AI’s Second Act
From buildout to outcomes
Jensen, the Spotlight, and Narrative Shifts
If there were a Mount Rushmore of modern technology leaders, Jensen Huang would stand alongside figures like Steve Jobs and Elon Musk. While long respected within the industry, Jensen’s prominence has risen dramatically alongside Nvidia’s extraordinary growth. He has entered a rare tier of cultural visibility where first names alone suffice, a status underscored by his recent appearance on one of the largest podcast platforms in the world, The Joe Rogan Experience.
That visibility, however, carries a risk of overexposure. Jensen’s frequent presence across earnings calls, conferences, podcasts, and media interviews provides transparency into Nvidia’s vision and business fundamentals, but it can also create the impression of a repetitive, hype-driven AI narrative at a time when investors and the public are eager for concrete evidence of measurable outcomes and impact. When specifics are scarce, skepticism fills the void, fueling concerns about overinvestment in AI infrastructure and the durability of returns on investment (ROI).
In a market environment focused on AI monetization and use cases, familiar talking points and speculative forecasts lose persuasive power without granular, quantified examples. A useful contrast comes from Sheryl Sandberg’s tenure as Meta’s COO, where in public appearances she consistently highlighted micro-level examples of small businesses using Meta’s ad tools and directly tied those campaigns to revenue growth. As the AI market evolves, investors are demanding a similar level of specificity.
Against this backdrop, at the center of investor discussions is whether the surge in AI infrastructure spend is a harbinger of a speculative “AI bubble.” This is a loaded phrase that often reflects uncertainty around timing and current capabilities rather than skepticism about AI’s long-term potential. Much of the debate has focused on speculative technological timelines, obscuring a more practical question: when and how will AI deliver broad based measurable economic value?
That question should come into sharper focus in 2026, as the news flow delivers more signal and less noise. We see a strong likelihood of a wave of evidence emerging that demonstrates AI agent and application use cases capable of tackling more complex, high-impact tasks, shifting attention away from AI vendors and infrastructure spenders toward enterprises, particularly outside the tech sector, that deploy AI agents with measurable results.
These deployments will reflect a pragmatic balance between system autonomy and acceptable fault tolerance, a relationship that will evolve toward greater autonomy as system capabilities advance, and error rates decline. This shift would represent an inflection point where vertical specific, micro-level deployments replace broad narratives and bring into focus AI’s near-term addressable market and reveal a new cohort of AI winners.
The Path Forward: Vertical, Domain-Specific AI
The current state of AI technology suggests that more compute power and more data alone are insufficient to overcome foundational technical challenges. While future breakthroughs may require an architecture beyond today’s large language model (LLM) paradigm, the clearest path to near-term, scalable deployment lies in domain specific applications. When paired with proprietary, first-party data, these systems can drive new AI enabled revenue opportunities and deliver productivity gains by compressing workflows and automating high-friction tasks, materially increasing operating leverage and revenue per employee. However, reliability, interoperability, and security concerns mean that most early deployments will remain within closed or controlled environments, with humans still in the loop.
Within these deployment constraints, AI’s near-term impact on the labor market is best understood as a task-level transformation rather than broad-based displacement. While roles centered on information retrieval, synthesis, and review do face genuine risk, early on most AI applications will function as force multipliers, widening performance gaps between organizations that effectively integrate AI tools and those that do not. In the near term rather than outright replacement, AI is more likely to act as a headwind to new employment growth by reshaping work and scaling human productivity.
These dynamics are already visible in practice.
The leading digital mortgage platform Rocket Companies provides a clear example. Following its 2025 acquisitions of Redfin and mortgage servicer Mr. Cooper, the company has deployed AI agents to drive cost efficiencies and support revenue growth. Management recently highlighted three agents: a pipeline manager, a purchase review agent, and an underwriting agent.
The pipeline agent prioritizes leads and customizes outreach and has driven a nine-point increase in client follow-up rates and a 10% lift in conversion for refinance applications. The purchase review agent reduced processing time by 80% across a workflow that previously required approximately 80 manual steps, exceeded the accuracy of the legacy process, and delivered more than 150,000 team member hours saved annually. The underwriting agent cuts task time from four hours to about 15 minutes by automating document verification, confirming eligibility, and task summarization.
Together, these deployments illustrate how AI agents operating in a closed environment with access to first-party data can remove layers of inefficiency, lower costs, improve margins, and enhance the customer experience, creating a stronger foundation for future revenue gains.
C.H. Robinson Worldwide, a freight and logistics company, is another good example of deploying AI agents in a vertical specific domain environment to drive greater operational efficiency and expand revenue opportunities. At a recent investor conference, management highlighted its use of AI agents to facilitate the automation of its freight quotation process. Previously, the company would respond to about 65% of those quote requests yet now, with the use of AI agents is able to respond to 100% of those requests in about 30 seconds and win more business.
Physical AI is also advancing fastest in fixed, defined environments with Amazon establishing a leadership position in physical AI and robotics. Amazon’s long-running automation of its fulfillment network reduced average staffing per facility from nearly 1,000 to roughly 670, and increased packages that Amazon ships itself per employee each year from about 175 in 2015 to around 3,870.
This evolution underscores a central theme of the next AI phase: durable value creation will come not from abstract intelligence claims, but from tightly integrated systems that measurably improve real-world productivity.
AI Divergent Trends: Adoption vs. Infrastructure Cyclicality
Even if 2026 brings a wave of AI agent deployments, that outcome does not eliminate the risk of a cyclical slowdown in AI infrastructure spending. At some point, AI compute capacity will exceed demand; it is not a question of if, but when. All shortages eventually give way to surpluses.
Importantly, a cyclical surplus in AI compute capacity can coexist with accelerating AI adoption. Falling compute costs lower barriers to deployment, while gains in resource efficiency and model capability expand viable use cases.
Timing an AI capex slowdown is challenging because effective compute supply expands through more than just new data centers. Advances in model efficiency, system optimization, GPU utilization, and algorithms all increase available compute. While these forces support adoption, they also expand effective capacity, obscuring near-term demand signals for GPUs and data center investment.
Although these efficiency gains are positive over the long term, they introduce timing risk. The shift from training heavy to inference-dominated workloads is unlikely to be linear. If infrastructure buildouts outpace realized demand, cyclical oversupply becomes inevitable, a risk amplified as a growing share of inference shifts to the edge and personal devices, bypassing centralized data centers and falling outside current capex assumptions.
A slowdown would pressure not only AI “picks-and-shovels” suppliers but also emerging data center operators with weaker balance sheets. A surplus driven faster than expected decline in AI compute pricing and lower capacity utilization would disproportionately impact highly leveraged players with Oracle and neoclouds, AI native cloud service providers such as CoreWeave, among the most exposed.
There are already signs that capital markets may be approaching the limits of their capacity to fund aggressive neocloud expansion. CoreWeave cut its 2025 capex guidance to $12-14 billion from $20-23 billion, yet still projects 2026 capex to be more than double 2025 levels. Free cash flow is expected to remain deeply negative in 2026 following an estimated rate of cash burn in 2025 of $12 billion, pushing year end 2025 gross debt toward $20 billion, up sharply from about $8 billion in 2024.
This financial profile underscores CoreWeave’s dependence on continued investor support, with the viability of the neocloud model hinging more on capital availability than underlying AI demand. While management cited delays in the delivery from a third-party developer of a “powered shell” for one of its data centers as the cause for its reduced capex guidance, this explanation appears incomplete, and other contributing factors cannot be ruled out.
CoreWeave management explained that the 2025 capex short fall will be mostly reflected in a corresponding increase in construction in progress, representing expenditures on infrastructure not yet in service. However, a review of the company’s financials suggests that while construction in progress will have increased in the last quarter of 2025, even after accounting for this increase, CoreWeave’s “effective” capex will still reflect a meaningful cut from previous guidance. Prior capex plans may have overestimated capital markets' capacity to finance such rapid balance sheet expansion for an unproven, loss-making business, making capital constraints a possible contributor to the cut in capex.
