For three years, the dominant AI infrastructure story was training: build the largest model, consume the most compute, acquire the most GPUs. That story produced a straightforward investment thesis. NVIDIA won. The hyperscalers spent. Colocation providers filled their data centers with power-hungry training clusters. The math was simple.
GTC 2026 is NVIDIA’s formal argument that the math has changed.
According to Data Centre News, NVIDIA is now projecting a potential $1 trillion opportunity in AI systems over the next few years, driven not by additional training compute, but by the deployment of AI into production at scale. The shift has a name in NVIDIA’s vocabulary: the move from training to inference, agentic AI, and what the company calls “AI factories.” The NVIDIA newsroom’s GTC 2026 coverage confirms that NVIDIA has put this positioning into production, with Dynamo, its inference operating system, entering general availability as the flagship product of this new phase.
This is a vendor projection, not a market consensus. But NVIDIA’s projections carry weight because they’ve historically preceded observable market movement. The question worth asking isn’t whether to believe the number. It’s what has to be true for the number to be correct, and who captures value if it is.
What changed: from training bursts to persistent inference
Training a large model is episodic. It happens in concentrated compute bursts, weeks or months of intensive GPU utilization, then a relatively quiet period before the next version. Infrastructure built for training can sit at lower utilization between runs.
Inference is the opposite. Once an AI model is deployed in production, serving users, powering agents, running automated workflows, it generates continuous demand. Every query, every agent action, every automated decision requires compute. The workload doesn’t end; it scales with usage.
That difference changes how infrastructure gets purchased and operated. Training clusters are built for peak capacity over a known period. Inference infrastructure is built for continuous, growing throughput. The buying decision looks more like a long-term utility contract than a capital equipment purchase. And the technical requirements are different: inference workloads prioritize latency and throughput efficiency over raw peak compute.
NVIDIA’s Blackwell and Vera Rubin platforms, cited by the company as demand drivers through the coming years, are designed with this inference-at-scale use case in mind. These are NVIDIA’s stated order-volume signals, and while the platform-specific pipeline figures weren’t independently confirmed in sources available for this brief, the platforms themselves are established GTC 2026 announcements.
The competitive landscape: why inference is contested in ways training wasn’t
NVIDIA’s near-monopoly on training compute was a function of software lock-in as much as hardware advantage. The CUDA ecosystem, the software stack that developers build on top of NVIDIA GPUs, created switching costs that custom silicon alternatives struggled to overcome. Training on anything other than NVIDIA hardware required re-engineering the software stack, and most organizations weren’t willing to pay that price.
Inference is a more open competition. The workloads are more varied. The latency requirements differ by application. And the custom silicon providers, Amazon’s Trainium and Inferentia chips, Google’s TPUs, Meta’s MTIA, have been designed specifically to compete at the inference layer, where the software lock-in is less entrenched. These providers have deployed their custom silicon at scale in their own production environments, which means they have real-world inference efficiency data that training-focused competitors don’t.
NVIDIA’s GTC 2026 positioning, and Dynamo’s release as an inference-specific operating system, is a direct response to this competitive pressure. The company is arguing that its inference stack, hardware plus software together, can match or exceed custom silicon efficiency at the scale that matters for production deployments.
Who captures value beyond NVIDIA
The training-era investment thesis was concentrated: NVIDIA, the hyperscalers, and the power and cooling infrastructure serving training clusters. The inference-era thesis is more distributed.
Consider the infrastructure chain for persistent inference at scale. Power and cooling infrastructure doesn’t go away, inference clusters generate heat continuously rather than in bursts, which may actually increase total thermal management demand. Colocation operators serving hyperscaler inference deployments benefit from longer-duration, more predictable contracts than training-phase tenants. Networking infrastructure, specifically high-bandwidth, low-latency interconnects, becomes more critical when inference workloads need to distribute across clusters in real time.
The context for this is significant. The Ohio data center campus backed by DOE and SoftBank, a 10-gigawatt facility representing $33.3 billion in confirmed funding, and Samsung’s reported $73 billion commitment to AI chip manufacturing in 2026 both reflect infrastructure investment decisions made in anticipation of sustained, growing AI compute demand. The inference pivot NVIDIA is describing at GTC 2026 validates the demand assumption underlying those commitments. Whether inference demand specifically materializes at the scale required to fill that capacity is the open question.
Beyond physical infrastructure, the inference era benefits a different software layer: orchestration, monitoring, and optimization tools that manage inference workloads at scale. In the training era, model developers were the primary software buyers. In the inference era, the buyers are operations teams managing production AI systems. Their tooling needs are different, and the market for inference-layer software is early.
The agentic AI multiplier
NVIDIA’s “AI factories” concept is most relevant when applied to agentic AI systems, AI that doesn’t just respond to single queries but executes multi-step tasks, uses tools, calls external APIs, and operates over extended sessions. Agentic workflows generate disproportionate inference demand because each step in an agent’s reasoning chain requires a compute call. A single user interaction that triggers a ten-step agentic workflow consumes ten times the inference compute of a single-query response.
If agentic AI adoption follows the trajectory that enterprise software adoption typically does, slow initial uptake, followed by rapid expansion once the first wave of deployments proves ROI, the inference demand multiplier kicks in with meaningful lag. The compute build-out happens before the demand is fully observable. That’s the gap NVIDIA is betting on: that infrastructure investment will precede adoption, and that the companies that build inference capacity now will be positioned to serve demand when it arrives.
What to watch
Hyperscaler earnings calls in the next two quarters are the most informative leading indicator. Meta, Microsoft, and Google have all signaled substantial AI infrastructure capex, Meta’s guidance for 2025 ranged in the $115-135 billion area, reflecting the scale of hyperscaler commitment to this phase. Whether that capital is being directed toward training or inference infrastructure will clarify whether NVIDIA’s GTC thesis is tracking with where customers are actually spending. Inference-specific capex line items in earnings commentary would be a meaningful signal.
Custom silicon performance data from Amazon, Google, and Meta, actual inference efficiency benchmarks at production scale, will also begin to surface as these providers publish more about their internal deployments. NVIDIA’s competitive claim rests on the assertion that Dynamo plus Blackwell/Vera Rubin outperforms custom silicon at the workloads that matter. Third-party benchmark data will test that claim.
Finally, enterprise adoption of agentic AI will be the demand-side indicator that matters most over a 12-24 month horizon. Enterprise software adoption typically trails infrastructure investment. When agentic deployments move from pilot to production at major enterprises, the inference demand curve steepens. That’s when the $1 trillion projection gets tested against observable market behavior.
TJS synthesis
NVIDIA’s GTC 2026 isn’t a product announcement dressed up as a market vision. It’s a structural argument about where value accrues in the AI economy as deployment overtakes development. The training era concentrated value. The inference era distributes it, across hardware, software, power infrastructure, networking, and orchestration, while still requiring NVIDIA to defend its position against custom silicon alternatives that are purpose-built for this phase. The $1 trillion figure is a bet size, not a guarantee. But the underlying shift it describes, from building AI to running AI, is already underway. The organizations that understand inference infrastructure as a distinct investment category, not just an extension of training infrastructure, are the ones positioned to allocate capital correctly in what comes next.