I have written many tweets about the AI industry chain over the past year or two, and previously detailed the investment logic of the AI computing power industry chain.
At the end of September last year https://x.com/qinbafrank/status/1973043276568781033?s=46&t=k6rimWsEbo2D2tXolYcM-A, I mentioned that the expansion of applications has led to greater demand for computing power, benefiting both computing and storage.
In mid-February this year https://x.com/qinbafrank/status/2023597321368547520?s=20, in the tweet about the war of capital expenditures, I discussed: 1) Key links in the computing power industry chain can still capture the greatest value: chips, packaging and testing, storage, optical modules, etc. Those whose capacity is not easily expanded quickly and those with high moats will enjoy the dividends of huge capital expenditures;
2) There is still significant room for efficiency optimization: distillation, quantization, MoE, specialized chips, liquid cooling, and nuclear fusion (long-term) on the inference side may reduce the energy consumption and cost per unit of computing power by another 10–100 times. Look for opportunities in these areas.
In mid-April this year, in https://x.com/qinbafrank/status/2043653387271712962?s=20, I detailed the underlying logic of how industry bottlenecks are transmitted among GPUs, storage, optical interconnects, power, and CPUs.
My recent thinking is: "AI capex spreading along the data center industry chain" remains the underlying map, but it is no longer an effective alpha framework. The truly effective framework now should upgrade from capex beta to:
Who owns the real bottleneck, who can convert the bottleneck into profit, who gains control points in the next-generation architecture migration, and whose valuation has already been overdrawn in advance.
In other words, in the past two years, you could buy "AI data center correlation"; going forward, you should buy scarce cash flow + architectural control + verifiable order sustainability. Why? The core logic is that after every link in an industry chain has been hyped once, you can no longer treat everything the same. Instead, you must truly distinguish which companies are irreplaceable in the industry chain and have strong core competitiveness, supported by massive orders and better-than-expected deliveries. It is essentially a shift from broad-brush investing to careful screening.
I. Three Key Variables of the Above Logic
First, it's not simply that "every link in the data center has been hyped, so the old logic is invalid," but rather that the industry chain mapping logic has not failed; it has just degraded from a "stock selection logic" to a "research entry point."
Big tech capex and demand have not weakened. Microsoft's FY26 Q3 Azure and other cloud services revenue grew 40% year-on-year, and it clearly stated that customer demand still exceeds available capacity; this shows that computing power supply constraints are still real. Meta also raised its 2026 capex guidance from $115–135 billion to $125–145 billion, citing higher component prices and future data center capacity costs. Alphabet's Q1 2026 Google Cloud revenue grew 63% year-on-year, with a cloud backlog exceeding $460 billion, indicating that cloud AI demand is entering the order/contract level.
So the problem isn't that the "capex logic is gone," but that: the capex logic has been fully learned by the market, and you can no longer indiscriminately buy all benefiting links.
Second, "earnings growth" must be broken down into "volume, price, mix, and sustainability."
Saying "earnings growth vs. expectation/valuation growth" is correct, but not detailed enough. Now the market will truly ask:
- Does earnings growth come from shipment volume or ASP?
- Is ASP a structural price increase or driven by short-term shortages?
- Does gross margin improvement come from product mix or a one-time supply-demand mismatch?
- Is the backlog cancellable or locked in by multi-generational platforms?
- Does customer concentration give the company bargaining power, or does it suppress profit margins?
For example, the storage sector is clearly not just about expectations now: Micron's FY26 Q2 revenue rose from $8.05 billion in the same period last year to $23.86 billion, with a gross margin of 74.4%. The company attributed the record performance to strong demand, tight supply, and the strategic value of memory in the AI era. This is completely different from many companies that "only tell CPO/800V stories but haven't confirmed large-scale revenue."
Third, "technology upgrades" must be broken down into "correct direction" and "correct adoption pace."
CPO, 800V DC, advanced packaging, liquid cooling, and power architecture upgrades are all real directions, but the easiest mistake in investing is: the direction is right, but the pace is wrong; the pace is right, but the company is not in the main chain; the company is in the main chain, but the profit is taken by the platform owner.
NVIDIA's 800V DC page has already called the traditional 54V architecture a bottleneck for the next-generation AI factory, emphasizing that 800V DC can reduce conversion stages, current, copper loss, and cable volume, improving tokens per watt. But this doesn't mean all 800V-related companies will scale up immediately; it is more likely to land first in the highest power density Rubin/Vera Rubin, NVL racks, and hyperscale AI factories, and then gradually spread.
II. Current AI Data Center Industry Chain Categories
The best assets are usually between the first and third categories: having both short-term performance and next-generation architectural control points. The worst is the fourth category: revenue follows capex, but competition barriers are low, gross margins are low, and customer bargaining power is strong.
III. Segment Analysis: New Changes in Investment Logic and Competitive Landscape
- AI Data Center Storage/Memory: Still the strongest "earnings realization" main line, but beware of supply backlash after 2027. Storage here should be divided into three layers: HBM, server DRAM/DDR5/SOCAMM, and enterprise SSD/eSSD.
1) HBM: From a GPU accessory to a core bottleneck of the AI system
The HBM logic is the hardest because it possesses three things simultaneously: strong demand, high technical barriers, and slow supply expansion. SK hynix cited market views showing that HBM3E is expected to still account for about two-thirds of HBM shipments in 2026, with HBM4 gradually increasing; meanwhile, cloud vendors' self-developed ASICs are also adopting HBM, indicating that HBM demand is no longer just tied to NVIDIA GPUs but has expanded to self-developed chip systems like Google and AWS.
In the competitive landscape, SK hynix remains one of the strongest. Its press release cited Counterpoint data saying SK hynix's HBM shipment share was 62% in Q2 2025 and its revenue share was 57% in Q3, stating it has established a mass production system for HBM4 and strengthened packaging cooperation with TSMC. Samsung stated in Q1 2026 that its Memory business has begun mass production sales of HBM4 and SOCAMM2 for the NVIDIA Vera Rubin platform and is developing PCIe Gen6 SSDs.
Investment Changes: In the past, people bought "total HBM demand"; now they buy HBM4 generational share, customer certification, base die/packaging synergy, 12-high/16-high yields, and multi-generational binding with NVIDIA/ASIC customers. HBM is no longer just a cyclical product but a strategic asset between storage, advanced packaging, and system architecture.
2) DDR5/SOCAMM: The undervalued CPU/inference side bottleneck
With the growth of Agentic AI and inference workloads, the importance of CPUs, main memory, KV cache, and data preprocessing has increased. Samsung clearly mentioned that server memory demand remains strong in H2 2026 and has made DDR5, SOCAMM2, and PCIe Gen6 eSSD/KV cache storage its AI product focus.
This means the storage logic has expanded from "training GPUs need HBM" to "inference and agents need a larger memory hierarchy." This benefits Micron, Samsung, and SK hynix, but the sources of elasticity differ: SK hynix leans toward HBM share, Micron toward price/profit elasticity, and Samsung toward HBM4 catch-up + DDR5/eSSD + foundry/base die combinations.
3) eSSD/NAND: From a common cyclical product to part of the inference infrastructure
In the past, NAND was more like a consumer electronics/enterprise storage cyclical product; now, AI data center checkpoints, embeddings, vector databases, KV caches, data lakes, and high-throughput inference will all drive eSSD demand. SK hynix also mentioned that AI training/inference expansion will drive up DRAM/HBM capacity per server and simultaneously pull up eSSD demand.
Risk: Storage is currently the segment with the strongest fundamentals, but it is also a classic cyclical industry. SK hynix itself warned that HBM prices might enter a correction phase after 2026 due to intensified competition and capacity expansion. Therefore, storage stocks cannot just be viewed as having cheap P/E ratios; one must look at whether HBM4 share is stable, whether ordinary DRAM/NAND prices are near their peak, whether customer long-term contracts are hard enough, and whether capex expansion disrupts supply and demand.
Personal Judgment: Storage remains one of the most certain segments in the AI data center, but after the second half of 2026, the strategy should switch from "comprehensive long" to "buying technical share and long-term contracts, selling pure price cycles."
2. Optical Interconnect/CPO: From "optical module volume and price increase" to "architectural war"
Optical interconnect is one of the segments with the most changes. In the past, the market bought into the scaling of 800G and 1.6T optical modules; the next phase is buying into the value migration among pluggable, LPO, CPO, silicon photonics, and switch ASICs.
NVIDIA has launched Quantum-X and Spectrum-X silicon photonics/CPO architectures, claiming its CPO can bring 5 times the energy efficiency and 5 times the sustained AI runtime compared to pluggable transceivers, while reducing latency. Spectrum-X Ethernet Photonics can reach up to 409.6 Tb/s and is planned to be available in the second half of 2026. NVIDIA also stated its photonics switches achieve 3.5 times the energy efficiency with 4 times fewer lasers, disclosing ecosystem partners including TSMC, Coherent, Corning, Fabrinet, Foxconn, Lumentum, SENKO, SPIL, Sumitomo, TFC, etc.
Broadcom is also advancing another path. Its Tomahawk 6 is already shipping, with a single chip at 102.4 Tbps, supporting 100G/200G SerDes and CPO, targeting AI networks with over a million XPUs. Broadcom's subsequent Davisson CPO Ethernet switch emphasizes that traditional pluggable optics face pressure in power consumption, latency, and system volume, while CPO can reduce optical interconnect power by 70% and support 512 XPU scale-up and 100,000+ XPU two-layer scale-out.
1. New Changes in the Competitive Landscape
This is not a simple "optical module demand continues to grow," but a competition between two routes:
- NVIDIA Full-Stack Network Route: InfiniBand + Spectrum-X + silicon photonics, tied to GPU/AI factory system design.
- Broadcom/Open Ethernet Route: Tomahawk + Ethernet ecosystem, making it easier for cloud vendors to avoid being locked into NVIDIA's full stack.
This is a double-edged sword for traditional optical module companies. In the short term, 800G/1.6T pluggable still has performance, but in the medium term, if CPO is adopted on a large scale, value may shift from front-panel optical modules to optical engines, lasers, silicon photonics, packaging, connectors, testing, and the switch ASIC ecosystem. Broadcom pointed out as early as the Bailly 51.2T CPO switch that traditional pluggable optical transceivers can account for about 50% of the power consumption and over 50% of the cost of a traditional switch system, which is exactly the profit pool CPO is trying to reconstruct.
2. Investment Changes:
Optical interconnect should no longer just ask "how many 800G units are shipped," but rather:
- Has the company secured design-ins with NVIDIA/Broadcom/top cloud vendors?
- Is the product a core capability for 200G/lane, or is it replaceable capacity from the previous generation?
- Will value be enhanced or replaced in CPO?
- Does gross margin come from technical scarcity or temporary supply shortage?
- Will the pace of 1.6T, LPO, and CPO lead to inventory/order mismatches?
Personal Judgment: Optical interconnect is still the major direction, but the market will shift from "total optical module volume logic" to "architectural alignment logic." CPO is the medium-term main line, but it shouldn't be understood as all pluggables being replaced immediately in 2026; a more reasonable pace is: 800G/1.6T pluggables realize first, and CPO is gradually verified in the highest density AI clusters.
3. 800V DC: The direction is very important, but the investment pace cannot be linearly extrapolated too early
The essence of 800V DC is not a simple power product upgrade, but a necessary change in power distribution architecture as AI rack power density rises. NVIDIA clearly stated that the traditional 54V standard has become a bottleneck, and 800V DC can reduce conversion and routing volume, lower distribution losses, reduce current/copper loss/cable volume, and improve tokens per watt. The same page also mentioned that data centers will gradually evolve from current AC distribution to 800V DC, with the future form being a single AC-to-800V DC conversion followed by direct distribution to the compute rack.
New Changes in the Competitive Landscape
800V DC will migrate value from traditional low-voltage power supply, AC distribution, and PSUs to:
- 800V rectifiers / power shelves;
- High-voltage DC/DC;
- Solid-state transformers;
- SiC/GaN power devices;
- High-voltage connectors, protection devices, and busbars;
- Rack-level battery backup / BBU;
- Collaborative design with liquid cooling/cabinets/server systems.
NVIDIA disclosed its 800V DC ecosystem partners cover ABB, Delta, Eaton, GE Vernova, Hitachi Energy, Infineon, LITEON, Schneider Electric, Siemens, STMicro, Vertiv, etc., indicating this is not a single-chip company story but a reconstruction of the entire electrical ecosystem.
Investment Changes:
800V DC is currently more like an "architectural option" and not all companies have entered the earnings realization phase. The most worth tracking is: who enters NVIDIA's reference design, who obtains mass production qualifications from cloud vendors/ODMs, whose products move from prototypes to rack-level shipments, and who can pass safety certification, reliability, thermal management, and maintenance convenience.
Main Risks:
- Standards are not fully unified; internal solutions from NVIDIA/OCP/cloud vendors may differ;
- Maintenance, safety, protection, and certification cycles for high-voltage DC may slow adoption;
- Once reference designs are standardized, profits for power supplies/connectors/protection parts may be squeezed by competition;
- Stock prices may trade on "800V penetration" first, but revenue realization may lag by 4–8 quarters.
My Judgment: 800V DC is a real direction, but it's not the most suitable segment for simple theme-chasing. It should be tracked by "who gets mass production share in subsequent high-power racks like Rubin/Vera Rubin," rather than buying just because a company mentions "800V."
4. Power Sector: Upgraded from "AI capex beneficiary" to "AI capacity licensing right"
Power is one of the most persistent bottlenecks in current AI data centers because it is not a single product but a comprehensive constraint of land, grid connection, transformers, switchgear, cables, UPS, gas turbines, energy storage, PPAs, and regulatory approvals.
IEA expects global data center power consumption to increase from about 415 TWh in 2024 to about 945 TWh in 2030, with AI being a major driver; in the US, data centers will contribute nearly half of the power demand growth by 2030. IEA also warned that about 20% of planned data center projects may face delay risks, as building new transmission lines takes 4–8 years in developed economies, and waiting times for key components like transformers and cables have doubled in the past three years.
This is why electrical equipment companies like Eaton still have strong orders and backlogs. Eaton Q1 2026 disclosed that its Electrical sector backlog grew 48% year-on-year, and Electrical Americas 12-month rolling average orders grew 42%, stating that data center momentum is driving demand.
New Changes in the Competitive Landscape
The power sector cannot be viewed as a whole. It is divided into at least four layers:
- Grid Equipment: Transformers, switchgear, circuit breakers, cables, protection and control.
- Internal Data Center Power Chain: UPS, PDU, busway, switchgear, rack power, BBU.
- Generation Side: Natural gas, gas turbines, nuclear power, SMRs, geothermal, energy storage, PPAs.
- Engineering and EPC/System Integration: Whoever can deliver available MW on time has pricing power.
In the short term, the hardest part is electrical equipment backlog + delivery capability; in the medium term, it's grid connection and generation resources; in the long term, it's regions/companies with low-cost, dispatchable power. Reuters also cited the EIA saying US power demand will continue to hit new highs in 2026 and 2027, driven by AI data centers and electrification.
Investment Changes:
Power is no longer just a "downstream supplier of data center capex" but a prerequisite for whether AI capacity can be deployed. For cloud vendors, if GPUs are insufficient, they can be bought; if HBM is insufficient, they can wait in line; but grid-connected MW and transformer lead times cannot be solved immediately with money.
Main Risks:
- Power equipment valuations are already quite crowded;
- Project delays will push revenue back;
- Utility stocks are constrained by regulation, interest rates, and return on capital expenditure;
- Gas turbine/nuclear/SMR narratives are often long-term correct but slow to realize in the short term.
Personal Judgment: Power is one of the "physical bottlenecks" with the best sustainability in AI data centers, but priority should be given to equipment and system deliverers, followed by generation assets, and finally ordinary utility beta.
**
5. CPU: Re-emerging as a critical path in AI systems from a "GPU accessory," but profit pools will be internalized by cloud vendors
The change in CPU logic is underestimated by the market. In the past, everyone understood AI servers as being dominated by GPUs/HBM, with the CPU just being a host processor. But agentic AI, inference, tool calling, data preprocessing, sandboxing, retrieval, and network/storage virtualization will all push the CPU back onto the critical path.
Google explicitly stated in its Axion introduction that general-purpose computing remains a key part of customer workloads; Amdahl's Law means that as accelerators continue to improve, general-purpose compute will become a cost and capability limit. Google claims Axion performs up to 30% better than the fastest Arm cloud instances at the time and up to 50% better than comparable x86 instances, with up to 60% better energy efficiency. A 2026 Arm article further pointed out that agentic systems will increase concurrency, latency sensitivity, and overall computing demand, putting the CPU on the critical path to success; Google TPU 8t/8i also uses Axion as a header to reduce data preparation latency and maintain TPU utilization.
AWS data also shows that CPUs/self-developed chips are not a side story. Amazon Q1 2026 stated its chips business, including Graviton, Trainium, and Nitro, has an annualized revenue run rate exceeding $20 billion and is growing at triple digits year-on-year; it also mentioned that OpenAI will start consuming about 2GW of Trainium capacity from 2027, and Anthropic will receive up to 5GW of Trainium.
New Changes in the Competitive Landscape
One must be very cautious with CPUs. Demand growth is real, but it doesn't mean all merchant CPU companies will benefit equally.
- Cloud Vendor Self-Developed Arm CPUs: AWS Graviton, Google Axion, Microsoft Cobalt will internalize part of the CPU profit within hyperscalers.
- AMD EPYC: Still has strong competitiveness in high core counts, memory bandwidth, PCIe lanes, and the x86 ecosystem, especially suitable for GPU hosts, general cloud, databases, and enterprise workloads.
- Intel Xeon: Still has a huge installed base, enterprise ecosystem, and AMX/acceleration instructions, but process, power consumption, share, and gross margin pressures remain key.
- NVIDIA Grace/Vera CPU: If system architecture becomes further rack-scaled, the value of NVIDIA CPUs is not just as a CPU but as part of the NVLink/NVL rack system.
Investment Changes:
CPUs are not the scarcest AI assets, and valuation elasticity is usually not as good as HBM, advanced packaging, power, and core networks. But the strategic importance of CPUs is rising, especially in inference/agentic AI. The real question is: will the profit from new CPU demand flow to merchant CPUs or be absorbed by hyperscalers' self-developed chips and system-level optimizations.
Personal Judgment: The CPU sector logic is improving, but it is more like an "AI system efficiency/cost optimization" main line rather than a pure shortage main line. The strongest is not a single CPU, but system control over CPU + accelerator + memory + network + software scheduling.
6. Advanced Packaging
AI chips are getting bigger, HBM is stacking more, and CoWoS continues to expand, but the pressure on efficiency, area, yield, and cost of "circular wafer + silicon interposer/organic substrate" will increase. CoPoS is a production form prepared for the next generation of ultra-large AI/HPC packaging.
TSMC's existing CoWoS is still the main force for AI/HPC. The official CoWoS page clearly states: CoWoS-S uses a silicon interposer for ultra-high performance computing like AI and supercomputing; CoWoS-S can reach about 3.3 times the reticle size, and for sizes beyond this, CoWoS-L or CoWoS-R is recommended. TSMC's 2026 technology forum further pushed the CoWoS roadmap to 5.5x and 14x reticle; 14x CoWoS is expected to be produced in 2028, integrating about 10 large compute dies and 20 HBM stacks.
CoWoS = The main battlefield for current AI GPUs/HPC.
It is already in mass production, with clear customers and tight capacity, making it the current performance main line.
CoPoS = A medium-to-long-term platform for next-generation ultra-large AI/HPC packaging.
It is not just a "packaging and testing plant business" but an extension of TSMC's 3DFabric system from wafers, RDL/interposers, chiplets, HBM, and substrates to system-level packaging.
FOPLP = Fan-out panel-level packaging, which is related to but not equal to CoPoS.
FOPLP can be used for mid-to-low to mid-to-high-end applications, such as RF, automotive, power, optical communication, and some AI server power/modules; CoPoS is more oriented toward large AI/HPC packaging, requiring stronger interposer/RDL/TGV/yield control. CommonWealth's comparison is very direct: CoWoS is mature and in mass production; both CoPoS and FOPLP shift from wafer to panel, but CoPoS is geared toward larger AI chips, while FOPLP is more used for thinner, high-density chips like mobile, automotive, and IoT.
CoPoS will reshape the AI advanced packaging supply chain, but what's truly worth tracking is not the words "glass substrate," but "who can help TSMC make ultra-large AI packaging, make it stable, and make it cheap."
From this perspective, the most worth studying are: TSMC platform, CoWoS/CoPoS equipment chain, RDL/TGV/wet processes, metrology and yield, ABF/glass core substrates, and key materials.
IV. "New Investment Conclusions" for Each Segment Now
V. How Your Investment Portfolio Should Be Positioned
You can screen companies with three questions:
Question 1: Is it a "real bottleneck" or "theme-related"?
Real bottlenecks are characterized by: lengthening lead times, customers willing to prepay or sign long-term agreements, rising ASP, expanding gross margins, and slow capacity expansion.
Theme-related ones are characterized by: rising revenue but stagnant gross margins, highly concentrated customers, high capex but unclear ROIC, or just "entering the supply chain" with a very small share.
Currently closer to real bottlenecks are: HBM, high-end server DRAM/eSSD, advanced packaging, transformers/switchgear/grid equipment, and some high-end optical interconnects.
Question 2: Will its value rise in architectural upgrades, or will it be replaced by architectural upgrades?
This is the most critical question for optical interconnects, 800V DC, and CPUs. CPO will compress the value of some traditional pluggable modules but will raise the value of silicon photonics, optical engines, lasers, connectors, packaging and testing, and switch ASICs. 800V DC will raise the value of the high-voltage power chain, SiC/GaN, protection devices, and rack power architecture, but may compress some traditional low-voltage power components.
Question 3: Can downstream cloud capex continue to be tolerated by the capital market?
This is the most important valuation ceiling for 2026. Amazon Q1 2026 trailing twelve-month free cash flow dropped to $1.2 billion, mainly because PPE purchases increased by $59.3 billion year-on-year, reflecting AI investment. Reuters reported on June 10 that Amazon obtained a $17.5 billion loan arrangement and stated that Big Tech's AI-related spending is expected to exceed $700 billion this year, with large tech companies increasingly financing through debt and equity markets. Oracle's case is more extreme: Reuters reported its FY2027 capex could reach up to $95 billion, with plans to finance nearly $40 billion through debt and equity, sparking market concerns about cash consumption and debt.
This means that the valuation of the upstream supply chain cannot just look at orders; it must also look at downstream customers' AI revenue conversion, token economics, depreciation drag, and the degree of free cash flow pressure. Once the market starts questioning AI capex ROI, the first to be hit will be segments with "pure expectations + high valuations + unrealized profits."
https://x.com/168X_Fortune/status/2062463349573718318
When I talked with 168X before, we also mentioned that the surge from April to May was driven by two core factors: first, cloud vendors verified that large capital expenditures brought better-than-expected growth in cloud business, and cloud business monetization comes from token consumption on the B-side and C-side; second, Anthropic's annualized revenue grew rapidly.
These two are the base points supporting the entire business logic now: AI commercialization has entered an inflection point and is growing rapidly.
V. Final Investment Logic and Main Lines
AI data center investment has moved from "where is capex flowing" to the stage of "where are the bottlenecks, how long can the bottlenecks last, who controls the next-generation architecture, and who ultimately keeps the profit."
My current relative judgment for each segment is:
- Storage/HBM/eSSD: The strongest earnings realization main line. But after 2026, pay attention to supply expansion and price cycles; it's not advisable to chase highs indiscriminately.
- Optical Interconnect/CPO: The technical direction is very strong, but the competitive landscape is being reconstructed. Future winners may not be the companies with the largest gains in 800G modules in the past, but those that secure platform positions with NVIDIA/Broadcom/cloud vendors.
- 800V DC: The general direction is certain, but short-term revenue pace is uncertain. This is a necessary evolution for high-power AI racks, but investment should wait for design-in, certification, and mass production pace to land.
- Power Sector: The most durable physical bottleneck. But distinguish between equipment manufacturers, generation assets, utilities, and EPCs; priority goes to electrical equipment manufacturers with backlogs, delivery capabilities, and pricing power.
- CPU: Importance is rising, but it's not a simple shortage logic. Agentic AI puts the CPU back on the critical path, but hyperscalers' self-developed chips will internalize part of the profit pool, and the valuation elasticity of merchant CPU companies needs more cautious evaluation.
The most dangerous investment is: buying a company just because it is "related to AI data centers"; giving high multiples just because it mentions CPO/800V/liquid cooling; looking only at revenue growth without looking at gross margin, share, customer concentration, and architectural replacement risks.
The most attractive investment is:
Bottleneck assets that have already seen earnings growth, but the market underestimates their sustainability;
Or control point assets that haven't fully realized earnings yet but have already been locked in by next-generation architectural platforms.
Simply put, the market has recognized the bottlenecks and choke points of the AI computing power industry chain, so investment must go a step further: it's not enough to just stand at the bottleneck or be in a choke point link. You also need to see if this enterprise is unique and irreplaceable at this bottleneck and choke point node, if its technical route is correct, and if it not only has a lot of orders but can also continue to exceed expectations in delivery. At the same time, you must consider how the pace of technical evolution and landing actually unfolds.