Viewed through this lens, neoclouds increasingly resemble late 1990s, early 2000s telecom overbuilders, debt-driven expansions scaling ahead of demand raising doubts about the durability of this source of AI compute infrastructure demand. While any industry-wide AI capex correction would pressure infrastructure suppliers, incumbent hyperscalers are far better positioned, leveraging scale, financial strength, and partnerships with AI leaders. In such an environment Meta, Amazon, and Google would be set to sustain strong cloud growth while moderating capex, a dynamic the market typically rewards as free cash flow growth accelerates amid reduced competitive intensity as neoclouds struggle to meet their debt obligations.
Unpacking AI ROIC
At a time of growing uncertainty over whether AI spending is approaching a cyclical peak and rising concern about the durability of infrastructure led growth, two questions anchor any assessment of the sustainability of AI capex growth and the scope and timing of AI’s near-term economic value creation. First, what returns on invested capital (ROIC) can AI compute infrastructure generate, and what does that imply for the sustainability of current spending levels? Second, given the state of AI technology, what applications are commercially viable today, and how do current technical limits shape where AI can and cannot deliver a measurable impact?
ROIC sits at the center of the AI debate. Some point to incumbents such as Meta, Google, and Amazon (AWS) as evidence that AI investments are already paying off, citing rising ROIC alongside surging AI capex. On closer inspection, however, much of this improvement reflects a transitory “sugar high,” masking a deterioration in marginal ROIC that reflects rising capital intensity and the computational demands of large language models.
Critically, the widely cited ROIC gains in 2023 and 2024 rely on a favorable and arguably cherry-picked comparison period. Meta and Amazon entered 2023 after a cyclically depressed 2022, making subsequent ROIC improvements appear more dramatic than they truly are. Much of the recovery was driven by aggressive operating expense reductions and rationalization rather than AI-driven productivity gains. From 2019 to 2022, Meta’s revenue increased 65%, but total costs and expenses rose 87%, while Amazon’s revenue grew 83% against an 112% increase operating expenses. The reversal of excess COVID-era spending, not AI, was therefore a primary driver of recent ROIC improvement, overstating AI’s true economic contribution.
ROIC headwinds are even more apparent at software-centric hyperscalers Microsoft and Oracle, where returns face persistent headwinds as capital-intensive infrastructure increasingly dilutes the contribution from higher-margin software revenues.
While incumbent cloud players may be accelerating revenue growth, they are doing so at the expense of free cash flow. Free cash flow returns on capital are deteriorating as ever greater amounts of capital are required to generate each incremental dollar of revenue. As a result, the cost structure of AI compute services, specifically cost per token, will become a defining competitive advantage in an increasingly computation intensive market.
Complicating matters further, a meaningful share of current hyperscaler AI compute demand is driven by loss-making AI leaders and VC-funded startups with elevated cash-burn rates, raising questions about the quality and durability of that demand. OpenAI and Anthropic alone account for a material portion of global AI compute consumption, with these loss-making enterprises’ spending effectively subsidizing hyperscaler returns. While this pattern is common in the early phases of major technological transitions, a durable AI ecosystem will ultimately require broader enterprise and consumer adoption supported by economically sustainable returns.
Taken together, this does not imply that AI infrastructure investments are value destructive. Rather, headline ROIC figures overstate early benefits, as recent improvements were heavily influenced by cyclical recovery and cost normalization. In the end, the AI investment cycle will be judged not by how much capital is deployed, but by whether returns, demand quality, cost discipline, and product innovation can keep pace with unprecedented capital intensity.
Defining the Boundaries of Near-Term AI Capability
The opportunity for AI-driven applications is enormous, but expectations are often inflated by imprecise language from industry leaders, particularly around concepts such as “reasoning” and “thinking.” Today’s large language models (LLM) do not reason or think in a human sense. They generate outputs by statistically predicting the next token based on learned patterns in data, rather than by understanding meaning, causality, or intent. What may appear to be reasoning is largely the reproduction of linguistic patterns that resemble logical steps and correlations learned during training, not true comprehension, abstraction, or the generation of genuinely novel ideas.
