The cost of running artificial intelligence in enterprise workflows is falling so fast — potentially 60 to 85 percent over two years — that it will push AI from optional pilot project to standard operating infrastructure across entire industries. That transition is already underway. The problem is that Wall Street's current trade is too narrow, too early in the supply chain, and almost entirely blind to the regulatory, power-grid, and labor-repricing shocks that will define who actually wins.
Five-Model Consensus
All five analysts agreed on the core mechanism: inference cost compression is the key variable that expands AI adoption from large enterprises into mid-market and SME deployments, and the market is still primarily pricing the infrastructure layer rather than the application and services layer where the next earnings revisions emerge. All five also agreed that the power demand story is real and underappreciated in regional specificity. The dissents were meaningful. Atlas stood largely alone in the depth and urgency of its regulatory enforcement argument — specifically that existing sectoral regulation, not future AI legislation, represents the first credible shock to enterprise AI deployment economics, and that the 12-to-18-month enforcement timeline is closer than consensus assumes. Grayline introduced a contrarian wrinkle the others did not: that the same cost curves enabling SME adoption also let large enterprises repatriate AI workloads away from public cloud to on-premises or colocation facilities, which could flatten the cloud revenue curve that sell-side analysts are currently using as the backbone of their AI earnings models. Meridian and Vantage were most aligned on the quantitative framing of the opportunity, though Meridian was more specific about the options market implications — noting that implied volatility, or expected price swings priced into options contracts, is already elevated in the crowded AI leaders while remaining too low in second-derivative beneficiaries like utilities, software integrators, and selected BPO names. Chronicle was the most conservative on adoption curve certainty, emphasizing that confirmed infrastructure facts are solid but that precise SME penetration timing remains unproven.
Contributing: Atlas, Meridian, Grayline, Vantage, Chronicle
Start with the number that changes everything: inference cost. Inference is what it costs to run an AI model on a real task — answering a question, drafting a document, flagging a suspicious transaction. Right now that cost is falling fast, driven by more efficient chips, better model architectures, and compression techniques that let smaller models do work that used to require much larger ones. When the cost per useful AI interaction drops below roughly two cents, the math changes for every mid-sized company in America, not just the Fortune 500. A firm with 3,000 knowledge workers spending on AI copilots — software tools that assist rather than replace workers — goes from spending $7.5 million a year on those tools to under $2 million, while the productivity gain stays roughly the same. At that price, AI moves from the IT budget to the operating budget. That shift is the story the market is still under-pricing.
The equity trade so far has been dominated by semiconductor companies and cloud platforms — the infrastructure layer. That trade is real and not over. But it is incomplete in two directions. First, the next earnings revisions will come from software companies with large installed customer bases. A platform with ten million paying subscribers that adds an AI assistant — and charges even a modest $25 a month for it — can generate hundreds of millions in new annual revenue with gross margins, meaning profit per dollar of revenue, that are materially higher than their core business. The market still treats this as a future option rather than a near-term revenue line. It is becoming the latter. Second, the market has almost entirely skipped the insurance, banking, and back-office services layer, where the productivity math is quietly devastating to old cost structures. A large insurance carrier with a 25 percent expense ratio — meaning 25 cents of every premium dollar goes to operations — that shaves five percent of claims-handling and underwriting labor hours does not save a little money. It improves its combined ratio, the industry's core profitability measure, by over 100 basis points. In insurance math, that is enormous.
But here is what the bull case is missing, and it is not abstract. Regulatory enforcement is not waiting for Congress to pass an AI law. It is already on the books. Financial regulators — the OCC, FDIC, and Federal Reserve — have had model risk management guidelines since 2011, written for statistical credit models, that apply directly to AI systems embedded in lending and compliance decisions. Regulators are beginning to apply them. The first major enforcement action against a bank or insurer for an AI-driven compliance failure — a discriminatory loan decision, a missed money-laundering flag — is likely within 12 to 18 months given current deployment pace. When it lands, it will not be treated as an isolated incident. It will be a sector-wide recalibration of what it costs to deploy AI in regulated workflows. The software vendors selling AI tools into banks, lenders, and insurers without built-in audit trails and bias testing are accumulating a liability that is not in their stock price.
