Where the Money Goes in AI: The Value Chain, the Bottlenecks, and the Next Investment Window

---

As a casual investor, I feel like we’re all drowning in AI information overload right now. One day a stock skyrockets by double digits; the next, another loses half its value. Honestly, I can’t remember the last time the market was this exciting. Driven by sheer curiosity, I decided to look at the AI value chain systematically—partly to make sense of what we're actually investing in, and partly to see what the future holds. I’m definitely no expert, so think of this as just a fun, casual deep dive!

The AI boom is often described as a software revolution, but the first large pools of money have actually gone into physical infrastructure. If your original instinct was that NVIDIA captured the first obvious wave, HBM suppliers like SK hynix and Micron captured a parallel memory wave, and server integrators like Dell captured the next deployment wave, you are broadly right. But the AI value chain is now widening into areas that used to look like slow-growth industrial markets: transformers, switchgear, substations, UPS systems, solid-state transformers (SSTs), silicon carbide (SiC) power devices, 800V DC power architecture, liquid cooling, optical interconnects, utilities, gas turbines, nuclear power, and grid construction.

Instead of a simple "chips, data centers, applications" mental model, I find it much more useful to view this through the lens of bottleneck migration. AI spending consistently moves toward whatever layer is most scarce at any given moment.

2023–2025: The bottleneck was mostly compute.

2025–2026: The bottleneck widened into HBM, packaging, servers, and networking.

2026 onward: The bottleneck is increasingly power, cooling, grid access, and cluster efficiency.

Over the longer run, the largest value pool will still live at the application layer, but applications aren't the immediate physical bottleneck because software can be copied faster than transformers, power plants, and data centers can be built.

Jensen Huang’s AI Stack: A Better Way to Organize the Competition

I've noticed that investors often ask a binary question: "Who wins AI?" A much better question to ask is: "Which layer is the bottleneck now, and which layer has the power to capture margin?" The answer changes over time.

Jensen Huang has repeatedly framed AI as a multi-layer infrastructure stack rather than a single product category. In his public discussions around the World Economic Forum and CSIS, he described the AI economy as a layered system beginning with energy, then chips and systems, then infrastructure and cloud data centers, then models, and finally applications.

This framing matters because it shows why NVIDIA's real advantage is not only its GPU chip; it is a system-level stack spanning accelerators, networking, CUDA, libraries, reference architectures, racks, and cloud partnerships.

Similarly, look at how China is responding. They aren't just trying to build a weaker replacement GPU; Huawei’s strategy shows a system-level alternative. They are using available process nodes, combining many chips, employing optical or high-bandwidth interconnects, implementing custom memory approaches, and driving software/toolchain localization to compete at the cluster level.

Here is how I visualize the current state of the stack:

AI stack layer	Main question	Current bottleneck?	Main winners today	Strategic point
Energy and power	Can enough reliable electricity be delivered?	Very high	Utilities, IPPs, grid equipment, turbines, substations, transformers, SST and SiC suppliers	The newest bottleneck; lead times can be measured in years
Chips and systems	Can enough compute be produced at the right cost per token?	High but evolving	NVIDIA, AMD, Broadcom, hyperscaler ASIC ecosystems, HBM suppliers, foundries, packaging vendors	Competition is moving from chip versus chip to system versus system
Infrastructure and cloud	Can chips be deployed, cooled, networked, and utilized?	High	Dell, ODMs, Arista, Broadcom, optics vendors, data-center developers, cooling vendors, cloud platforms	Deployment capacity is becoming as important as chip allocation
Models	Can model quality improve efficiently?	Medium	OpenAI, Anthropic, Google, Meta, xAI, DeepSeek-like open models	Model efficiency can reduce pressure on hardware, but frontier competition still hot
Applications	Can AI generate durable revenue and productivity?	Low physical bottleneck, high business uncertainty	Enterprise software, vertical AI, agents, cybersecurity, productivity suites	Long-term value pool may be largest, but winners are harder to identify

Mapping the AI Value Chain: A Comprehensive View

To understand where the money is flowing, I find it useful to break down the entire chain of industries. Some have already benefited visibly; others are only starting to see a re-rating because their order books, pricing power, or lead times are changing radically.

