GearFlow Lab

recent posts

about

Just an analyst looking at the numbers and the gear. Breaking down small-cap tech plays, defense algorithms, and the systems that drive them. No noise, just data.
The Ultimate 30-Year AI Supercycle Guide: Capital Flow

The Ultimate 30-Year AI Supercycle Guide: Capital Flow

Mapping the 3 Phases of the AI Supercycle
and Infrastructure Megatrends

Part 5 To master the AI Supercycle, we must look at the big picture. [Part 5] closed with a warning: the map of the semiconductor empire… drawn across this series was built for the relatively peaceful era led by mobile and PC — and a tidal wave was rewriting every rule at once. This installment maps that wave. The AI Supercycle is not a theme trade. It is a 30-year structural repeat of the PC/internet/smartphone buildout, compressed and overlapped, and it changes where capital flows on every district of the map built in Parts 1 through 5.

November 2022: The iPhone Moment of AI

The starting gun fired in November 2022. The same link circulated through investor and tech group chats several times a day, and the initial reaction was dismissive. Face recognition worked. Translation worked. Phones already shipped plenty of features labeled “AI.” This looked like one more chatbot — and in late 2022, “chatbot” was in fact the more common label than “LLM” (large language model).

Two weeks in, then a month in, the perception shifted: something qualitatively different had arrived. Programmers used it as a debugging assistant. Marketers pulled ad copy from it. Students drafted papers. Investors had it summarize industry reports.

Chart explaining the 30-Year AI Supercycle and infrastructure megatrends

What made ChatGPT different was not raw capability but access. Prior AI existed as equations in research papers, as functions in API documentation, as technology buried somewhere inside a product. ChatGPT required no installation, worked immediately after login, and ran on the one interface every human already knows: conversation. For the first time, AI was not an app — it was a counterpart you talk to.

That distinction matters enormously to investors, because technology reorders industries not when performance improves, but when ordinary people adopt it at scale. Touchscreens existed before 2007; the industry reorganized only when the iPhone and the App Store put them in everyone’s hands. ChatGPT was AI’s iPhone moment — the point where the technology left the lab and entered mass perception.

Scaling Laws: When AI Performance Became a Capital Allocation Problem

Early users noticed something unsettling: a “language model” was writing working code and structuring tables and project plans. It seemed to understand, not merely predict. Researchers labeled this emergent ability — capabilities that appear abruptly past certain thresholds. The mechanism underneath carries the investment thesis: the scaling law. Grow a model’s parameter count, training data volume, and compute in a disciplined ratio, and performance improves along a clean, predictable curve.

Read that sentence as an allocator and the implication is explicit: AI performance stopped being a matter of research luck and became a function of capital and infrastructure. Pour in more compute and data, and improvement follows on schedule. That predictability is what unlocked the CAPEX flood — corporations could now underwrite the investment against expected capability gains.

And past certain scale thresholds, abilities beyond next-word prediction — reasoning, coding, summarization, translation — kept emerging. Whether this constitutes “real” intelligence remains fiercely debated among experts. For investors, the debate is secondary to the fact: scaling alone keeps unlocking commercially useful capabilities nobody fully predicted.

The architecture of the industry changed with it. Pre-ChatGPT AI was a collection of single-purpose tools — face recognition, spam filtering, recommendation engines, speech recognition — each a component doing one job. Post-ChatGPT, one model writes, codes, plans, summarizes meetings, drafts product documentation, generates support scripts, and personalizes itself to a user’s style.

AI stopped being a function bolted onto Excel and became something closer to an operating system layer — the way iOS and Android became platforms carrying thousands of apps, LLMs are becoming the intelligence platform carrying industry-, role-, and company-specific agents on top.

The Boundary Conditions: AGI, ASI, and LeCun’s Objection

Two terms flooded the press after ChatGPT: AGI (artificial general intelligence — AI that understands and handles diverse tasks at roughly human level, flexibly across domains) and ASI (artificial superintelligence — intellect far beyond human, the sci-fi super-AI). Some argue AGI is imminent; others argue today’s language models are nowhere near it.

Meta’s ($META) chief AI scientist Yann LeCun has repeatedly staked out the skeptical position: today’s LLMs are remarkable tools, but genuine intelligence requires understanding the physical world, internalizing common sense and physics, and planning long-horizon actions — none of which LLMs alone deliver.

