By Hamza Faisal

For the first time in the history of the AI industry, the world is spending more money running artificial intelligence models than building them. Gartner forecasts that global spending on AI inference — the cost of running a trained model every time someone asks it a question — will reach $23.3 billion in 2026, overtaking the $19 billion spent on training new models. The crossover marks a turning point: AI is no longer mainly a race to build bigger models, but a business of serving them cheaply, constantly, and close to the people using them.

The figures come from Gartner's August forecast for AI-optimized cloud infrastructure, which puts total spending in the segment at $42.3 billion in 2026, up 96 percent on the year before, and $66.1 billion in 2027. Inference will account for 55 percent of that spending this year and 59 percent next year. Over the lifetime of a deployed model, analysts estimate inference eats 80 to 90 percent of the compute bill.

The edge hosts the agents

The shift was on full display at the China International Import Expo in Shanghai, which opened on November 5. Qualcomm, Samsung, Intel, AMD, SK Hynix and Synopsys all put real-world AI deployment — not raw model size — at the centre of their booths. Qualcomm showed a physical 6G terminal prototype, the first of its kind shown publicly, and ran a 30-billion-parameter mixture-of-experts model entirely on a smartphone. AMD launched an "agent host" built on its Ryzen AI Max+ PRO 495 platform, with 192GB of unified memory, designed to run large models locally and juggle several AI agents at once without sending data to the cloud.

Intel's China chairman, Wang Zhicong, told reporters at the expo that more AI compute will move to the inference side and the edge over the next five years, because "compute needs to be closer to the data, closer to the scenarios and closer to the users." Two years ago, he said, the industry talked mostly about training bigger models; today the conversation is about agents — "which is really about inference."

Memory giants double down

The memory makers are following the money. Samsung showed LPDDR6, LPDDR5X-PIM — the industry's first LPDDR memory with processing built into the memory itself, delivering up to eight times the effective bandwidth of LPDDR5X — and UFS 5.0 flash storage, with bandwidth up to 10.8GB/s. SK Hynix brought high-capacity server memory and enterprise SSDs. But the memory shortage that has pushed prices up shows no sign of easing: SK Hynix staff at the expo said supply and demand will only rebalance after capacity increases land in 2028.

From a training race to a scenario race

IDC and Inspur forecast that China's intelligent compute capacity will reach 2,576.5 EFLOPS in 2026, up 87.9 percent year on year, and that by 2030 half of all enterprise AI inference work will happen on edge devices rather than in the cloud. The AI compute market in China alone is expected to reach $51.5 billion this year.

For the chipmakers, the competitive dimension is moving from peak compute to energy efficiency, deployment flexibility and ecosystem partnerships. Training remains a must-win battleground, but the growth is now at the edge — in AI PCs, agent hosts, robots and cars. Whoever can run inference cheapest and closest to the user is likely to win the next round of orders.

By Hamza Faisal