Nio has increased its investment in AI training compute roughly fivefold this year and plans an even larger wave of spending, founder and CEO William Li told reporters in Beijing on Thursday, a day after the launch of the company’s ES9 flagship SUV.
The step-up marks a sharp turn from a deliberately restrained 2025, when Li said the company spent only about a tenth of what rivals put into training their assisted-driving systems.
Li framed last year’s caution as a strategic choice rather than a budget constraint, arguing that heavy spending made little sense before the underlying architecture was complete.
“Last year, most of our effort went into validating and building the entire world-model and closed-loop reinforcement learning framework,” Li said before comparing the investment to Nio’s rivals.
“Our investment in training compute last year was actually quite small,” he stated. “Based on our understanding, it was probably only around one-tenth of what competitors spent. This wasn’t simply about cost control.”
The Framework First, Then the Spend
Nio‘s assisted-driving system, the Nio World Model, pairs a world model with closed-loop reinforcement learning, an approach Li said the company spent last year validating and building rather than scaling.
With the first version now in users’ hands, he said the architecture had demonstrated its potential even on relatively little training data, justifying the larger investment now underway.
“At that stage, we believed massive investment wasn’t useful yet because the model architecture and framework weren’t fully established,” Nio’s founder told reporters. “Throwing in huge amounts of computing power too early would be like adding firewood before the fire is ready.”
The relationship is direct, Li argued, as within the world-model-and-reinforcement-learning framework, feeding in more data and more training compute improves the user experience accordingly, a dynamic the founder suggested could drive rapid gains throughout the year.
“This year, however, we’ve already begun seeing results from the first version,” he stated. “Even without using enormous amounts of training data yet, people can already see the system’s potential. Compared with last year, we’ve increased compute investment roughly fivefold this year — and an even larger wave of investment is still coming.”
A Claimed Efficiency Edge
The spending ramp builds on an efficiency claim Nio has made repeatedly.
Li has said the system matches or beats rivals while running on about a fifth of their cloud computing power.
The operating efficiency is distinct from the training investment Li addressed on Thursday, which concerns the compute used to develop and improve the model rather than to run it.
“The correlation is extremely direct,” Nio‘s chief said on Thursday. “Under the framework of world models plus closed-loop reinforcement learning, the more data and compute power we feed into the system, the stronger the results become.”
Nio has credited the system with sharp usage growth.
Urban Navigate on Pilot mileage rose 92% quarter-over-quarter in the first quarter, and the company said total mileage driven on the system topped 200 million kilometers in a single month for the first time in February.
A major upgrade to the Nio World Model is due in China in June, with two further upgrades planned later in the year.
Spending More on Compute, Less Overall
The compute push comes even as Nio has cut its overall research and development budget.
R&D expenses fell to 1.89 billion yuan in the first quarter, down 40.7% from a year earlier and 7% from the fourth quarter, the company reported.
Nio attributed the decline to lower personnel costs following what it called organizational optimization, along with reduced design and development costs.
Finance chief Stanley Qu told analysts the company aims to hold quarterly R&D spending at roughly 2 billion to 2.5 billion yuan, a level he said was sufficient to support investment in core technologies including chips and operating systems.
Nio does not disclose how much of its R&D goes to AI training compute, so the fivefold increase Li described cannot be reconciled directly against the total.
The two figures point in opposite directions, however, suggesting the company is concentrating a larger share of a shrinking budget on developing the world model.
Synchronized Software Releases
Li also pointed to the company’s recent ability to update its three software platforms at once.
For the first time, the Cedar, Banyan and Coconut systems, which run across Nio‘s namesake brand and its sub-brands, were able to release simultaneously rather than on staggered timelines.
“This time, several of our development baselines — Cedar, Banyan, and Coconut — were all able to release simultaneously — that was extremely difficult,” Li stated.
Li said the synchronization is related to improved data-reuse capabilities and the platformization of the company’s underlying engineering, presenting it as evidence the development pipeline is maturing.
“Previously there were timing gaps between them,” he stated. “Achieving synchronized releases means our data reuse capabilities and our underlying engineering platformization are working very well.”
