Goldman Lifts China AI Revenue Forecast 30% to US$13B

Goldman Lifts China AI Revenue Forecast 30% to US$13B

Beijing — Wall Street is sharply revising up how much money China's artificial intelligence industry can make. Goldman Sachs has raised its year-end annualised recurring revenue (ARR) forecast for mainland Chinese AI models by 30 per cent to US$13 billion, in a research note published Monday that cites aggressive price cuts, technical breakthroughs and accelerating corporate adoption across the country's model-making sector.

The upgrade is the latest sign that China's AI sector has shifted from a story about model quality alone to a story about revenue — and that the pricing war now defining the global industry is being fought on Chinese terms. For companies across Asia-Pacific that buy AI services by the token, the revision is a direct signal that the cost curve is still falling.

Tags: China AI, Goldman Sachs, DeepSeek V4 Flash, MiniMax H3, Qwen3.8 Max, AI pricing war, open-weight models, Zhipu AI, AI revenue forecast, OpenAI competition, Alibaba Qwen, Tencent WorkBuddy, AI market China, AI token consumption

What Goldman's New Numbers Say

The revision lifts Goldman's projection from an earlier US$10 billion year-end ARR estimate, reflecting a view that China's model developers are converting technical momentum into paying workloads faster than the market had expected. The bank singled out two Hong Kong-listed names — Zhipu AI, whose projected year-end ARR was raised to US$2.5 billion, and MiniMax, now expected to reach US$1 billion.

"We expect competition for the best performance-to-price balance to intensify among Chinese AI models," the report said, noting that recent releases from domestic pioneers had reached new frontiers in performance per dollar. The forecast covers run-rate revenue from model APIs, enterprise deployments and related services, rather than the broader cloud and hardware businesses that have historically dominated Chinese AI earnings. The distinction matters: it measures the portion of the sector that is most exposed to global price competition and most indicative of real commercial adoption.

MiniMax H3 and DeepSeek V4 Flash: A New Performance-Price Frontier

The upgrade landed days after two landmark releases reset the pricing conversation. MiniMax launched its H3 model last Friday under an open-weight approach — a multi-modal system capable of processing text, image, video, audio and music, priced at just 30 to 50 per cent of incumbent market levels, according to Goldman Sachs. The open-weight decision is significant: it lets developers self-host the model, audit it and fine-tune it, stripping away the margin cushion that proprietary API providers have long enjoyed.

On the same day, DeepSeek officially opened API access to its V4 Flash model, with front-end coding capabilities that rival Zhipu's flagship GLM-5.2. Arena AI's latest leaderboard places GLM-5.2 seventh globally in front-end coding, with DeepSeek V4 Flash a close eighth — a sign that China's second-tier models now sit within striking distance of the frontier on specialised tasks. Two years ago, a Chinese model ranking in the global top ten for a core software-engineering capability would have been dismissed; today it is the baseline expectation.

Alibaba's Qwen3.8 Max Tests the Frontier

The competitive shock is not limited to the challengers. E-commerce giant Alibaba has rolled out its massive 2.4-trillion-parameter Qwen3.8 Max model, built on a mixture-of-experts architecture that activates only a fraction of its parameters per query. On Arena AI's front-end coding rankings, Qwen3.8 Max sits fourth globally — ahead of OpenAI's top-tier GPT-5.6 Sol in sixth place — while undercutting it on price.

Alibaba prices Qwen3.8 Max at US$2 per million input tokens and US$6 per million output tokens, with a 1-million-token context window. OpenAI's flagship GPT-5.6 Sol, by comparison, is billed at US$5 per million input tokens and US$30 per million output tokens. The gap — roughly 60 per cent cheaper on input, 80 per cent cheaper on output — is the arithmetic driving Goldman's revised thesis. When a model that outperforms the incumbent on a key benchmark also costs a fraction of the price, enterprise procurement teams do the math quickly.

OpenAI's Counter-Offensive and the Global Token Shift

OpenAI has not stood still. Last Friday it announced an 80 per cent price cut for its lightweight GPT-5.6 Luna model, slashing API fees to 20 US cents per million input tokens, alongside a 20 per cent reduction for its mid-tier GPT-5.6 Terra. The counter-move is a direct response to a structural shift: Chinese open-weight models now command nearly 70 per cent of global token consumption on developer platforms like OpenRouter, according to Goldman's tracking.

That figure matters beyond market share. It signals that developers worldwide are routing real production workloads through Chinese models — not just experimenting with them — a behavioural change that Western incumbents have rarely faced at this scale. For the US-China AI race, the pricing war is redrawing the economics of frontier AI faster than the model-quality gap is closing. OpenAI's discounting is effectively an admission that performance per dollar, not raw capability, has become the battlefield.

The Monetisation Squeeze and the Shift to Community Licences

The aggressive pricing has a cost. Goldman expects API pricing and gross margins to remain under pressure through the second half of the year, with well-funded model makers using post-fundraising cash to subsidise aggressive pricing in a bid for market share. The squeeze is pushing Chinese developers toward a new commercial model: "community licences" for open-weight models, under which commercial users pay fees while researchers and hobbyists use them freely. It is a deliberate strategy to preserve the open ecosystem's credibility while still building a revenue base.

Workplace agentic tools are emerging as the next battleground. Tencent's WorkBuddy topped China's desktop AI agent rankings in June with a 34 per cent market share by monthly visits, followed by ByteDance's TRAE at 21 per cent, according to Goldman's research. The bank also flagged an emerging geopolitical constraint, noting reports that Chinese authorities were considering restrictions on foreign downloads of model weights and overseas transfers of training data — a development that could complicate the global open-weight ecosystem China itself has built.

What It Means for Japan and Asia-Pacific

For Japanese enterprises, the price-performance equation has changed overnight. A Japanese developer can now access frontier-adjacent coding models from DeepSeek, Qwen or GLM at a fraction of Western API prices — a factor that procurement teams at Tokyo's banks, manufacturers and software houses are beginning to price into their AI budgets. Japan's own AI ambitions, anchored by domestic players such as Sakana AI, Preferred Networks and SoftBank-backed ventures, now face a strategic choice between building, buying Western, or adopting Chinese open-weight foundations.

The reported export-control discussions in Beijing add a layer of risk to that calculus. If Chinese authorities restrict foreign downloads of model weights or cross-border training-data transfers, Japanese users of Qwen, DeepSeek and GLM models could face sudden supply constraints — the same dependency risk Tokyo is trying to engineer out of its semiconductor supply chain. For the Japanese government, the episode also sharpens the debate over its own AI strategy: METI has pushed for domestic compute and sovereign model development, but the private sector's cost calculus is pulling in the opposite direction.

Watch for how Japan's Ministry of Economy, Trade and Industry and its evolving AI safety regime respond, and whether the September Xi-Trump summit produces any stabilisation of the broader tech-trade environment. The pricing war is not a blip; it is a structural repricing of global AI that Asia-Pacific companies must now navigate, with Japan caught between the world's two AI superpowers and the falling cost of intelligence working in everyone's favour.

By Kenji Tanaka, Staff Writer

This article was produced with AI-assisted research and editorial support. Reporting is based on sources cited in the article.