The power story has the same problem: it is real but being analyzed at the wrong level. National data-center power demand figures are accurate and large. They are also nearly useless for investment decisions. The consequential version of this story lives at the regional grid level — in PJM, ERCOT, and MISO, the regional organizations that manage electricity transmission across large parts of the country. Those grids already have interconnection queues — the waiting list for new facilities to connect to the power grid — stretching three to seven years. Announced AI data center capacity and actually powered AI data center capacity are going to diverge sharply by 2026 and 2027. That gap is not in hyperscaler capital expenditure plans. It is not in utility valuations. And when state utility commissions start receiving rate cases asking residential customers to help pay for grid upgrades primarily serving large tech campuses, the political backlash will generate new cost-allocation rules that shift the bill back to the data centers. That compresses return assumptions for hyperscalers in ways the market has not modeled.
The labor story is being told as a jobs debate when it is actually an earnings story in the near term and a fiscal story in the medium term. Companies are not going to announce mass layoffs. They are going to slow hiring while revenue grows. The margin expansion shows up in quarterly earnings before unemployment statistics move at all. That is the quiet near-term signal. The larger and less-discussed risk is geographic concentration: New York, California, Illinois, and Connecticut run their state budgets heavily on income tax revenue from high-density professional services employment — lawyers, analysts, insurance professionals, compliance staff, back-office finance workers. If legal, financial services, insurance, and customer operations all face the same labor repricing simultaneously rather than sequentially — as spreadsheet software hit accounting alone, over a decade, in the 1980s — the tax revenue effect on those state governments is a fiscal exposure that appears in no current budget projection. That is not a next-quarter problem. It is an 18-to-36-month problem. But the clock is running.
Model Perspectives — Original Analysis
The regulatory and historical framing on AI infrastructure deployment is almost entirely wrong in mainstream coverage, and the error is not one of omission but of category. Reporters and analysts are treating this as a technology adoption story with downstream labor effects, when the correct historical analogy is the buildout of electrical grid infrastructure in the 1920s and early interstate highway system in the 1950s — both of which triggered regulatory frameworks that lagged the buildout by a full business cycle, then arrived suddenly and reshaped competitive dynamics in ways that punished first movers who had optimized for the unregulated environment. The second and third-order regulatory effects are not being modeled at all.
On the regulatory front, the operative framework being missed is not the AI safety legislation everyone is watching — EU AI Act implementation, potential US federal AI bills — but rather the intersection of AI deployment with existing sectoral regulation that is already on the books and is about to be enforced with new intensity. Financial services regulators (OCC, FDIC, Fed, SEC) have model risk management guidance (SR 11-7) that was written for statistical models but will be applied to generative AI systems embedded in credit, trading, and compliance workflows. The enforcement gap between deployment and regulatory catch-up is closing faster in finance than anywhere else, and firms rushing AI copilots into underwriting and AML functions are accumulating model governance liabilities they have not provisioned for. The first major enforcement action against a bank or insurer for an AI-driven compliance failure will be a sector-wide shock, and it is closer than consensus assumes — likely within 12 to 18 months given the deployment pace and the current regulatory posture of the CFPB and OCC under any administration that inherits ongoing supervision cycles.
The labor market displacement story is being told entirely wrong. Every mainstream piece frames this as a jobs-versus-no-jobs debate borrowed from previous automation waves. The historically accurate precedent is not manufacturing automation but the introduction of spreadsheet software in the 1980s, which did not eliminate accounting jobs in aggregate but radically restructured the skill premium within the profession, compressed middle-tier roles, and created a decade-long transition period during which firms that moved first captured outsized margin before the new wage equilibrium settled. The difference this time is speed and breadth: the compression cycle will be 24 to 36 months rather than a decade, and it will hit multiple cognitive-task sectors simultaneously rather than sequentially. This simultaneity is the key variable no one is modeling. When legal, insurance, financial services, and customer operations all face the same labor repricing at the same time, the macroeconomic feedback loop through consumer spending and tax revenue is materially different from historical single-sector transitions. State governments reliant on income tax revenue from high-density professional service employment — New York, Illinois, California, Connecticut — face a fiscal exposure that is not in any budget projection.
