Scarcity Becomes the New AI Advantage

Trend Spotlight

Half the story in construction AI this week is what's failing. Enterprise AI usage dropped from 46% to 37% in a single quarter. Forty-two per cent of initiatives were shelved in 2025, more than double the year before. Only 28% of AI projects meet their ROI targets, according to a Gartner survey of 782 leaders. The pattern is clear: companies that wrapped an API around a foundation model and called it a product are discovering that access to GPT-4 is not a moat. Training a single frontier model now costs upward of $191 million, and the infrastructure layer has consolidated around a handful of giants while the application layer drowns in near-identical tools.

The other half of the story is who is surviving. Firms with proprietary data are pulling ahead. Construction companies that control field data, machine trajectories, soil conditions and historical project records can train tools nobody else can replicate. The competitive advantage has shifted from "who has the best model" to "who has the best data to feed it." Physical AI systems that rely on sensor inputs and real-world interaction are inherently defensible, because the data they need does not exist on the public internet.

Meanwhile, the physical constraints keep tightening. Hyperscaler capital expenditure has reached levels that strain credibility, with multiple analysts projecting $600–700 billion in 2026 alone, and Goldman Sachs projecting the 2025–2027 cumulative total will exceed $1.15 trillion. But the bottleneck is no longer chips — it is power, grid capacity, and the five-year lead time on electrical infrastructure for a data centre that needs to break ground next quarter. "Time-to-Power" is the metric that determines whether trillion-dollar capex commitments translate into actual computing capacity or just press releases. The construction firms capable of executing high-voltage, high-complexity power builds hold the keys to the AI economy.

This Week's Headlines

🟡 AI infrastructure shortages now span eight component categories at once

The component shortage feeding AI infrastructure has broken out of the semiconductor silo. As of Q1 2026, eight distinct categories of inputs — helium, substrate materials, power components, optics, thermal infrastructure and others across the hardware stack — are in active shortage simultaneously, with lead times ranging from 20 to 128-plus weeks, according to Supply Chain Connect [1]. Unlike the 2021–2022 cycle, there is no inventory overhang to self-correct: every constrained component is consumed in production as fast as it arrives. An AI accelerator rack draws 10 to 15 times the power of a traditional server and needs roughly 10 times the multi-layer ceramic capacitor count, and optical networking components that used to ship in 12 to 16 weeks now take 36 to 56 weeks. Omdia forecasts a new allocation hierarchy in which Tier 1 customers — hyperscalers and model builders — receive roughly 70% of component supply, leaving everyone else to compete for the remainder [14]. Intel's chief executive has said memory shortages will persist until 2028, since new fabrication capacity takes three to five years to build [14].

Why it matters: The data centres driving hyperscaler capex are also driving the largest nonresidential construction boom in recent memory. If component shortages delay the servers that fill those buildings, the buildings themselves risk becoming stranded assets. Standard construction IT is not immune either — the same shortages crowding out enterprise servers affect building management systems, IoT sensors and edge computing devices that construction firms rely on for digital workflows. Firms planning technology refreshes should budget for extended lead times.

🟢 42% of enterprise AI projects were shelved in 2025

Contrary Research reports that 42% of enterprise AI initiatives were discontinued in 2025, up sharply from 17% in 2024 [2]. Enterprise AI usage fell from 46% to 37% between June and September 2025 alone. The reason is structural, not cyclical: foundation model providers have commoditised the API layer, making it trivially easy to launch an "AI-powered" tool — in one week alone, more than 73 companies launched identical PDF-chat products [2]. Training GPT-4 required roughly $78 million in compute; Google's Gemini Ultra cost an estimated $191 million [2]. Those costs have risen 4,300% since 2020, consolidating the foundation model market into a handful of players — but that consolidation means differentiation has to come from elsewhere: proprietary data, domain-specific workflows, and deep integration into how a particular industry actually works.

Why it matters: For construction, the implication is clear. Tools that understand drawings, contracts, schedules and field conditions because they were built with that data have a future. Tools that bolt a chat interface onto a generic model and call themselves "construction AI" will go the way of the 42%. Forrester predicts enterprises will defer 25% of planned 2026 AI spending into 2027 for lack of visible value, with only 15% of AI decision-makers reporting a positive impact on profitability [16]. The market is sending a clear signal: show measurable workflow improvement or get shelved.