1. AI Accelerators and GPUs

Representative segments/companies: NVIDIA, AMD, Intel Gaudi, Huawei Ascend.

2. Custom AI ASICs

What they provide: Lower-cost, workload-specific chips.

Representative segments/companies: Google TPU, AWS Trainium/Inferentia, Microsoft Maia, Meta MTIA, Broadcom ASIC ecosystem.

Why AI boosts them: Hyperscalers want lower inference costs and less dependence on merchant GPUs.

3. High-Bandwidth Memory (HBM)

What they provide: HBM3E, HBM4, and advanced server DRAM.

Representative segments/companies: SK hynix, Micron, Samsung.

Why AI boosts them: AI accelerators are fundamentally memory-bandwidth constrained.

4. Foundry and Advanced Nodes

Representative segments/companies: TSMC, Samsung Foundry, Intel Foundry, SMIC (for China-constrained nodes).

5. Advanced Packaging

What they provide: CoWoS, chiplets, and 2.5D/3D integration.

Representative segments/companies: TSMC, ASE, Amkor, substrate suppliers, packaging equipment.

Why AI boosts them: Logic and HBM must be integrated tightly to achieve necessary performance. The importance and complexity of the advanced packaging have been rising day by day.

6. Semiconductor Equipment

What they provide: Lithography, etch, deposition, and inspection tools.

Representative segments/companies: ASML, Applied Materials, Lam, KLA, Tokyo Electron, Naura, AMEC.

Why AI boosts them: self explanatory, but what's often missed out is that new equipment is only needed when there is new fab, the economic of equipment business is very different from "consumables".

7. EDA and Design IP

What they provide: Chip design software and foundational IP blocks.

Representative segments/companies: Synopsys, Cadence, Siemens EDA, Arm, RISC-V ecosystems.

Why AI boosts them: The explosion of custom chips requires significantly more design tools.

8. AI Servers and ODMs

What they provide: GPU servers, rack-scale systems, and hardware integration.

epresentative segments/companies: Dell, Supermicro, HPE, Quanta, Wistron, Wiwynn, Foxconn Industrial Internet.

Why AI boosts them: Massive amounts of compute must be physically assembled and delivered into functioning units. Think not just CSPs but also the enterprises with their own on-prem setup, these are two very different segments.

9. Networking and Switching

Representative segments/companies: NVIDIA, Broadcom, Arista, Cisco, Marvell.

10. Optical Interconnects

What they provide: 800G/1.6T transceivers, lasers, and silicon photonics.

Representative segments/companies: Coherent, Lumentum, Innolight, Eoptolink, Fabrinet, optical chip vendors.

Why AI boosts them: Data movement across clusters has become a major physical bottleneck.

11. Data-Center Developers and REITs

What they provide: Land, powered shells, and colocation services.

Representative segments/companies: Equinix, Digital Realty, private hyperscale developers.

12. Construction and Engineering

What they provide: Data-center design, buildout, and commissioning.

Representative segments/companies: Engineering contractors, electrical contractors, EPC firms.

13. Power Equipment

What they provide: Transformers, switchgear, UPS systems, PDUs, and busways.

Representative segments/companies: Schneider Electric, Eaton, ABB, Siemens, GE Vernova, Vertiv, Hitachi Energy.

Why AI boosts them: Exponentially higher rack density requires far more power conversion and electrical protection.

14. Advanced Power Electronics

What they provide: SiC, GaN, SSTs, 800V DC architectures, and medium-voltage conversion.

Representative segments/companies: Wolfspeed, Infineon, STMicro, onsemi, Rohm, Navitas, Power Integrations, Heron Power, DG Matrix.

Why AI boosts them: AI data centers require denser, vastly more efficient power delivery to prevent massive energy loss.

15. Liquid Cooling

What they provide: Direct-to-chip cooling, immersion cooling, CDUs, cold plates, and fluids.

Representative segments/companies: Vertiv, Schneider, CoolIT, Boyd, Johnson Controls, Carrier, Modine, nVent, dielectric fluid suppliers.

Why AI boosts them: Rack densities reaching 60–100kW (and up) physically challenge the limits of traditional air cooling.

16. Grid Interconnection

What they provide: Substations, transmission lines, controls, and metering infrastructure.

Representative segments/companies: Utilities, transformer makers, grid automation suppliers.