The power infrastructure story is being told at the wrong level of granularity. The national-level data center power demand numbers being cited are accurate but analytically inert. The consequential story is at the ISO and utility commission level. PJM, ERCOT, and MISO are already seeing interconnection queues dominated by data center load requests, and the queue processing timelines — three to seven years in many cases — mean that announced AI data center capacity and actual grid-connected capacity will diverge sharply by 2026 to 2027. This creates a regulatory chokepoint that is not priced into either hyperscaler capex plans or utility equity valuations. The historical precedent is the 2000s LNG import terminal buildout, where billions were committed to infrastructure premised on regulatory and interconnection timelines that proved systematically optimistic, and early movers faced stranded asset risk when the environment shifted. The politically salient flashpoint will arrive when a state public utility commission faces a rate case in which residential customers are asked to subsidize grid upgrades primarily serving hyperscaler data centers — this is a populist regulatory flashpoint that will generate legislative intervention, likely in the form of large-load customer interconnection cost allocation rules that shift capex burden back to the data center operators and compress their return assumptions.
The SME AI adoption expansion story — correctly identified in the source brief as undercovered — has a specific regulatory implication that no one is discussing: when AI tools reach the sub-500-employee firm tier at scale, the compliance infrastructure those firms have in place to govern model use is essentially zero. This creates a systemic consumer protection exposure in sectors like mortgage origination, auto lending, employment screening, and healthcare billing that existing regulators have enforcement jurisdiction over today, without any new legislation. The FTC's unfair and deceptive practices authority, ECOA and Fair Housing Act enforcement by CFPB and DOJ, and HIPAA privacy rules all apply to AI-assisted decisions at SMEs right now. The gap between deployment velocity and compliance capability at the SME tier is the single largest unpriced regulatory risk in the current AI investment thesis.
Looking six months out: the most likely visible regulatory event is not a new AI law but the first wave of sectoral enforcement actions — a CFPB supervisory finding against a mid-tier lender using AI in adverse action, an OCC matter against a regional bank's AI-assisted BSA program, and potentially an EEOC or state AG action against an employer using AI hiring tools without disparate impact validation. Each of these will be treated as isolated incidents by the press, but they represent the leading edge of a systematic enforcement posture that will force a compliance cost recalculation across the enterprise AI rollout thesis. The firms that will be hurt most are not the hyperscalers but the SaaS vendors selling AI-embedded workflow tools to regulated industries without having built the auditability, explainability, and bias testing infrastructure that regulators will demand. Valuations for that tier of the AI software stack do not reflect this risk.
The core market mistake is treating AI as a single-theme semiconductor trade rather than a three-stage earnings and cash-flow transmission mechanism: (1) 0-12 months, capex and infrastructure suppliers win; (2) 6-24 months, software and workflow owners convert lower inference cost into gross-margin expansion and seat growth; (3) 12-36 months, labor-heavy service sectors see a measurable change in unit labor costs, pricing power, and wage mix. Quantitatively, the key variable is not model quality alone but cost per useful task completed. If accelerator performance per watt improves ~1.5-2.0x every 12-18 months and model-side efficiency plus distillation/quantization cut inference cost another ~2-4x over the same window, the delivered cost per enterprise AI interaction can plausibly fall 60-85% over two years. That is the threshold where AI usage moves from pilot budgets to operating budgets.
A practical enterprise model: assume a mid-sized firm with 5,000 knowledge workers, 60% addressable for copilots, and 20 AI interactions per worker per day. At $0.08 fully loaded cost per interaction, annual spend is ~$7.5M; at $0.02, it falls to ~$1.9M. If those interactions save only 10 minutes per worker per day for 3,000 workers, annual labor capacity released is ~125,000 hours, equivalent to roughly 60 FTEs at 2,080 hours each. At a fully loaded cost of $120k/FTE, value creation is ~$7.2M annually. At the lower inference cost, ROI rises from roughly 1.0x to nearly 4.0x before second-order benefits. That is why cost compression matters more than benchmark scores for market impact.
Sector translation: software vendors with large installed bases can monetize AI in three ways: attach uplift ($15-60/user/month for copilots), lower service/support cost, and improved retention from workflow lock-in. For a SaaS platform with 10M paid seats, just 8% copilot attach at $25/month yields $240M ARR; 15% attach at $35/month yields $630M ARR. If inference costs drop 70%, gross margin on that AI layer can move from low-40s to 60-75%, materially changing valuation. The market still under-models attach-rate optionality in horizontal software and over-focuses on pure-play model vendors.