🟢 Only 28% of AI projects meet their ROI targets, Gartner finds

A Gartner survey of 782 business leaders found that only 28% of AI use cases fully succeed and meet their ROI expectations [3]. Twenty per cent fail outright; the remaining 52% land in a grey zone — technically operational but failing to deliver the value that justified the investment. The primary cause, cited by 57% of leaders, is unrealistic expectations: companies expected AI to deliver dramatic results without investing in the prerequisites — clean data, defined workflows, measurable success criteria. Organisations whose AI projects succeed invest up to four times more in data and analytics foundations than those whose projects fail [3]. In commercial real estate specifically, the barrier is concrete: data sits in fragmented Excel rent rolls, scanned PDFs and proprietary formats that machine-learning models cannot parse. The AI does not fail because the model is weak — it fails because the data was never structured for consumption in the first place.

Why it matters: The question for construction firms is not "which model should we use" but "is our data ready for any model to use." Firms that have invested in structured BIM data, digitised historical project records and standardised their document workflows are positioned to capture AI value. Firms sitting on decades of paper drawings and ad-hoc spreadsheets will find that no amount of model sophistication can compensate for data that was never captured in a usable form.

🟡 Power, not chips, is now the industry's real AI bottleneck

Satya Nadella, Sam Altman and Jensen Huang agree on something: the primary constraint on AI growth is no longer chips, it is power [4]. Nadella put it bluntly: "The biggest problem we face right now is not a shortage of compute, but a shortage of power. My issue today isn't chip supply. It's that I don't have facilities with sufficient power and cooling to deploy those chips." Altman has warned that without major breakthroughs in energy technology, AI will not reach its next stage. The US power grid was not built for AI — interconnection backlogs, hardware shortages (particularly power transformers), and construction-time mismatches mean capital alone cannot solve the problem near-term [4]. "Time-to-Power" will determine AI capex returns, according to MacroMicro's analysis; the physical choke points have shifted to companies sitting on power-delivery infrastructure, memory (HBM), packaging capacity (CoWoS) and regional deployment capability [13]. Accuris reports the top five hyperscalers are projected to spend over $600 billion on infrastructure in 2026, a 36% increase on 2025 [5]; William Blair places the figure near $700 billion [6]; Goldman Sachs projects the cumulative 2025–2027 hyperscaler capex will reach $1.15 trillion, more than double the previous cycle [5].

Why it matters: Management teams at these companies describe the spending as multiyear commitment, not discretionary [6] — but the gap between announced capacity and energised capacity is where construction firms live or die. The firms capable of delivering energised capacity fastest, not just winning the contract, will capture the disproportionate value in this cycle.

🟡 BIM has become a contractual requirement on more projects

BIM has crossed a threshold. Industry reporting indicates that 66% of owners now contractually require digital documentation on their projects [9]. This is not a technology preference; it is a procurement standard. If a contractor cannot deliver a structured digital model, it cannot bid the work. AI tools now automate clash prioritisation, taking the traditional flood of 500-plus detected conflicts and filtering them down to roughly 20 high-priority issues that need human attention [10]. Model validation has moved beyond visual checking: AI can automatically verify naming standards, model completeness and Level of Development (LOD) compliance against the BIM execution plan [10], and IFC-standard compliance is being automated through machine-learning algorithms that classify building elements against Industry Foundation Classes protocols [12]. BIM's scope is expanding too — the term now covers predictive decision-making across cost, sustainability and timeline scenarios, not just geometric modelling, and digital twins are extending BIM into real-time operational use across the asset lifecycle [8].

Why it matters: For contractors still treating BIM as an add-on, the implication is stark: digital delivery is becoming a condition of entry, not a competitive advantage. Firms that have invested deeply in BIM are reporting higher ROI, better risk control and improved collaboration outcomes; firms that have not are finding themselves locked out of an expanding share of the market [8].

🟡 Carbon rules are turning AI-BIM tools into compliance instruments

A peer-reviewed study published in Springer Nature's Journal of Umm Al-Qura University links AI-BIM integration directly to the EU's 2030 Climate Target Plan, which mandates a 55% reduction in carbon emissions [7]. The study reports AI-BIM integration achieving an 89.2% error-detection rate and potential carbon-emission reductions of 37%. Economic pressure reinforces the regulatory push: global construction material prices rose 23.1% between 2020 and 2023, according to Turner & Townsend data cited in the same study [7]. Firms that can model material quantities precisely, optimise structural design for minimal waste, and track embodied carbon through BIM data have a cost advantage that compounds with every regulatory tightening.

Why it matters: The regulatory landscape is still taking shape — EU member states are translating the 2030 targets into national plans, UK local authorities are attaching sustainability conditions to planning permissions, and US jurisdictions are beginning to mandate carbon disclosure on large projects. The direction is consistent even where the mechanisms differ: AI-BIM tools that can demonstrate measurable carbon reduction will increasingly be bought not because they are new, but because they are how firms stay compliant.

🟡 Construction AI is splitting into software copilots and "physical AI"

The construction AI market is splitting in two. On one side: software copilots that read drawings, generate schedules and draft documents — commoditising fast. On the other: physical AI systems that use sensor networks, autonomous equipment and real-time monitoring to act on physical environments. The data physical AI needs — machine motion trajectories, soil-interaction patterns, operator micro-adjustments — is scarce and does not exist at scale on the public internet [11], which is precisely what makes it a moat. CompScience deployments have reported a 35% decrease in incidents using AI-powered safety monitoring [11], and broader industry data suggests firms with mature AI safety deployments see 40–50% incident reductions. The physical-AI segment of the construction AI market is projected to grow at a 17% CAGR, while the broader construction AI application market grows at 24.7% annually [11], [15] — the two rates diverge because they measure different things: software AI is growing through wider adoption of existing capability, physical AI through new capability creation.

Why it matters: Software copilots will keep improving, but the differentiation ceiling is low because the underlying models are shared. Physical AI investment — sensor-equipped equipment, wearable safety devices, autonomous machinery — generates proprietary data that compounds in value over time. A firm that has collected equipment telemetry for three years has a training dataset no competitor can buy, one that becomes the foundation for predictive maintenance, autonomous operation and, eventually, closed-loop construction automation.

Data Point of the Week

42% of enterprise AI initiatives were discontinued in 2025, up from 17% in 2024.

Source: Contrary Research, January 2026 [2]. The discontinuation rate more than doubled in a single year, and the trend is accelerating — enterprise AI usage dropped from 46% to 37% in just one quarter (June to September 2025). This is not a market correction; it is a market verdict. Companies have spent two years experimenting with AI tools and are now systematically killing the ones that do not produce measurable value, and the Forrester prediction that 25% of planned 2026 AI spending will be deferred to 2027 suggests the culling is far from over. The question for construction firms is simple: which of your AI tools would survive a rigorous ROI audit tomorrow?

The Longer View

The capacity gap: capital versus energised power

Hyperscalers are committing hundreds of billions in capital, but the physical infrastructure to absorb it cannot be built fast enough. The gap between announced data-centre capacity and energised capacity is the single biggest variable in the AI infrastructure equation. Multiple analysts project hyperscaler spending at $600–700 billion for 2026 alone [5], [6], with Goldman Sachs projecting $1.15 trillion cumulative for 2025–2027 [5]. Electrical infrastructure lead times extend to five years, transformer shortages are acute, and grid interconnection queues are growing.

Whether hyperscalers and their construction partners can close that gap is the open question. Alternative power solutions — on-site generation, nuclear SMRs, fuel cells — are being tried, but it is not yet clear they can scale fast enough to matter. The construction firms that can demonstrably deliver high-voltage projects on schedule are sitting on one of the largest arbitrage opportunities in the industry right now.

Data as the real competitive moat

The Gartner finding that successful AI projects invest four times more in data foundations [3] is not a one-off observation; it is a consistent pattern across every sector studied. In construction, the firms best positioned for the AI decade are the ones building structured, machine-readable data repositories now — not models, but data: drawings with consistent naming conventions, historical project costs in queryable formats, standardised BIM libraries with proper IFC classification, sensor data from field operations.

How far along that curve the industry actually is remains unclear — data readiness varies enormously across contractors, and there is little public benchmarking of where large firms stand. The firms that solve this problem first are likely to compound their advantage faster than any technology-adoption curve would suggest.

Who is liable when AI misses a clash?

AI tools now claim an 89.2% error-detection rate in BIM coordination [7] — which also means 10.8% of errors go undetected. When a human reviewer misses a clash, professional indemnity insurance covers the liability. When an AI tool misses a clash a human reviewer would have caught, the liability allocation becomes legally murky: does the software vendor share responsibility, or does the firm that deployed the AI carry enhanced liability for over-relying on automated checking?

Standard-form contracts — AIA, JCT, NEC — have not yet caught up with AI-assisted design, and no court case has yet established precedent for shared liability between AI vendors and design professionals. The legal framework is lagging the technology by several years, and the gap looks likely to close through litigation before it closes through legislation.

Sources

[1] Supply Chain Connect. "The 2026 Supply Chain Supercycle: Why AI Infrastructure Shortages Run Deeper Than the Chip." ~June 23, 2026. https://www.supplychainconnect.com/procurement/article/55384522/the-2026-supply-chain-supercycle-why-ai-infrastructure-shortages-run-deeper-than-the-chip

[2] Contrary Research. "The Case for Specialized AI." ~January 9, 2026. https://research.contrary.com/report/the-case-for-specialized-ai

[3] Gartner via AI Consulting Network. "AI Project Success Rates and Data Foundation Investment." ~2026. https://www.aiconsulting.network/ (specific URL not captured; data from Gartner survey of 782 leaders)

[4] MacroMicro. "Outlook 2026 Series IV: The AI Power Endgame: The Infrastructure Race from Chips to the Grid." ~December 17, 2025. https://en.macromicro.me/blog/outlook-2026-series-iv-the-ai-power-endgame-the-infrastructure-race-from-chips-to-the-grid

[5] Accuris. "How AI Data Centers Are Reshaping Electronic Component Supply in 2026." ~May 27, 2026. https://accuristech.com/blog/ai-data-center-electronic-component-supply/

[6] William Blair Investment Management. "The AI Infrastructure Supply Chain: AI Enablers Growing Alongside Hyperscalers." ~June 23, 2026. https://im.williamblair.com/insights/articles/the-ai-infrastructure-supply-chain-ai-enablers-growing-alongside-hyperscalers

[7] Springer Nature (Journal of Umm Al-Qura University). "The Impact of Integrating Artificial Intelligence and Building Information Modeling (BIM) Systems." ~2025. https://link.springer.com/article/10.1007/s43995-025-00193-2

[8] United BIM. "5 BIM Trends in 2026 That Will Shape the Future of AEC." ~February 2, 2026. https://www.united-bim.com/5-innovative-trends-shaping-the-future-of-bim-technology/

[9] Tesla Outsourcing Services. "The 2026 AEC Technology: BIM, AI, and Digital Twins." ~2026. https://www.teslaoutsourcingservices.com/blog/the-2026-aec-technology-bim-ai-digital-twins

[10] BIM Home Studio. "BIM Trends 2026: AI and Digital Twins in Construction." ~May 31, 2026. https://www.bimhomestudio.com/2026/05/31/bim-trends-2026-ai-digital-twins-in-construction/

[11] Buildcheck. "Physical AI Transforms Construction in 2026." ~March 18, 2026. https://buildcheck.ai/insights-case-studies/physical-ai-transforms-construction-in-2026

[12] MyArchitectAI. "AI in BIM: Tools, Workflows, and Real-World Use Cases." ~July 17, 2026. https://www.myarchitectai.com/blog/bim-ai

[13] Sesame Disk. "AI Infrastructure Capex in 2026: Physical Buildout and Supply Chain Constraints." ~May 27, 2026. https://sesamedisk.com/ai-infrastructure-capex-2026/

[14] In Balance IT. "Navigating the 2026 IT Supply Chain Crisis." ~2026. https://inbalanceit.com/supply-chain-crisis

[15] MarketScale (citing Chosun Ilbo). "AI Moves from Pilot to Platform Across Global Construction Operations." ~July 9, 2026. https://www.marketscale.com/industries/engineering-and-construction/ai-moves-from-pilot-to-platform-across-global-construction-operations

[16] Forrester. "Enterprise AI Profitability and Spending Deferral." ~2026. (Cited via industry compilation; specific URL not directly captured.)