In 1945, Dr. Vannevar Bush published his landmark essay As We May Think, envisioning a future where machines would augment human intellect, compressing the vast accumulation of human knowledge into tools that could extend the reach of the mind. Eight decades later, that vision is not merely being realized — it is being surpassed at a pace that would have staggered even Bush’s formidable imagination. As Eric Jang, a leading AI researcher and roboticist, wrote in a sweeping February 2026 essay: “If we consider life to be a sort of open-ended MMO, the game server has just received a major update.” The update he’s referring to is the convergence of AI agents, thinking models, and an exponential surge in inference compute that together represent the most consequential shift in the technology industry since the advent of the internet itself.
The metaphor is apt. In massively multiplayer online games, a server update changes the rules for every player simultaneously — new capabilities emerge, old strategies become obsolete, and the entire economy of the game rebalances. That is precisely what is happening across the global economy as AI systems transition from impressive but passive tools into autonomous agents capable of conducting research, writing code, managing workflows, and even designing new AI systems. The implications for capital allocation, energy infrastructure, labor markets, and scientific discovery are profound, and the speed at which these changes are arriving has caught even seasoned industry observers off guard.
From Chatbots to Agents: The Quiet Revolution in AI Capability
The most important conceptual shift in AI over the past eighteen months has been the move from static language models to what the industry now calls “thinking models” and “AI agents.” A thinking model doesn’t simply pattern-match against its training data; it engages in extended chains of reasoning, breaking complex problems into sub-steps, evaluating its own outputs, and iterating toward better solutions. OpenAI’s o-series models, Google DeepMind’s Gemini with extended thinking, and Anthropic’s Claude with its agentic capabilities all represent variations on this theme. The result is a qualitative leap: models that can not only answer questions but solve multi-step problems that previously required teams of human experts working over days or weeks.
Eric Jang’s essay, published on his personal site evjang.com, provides one of the most detailed technical and philosophical accounts of this transition. Jang argues that the combination of thinking models with agentic scaffolding — the ability for AI systems to use tools, browse the web, execute code, and call other models — has created a fundamentally new kind of intelligence infrastructure. These agents don’t just respond to prompts; they pursue goals over extended time horizons, maintaining context, recovering from errors, and adapting their strategies in real time. The distinction matters enormously. A chatbot is a tool. An agent is a worker. And the economy treats tools and workers very differently.
The Inference Compute Wall: Why the Next Bottleneck Is Already Here
If training large language models was the defining infrastructure challenge of 2023 and 2024, the defining challenge of 2025 and beyond is inference compute — the processing power required to actually run these models at scale as they serve billions of queries, many of which now involve extended chains of thought that consume orders of magnitude more computation than a simple chatbot response. A single complex reasoning query on a thinking model can require 10 to 100 times the compute of a standard prompt completion. Multiply that by the hundreds of millions of users and autonomous agents that are now hitting these systems, and the numbers become staggering.
Andrew McCalip, a hardware and infrastructure commentator, highlighted this dynamic in a post on X, pointing to the enormous capital expenditures being planned by hyperscalers to meet inference demand. Microsoft, Google, Amazon, and Meta have collectively committed hundreds of billions of dollars to data center construction over the next several years, with a significant and growing share of that investment directed specifically at inference workloads rather than training. The shift reflects a fundamental economic reality: the value of AI is realized not when a model is trained but when it is deployed, and deployment at the scale now being contemplated requires an infrastructure buildout rivaling the construction of the electrical grid in the early twentieth century.
Wall Street Takes Notice: The Capital Cycle Accelerates
The financial markets have begun to price in this reality, though opinions differ sharply on whether current valuations adequately reflect the magnitude of the opportunity — or the risks. Nvidia, whose GPUs remain the dominant hardware for both training and inference, has seen its market capitalization fluctuate wildly as investors attempt to model demand curves that are, by any historical standard, unprecedented. But the story extends far beyond a single chipmaker. The entire supply chain — from TSMC’s advanced packaging to the rare earth minerals required for high-performance magnets in cooling systems to the natural gas turbines being installed at data center sites — is being reshaped by the inference compute buildout.
Packy McCormick’s Not Boring newsletter has been tracking these developments with characteristic enthusiasm, framing the current moment as one of the great capital deployment cycles in economic history. McCormick argues that the pessimists who compare AI spending to previous technology bubbles are making a category error: unlike the dot-com era, where revenue models were speculative, AI inference is already generating enormous and rapidly growing revenue for the companies deploying it. Microsoft’s Azure AI revenue, Google Cloud’s AI-driven growth, and Amazon Web Services’ Bedrock platform are all posting numbers that justify continued investment. The question is not whether demand exists but whether supply can be built fast enough to meet it.
Automated Research: When AI Begins to Improve Itself
Perhaps the most consequential — and most debated — development in the current AI cycle is the emergence of automated research capabilities. AI agents are now being used not just to assist human researchers but to conduct independent research workflows: formulating hypotheses, designing experiments, writing and executing code, analyzing results, and drafting papers. Jang’s essay on evjang.com describes this as a critical inflection point, arguing that when AI systems become capable of meaningfully contributing to AI research itself, the feedback loop accelerates in ways that are difficult to model with conventional forecasting tools.
