory compliance, data-residency options, developer tools and service-level guarantees.

Amazon Web Services, Microsoft Azure and Google Cloud have spent years creating those capabilities.

Meta’s competitive advantage is its massive infrastructure and access to billions of users. Its weakness is that it has not historically operated a general-purpose enterprise cloud platform.

The company’s most credible entry may therefore be specialised rather than universal.

Meta could focus on model access, training clusters, inference, advertising agents, business messaging tools and AI services connected to Facebook, Instagram and WhatsApp.

It does not need to recreate every feature of AWS to build a meaningful business.

The central question is whether Meta can turn internal infrastructure into a reliable, differentiated and commercially trusted platform.

Idle Capacity Does Not Necessarily Mean Overbuilding

The phrase “excess compute” can create the impression that Meta has already built more capacity than it needs.

That conclusion may be premature.

Large infrastructure operators deliberately maintain reserve capacity for demand surges, redundancy, testing, migrations and future product launches. Temporary spare capacity can be operationally prudent.

The distinction is between temporary reserve capacity and structural overcapacity.

If Meta sells capacity opportunistically while maintaining strong internal utilisation, the strategy may improve returns.

If it repeatedly sells large quantities of infrastructure below the expected internal return because original demand forecasts were wrong, that would indicate overinvestment.

Until Meta discloses capacity, utilisation and pricing data, neither interpretation can be treated as proven.

Palantir’s Open-Model Push Is Really About Control

Palantir chief executive Alex Karp has argued that enterprises should not surrender their proprietary knowledge to external model providers whose primary objective is selling access to tokens and centralising control.

His alternative is an architecture in which organisations retain greater authority over data, computing environments and operational workflows.

This is not simply an ideological argument about open source.

It is an argument about sovereignty.

Enterprises must ask who owns their data, where it is processed, which laws apply, whether a provider can use interactions to improve other products, whether sensitive workloads can operate offline, and whether the organisation can switch vendors without rebuilding its entire system.

Open-weight models can help address some of these concerns. They may be deployed locally, customised and operated within controlled environments.

Palantir’s collaboration with Nvidia seeks to combine open models with Palantir’s ontology, governance and operational software. The objective is to connect AI with sensitive institutional data while preserving permissions, auditability and control.

This model is particularly relevant to governments, defence agencies, financial institutions, healthcare providers and critical infrastructure operators.

Open Models Do Not Automatically Guarantee Security

Open deployment can improve transparency, portability and control. It does not remove risk.

An organisation running a model internally remains responsible for cybersecurity, access controls, model vulnerabilities, monitoring, patching, data leakage and misuse.

NIST’s Artificial Intelligence Risk Management Framework emphasises that trustworthy AI requires governance throughout the entire system lifecycle. Selecting an open model or keeping data within a private environment is only one component of responsible deployment (National Institute of Standards and Technology, 2024).

The healthcare example is particularly important.

Healthcare organisations are not categorically prohibited from using cloud infrastructure. US guidance permits HIPAA-regulated entities to use cloud services when appropriate contracts and security safeguards are established (US Department of Health and Human Services, 2022).

The real decision is not cloud versus no cloud.

It is which combination of public cloud, private infrastructure, sovereign cloud and on-premises systems satisfies legal, operational and security requirements.

Most large enterprises will probably adopt hybrid architectures.

Open and Proprietary Models Will Coexist

The AI market is unlikely to produce a single model category that serves every requirement.

Open models may become dominant for repetitive, cost-sensitive, highly customised or private workloads. Enterprises can deploy them locally and avoid dependence on one provider.

Proprietary frontier models may remain attractive when users need the best available reasoning, multimodal capability, advanced tools, rapid innovation or professional support.

The risk of relying on a small number of foundation models is that defects, biases and governance failures can propagate across many downstream applications (Bommasani et al., 2021).

Open releases create a different set of concerns. They can promote competition and research, but capable models may become difficult to monitor, restrict or withdraw once widely distributed (Seger et al., 2023).

The rational enterprise strategy is therefore model pluralism.

Companies should use different models for different tasks, maintain portability and continuously compare cost, performance, security and reliability.

Cost per Token Is the Wrong Final Metric

Much of the model debate focuses on token pricing.

That is useful but incomplete.

A cheaper model that requires extensive human correction may be more expensive than a costly model that completes the task accurately. An expensive frontier model may also be unnecessary for routine classification or summarisation.

The proper metric is cost per successful outcome.

Enterprises should measure task completion, error rates, human-review requirements, latency, customer satisfaction and the financial result of the workflow.

Research by Brynjolfsson, Li and Raymond found that access to a generative AI assistant increased productivity among customer-service workers by approximately 15 percent on average, with greater gains among less experienced employees. The research also demonstrated substantial variation across workers and tasks, showing that implementation determines value (Brynjolfsson et al., 2025).