LLMs also continue to struggle with generalization, the ability to reliably apply knowledge learned during training to novel, previously unseen situations. This limitation is structural and reflects several persistent challenges. Current models struggle to support continual learning, making it difficult to incorporate new information over time without degrading performance on previously learned tasks. This challenge is closely related to catastrophic forgetting, in which neural networks, including LLMs, overwrites existing knowledge as new data is introduced. In addition, LLMs are also prone to overfitting, performing well on data that closely matches their training distribution, but failing to generalize when encountering previously unseen out of distribution data. Causal confusion remains another constraint, as models frequently conflate correlation with causation, leading to errors when underlying conditions change.
These constraints do not undermine AI’s long-term potential, but they do define the boundaries of near-term commercial viability. A useful way to understand these limits is through the distinction between closed-domain and open-domain problems.
Closed-domain problems are more structured and comparatively tractable. They involve a well-bounded set of rules, scenarios, and solutions. A classic example is chess, where the rules are fixed, and the space of possible moves is well defined. A commercial example is Netflix’s recommendation system, which operates within the constrained universe of its content catalog and user interaction data. In such environments, abundant data and well-defined objectives allow AI systems to excel through sophisticated statistical approximation and pattern recognition.
In contrast, open-domain problems are defined by their breadth and uncertainty. The effectively unbounded range of possible scenarios makes these problems far more difficult for current AI systems to handle reliably or sustain consistent performance. Addressing open-domain challenges requires capabilities beyond statistical function approximation, including robust generalization, contextual awareness, and adaptive behavior in response to novel and previously unseen situations and data.
Driving exemplifies an open-domain challenge characterized by an infinite long tail of edge cases and unknown conditions. Real-world operation spans a constantly shifting set of circumstances, from variable weather and road conditions to unforeseen obstacles and the inherently uncertain behavior of other drivers. Addressing this level of complexity extends well beyond simple patter recognition, requiring capabilities associated with general intelligence, including abstract reasoning, adaptive decision making, and a nuanced understanding of physical and environmental dynamics.
Some AI system architectures erode the distinction between closed and open-domain problem-solving, taking a hybrid approach. Waymo’s autonomous vehicle tech stack has overcome some of the inherent challenges of solving an open-ended problem by combining modular, domain-specific deep learning models with end-to-end learning techniques in a compound AI architecture. Within a defined operational domain, this hybrid architecture has demonstrated stronger real-world performance and lower disengagement rate than end-to-end centric systems, highlighting the practical advantages of this approach.
The key implication is that near-term AI winners will be defined less by abstract notions of “general” or “super” intelligence and more by disciplined alignment between model capability and problem structure. In the immediate future, the most commercially viable deployments will be vertical specific applications built on first-party data and designed to deliver measurable outcomes, rather than open-ended systems that generate unacceptable error rates and exceed the limits of today’s technology.
While 2026 will not resolve every debate about AI’s ultimate trajectory, it should bring greater clarity on where real value is being created, who is capturing it, and which parts of the ecosystem are built on sustainable foundations.
- Stan Shpetner
Research Analyst
Hurricane Capital
If you would like to discuss your evolving portfolio needs, please connect with us here:
mailto:info@hurricanecap.com
Not investment advice: “This commentary is for informational purposes only and does not constitute investment advice, an offer, or a solicitation to buy or sell any security.”
Opinions & forward-looking: “Statements reflect the author’s opinions as of the date published and are subject to change. Forward-looking statements are inherently uncertain; actual outcomes may differ materially.”
Sources & verification: “Company metrics cited herein are based on public statements by company management and/or public filings; the author has not independently verified all third-party data.”
No performance claims: “References to companies are for illustrative purposes and are not indicative of future performance.”
Conflicts (if applicable): “The author/firm may hold positions in securities referenced.” (Only include if true—otherwise omit or add “does not hold positions” if that’s accurate and you’re comfortable stating it.)