These past few days, three things have simultaneously come to the fore: Oracle laid off 21,000 people in the last fiscal year, explicitly stating that AI replacement is the core reason; Doubao's daily average token call volume has exceeded 180 trillion, and the model of hundreds of intelligent agents collaborating synchronously is already operational; Zhipu's Hong Kong stock market value has surpassed 1 trillion Hong Kong dollars, having increased 18 times in the six months since its listing.
These three things seem to be talking about different topics, but for business managers, they are actually about the same thing—the impact of AI on enterprises has shifted from "whether to use it" to "what will happen after using it."
Signal 1: Oracle lays off 21,000 people — AI substitution effect is no longer a prediction
Oracle disclosed that it laid off 21,000 employees in the last fiscal year, explicitly stating that large-scale AI deployment is the core reason for job reductions, and this trend will not stop.
This is not layoff news; it is a notice. Oracle is the world's second-largest enterprise software company, and its layoffs are not due to poor management—it explicitly states that AI deployment has replaced these positions. When an enterprise software giant with annual revenue exceeding $50 billion openly admits that AI substitution is the core reason for layoffs, other business managers need to consider one thing: how many positions in your organization are being pushed by AI from "efficiency improvement" to "direct replacement"?
I've thought about this matter repeatedly. Oracle's layoffs are likely concentrated in several areas: customer service support, basic operations, data entry, and junior development. The common characteristic of these positions is that they are highly procedural, standardizable, and do not involve complex decision-making.
| Job Type | AI replacement risk | Alternative method |
|---|---|---|
| Customer Service/Technical Support | High | AI agent 24/7 response + automatic tiered processing |
| Basic Operations and Maintenance | High | Agent automatic inspection + fault self-closure |
| Data Entry/Report | High | RPA+AI extraction+automatic generation |
| Junior Developer/Tester | Medium-high | AI programming assistant end-to-end delivery |
| Compliance/Audit | 中 | AI audit system coverage (e.g., EY's 130,000 auditors) |
| Strategic Decision/Innovation | Low | AI assists but does not replace judgment |
The key is not "who will be laid off," but "how to operate after the layoffs." Oracle laid off 21,000 people, but its products and services did not shrink—indicating that AI has already filled the output of those positions internally. This is what business managers need to take seriously: it's not about whether AI will replace your employees, but whether your organization can run better after the replacement.
Signal 2: 180 Trillion Tokens — A New Paradigm of Agent Collaboration Is Already in Motion
The Doubao 2.1 Pro launched at the Volcano Engine FORCE Conference has several numbers worth examining carefully:
| Indicator | Data | What does it mean for enterprises |
|---|---|---|
| Average daily token call volume | 180 trillion | Two years ago, 140 trillion was the annual figure; now it has been reached in a single day. |
| Number of agent collaborations | Hundreds of synchronous collaborations | A springboard from "one AI assistant" to "a group of AI employees" |
| New Capabilities | Multi-agent collaboration + multimodal | Cross-system, cross-task orchestration capability |
| Life Services | Ride-hailing + lifestyle services launched | AI transforms from "tool" to "portal" |
The key point is the synchronized collaboration of hundreds of agents. Over the past two years, the mainstream model for enterprises deploying AI has been one AI assistant serving one employee. The new paradigm demonstrated by Doubao 2.1 Pro is: a single task is broken down and assigned to multiple agents, each responsible for one step, executing synchronously and coordinating automatically. Isn't this a digital mirror of enterprise workflows? The procurement agent receives requirements, the finance agent approves budgets, the logistics agent schedules deliveries—once the process is complete, humans only need to review at critical nodes.
Zhipu GLM-5.2's performance also confirms the same direction: its coding capability surpasses multiple mainstream overseas models, and the Hong Kong stock market value surged 18 times in half a year to exceed 1 trillion Hong Kong dollars. Domestic large models have shifted from "catching up" to "running alongside" and even "leading in certain areas," meaning the underlying logic of enterprise model selection has changed—no longer are OpenAI and Claude the only two options. The scenario adaptability and localized compliance support of domestic models are genuine advantages.
Signal Three: Computing Localization — Enterprise AI No Longer Relies Solely on the Cloud
The Jetson AGX Thor robot chip released by NVIDIA has a 7.5x increase in local computing power and can run large models without relying on the cloud. It is now available for mass supply to industrial and service robots.
The corporate significance of this matter is greater than it appears:
Cloud AI Mode
Data upload → Cloud inference → Result return
Dependence on network stability
Delay 0.5-2 seconds
Data security boundaries are blurred
Continue to pay by Token
Local AI Mode
Local inference → instant response
Can run even with network interruption
Delay
Data does not leave the enterprise boundary
One-time hardware investment
For industries with high real-time requirements such as manufacturing, warehousing and logistics, and energy, local computing power is not a "nice-to-have" option but a prerequisite for "whether it can be used." If a quality inspection AI on a production line relies on cloud-based inference, network fluctuations could bring the entire line to a halt.
But don't rush to move all AI back on-premises either. On-premises computing power is suitable for scenarios with high real-time requirements and data sensitivity (quality inspection, control, security). For tasks requiring complex reasoning by large models (strategic analysis, contract review, R&D design), the cloud remains superior. A hybrid architecture is the pragmatic choice—deploy critical scenarios on-premises, handle complex tasks with cloud reasoning, and use data desensitization as the bridge in between.
How should corporate managers understand these three signals
The image pieced together from three things looks like this:
1. AI replacement has shifted from "possibility" to "reality." Oracle's 21,000 layoffs are not an isolated case but a signal. Business managers need to reassess the AI replacement boundaries for each position—not by panicking and cutting jobs, but by systematically transferring the functions of "replaceable positions" to AI, while reallocating the freed-up human resources to "irreplaceable" judgment and innovation roles.
2. "A group of AI employees" is closer to real enterprise workflows than "a single AI assistant."The model of hundreds of intelligent agents collaborating is, in essence, digitizing enterprise workflows. If your ERP system is still at the stage of "a single AI button," you need to seriously consider upgrading to "process-level intelligent agent orchestration"—this is the entry point where AI becomes useful for enterprise management.
3. Hybrid computing architecture is the next infrastructure decision. Don't just look at cloud costs; also consider real-time performance, security, and determinism. Manufacturing and logistics companies especially need to evaluate: In which scenarios must AI inference be localized? How large will the gap be between local computing hardware costs and cloud service fees three years from now if we invest in local computing hardware now?
