Jensen Huang Claims AI Has Crossed the Commercialization Turning Point
NVIDIA CEO Jensen Huang stated that AI has reached a turning point: transitioning from experimental technology to systems capable of performing productive, valuable, and profitable tasks. His core judgment is that agent-based AI allows inference computing to directly generate revenue and profit.
Huang referred to this change as the "inference turning point." Previously, AI investments were primarily focused on training model capabilities; now, businesses need to continuously purchase computing power for online inference, code agents, customer service, research analysis, content generation, and automated workflows, as each model invocation may correspond to billable tasks or productivity improvements.
He summarized this narrative as "In AI, compute is revenue": AI chips, servers, data centers, and electricity are no longer just IT costs but are defined as revenue-generating "AI factory" assets. NVIDIA is attempting to drive long-term capital investment in Blackwell, Vera Rubin, and related networks and systems based on this logic.
NVIDIA has indicated that its next-generation chip platforms and systems could bring at least $1 trillion in revenue opportunities by 2027; this figure reflects the company's expectations for market demand and platform sales prospects, not signed or realized revenue.
"AI is profitable" should not be interpreted as all enterprise AI projects generating positive returns. Model providers, cloud vendors, and chip companies can benefit from GPU sales, cloud inference, and subscriptions, but this does not mean that all traditional enterprises adopting AI have converted AI into quantifiable profits; the market is still debating whether AI capital expenditures can generate terminal revenue that matches the scale of investment.
From a market mechanism perspective, if enterprises continue to shift from pilot projects to production deployment, budgets will move from experimental SaaS, consulting, and short-term POCs to GPUs, cloud inference, data centers, electricity, model security, data governance, and agent integration. NVIDIA, cloud service providers, network equipment, HBM memory, liquid cooling, and power generation infrastructure will benefit; application layer companies that cannot prove AI can reduce costs, increase revenue, or shorten delivery cycles will face stricter ROI scrutiny.
Source: Public Information
ABAB AI Insight
Huang's concept of the "turning point" essentially signifies the shift of AI from training economics to inference economics. Training involves one-time or periodic large capital expenditures, while inference is a continuous consumption of computing power occurring with each user, each agent task, and each tool invocation. As long as agents can handle economically valuable work, the volume of inference is no longer just a cost but can be monetized through subscriptions, APIs, task fees, or labor time savings.
Thus, the capital path changes: model companies and enterprise clients no longer just purchase a round of training clusters but need to secure long-term online capacity, GPUs, network bandwidth, low-latency storage, and electricity. The $45 billion, 460 MW long-term computing power agreements between Anthropic and Nscale illustrate that leading model companies are converting future inference demand into multi-year infrastructure commitments; however, such deals also shift business risk from "can the model be trained successfully" to "can the actual task volume fill the reserved data center."
Historically, this mirrors the evolution of cloud computing from "renting servers" to the on-demand infrastructure market of AWS, Azure, and Google Cloud. Initially, cloud vendors sold elastic computing, later locking in enterprise workflows through databases, storage, security, and development tools; AI factories aim to turn inference capabilities into the next layer of general infrastructure. The difference is that AI inference costs are higher, demand is more volatile, and model capabilities change more rapidly, thus concentrating the risks of hardware depreciation and capacity utilization.
This represents a concentration of capital. Once inference becomes a long-term revenue source, the winners will not only be companies with good models but also those that can continuously provide large-scale inference at the lowest cost. The mechanism is that model capabilities generate task demand, task demand leads to token consumption, and token consumption supports GPU and data center investments; scaling up further reduces unit inference costs. NVIDIA aims to occupy the hardware and system layer of this cycle, but whether the cycle holds depends on whether end enterprises are willing to continuously pay for the work done by AI.
ABAB News · Cognitive Laws
- The true turning point for AI is not the ability to answer questions but the ability to continuously complete paid work.
- Training creates capabilities, inference determines cash flow.
- When computing power becomes revenue, chips transform from equipment into production materials.