The largest opportunity is therefore not the chatbot.

It is the redesign of real business processes.

Falling Token Costs Could Increase Total Compute Demand

A common bearish argument is that more efficient hardware and cheaper models will reduce infrastructure revenue.

That assumes demand remains fixed.

When the cost of inference falls, companies can apply AI to more tasks. An organisation that previously used AI occasionally may begin deploying it continuously across coding, customer service, compliance, logistics, forecasting, research and sales.

This creates a version of the Jevons paradox. Greater efficiency can increase total resource consumption because lower costs expand usage.

The relevant equation is not simply the price of one token.

It is total inference volume multiplied by revenue per unit of inference.

Model providers can reduce unit prices and still generate more revenue if usage expands faster than prices decline.

Agentic systems could accelerate this trend. Unlike a chatbot that responds only when a human enters a prompt, an agent can monitor information, call tools, update systems and perform workflows continuously.

However, current agents remain imperfect. Many require supervision, make errors or become expensive when repeated model calls accumulate.

The infrastructure thesis will depend on whether agents move from demonstrations into reliable production systems.

Corporate Expenses Are an Opportunity, Not a Guaranteed Savings Pool

Large corporations spend enormous sums on selling, general and administrative expenses. AI could improve fraud detection, document processing, coding, customer service, compliance and decision-making.

However, SG&A should not be treated as entirely automatable waste.

It includes salaries, marketing, rent, professional services, regulatory functions and activities requiring human judgement and accountability.

AI savings will also be offset by spending on models, data preparation, systems integration, cybersecurity, monitoring and employee training.

The strongest case is not that AI eliminates every administrative worker.

It is that AI allows employees to handle greater workloads, automates repetitive processes, reduces errors and improves response times.

Companies that redesign workflows carefully may achieve substantial productivity gains. Companies that purchase AI tools without organisational change may achieve very little.

Geopolitical Risk Still Runs Through Energy

The Strait of Hormuz remains one of the most economically important maritime chokepoints in the world.

Approximately 20 million barrels per day of oil passed through the strait in 2024, representing roughly 20 percent of global petroleum-liquids consumption. Approximately one-fifth of global liquefied natural gas trade also travelled through the route (US Energy Information Administration, 2025a, 2025b).

A sustained interruption would affect oil, natural gas, freight, insurance, inflation and economic growth.

However, not every geopolitical escalation produces the same market outcome.

A temporary risk premium may increase oil prices without creating a recession. A prolonged physical disruption could reduce supply, weaken consumer purchasing power and force central banks to confront higher inflation.

Research by Kilian demonstrates that oil-price shocks caused by supply disruptions can have different economic consequences from price increases caused by stronger global demand (Kilian, 2009).

The equity market’s resilience does not prove that geopolitical risk is irrelevant.

It indicates that investors currently assign a relatively low probability to a prolonged interruption.

Markets can be wrong.

They can also rise because of strong earnings, hedging, liquidity, short covering or expectations of government intervention.

Responsible analysis should therefore consider multiple scenarios rather than treating one day’s market movement as proof.

The Federal Reserve Faces a Genuine Conflict

The Federal Reserve maintained its policy rate at 3.50 to 3.75 percent in June 2026.

The data provide arguments both for caution and against further tightening.

Headline consumer inflation rose 4.2 percent over the 12 months ending in May, while energy prices increased sharply. Core inflation was lower but remained above the Federal Reserve’s objective. At the same time, nonfarm payrolls increased by only 57,000 in June and unemployment remained at 4.2 percent (Bureau of Labor Statistics, 2026a, 2026b).

This creates a genuine policy dilemma.

Persistent inflation may require restrictive policy.

A weakening labour market argues against making borrowing conditions even tighter.

The fact that policymakers revise their projections does not prove political manipulation. Forecasts change because inflation, employment, fiscal policy, energy prices and financial conditions change.

The Federal Reserve can still be criticised for relying too heavily on delayed or revised indicators. Shelter inflation, for example, can respond slowly to changes in newly signed rents.

High-frequency information from rental listings, freight rates, commodities, job advertisements and payment systems can improve policymaking. Orphanides demonstrated that policy recommendations based on real-time information can differ materially from those calculated later using revised data (Orphanides, 2001).

The best framework would combine stable official statistics with high-frequency indicators and clearly acknowledge uncertainty.

Rate Hikes Are Possible, but Not Predetermined

The evidence does not justify certainty that the Federal Reserve will raise rates.

A hike would become more plausible if energy prices remained elevated, inflation broadened beyond energy, wage pressure accelerated or inflation expectations became less anchored.

A hike would become less likely if geopolitical tensions eased, employment weakened, inflation slowed or financial conditions tightened independently.