Professional services and BPO are the next major P&L transmission channel. In customer service, claims handling, basic legal review, and underwriting support, AI can realistically automate or compress 15-35% of routine cognitive workflow time over 24 months, not by replacing whole jobs immediately but by reducing hours per transaction. In industries where SG&A/labor is 20-40% of revenue, a 3-8% reduction in labor hours can lift EBIT margins by 100-300 bps even after reinvestment. Insurance carriers and brokers are especially exposed because underwriting ops, claims triage, document review, and service centers are text-heavy and rules-rich; a 5% reduction in operating expense on a carrier with a 25% expense ratio improves the combined ratio by ~125 bps, which is highly material to equity value. Banks with large back-office operations can see similar effects: for a universal bank with a 55-60% cost/income ratio, a 2-point reduction from AI-assisted servicing and compliance work can translate into 3-6% EPS upside depending on buyback intensity and credit assumptions.
What coverage misses on semis: investors discuss GPU unit demand but not the shape of the bill of materials. Every $1 of accelerator silicon drives additional spend on high-bandwidth memory, advanced packaging, optical/electrical interconnect, switching, cooling, and power equipment. Depending on architecture and cluster design, non-GPU infrastructure can represent 35-70% of system-level AI capex. That means the more durable trade may not be only in compute but also in memory, networking, data-center electrical gear, and thermal management. If hyperscaler AI capex grows from, for example, ~$150B aggregate toward $220-280B over the next 12-24 months, the incremental pool available to these adjacent suppliers is large enough to support earnings revisions even if GPU lead times normalize. Equity markets are still too concentrated in one or two beneficiaries.
Utilities and power are the most underappreciated cross-asset consequence. A large AI data center campus can require 100-500 MW; the top end is equivalent to a mid-sized industrial load. If US AI-related data-center demand adds even 15-25 GW over 3-5 years, that is a nontrivial share of load growth versus a grid that had expected much flatter demand. Regional implications matter more than national averages: Northern Virginia, Texas, Arizona, the Midwest, and parts of the Pacific Northwest could see tighter reserve margins and materially higher nodal congestion. In regulated utilities, that can support rate-base growth via transmission, substations, and generation interconnection spending. In merchant power markets, scarcity pricing risk rises. Mainstream equity coverage mentions “higher power demand” but rarely translates it into EPS or allowed-return math. A utility able to add $5B of transmission and distribution rate base at a 9.5-10.5% allowed ROE creates roughly $475-525M of pre-tax earnings opportunity over time, highly relevant for valuation.
On labor economics, consensus is also too binary. The first-order effect is not mass layoffs; it is slower hiring and lower outsourcing intensity. That matters because the earnings benefit appears before unemployment does. Software, banks, insurers, consultancies, and large corporates can hold headcount roughly flat while revenue grows, producing operating leverage. If a company growing revenue 6% would otherwise need 4% headcount growth, but AI cuts that to 1%, labor-cost savings can equal 100-250 bps of margin depending on wage inflation. This is especially important in customer care, finance ops, coding support, paralegal work, and compliance review. Articles tend to discuss social disruption but miss the nearer-term accounting reality: lower net hiring can drive earnings upgrades without visible macro labor stress for several quarters.
SME adoption is the ignored TAM unlock. Large enterprises can justify bespoke deployments at higher inference costs; SMEs cannot. If cost per task falls below rough thresholds like $0.01-0.03 for lightweight text workflows and subscription copilot pricing lands in the $10-30/user/month range, penetration can broaden dramatically into businesses with 50-500 employees. That expands the revenue base for cloud and SaaS providers beyond current enterprise-centric forecasts. A simple scenario: 20M addressable SME knowledge workers adopting at $15/month is a $3.6B annual revenue pool before API consumption and implementation services. Consensus numbers often do not reflect this because analysts anchor on Fortune 500 purchasing cycles.
Options market implications: the listed options market in the major AI infrastructure names has generally priced elevated realized volatility and strong upside skew, but the more interesting signal is where implied volatility remains too low relative to potential estimate revisions in the second-derivative beneficiaries. Semis and hyperscale cloud often trade with front-end implied vol already reflecting event risk; software integrators, IT services, utilities with data-center exposure, and selected staffing/BPO names frequently do not. In practical terms, for the crowded AI leaders, call skew often implies the market already pays heavily for upside convexity, reducing expected value of chasing short-dated upside. Better asymmetry may sit in 6-18 month calls or call spreads on companies where AI margin lift is not yet in consensus. Conversely, some labor-intensive service companies may have too little downside skew despite credible risk of 200-500 bps medium-term margin pressure from AI-enabled insourcing or automated delivery.
