Introduction: Not Infinite Prosperity, Written after Nvidia’s Q3 2025 financial report.
AI Hardware’s ‘Long Cycle’ Strategic Research Report
Insight: From Cyber’s guidance on Nvidia’s next quarter growth rate, the market’s enthusiasm for AI hardware masks a calm and objective rule: The future AI hardware is likely to be a truly ‘long cycle’ industry, but certainly not an ‘infinite cycle.’
When we strip away short-term emotions, Cyber’s deep predictions at the level of primary markets and industrial strategic thinking, combined with the boundaries of physical laws and economic rules, reveal a clear timeline: before 2027, the industry will maintain high prosperity; and after the turning point in 2028, growth will face a stepwise decline.
During this process, the industrial logic will undergo three profound shifts: from Capex-driven to ROI-driven, from single funding to infrastructure financing, and from GPU dominance to heterogeneous computing.
1. Cycle Theory: The Revelry of 2027 and the Awakening of 2028
The current AI expansion wave is essentially a race against the ‘Scaling Laws.’
The ‘high prosperity’ logic before 2027:
The next three years will be a concentrated period for the deployment of GW-level data centers. To position themselves for the next generation of models (such as GPT-5/6 and their counterparts), tech giants cannot stop the arms race. The advancement of Moore’s Law at the 2nm/1.6nm nodes, along with the release of HBM (High Bandwidth Memory) capacity, will support this phase of ‘irrational prosperity.’
The ‘stepwise decline’ of 2028:
The turning point will appear around 2028. By then, the first large-scale deployed H100/H800 clusters will enter the later stages of depreciation, while the demand for computing power on the training side may show marginal diminishing returns due to model architecture optimization. The market will shift from ‘panic hoarding’ to ‘stock replacement and optimization.’ Growth will not return to zero, but will step down to a rational growth range similar to traditional semiconductor cycles.
2. Profit Model: From ‘Stacking Capex’ to ‘Exchanging Efficiency for Returns’
The current prosperity of the AI industry is built on a fragile assumption: as long as the computing power is large enough, applications will explode. But the reality is, the profit model must shift from ‘stacking Capex’ to ‘exchanging efficiency for returns.’ The explosion of the Google model is a testament to this, which has, to some extent, rapidly driven industry development and new expectations in the short term. ‘Google: AI computing power doubles in six months’ implies an investment treasure map.
The urgency of cash flow gaps:
Currently, the cash flow (Revenue) of model manufacturers and application sides far exceeds the depreciation (Depreciation) and operating costs (Opex) on the hardware side. This cash flow gap must be bridged by the current business model; otherwise, the turning point in 2028 will evolve into a collapse triggered by an inability to pay electricity bills and depreciation.
Efficiency is paramount:
The core metric in the future will no longer be ‘how many cards are owned,’ but ‘the cost of generating per Token.’ This requires the entire industry chain to shift focus from peak performance to performance per watt and return per dollar. Those who can help customers save money will be the ones to survive in the second half.
3. Capital and Energy: ‘Infrastructure’ under Tight Constraints
Trillions of AI capital expenditures and electricity investments will continue to expand, but marginal returns are constrained, with the core bottleneck being capital costs and electricity supply constraints.
Generational shift in financing structure:
Early AI investments were driven by VC and the free cash flow of tech giants. As AI data centers evolve into ‘digital infrastructure’ similar to power grids and dams, their funding sources will undergo a qualitative change.
Private equity (PE), infrastructure funds, and even the bond market will enter the scene, collectively footing the bill for this data center frenzy. These funds are more sensitive to risk and demand more stable long-term returns, which will in turn compel AI hardware manufacturers to provide longer-lasting and lower-maintenance solutions.
The backlash from the physical world:
Electricity is no longer a cheap resource but a scarce strategic resource. The construction cycle of electricity infrastructure (3-5 years) is much slower than the iteration cycle of GPUs (1-1.5 years), and this mismatch will become a hard ceiling to curb the unlimited expansion of computing power.
4. Hardware Structure: From ‘GPU Dominance’ to ‘Hybrid Ecology’
Under the pressure of costs and efficiency, the hardware architecture will undergo structural changes: shifting from ‘GPU single-core dominance’ to a hybrid ecology of ‘GPU + ASIC + Edge NPU.’
Training Side: The Last Bastion of Top GPUs
Large model training involves extremely complex parameter synchronization and communication, and this field will still be dominated by a few leading GPU platforms (such as Nvidia). The moat here (CUDA ecosystem, NVLink) remains profound.
Inference Side: The Rise of ASICs
Once model training is complete and enters the large-scale service phase, the cost-effectiveness of general-purpose GPUs becomes too low. To achieve extreme ‘exchanging efficiency for returns,’ inference tasks will largely migrate to self-developed ASICs (such as Google TPU, AWS Inferentia, Groq, etc.).
Edge Side: The Popularization of NPUs
Many enterprise applications and personal assistant functions will no longer rely on the cloud but will be pushed down to the terminal. NPUs (Neural Processing Units) in mobile phones, PCs, and automotive terminals will take on the tasks of processing lightweight models and privacy-sensitive data.
Conclusion and Outlook
Conclusion: Nvidia’s market value myth may continue in the training side, but in the inference side (which is the future’s largest incremental market), it will face strong competition from Google, Amazon, Broadcom (designing ASICs for giants), and edge chip manufacturers. In summary, the AI hardware industry is undergoing a transformation from ‘gold rush’ to ‘industrialization.’ 2027/2028 will be the peak of this revelry, while 2028 will mark the beginning of maturity. For investors and practitioners, understanding the ‘long cycle’ is not difficult; the challenge lies in acknowledging the ‘non-infinite cycle’ and preparing in advance for a future that values efficiency, ASICs, and diversified architectures.
Outlook: The ‘Second Half’ Script of AI Hardware Investment:
Beta Returns (-2028): Still come from the inertia of infrastructure construction, buying from ‘those who sell shovels’ (Nvidia, optical modules, cooling, power equipment). Alpha Returns (2027-2028+): Come from efficiency enhancers and vertical application winners. Who can reduce inference costs by 50% with ASICs? Who can solve the cooling and power issues of data centers? Who can make small models perform like large models at the edge?
This is a critical turning point from ‘building to it’ to ‘profiting from it.’
Related articles: Cyber will delve into the current competitive landscape of ‘inference-side ASICs.’ (For example, how Broadcom/Marvell assists major manufacturers in self-developing chips and the specific erosion of Nvidia’s long-term moat) Anti-Nvidia Alliance: Analyzing the current competitive landscape of ‘inference-side ASICs,’ Gemini 3 emerges: Bubble explosion reconstructs, AI productivity undergoes major changes, Cyber’s in-depth assessment | The specific impact of ASIC order spillover on various links in the supply chain (recommended)