May 20, Beijing. At the IDC Directions 2026 forum, one number left many CEOs and CIOs in the audience silent: by 2026, 50% of enterprise AI projects will fail to meet ROI targets.
On the same day, at another closed-door meeting, Huawei released the 'Intelligent World 2035' white paper, painting a different picture: in the next decade, the computing power of society as a whole will increase 100,000-fold, and spending on embodied intelligence will surge from $1.4 billion to $77 billion within five years.
Looking at these two things together, it's quite contradictory, isn't it? On one hand, it says the opportunities are unimaginably huge; on the other, it says half the projects will fail. I carefully compared the two reports and found they are not contradictory—they are two sides of the same coin.
Global enterprise AI spending will be $940 billion in 2026 (IDC) and is expected to reach $2.1 trillion by 2029, growing 2.2 times in four years
Two Reports, One Judgment: The Problem of Keeping Pace
First, let's look at the numbers. IDC's data granularity is very fine, broken down into several rows:
- Enterprise AI spending growth rate: $940 billion in 2026 and $2.1 trillion in 2029—a 2.2x increase over four years, with a compound annual growth rate of about 22%. This pace would be rapid in any technology field, but in AI, it may only be the beginning.
- MaaS (Model as a Service): In 2026, token calls reached 400 trillion, with market revenue of about RMB 18.6 billion. Large models are moving from "flexing muscles" to commodities with infrastructure-level pricing.
- Embodied intelligence: currently US$14 billion, projected to reach US$77 billion within five years, with a compound annual growth rate exceeding 120%. This is the only track in both reports that is simultaneously described as “explosive growth”.
Huawei's vision is even grander. By 2035, society's total computing power will increase one hundred thousandfold (from 10^18 to 10^22 FLOPS), robot density will increase from 300 units per 10,000 people to more than 1,000 units, and manufacturing defect rates will drop below 0.05%. Their logic is: artificial intelligence is moving from "generation" to "action," and the physical world is the main battlefield.
But Huawei's report also has a very interesting gap—there is no mention of the failure rate at all.
IDC's Warning: Four Ways AI Could Die
IDC has broken down the reasons why AI projects fail into four traps:
| Cause of failure | Key Data | My understanding |
|---|---|---|
| Unclear return | Projects without a clear return on investment path have a failure rate exceeding 70%. | This isn't because the technology isn't good enough; it's because they haven't figured out what problem they need to solve. Many projects start out simply saying, "We need to adopt AI," but when it comes to how much money was actually saved or how much efficiency was improved after adoption, they can't say. |
| Weak data foundation | 60% of enterprises are in a state of decentralized AI management | Data silos, poor quality, and lack of unified governance. This is almost a replay of every ERP implementation project—without master data governance, upper-level intelligence is nothing but a castle in the air. |
| Weak human-machine collaboration | 40% of enterprises will move to unified governance by 2027 | The organization is not yet ready; employees are either unable or unwilling to use them. Purchasing tools does not mean those tools are already in place. |
| Each operates independently. | By 2027, 40% of enterprises will replace decentralized management with unified governance. | Various departments are running their own AI pilot projects, lacking unified data standards, security boundaries, or evaluation mechanisms. The situation is chaotic. |
Here is a figure worth considering: IDC points out that 60% of enterprises are still in a state of decentralized AI management, but at the same time predicts that by 2027, 40% of enterprises will move toward unified governance. This means that within the next 18 months, CIOs at many enterprises will need to transform "AI governance" from a PowerPoint presentation into an actual organizational structure.
Huawei sees the direction, IDC sees the trap
I made a comparison:
| Size | Huawei "Smart World 2035" | IDC Directions 2026 |
|---|---|---|
| Perspective | Technology Vision (2035) | Current landing page (2026) |
| Failure risk | completely unmentioned | Clear warning: 50% of projects will fail |
| Embodied intelligence | Emphasize technological breakthroughs | Emphasize the speed of market adoption |
| Industry AI | The technical path is clear. | The data foundation is the biggest bottleneck |
| Core narrative | How big is the chance? | How many people will die on the road |
This comparison is not to say that Huawei's report is problematic—they are inherently focused on technological vision. But for corporate CIOs, having both reports together provides a complete decision-making map: The direction is right, but the road is full of pitfalls.
Three practical suggestions for enterprise CIOs
Based on these two reports and the actual situations of the enterprises we have engaged with, I have summarized three points:
First, get the data in order before talking about AI. Two of the four types of failure are directly related to data governance. Among the manufacturing AI projects I have seen, six out of eight get stuck on data quality—either material codes in the ERP system are inconsistent, or equipment data collection is incomplete. The 100,000x computing power Huawei talks about sounds very distant, but if you don't manage master data well now, even if that 100,000x computing power really arrives, you won't be able to use it.
Second, clarify ROI before taking action. IDC points out that more than 70% of projects without a clear ROI path ultimately fail—and this is no exaggeration. In my enterprise AI research conducted last year, fewer than 20% of companies could clearly state "how much money this AI project saved in a year." Most answered "improved efficiency" or "enhanced experience." It's not that these statements are wrong, but that they cannot be verified. For things that cannot be verified, budgets will eventually be cut.
Third, do not wait. Decentralized management is the reality; you do not need to achieve unified governance in one step.You can first run a pilot in one department—for example, use an AI assistant to handle financial reimbursement reviews, or use agents to perform quality checks on customer service scripts—calculate the return on investment, and then roll it out. This approach is far more practical than establishing a 'Group AI Governance Committee' and issuing a few documents.
Huawei says the AI supercycle has arrived, while IDC says 50% of projects will fail. Both statements are true. The difference lies in whether you are the former or the latter.
IDC Directions 2026 Beijing Forum (May 20, 2026), Huawei’s “Smart World 2035” White Paper
Disclaimer: This article represents only the author's personal views and does not constitute investment advice. Reproduction must indicate the source.