Specific thresholds to watch because they change valuation frameworks: (1) enterprise inference cost below $0.02 per high-value workflow interaction, which makes broad deployment economic; (2) copilot attach rates above 10% in installed SaaS bases, which starts to matter to revenue estimates; (3) data-center power contracts above 100 MW signed in nontraditional markets, signaling regional utility upside and possible power-price stress; (4) evidence of SG&A per employee flattening while revenue per employee rises, which is the cleanest accounting signature of AI productivity; (5) cloud AI revenue run-rates exceeding associated AI depreciation growth, which indicates the capex cycle is monetizing rather than merely inflating. If those thresholds are crossed, current sector-level earnings estimates are too low for software, selected services, and utilities, while too high for some labor-arbitrage businesses.
The market is also underestimating the duration mismatch between capex burden and software monetization. Hyperscalers can absorb heavy depreciation today because they monetize through a portfolio of cloud services over time; software vendors and enterprise adopters get the benefit later and with less capital intensity. That creates an intertemporal transfer: near-term free cash flow pressure on infrastructure owners, medium-term margin expansion for application owners. If investors only screen current FCF yields, they may miss where estimate revisions emerge next.
Bottom line by instrument: overweight enabling semis beyond just compute GPUs, especially memory/networking/power-thermal supply chains; selectively overweight hyperscalers where AI revenue is beginning to absorb depreciation; increasingly overweight software platforms with installed-base distribution and workflow data; overweight utilities and electrical equipment names tied to data-center interconnection and transmission; underweight labor-intensive outsourcing models with weak proprietary data and low switching costs; be cautious on short-dated upside options in crowded AI leaders because implieds already discount much of the bull case. The better trade is dispersion: long beneficiaries of lower unit labor cost and long regional power-build beneficiaries, against shorts in routine cognitive labor providers and software vendors without either distribution or data moats.
Private signals from GPU allocation calls and energy procurement desks show hyperscale CFOs accelerating 2025-26 power purchase agreements at 2-3x prior volumes while simultaneously signaling to analysts that inference ASP declines will compress their own AI margins faster than modeled. Sell-side notes continue to anchor on visible capex, missing that the same cost curves enabling SME copilots also let large incumbents repatriate workloads to on-prem or colocation, flattening the very cloud revenue curve they forecast. Trader chatter in listed utilities and transformer names has already repriced regional grid tightness through 2027, a move not yet reflected in software or semiconductor multiples that still price perpetual hyperscale demand.
The pervasive narrative of AI accelerating economic shifts is fundamentally accurate, but the market's appreciation for the granular, technical drivers and cascading operational consequences remains nascent. The headline figures for hyperscaler capex and semiconductor revenue obscure the underlying mechanism: a relentless, technically-driven compression of AI inference costs that is rapidly expanding the total addressable market (TAM) beyond enterprise giants into the mid-market and small-to-medium enterprises (SMEs). This is not merely an incremental improvement; it's a structural shift driven by advanced hardware architectures (e.g., Nvidia's Blackwell offering up to 30x inference performance uplift over Hopper for new models, AMD's MI300X competing effectively with H100) and model efficiency techniques (quantization down to FP4, Mixture-of-Experts, compact yet performant models like Llama 3 8B). These advancements translate directly to dramatically lower 'cost-per-token' for running AI, making custom AI copilots and automation economically viable for a broader range of businesses. For instance, self-hosting a performant 8B parameter model on a consumer-grade GPU can yield thousands of tokens per minute for pennies an hour, fundamentally altering the economics of AI adoption compared to API-based pricing for larger models like GPT-4 Turbo ($30/M output tokens). This technical reality implies a much wider, deeper operational transformation than what is often reflected in high-level market cap discussions. The direct consequence is an unprecedented surge in data center power demand, as evidenced by major utilities like Constellation Energy reporting 20-25 GW of new data center load requests through 2030 in their service territories – a demand shock equivalent to adding multiple medium-sized states to the grid, requiring hundreds of billions in new generation and transmission investment. The market underestimates the non-linear impact of this power demand on regional grids, resource availability (especially water for cooling), and the resultant need for dispatchable, reliable power, not just 'clean' power. Simultaneously, the accelerating rollout of AI tools into day-to-day workflows, fueled by these reduced costs, moves labor market shifts from theoretical to tangible. While widespread layoffs are not yet a macroeconomic fact, the 'redeployment risk' for routine cognitive tasks is becoming acute across back-office functions in finance, legal, and customer service. The current discussion often frames this in broad strokes rather than specific impacts on unit labor costs and wage dynamics within particular industries, failing to quantify the true operational productivity dividend and the associated workforce transition challenges.
