For the past three years, the AI industry's obsession has been training: throwing vast quantities of computing power at raw data to build ever-larger models.
Nvidia Corp's (NASDAQ:NVDA, XETRA:NVD) graphics processors dominated that phase so completely that the company briefly became the world's most valuable business. But training is maturing. The frontier labs have their models. Now comes the harder, more commercially urgent challenge of actually running them at scale, and that is where Nvidia's next chapter is being written.
At its annual GTC developer conference in San Jose on Sunday, CEO Jensen Huang raised the company's estimate of the addressable revenue opportunity for its AI chips to at least $1 trillion through 2027.
That figure, up from the $500 billion forecast Nvidia gave for its Blackwell and Rubin chips just last month, signals something more significant than bullish marketing. It reflects a genuine restructuring of where money is flowing across the industry.
The inference bet takes shape
The announcement that drew most attention was a new AI system built on technology from Groq, the chip startup from which Nvidia licensed intellectual property for $17 billion in December. Huang also unveiled a new central processor, the Vera CPU, marking a push into territory long dominated by Intel.
The architecture Huang described splits inference, the process by which an AI system answers a query or completes a task, into two distinct stages. Nvidia's Vera Rubin chips handle the first step, called prefill, which converts a user's request into the numerical tokens that AI systems process internally. Groq's chips then take over for the decode stage, generating the actual response.
The division of labour matters because each step has different computational demands. Prefill is intensive and parallel; decode is sequential and latency-sensitive. By pairing its own hardware with Groq's specialised architecture, Nvidia is arguing it can optimise for both, rather than asking a single chip design to compromise on either.
Why inference has become the battleground
The timing of this pivot reflects a shift across the industry's biggest spenders. Companies such as OpenAI, Anthropic and Meta have spent hundreds of billions of dollars building and training their models. Their focus is now shifting toward serving the hundreds of millions of users who are querying those systems daily.
That shift changes the competitive map. Training was a market Nvidia effectively owned, with its H100 and A100 GPUs becoming the default infrastructure for every major lab. Inference is more contested. Central processing units, which Intel dominates, are increasingly viable for deploying AI models.
Google and other hyperscalers have invested heavily in custom silicon designed specifically for serving workloads. Nvidia's margin advantage is narrower here.
Huang acknowledged the CPU opportunity directly. "We are selling a lot of CPU standalone," he said, describing the Vera CPU as already certain to become a multi-billion-dollar business. That statement would have seemed unlikely from an Nvidia CEO even 18 months ago.
A roadmap extending to 2028
Beyond the immediate product announcements, Huang outlined a forward architecture called Feynman, expected in 2028 and following the company's Rubin Ultra chips. Details were sparse, but the roadmap communicates continuity: Nvidia is signalling to hyperscalers and enterprise customers alike that its technology generation cycle will remain reliable enough to plan around.
The company also introduced NemoClaw, a tool targeting the market for autonomous AI agents. It integrates with the OpenClaw platform and adds privacy and safety controls to agent systems capable of executing tasks with limited human oversight.
Bob O'Donnell of Technalysis Research captured the broader shift in Nvidia's product presentation: "He used to come out with a new GPU chip and say, look, here's my new chip. Now he's got five racks of equipment that make up these systems." That evolution, from component supplier to systems architect, is central to how Nvidia is trying to defend its position as competition intensifies.
What the market made of it
Nvidia shares closed up around 1.2% on the day, having briefly spiked higher before retreating. The muted reaction reflects a tension that has been building around the stock since it hit a $5 trillion valuation last October. Investors are weighing Huang's vision of durable, expanding demand against questions about whether the company's practice of reinvesting profits back into the AI ecosystem will generate sustainable returns.
Analysts argued the $1 trillion forecast addresses those doubts directly, describing it as evidence that Nvidia is sustaining leadership as the AI industry matures beyond early experimentation into large-scale deployment.
The more substantive question is whether Nvidia can hold that leadership in inference as effectively as it did in training. GPU's dominance in training was, in part, a function of timing and ecosystem lock-in.
The CUDA software platform created switching costs that made alternatives difficult to adopt, even when they existed.
Inference may prove more amenable to competition precisely because workloads are more varied, latency requirements differ by application, and the cost pressures on deploying AI at consumer scale are more acute than those on training runs done in private data centres.
Huang's answer to that challenge, judging by GTC, is to move up the stack. Rather than selling chips, Nvidia is selling systems, software and roadmaps. Whether that is sufficient to hold the line will become clearer as inference spending accelerates through 2026 and 2027.