The most likely immediate risk may be that rates remain restrictive for longer, rather than an aggressive new hiking cycle.

This still matters for AI infrastructure.

Higher rates increase the financing cost of data centres, reduce the present value of distant cash flows and place pressure on companies with large debt or negative free cash flow.

The AI investment cycle cannot be analysed separately from monetary policy.

Beyond Technology: Commercial Property, Power and Finance

The concentration of equity-market performance in technology creates a strong incentive to examine other sectors.

Commercial real estate is one area, but it must be analysed by quality.

Remote and hybrid work have weakened demand for older office buildings. Gupta, Mittal and Van Nieuwerburgh estimated substantial long-run value destruction across US office markets under their model (Gupta et al., 2026).

Yet high-quality, well-located and energy-efficient buildings may benefit from a flight to quality. Major tenants may consolidate into modern properties with strong transport connections and better amenities.

Even premium assets face refinancing risk when rates remain elevated. Investors must examine debt maturities, tenant concentration, leasing incentives and capital requirements.

A large dividend yield is not protection against an unsustainable balance sheet.

AI also creates demand outside traditional technology.

Data centres require electricity generation, transmission, transformers, cooling, engineering, construction and land. In some regions, access to power is becoming more constrained than access to capital.

Utilities and industrial suppliers may benefit, but they also face regulatory, financing and execution risks. Local communities may resist projects because of water consumption, electricity prices, land use or environmental effects.

Financial institutions may benefit from AI through fraud detection, underwriting, compliance, customer service and software development. Yet regulated decisions require auditability, cybersecurity and human accountability.

The institutions that create the greatest value will not necessarily be those that buy the most AI. They will be those that implement it within high-volume workflows without undermining trust.

The AI Trade Requires a More Disciplined Framework

The AI boom is supported by real evidence.

Semiconductor earnings have surged.

Hyperscalers are investing extraordinary amounts.

Infrastructure providers have accumulated major contracted backlogs.

Enterprises are deploying AI across a widening range of functions.

This is materially different from a speculative cycle built entirely on unprofitable companies with no customers.

Yet real demand does not guarantee attractive returns for every participant.

Investors should ask six questions.

First, is demand contracted, speculative or already producing revenue?

Second, who captures the economics: processors, memory, clouds, software providers or end users?

Third, what happens when present shortages become future supply?

Fourth, can expansion be financed without excessive debt or dilution?

Fifth, does the business possess software and workflow differentiation, or is it selling commoditised capacity?

Sixth, what evidence would prove the investment thesis wrong?

The bullish AI thesis would weaken if utilisation declined, contracts were cancelled, customers failed to monetise applications or infrastructure providers could not finance construction.

The bearish thesis would weaken if enterprise adoption broadened, inference volume expanded faster than unit costs fell, and providers converted backlogs into profitable cash flow.

An investor who cannot identify what would change a view is not analysing. That investor is defending an identity.

Conclusion: The Infrastructure Is Real, but the Returns Must Still Be Earned

The AI economy is neither a simple bubble nor a guaranteed path to limitless wealth.

Micron’s results, Samsung’s profitability and SK Hynix’s market reception demonstrate that memory has become strategic infrastructure.

Meta’s reported examination of a cloud business suggests that compute is becoming a monetisable asset, but success will depend on building a credible platform above the hardware.

CoreWeave, Nebius and Oracle possess substantial contracted demand, but they must deliver complicated infrastructure while managing financing, customer concentration and technological change.

Palantir’s argument for data sovereignty addresses a real enterprise concern. Open models and proprietary frontier systems will probably coexist because organisations require different combinations of capability, security, control and cost.

Geopolitical instability remains important because energy links foreign policy with inflation and monetary policy.

The Federal Reserve faces conflicting signals rather than an obvious decision. Restrictive rates may persist, creating greater pressure on capital-intensive businesses.

The correct investment framework is disciplined, not ideological.

Bulls must acknowledge cyclicality, competition, financing risk and valuation.

Bears must acknowledge real revenue, real contracts, productivity gains and the strategic commitments of some of the world’s largest companies.

The next phase of the AI cycle will not be determined by the loudest prediction.

It will be determined by utilisation, cash conversion, productivity and return on invested capital.

Artificial intelligence has already created a historic infrastructure build-out.ƒ

The remaining question is whether that infrastructure becomes one of history’s greatest capital-allocation successes, or an extraordinary case of too much capital chasing a technological revolution whose commercial value took longer to arrive than investors expected.

---

Author’s Disclaimer

This article is intended solely for general education, market commentary and informational purposes. It does not constitute financial, investment, legal, tax or geopolitical advice, nor an offer, solicitation or recommendation to buy or sell any security or financial product.