Why AI boosts them: New data centers require massive, immediate upgrades to regional electrical grids, especially when renewable energy is increasingly onboarded to the grid.

17. Power Generation

What they provide: Nuclear power, gas turbines, renewables, utility-scale storage, and PPAs.

Representative segments/companies: Constellation, Talen, Vistra, NextEra, GE Vernova, Siemens Energy, Caterpillar backup power.

Why AI boosts them: self explanatory, but do not underestimate carbon-free electricity mandate in many countries.

18. Cloud and Neo-Cloud

What they provide: GPU/ASIC capacity delivered as a service.

Representative segments/companies: AWS, Azure, Google Cloud, Oracle, CoreWeave.

Why AI boosts them: They convert raw physical infrastructure into accessible, on-demand customer services. Though personally, I have strong negative biases towards BMaaS operators.

19. AI Software and Applications

Representative segments/companies: Model labs, software companies big and small, cybersecurity providers, data platforms.

The Chip Layer: Turning Multi-Dimensional

The chip layer captured the first wave because AI capacity began with accelerator scarcity. The numbers here are mind-boggling: NVIDIA reported FY2025 revenue of $130.5 billion (up 114% year over year), with its Data Center segment hitting $115.2 billion (up 142%). I read some reports saying that AI chips are approaching half of global chip sales, despite representing less than 0.2% of total chip unit volume.

However, the competition inside the chip layer is becoming multi-dimensional. It’s no longer a simple story of "NVIDIA vs. AMD." The dynamics have fractured into several key battles:

Training vs. Inference

Training frontier models still favors the most capable GPU clusters because software maturity, memory bandwidth, and networking are everything. But inference is a different game. Inference is about cost per token, latency, throughput, power efficiency, and capacity utilization.

This explains why hyperscalers are aggressively building their own chips. Google’s 2026 AI infrastructure announcements include the TPU 8t for training and the TPU 8i for inference and reinforcement learning. Google notes that the TPU 8i targets 80% better performance-per-dollar ratio for inference than its prior generation. Similarly, AWS positions Trainium as purpose-built accelerators, with Trainium2 offering 30–40% better price performance than certain GPU-based EC2 instances, and Trainium3 explicitly targeting agentic, reasoning, and video-generation workloads.

NVIDIA can remain utterly dominant in frontier training while simultaneously facing market share pressure in internal hyperscaler inference. AI compute demand can skyrocket while the mix of who captures the margin changes completely.

Chip vs. System: Huawei’s Alternative Path

The unit of competition is moving from the chip to the rack, the cluster, and the entire data center. Look at how Huawei is playing the game under heavy constraints. Their CloudMatrix 384 system uses 384 Ascend 910C chips, a supernode architecture, high-speed internal interconnects, and optical links for low-latency communication. The claim isn't that a single Ascend chip beats NVIDIA's top silicon; the claim is that a larger, tightly interconnected system can compete on system-level metrics. The trade-off is power efficiency, but China can then play to its strength in power dominance.

HBM and Advanced Packaging: The Hidden Choke Points

The stock price of HBM suppliers have been doubling, and doubling, for the last few months. They are direct beneficiaries of the AI boom because modern accelerators are limited not just by processing power, but by how quickly they can access data.

Advanced packaging is the twin bottleneck. AI chips require the incredibly tight integration of logic, HBM, substrates, and interposers. TSMC’s CoWoS ecosystem became highly critical because packaging determines whether the chip and memory can actually work together at acceptable bandwidth and power levels. The future of AI hardware isn't simply smaller transistors; it's memory proximity, chiplets, packaging yield, substrates, thermals, and system-level design.

Semiconductor Equipment and China Localization

I have never believed in export controls being anywhere effective for a market of 1.4 billion people. They are leaky at best, and most of time, they serve to surrender a huge market for domestic players to catch up "safely and patiently" with a captured customer base.

Chinese equipment suppliers like Naura, AMEC, ACM Research, and Piotech posted record 2025 revenues as local fabs expanded and sourced around export controls. Industry reports from eeNews Europe note that while direct U.S.-origin chipmaking equipment shipments to China fell, imports through alternative routes and foreign toolmaker exposure to China remained highly significant.

To route around hardware restrictions, Huawei even proposed a "Tau Scaling Law" and a related LogicFolding architecture. This tells me that scarcity is always the mother of great invention. The restrictions are pushing China to compete across a massive surface area, and grow domestic eco-system aggressively.

Servers, Networking, and Optics: The Deployment Wave

Once chips are secured, the next massive hurdle is turning them into deployed capacity. Dell illustrates this beautifully: stock overnight jumped 40%. What I did not foresee is that Dell is able to increase its profit margin despite rising component cost. It shows that their business is much more than commoditized box moving.

Optical interconnects are emerging as a prime high-growth bottleneck. TrendForce projects the global AI optical transceiver market will jump from $16.5 billion in 2025 to $26 billion in 2026 (a 57%+ YoY growth rate), driven by shortages in optoelectronic chips, alignment capacity, power consumption, and thermal management. I have been monitoring several China-based companies in this area with strong order book globally.

The Industrial Supercycle: Data Centers, Power, and Cooling

AI data centers are not ordinary real estate assets; they are industrial factories designed to convert raw electricity into intelligence. My family is intimately familiar with this line of business, and it deserves a post on its own.

Power Equipment: The Underappreciated Profit Pool

As AI racks become denser, the bottleneck shifts from digital scarcity to electrical scarcity.

This is one of the clearest next-window investment opportunities because it combines extreme demand visibility with long manufacturing lead times and limited substitution. Look at Schneider Electric: their Q1 2026 revenue grew 11.2% organically to €9.77 billion, explicitly citing AI data-center demand for power and cooling as the key driver. But again, power business is notoriously local, fragmented and regulated, and it does take courage to throw your investment dollar into this industry.

Why SST, SiC, GaN, and 800V DC Architecture Matter

The power story is undergoing a fundamental technological shift. AI racks are moving to massive power densities, prompting the adoption of NVIDIA's 800V HVDC architecture to reduce wiring, free up rack space, and improve efficiency in megawatt-scale racks. Wolfspeed argues that these 800V DC architectures require 1200V silicon carbide (SiC) MOSFETs for AC-DC rectification and DC-DC conversion, which can slash conversion losses by 25–40%.

Solid-State Transformers (SSTs) are also entering the chat. Unlike traditional copper-and-iron transformers, SSTs use power semiconductors, high-frequency switching, and digital controls to convert medium-voltage grid power directly into high-voltage DC distribution. Wide-bandgap semiconductors like SiC and gallium nitride (GaN) are crucial here because they support higher voltages, faster switching, and vastly superior power density.

While still early, I believe SST market will grow significantly. Emerging startups like Heron Power (with its Heron Link medium-voltage AC to 800V DC system) and DG Matrix (with its Interport multi-port architecture) are starting to compete alongside massive incumbents like Hitachi Energy, ABB, Siemens, Eaton, and Schneider. If AI data centers are factories, the factory floor needs a completely new electrical architecture.

A Personal Note on Investing: Know the Bottleneck, but Know Yourself First

There is another layer to this discussion that is far more personal than industrial. Finding the next AI bottleneck is only half the battle; the other half is knowing whether you have the temperament to invest in it rationally.

A fascinating article summarizing Swedish twin research noted that roughly 45% of investing behavior may be genetically influenced—including tendencies toward excessive trading, performance chasing, home bias, and the disposition effect.

That finding resonates deeply with me. Investing isn't simply about reading a full shelf of books, memorizing valuation formulas, or ingesting behavioral-economics papers. Those things help, but they don't automatically rewrite who we are under acute stress. When prices plummet, when a popular theme runs ahead without us, or when a position doubles and then drops sharply, our real character appears.

The encouraging part of the research is that genetics is not destiny. The variation also comes from your environment, your education, your habits, and your systems. The research found that direct work experience in finance can significantly reduce genetic predispositions to investment biases, even though general education alone doesn't seem to wipe them out.

The practical lesson here isn't fatalism; it’s humility. If part of our investing behavior is hardwired, the rational response is to design an investment process that works with our nature rather than pretending we are perfectly rational, emotionless machines.

My core investment principle is this: only invest in things that I know.

In that sense, the most critical investment bottleneck may not be chips, HBM, transformers, or electricity. It might just be our own behavior.