Has the AI industry “changed direction”? Don’t panic—it’s not about braking, but about shifting tracks
Hello everyone, I’m your financial journalist and friend economist.
These past two days have been a bit tumultuous in the tech world. On one hand, OpenAI has paused some of its work due to security concerns, and Anthropic has called for an industry-wide “emergency braking” mechanism. On the other hand, Jensen Huang (the CEO of NVIDIA) is still pushing hard for more computing power, and major companies have not stopped buying chips.
Many people panicked upon reading the headlines: “Is AI on the decline?” “Is the computing power bubble about to burst?”
Let me reassure you: AI is not slowing down, nor is it receding.
What’s really happening is that the AI industry is undergoing a shift in valuation methods. In the past, we focused on AI’s imagination and the scale of its parameters: the bigger the model, the more powerful it seemed. Now, we’re looking at the financials and the efficiency of its use. In other words, we’ve moved from a phase of “spending money to create the future” to a more pragmatic phase where every penny spent must generate a return.
I’ll break down this in-depth analysis into five easy-to-understand points to help you see beyond the surface.
---
1. Rising security discussions ≠ A collective slowdown in research and development
Core logic: The temporary halts by individual companies do not mean the entire industry is coming to a standstill.**
OpenAI has paused certain code executions and network access for research purposes, and Anthropic has suggested the need for a “slowdown option.” This might sound alarming, as if the industry is coming to a halt. But we need to distinguish between facts and proposals:
- OpenAI’s actions: These are temporary, targeted measures for specific models and risks. It’s like a car stopping to check the brakes; once fixed, it will continue on its way.
- Anthropic’s proposal: This is a call for a new system. They want the industry to establish a credible mechanism for slowing down, but it’s just a suggestion, not a mandatory stop. In fact, Anthropic itself acknowledges that AI research is accelerating, and their internal efficiency has increased.
Looking at the actions of major companies: Microsoft’s latest financial report shows that they are still aggressively expanding their AI infrastructure. Money is still flowing into computing power, but management is being questioned by the board and investors: “Where will all this spending on chips lead to revenue?”
Jensen Huang’s statement is crucial: He mentioned that China talks about AI in a more pragmatic way, rarely discussing an “AI apocalypse.” This highlights a truth: Security discussions can be intense, but they won’t stop the industry’s progress. A pragmatic approach is neither ignoring risks nor prematurely declaring the end of a cycle. Security is a safety measure, not a command to stop.
---
2. Beyond model limits, financials matter
Core logic: Customers don’t pay for “rankings” but for cost savings, time efficiency, and fewer errors.
For years, the AI industry’s narrative revolved around the Scaling Law: more data, larger models, and more powerful computing power equal greater capabilities. This rule still holds, but commercial value doesn’t equate to technical metrics.
Imagine you own a restaurant:
- In the past: As long as your chef (the model) has a Michelin three-star rating (large parameters, strong capabilities), customers would queue up.
- Now: Customers care if the food is served quickly (response time), if the taste is consistent (accuracy), and if the price is reasonable (cost).
Business customers won’t pay just because a model has a high ranking. They’ll pay if:
- The wait time for service is reduced,
- Code delivery is faster,
- Quality checks are more accurate,
- Development cycles are shorter.
The key point: GPUs, data centers, and electricity are not intangible assets; they are tangible fixed costs. They depreciate over time, and even if they’re not in use, they still incur expenses. If usage rates don’t increase, larger scales become a burden.
So, the focus is no longer on whether we can create stronger models but on whether the new capabilities can lead to higher customer prices, increased usage, or lower delivery costs. Model capabilities must be integrated into business processes to generate real revenue; otherwise, it’s just self-indulgence.
---
3. In the era of inference, scarcity lies in **utilization**
Core logic: Having a GPU is just an entry ticket; the real skill is using it efficiently.**
Many people think that “training power” will be replaced by “inference power.” However, they complement each other:
- Training power determines the limits of a model’s capabilities.
- Inference power determines whether those capabilities can be reliably used millions of times.
Inference directly connects computing power with user needs. Every user query and code generation tests the system’s concurrency, latency, energy consumption, and cost per token.
For example: With the same set of chips (e.g., 1000 GPUs):
- Company A: Poor scheduling, bulky models, lots of idle time.
- Company B: Optimized scheduling, lightweight models, efficient batch processing.
As a result, Company B can serve more users with the same hardware at a lower cost.
The most valuable computing power in the future won’t be the most expensive clusters but systems that can handle high loads and generate continuous revenue. This is why we’re now focusing on less glamorous technologies like quantization, distillation, MoE (Mixed Expert Models), caching, and hardware-software integration. Their role is to reduce the cost per use and make previously unfeasible scenarios scalable.
In this phase, simply having GPUs is not enough; you need to efficiently manage different models and hardware and ensure customers continue to use your services, thus gaining control over the pricing of your computing assets.
---
4. Investors’ perspective: From “counting GPUs” to “calculating returns”
Core logic: A cluster of thousands of GPUs without proper utilization is just expensive inventory.
From an investor’s standpoint, the criteria for evaluating AI projects have changed significantly.
In the past: Questions were about:
- “How many GPUs do you use?” (Sounds impressive.)
- “How large are your model parameters?” (Sounds cutting-edge.)
Now: We ask:
1. Utilization: How much of the time do these GPUs spend serving paid requests? A cluster of thousands of GPUs in a warehouse full of expired milk is just a waste of money. A smaller, well-used cluster can still be profitable.
2. Unit economics: Does revenue growth come with slower increases in model usage, memory, bandwidth, and costs? If more users lead to faster losses, scale becomes a problem rather than a benefit (this is called “diseconomies of scale”).
3. Engineering and scenarios: Can the company dynamically match models, chips, and tasks? Does the customer’s data and processes create stickiness (i.e., making it difficult for them to switch to another provider)?
In summary: The return on AI computing power comes from technical efficiency multiplied by real demand. Without either, it won’t work.
---
5. Conclusion: The “coming-of-age” of AI commercialization
Core logic: The industry hasn’t left the table; it’s just no longer rewarding grandiose promises.**
Seeing these security discussions as a kind of “coming-of-age” for AI commercialization is apt.
- In the early stages: Everyone competed on speed, volume, and grand visions. Good storytelling drove high valuations.
- Now: You must be responsible for your actions and results. The industry is still growing, but the rules have changed: grandiose promises without a clear path to realization are no longer rewarded.
Increasing security requirements are to ensure stability; continued investment in computing power reflects real demand; and the shift in valuation methods indicates market maturity.
Finally, here’s a question for you: In your industry, what’s holding back AI from generating returns?
- Are the models not powerful enough to solve core problems?
- Are the costs too high, making it unprofitable?
- Or have you simply not found a valuable use case?
Feel free to share your thoughts in the comments. Understanding these aspects will help you determine the next steps.