Blindly betting on computing power is extremely dangerous, and if the bottleneck of computing power is overcome, it is still unknown what the industry will look like In 2023, we are betting on shovels, computing power, memory, and electricity. The logic is simple: you can't guess what AI will look like in ten years, but you know that no matter what it looks like, there will be more of this stuff. Three years have passed, and theoretically, it should have converged. The result not only failed to converge, but also widened the odds. In terms of AI infrastructure. SK Hynix、 Samsung publicly warned in April that the shortage of AI driven memory will continue at least until after 2027. HBM's 2026 production capacity has been sold out, DRAM will increase by 60% in 2025, and there will be another 30-40% increase in 2026. This is a "ten-year drought" script, but on the other hand, DeepSeek cut a large part of the inference FLOPs with MLA+MoE this year, triggering architecture innovation due to requirement disruption, and directly bypassing bottlenecks is completely feasible. It is still uncertain where the profit will ultimately fall, whether it will be from computing infrastructure, large models, or cloud providers SemiAnalysis directly announced at the end of April that all value will be consumed by the big model infrastructure layer from 2023 to 2025, and will shift to big model companies and inference suppliers starting in 2026. The evidence is that Anthropic's revenue has skyrocketed from $9 billion to an estimated $30 billion to $44 billion by 2026. At the same time, it signed the Google+Broadcom multi GW computing power agreement, signed the SpaceX computing power agreement, announced its self-developed chip plan, and pushed Claude Cowork to cut the application layer. Three years ago, no one dared to write the script for a company that can handle everything from silicon wafers to application layers on the full stack. The attack was launched from the big model and ultimately swallowed up everything. But you can also believe another story. By 2026, the capital expenditure of the giants will reach $700 billion (only $200 billion by 2024), and the monopoly of computing power will deepen. Ultimately, the laboratory will still be a tenant of computing power. On the application layer side, Salesforce, ServiceNow, and Adobe's stock prices hit a new low for the year in 2026, and the market is concerned that Agentic AI will directly eat up SaaS. The application layer may be both the biggest winner and the biggest loser. Just like in the early days of the Internet, everyone felt that everything should be subverted. At that time, people were optimistic about telecom operators (broadband infrastructure), portal websites (super portals), and B2C celebrities (offline business moving online). The software company that ultimately survived was Amazon eBay、 Microsoft, Google, and most others have gone bankrupt, and even the business model of the Internet has undergone tremendous changes. The survivors have become giants. Although most hardware companies have not gone bankrupt, they have been in a semi dead state for a long time until 2018, when it took about twenty years to recover. Previously, servers, network routers, and optical communication, like today's GPU computing power, HBM memory, and optical modules, were the most deterministic shovels. During the frenzy of infrastructure construction, the performance of hardware sellers is often astonishingly good, as downstream investors buy hardware at no cost. But at the same time, we must also be vigilant that once downstream business models fail to generate sufficient positive cash flow loops, hardware orders will evaporate instantly, followed by a brutal killing logic and valuation. So the current situation is 1. Continue to take the shovel compartment. HBM、 Electricity and ASIC design, which do not require predicting the final outcome, are still the most comfortable bottom positions, but must always pay attention to downstream demand changes and cash flow loops. Hardware companies with computing power infrastructure are not one-time targets. 2. Do not re invest between AI big models/ultra large scale cloud service providers/application layers. The three layers erode each other, whoever eats has not finished yet, and heavy warehousing on any layer is a gamble on a single script. 3. Keep the cash. Don't check-in too early, market divergence itself is an opportunity, and the biggest loss often comes from chasing after the consensus stage.