His most useful comparison for investors: the world humans perceive is overwhelmingly video, image, and sound; text is a tiny fraction. Yet AI to date has been built almost entirely on humanity’s text archive. The next stage, LeCun and others argue, must fuse multi-modal data — video, image, audio, sensor streams — and embody itself through robotics and autonomous systems, i.e., physical AI.

A scatter plot matrix titled 'Attacking Humanity's Scarcest Resources' with axes for 'Saving Time (Labor Constraint)' and 'Saving Distance & Cost (Capital Constraint)'. It features four colored bubbles moving from bottom-left to top-right: PC (1984), Internet, Smartphone, and a large purple bubble at the peak for Artificial Intelligence.

The investment translation: the LLM revolution is not the end of the AI journey. It is closer to the starting line. Today’s AI lives as text on screens, apps on phones, software on PCs. The defining keyword of the coming decade is physical expansion — AI that interacts with the real world through a body — and that is a co-evolution of software and hardware with semiconductors at dead center.

On the technology-history ladder — movable type, the steam engine, electricity, the internet, the smartphone — AI plausibly slots somewhere between movable type and the internet in impact: printing collapsed the cost of copying knowledge, the internet collapsed the cost of transmitting it, and AI collapses the cost of generating, analyzing, combining, and applying it.

The 30-Year Pattern: 1984 and 2016

Some compare this AI cycle to the smartphone revolution, others to the internet. The more accurate frame: the PC/internet revolution and the smartphone revolution arriving stacked on top of each other. Which means the cycle is longer than it looks, and still early.

Set the historical anchors. PCs entered homes and offices in the mid-1980s — symbolically, around 1984. Computers existed before, but as institutional machinery; the PC was the first computer an individual touched. By today’s standards a mid-80s PC barely outperformed a calculator, useful for word processing, simple spreadsheets, primitive games. That modest start launched the megatrend running through Microsoft Windows, Intel CPUs, the software industry, the internet, and today’s cloud and mobile.

A stacked area chart titled 'Megatrends: Stacked, Not Replaced' showing the geometric compounding of technology eras from 1980 to 2040. The colored layers stack sequentially from the bottom up: PC (Computation) in gray, Internet (Connection) in blue, Smartphone (Extensibility) in cyan, and AI (Understanding/Thought) in purple.

AI’s equivalent moment for the general public arrived in 2016 with AlphaGo — the match against Lee Sedol, then the world’s top Go player. Before that broadcast, AI was a game character or a movie premise. AlphaGo planted AI in mass consciousness the way 1984 planted the PC. Place the two events side by side and you get the first signal of a 30-year pattern repeating on a 30-year interval.

Almost nobody standing in 1984 predicted the next three decades would run through the internet, smartphones, and cloud. The symmetric claim: we are almost certainly under-imagining how far AI’s 30-year arc extends.

One framework keeps megatrends legible. Humanity’s two scarcest resources are money and time — in economic language, capital and labor. Every transformative technology attacked both at once. The PC multiplied calculation and document work hundreds to tens of thousands of times past human hands. The internet erased geographic distance from information exchange. The smartphone removed even the need to sit down — services became reachable on the subway, in bed, mid-stride. PC = computation revolution; internet = connection revolution; smartphone = always-on extensibility.

A dual timeline wave chart titled 'The 30-Year Repeating Interval' showing two symmetric arcs. The first blue wave marks the '1984-2014: PC / Internet / Mobile Arc' starting from the 1984 PC Era. The second purple wave marks the '2016-2046: The AI Supercycle Arc' starting from AlphaGo in 2016.

AI now moves past retrieving information to understanding, combining, summarizing, and restructuring it — a direct extension of human thought — and it will compound in reach as it lands in devices: phone assistants, PC copilots, cars, robots, wearables, AR glasses. The internet handled connection; AI assists — and increasingly substitutes for — the thinking that runs on top of the connection.

The Investor’s Replay of 1984–2024

Run the last 30 years through an allocator’s lens, because the sequence is the template:

Platform + silicon first. Microsoft (MSFT)took the operating system and office software; Intel(INTC) took the CPU. “Wintel” owned the PC era’s software platform and hardware brain simultaneously.

Then the network buildout. Installed PCs needed connecting, and Cisco ($CSCO)-class network equipment makers — routers, switches, optical gear — became the bridges and tunnels of the internet highway, harvesting global telecom infrastructure CAPEX.

Then software and services on top. Google (GOOGL)invented the search−plus−advertising monetization model; