CSC: AI Computing Hardware Remains Highly Prosperous, Software Recovery Signals Increasing
Claire Weston
China Securities Construction (CSC) flags surging profits across the AI hardware chain in its latest A-share mid-year preview, while software firms show early signs of operational recovery; three major overseas model launches in a single week further validate sustained demand for computing power.
How much are hardware leaders making — and what do the numbers tell us?
Inspur expects H1 2026 net profit of RMB 2.6–3.1 billion, up 226%–288% year-on-year. H3C parent Unisplendour expects RMB 1.91–2.32 billion, up 83.5%–122.89%.
This means → across AI servers and smart-computing infrastructure — the data centres and equipment that power AI models — revenue and profit growth is no longer a forecast; it is already showing up in earnings.
In plain terms = whoever is building the "power plants" for AI is the first to collect real cash this cycle.
Are software firms still losing money — and how strong is the recovery signal?
Saiyi Information expects H1 net profit of RMB 53–65 million, up 188%–257%, driven by better delivery efficiency, cost control, and overseas expansion.
Venustech expects RMB 26–38 million, a swing to profit, with visibly improved operating cash flow. Qihoo 360 expects RMB 180–260 million, also swinging to profit.
This means → the software signal is "bleeding has stopped," not "breakout" — absolute profit is still far smaller than hardware leaders, but the direction has flipped from loss to gain. Cost-cutting and efficiency gains are starting to show.
Three models in one week overseas — who is competing on price, who on performance?
OpenAI released the GPT-5.6 series on July 9, spanning Sol / Terra / Luna tiers. Sol shows major gains in Coding Agent — a system that lets AI autonomously write code to complete tasks — with a Coding Agent Index score of 80.
xAI released Grok 4.5 on July 8. API pricing: $2 / $6 per million tokens (input / output), well below GPT-5.5's $5 / $30. Its Coding Agent Index hit 76, roughly matching GPT-5.5, but per-task cost is just $2.49 — about half of GPT-5.5 Codex's $5.07.
In plain terms = xAI's play is "match the performance, halve the price." For any company using AI to write code or automate workflows, the barrier to adoption just dropped.
Why does Meta's move deserve its own section?
Meta released Muse Spark 1.1 on July 9, supporting a one-million-token context window — the amount of text the model can "read" in a single pass — at API prices below both Grok 4.5 and GPT-5.6's flagship tiers.
Separately, Meta raised its 2026 capex guidance to $125–145 billion and announced a $9.1 billion investment in its largest AI data centre outside the US, located in Canada.
This means → Meta is not just releasing models; it is scaling capacity with real money. Demand for GPUs, networking, storage, and data-centre space is still accelerating. This reflects confidence among top-tier players that compute-as-a-service has genuine commercial runway.
What are China's model builders doing — and where is the money coming from and going?
Zhipu AI founder Tang Jie issued an internal letter on July 11 announcing the "Touch High" initiative — a two-year strategic push into long-horizon task capability and autonomous agents (AI systems that complete complex multi-step tasks on their own). Zhipu recently raised approximately HK$31.4 billion via a placement.
MiniMax CEO Yan Junjie announced he will forgo his salary and contribute personal shares equal to 4% of total equity for team incentives plus 1% to support the open-source community. The company completed a HK$16 billion raise on the same day, with 80% earmarked for AI infrastructure and model R&D.
This means → China's leading model firms are concentrating investment on AGI (artificial general intelligence), coding agents, and infrastructure expansion. The fundraising scale signals that capital markets are still backing this trajectory.
Can this boom last — and what is the key checkpoint CSC identifies?
CSC notes that earnings divergence across the sector remains stark: hardware is already delivering, software is in early-stage recovery.
The report identifies a key validation point: whether inference-side compute consumption — the computing power consumed when AI models are actually used — infrastructure investment, and AI application monetisation can sustain mutual reinforcement.
In plain terms = hardware profits are confirmed. The next question is whether the number of people actually using AI keeps growing and whether they keep paying. If adoption stalls, even the biggest data centres become idle capacity.
Content is for reference only, not financial advice.