References to public companies are included for analytical purposes and should not be interpreted as endorsements. Corporate guidance, capital expenditure plans, backlogs, valuations and geopolitical conditions may change materially. Readers should conduct independent due diligence and consult appropriately licensed professionals before making financial decisions.

Discussion of regulated gaming businesses is limited to corporate and market analysis and does not promote gambling. Discussion of defence companies and geopolitical conflict does not endorse violence or military escalation. The humanitarian consequences of conflict must remain central to responsible analysis.

References

Armbrust, M., Fox, A., Griffith, R., Joseph, A. D., Katz, R. H., Konwinski, A., Lee, G., Patterson, D. A., Rabkin, A., Stoica, I., & Zaharia, M. (2010). A view of cloud computing. Communications of the ACM, 53(4), 50-58. https://doi.org/10.1145/1721654.1721672

Bommasani, R., Hudson, D. A., Adeli, E., Altman, R., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E., et al. (2021). On the opportunities and risks of foundation models. Stanford Center for Research on Foundation Models.

Brynjolfsson, E., Li, D., & Raymond, L. R. (2025). Generative AI at work. The Quarterly Journal of Economics, 140(2).

Bureau of Labor Statistics. (2026a). Consumer Price Index: May 2026. US Department of Labor.

Bureau of Labor Statistics. (2026b). The employment situation: June 2026. US Department of Labor.

CoreWeave, Inc. (2026). CoreWeave reports first-quarter 2026 results.

Federal Open Market Committee. (2026). Minutes of the Federal Open Market Committee, June 16-17, 2026. Board of Governors of the Federal Reserve System.

Gupta, A., Mittal, V., & Van Nieuwerburgh, S. (2026). Work from home and the office real estate apocalypse. American Economic Review, 116(2), 674-709. https://doi.org/10.1257/aer.20231619

Kilian, L. (2009). Not all oil price shocks are alike: Disentangling demand and supply shocks in the crude oil market. American Economic Review, 99(3), 1053-1069. https://doi.org/10.1257/aer.99.3.1053

Meta Platforms, Inc. (2026). First-quarter 2026 financial results.

Micron Technology, Inc. (2026). Micron Technology reports record results for the third quarter of fiscal 2026.

Microsoft Corporation. (2026). Fiscal year 2026 third-quarter earnings conference call.

National Institute of Standards and Technology. (2024). Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile. US Department of Commerce.

Nebius Group. (2026). Nebius signs new AI infrastructure agreement with Meta.

Oracle Corporation. (2026). Fiscal 2026 fourth-quarter and full-year financial results.

Orphanides, A. (2001). Monetary policy rules based on real-time data. American Economic Review, 91(4), 964-985. https://doi.org/10.1257/aer.91.4.964

Samsung Electronics. (2026a). Samsung Electronics announces first-quarter 2026 results.

Samsung Electronics. (2026b). Samsung Electronics announces earnings guidance for the second quarter of 2026.

Seger, E., Dreksler, N., Moulange, R., Dardaman, E., Schuett, J., Wei, K., Winter, C., Arnold, M., Ó hÉigeartaigh, S., & Anderljung, M. (2023). Open-sourcing highly capable foundation models: An evaluation of risks, benefits and alternative methods for pursuing open-source objectives.

US Department of Health and Human Services. (2022). Guidance on HIPAA and cloud computing.

US Energy Information Administration. (2025a). Amid regional conflict, the Strait of Hormuz remains critical to global oil flows.

US Energy Information Administration. (2025b). About one-fifth of global liquefied natural gas trade flows through the Strait of Hormuz.

Meta’s Cloud Push and Micron’s Memory Boom Redefine the AI Trade

Artificial intelligence is no longer only a technology story. It is reshaping global capital flows, data centre demand, energy infrastructure, financing costs, business confidence and investor risk appetite. These forces also influence Singapore property decisions.

For buyers, higher interest rates and changing capital allocation affect affordability, loan structures and entry timing. For sellers, market positioning, pricing discipline and buyer sentiment become increasingly important. For landlords and tenants, shifts in employment, business expansion and operating costs can influence rental demand. For investors, the key is to identify properties supported by genuine scarcity, sustainable rental income, strong connectivity and long-term economic relevance.

The lesson is clear: property decisions should not be made in isolation. They should be evaluated alongside interest-rate cycles, liquidity, infrastructure investment, supply pipelines and broader economic trends.

As a Singapore real estate salesperson with a strong focus on macroeconomics, capital markets and property strategy, I help clients translate complex global developments into practical decisions for buying, selling, renting and investing.

For personalised Singapore property advice, contact Zion Zhao Real Estate at 8884 4623 or wa.me/6588844623.

Like, save, follow and subscribe for more professional market and property insights.