The documented record supports the core premise that AI capability gains are now constrained less by model novelty and more by infrastructure economics: power delivery, networking, memory bandwidth, and inference efficiency. NVIDIA has publicly emphasized the scaling problem in AI infrastructure and has invested heavily in photonics-related companies to address data-movement and energy constraints, which is a concrete signal that the bottleneck is shifting from raw FLOPS to systems-level efficiency[1]. That matters because it validates a broader industrial pattern: the next wave of AI adoption is not primarily about training the biggest model, but about making inference cheap, reliable, and deployable inside enterprise workflows at scale[1].
The main analytical point is that mainstream coverage often misframes AI as a valuation story when it is also a cost-structure story. Once inference cost falls and model capability rises, the adoption frontier moves from “a few pilot projects in large firms” to “embedded automation across mid-market firms and departments,” which is the economically consequential transition. The cited materials show this in infrastructure terms, but the market narrative usually stops at GPU demand, headline capex, or market caps; it underweights the operational diffusion mechanism that turns infrastructure into productivity gains. The specific undercoverage is not whether AI is important, but how quickly cheap inference can move from discretionary experimentation to standardized enterprise process design.
What can be stated as confirmed fact is narrower than the bull case, but still substantial. Confirmed facts include: large AI vendors are spending billions on infrastructure adjacency such as photonics to relieve AI scaling constraints[1]; enterprise AI use cases are moving toward real-time inference and autonomous systems rather than only generalized computing[3]; and the commercialization opportunity increasingly depends on model compression and deployment efficiency, not only model size[2]. The important inference is that when infrastructure investments and model-efficiency improvements happen simultaneously, the total addressable market expands downward to smaller customers and laterally into more workflows, because the unit economics of inference improve. That inference is consistent with the cited sources, but the precise adoption curve is still unproven.
The strongest market implication is that enabling infrastructure should retain the most direct near-term benefit, while software and services beneficiaries will lag but potentially compound more durably. Semiconductor vendors, optical-interconnect firms, memory suppliers, and cloud platforms are captured early because they sell into the capex cycle and the bottlenecks are still physical[1][3]. The second-order effect is broader: as inference becomes cheaper, firms can redeploy labor rather than simply add headcount, which could pressure routine cognitive labor markets in finance, insurance, legal, customer service, and back-office operations. That labor effect is often missing because analysts discuss AI as an IT budget item rather than as a substitute for workflow labor. The more accurate framing is that AI efficiency gains are a de facto productivity technology with distributional consequences.
What regulatory filings, legislative documents, or institutional reports are directly relevant is best framed around capacity, power, and disclosure obligations rather than around “AI hype” documents. The most relevant institutional record would be utility interconnection filings, regional transmission operator load forecasts, public utility commission proceedings, power purchase agreements, and data-center siting disclosures, because those documents are where incremental AI electricity demand becomes observable and priceable. On the corporate side, 10-K risk factors, capex guidance, and cloud-infrastructure disclosures are the filings most likely to reveal whether AI demand is translating into durable spend. On the labor side, Bureau of Labor Statistics occupational data, productivity releases, and industry output tables are the right institutional series for testing whether unit labor costs and employment elasticities are moving. The prompt’s thesis is therefore more credible when tied to these documentary sources than when inferred from equity narratives alone.
The market is missing three things. First, it underestimates how much inference-cost compression broadens the customer base beyond mega-enterprises; that matters because the largest revenue step-change may come from SMEs and departmental deployments, not just flagship enterprise contracts. Second, it lacks regional power-market accounting: AI data-center load is not a national abstraction, it is a local transmission, generation, and permitting problem that can change utility economics and renewable buildout economics unevenly by geography. Third, it does not yet model labor substitution at the task level, so it misses how the same AI tool can simultaneously raise margins, flatten wage growth in routine occupations, and expand output without proportional hiring.
My view is that the correct base case is not “AI destroys jobs quickly” or “AI boosts GDP broadly and evenly,” but rather a staged diffusion: infrastructure providers capture immediate earnings power; enterprises that redesign workflows capture margin expansion over 6–24 months; and labor-market effects appear selectively where tasks are standardized, high-volume, and text-heavy. The investment implication is that the most durable alpha likely sits in the picks-and-shovels layer and in incumbents with distribution, data, and process integration advantages—not in generic AI branding alone.