This is not a theoretical concern. Google DeepMind has published results showing AI systems contributing to the design of more efficient neural network architectures. Anthropic has described using Claude to help identify and fix alignment issues in its own models. OpenAI has built internal tools where AI agents assist researchers in running experiments and interpreting results. Sholto Douglas, a researcher at Google DeepMind, noted on X the remarkable pace at which these capabilities are improving, suggesting that the effective research output of top AI labs is being multiplied by the very systems they are building. The recursive nature of this dynamic — AI making AI research faster, which produces better AI, which further accelerates research — is what makes the current moment so qualitatively different from previous technology waves.
The Human Element: Displacement, Augmentation, and the New Division of Labor
Kevin Roose of The New York Times captured the public mood around these developments in a widely shared post on X, reflecting on the speed at which AI capabilities are advancing and the growing unease among knowledge workers about what these advances mean for their careers. Roose has been one of the most thoughtful mainstream journalists covering AI’s impact on work, and his observations resonate because they reflect a tension that is playing out in offices, newsrooms, law firms, and research labs around the world: the tools are getting better faster than most people’s ability to adapt to them.
The data supports the anxiety. Coding assistants like GitHub Copilot and Cursor are now writing a significant percentage of new code at many software companies. Legal AI tools are drafting contracts and conducting due diligence at speeds that would have been unthinkable two years ago. Financial analysts are using AI agents to build models, scrape filings, and generate investment memos. In each case, the pattern is the same: tasks that once required hours of skilled human labor are being compressed into minutes. This does not necessarily mean mass unemployment — historical technology transitions have generally created more jobs than they destroyed — but it does mean a profound and rapid restructuring of how work is organized, who does it, and what skills command a premium.
Energy, Infrastructure, and the Physical Constraints on Digital Ambition
The AI inference buildout is increasingly bumping up against physical constraints that no amount of software innovation can fully overcome. Data centers require enormous quantities of electricity, water for cooling, and physical space. In many regions, the limiting factor on AI deployment is not the availability of chips but the availability of power. Utilities across the United States are reporting unprecedented demand growth after two decades of essentially flat electricity consumption. New natural gas plants, nuclear restarts, and massive solar and battery installations are all being fast-tracked to meet the demand, but permitting, construction, and grid interconnection timelines remain stubbornly long.
Andrew McCalip’s commentary on X has been particularly incisive on this point, noting that the physical infrastructure required for the inference era dwarfs what was needed for the training era. Training a frontier model is a discrete project: you assemble a cluster, run the training job, and move on. Inference, by contrast, is a continuous, always-on workload that scales with the number of users and agents. As AI agents become more prevalent and their reasoning chains grow longer and more complex, the compute — and therefore the energy — required per unit of economic output will continue to climb. This creates a multi-decade investment cycle in energy infrastructure that will have profound implications for utilities, grid operators, real estate developers, and policymakers.
The Geopolitical Dimension: Compute as Strategic Asset
The concentration of advanced AI compute in a handful of countries — primarily the United States, with critical supply chain dependencies on Taiwan, South Korea, the Netherlands, and Japan — has turned inference capacity into a geopolitical asset of the first order. The U.S. government’s export controls on advanced chips to China, the CHIPS Act subsidies for domestic semiconductor manufacturing, and the growing diplomatic efforts to secure supply chains for critical minerals all reflect a recognition that compute is becoming as strategically important as oil was in the twentieth century.
This geopolitical dimension adds another layer of complexity to the investment calculus. Companies building data centers must consider not only the economics of power and real estate but also the regulatory environment, the security of their supply chains, and the potential for government intervention. The AI inference buildout is not just a corporate capital expenditure cycle; it is a national security priority, an energy policy challenge, and an industrial policy experiment all rolled into one. As Jang observed in his essay on evjang.com, the scale of the transformation underway is difficult to overstate — and we are still in the early innings.
What Comes Next: The Open Questions That Will Define the Decade
Several critical uncertainties will determine how the next phase of the AI revolution unfolds. First, will the scaling laws that have driven progress in AI capabilities continue to hold, or will diminishing returns set in? The evidence so far suggests that both training compute and inference-time compute continue to yield improvements, but the curve could flatten. Second, will the energy and infrastructure buildout keep pace with demand, or will physical constraints impose a binding ceiling on AI deployment? Third, how will labor markets adjust? The speed of AI capability improvement is outpacing the speed of institutional adaptation — in education, in corporate management, in regulation — and the gap between what AI can do and what organizations are prepared to let it do is one of the most important variables in the near-term economic outlook.
Finally, there is the question that Vannevar Bush himself would have found most compelling: what happens when these tools are used not just to automate existing work but to enable genuinely new forms of discovery and creation? The early evidence from automated research, from AI-assisted drug discovery, from materials science breakthroughs accelerated by machine learning, suggests that the augmentation of human intellect that Bush envisioned is not just possible but already underway. The game server has indeed received a major update. The players — all of us — are still figuring out the new rules. But the pace of play has accelerated beyond anything the previous generation of technologists imagined, and the stakes, measured in trillions of dollars of economic value and potentially transformative scientific breakthroughs, have never been higher.
WebProNews is an iEntry Publication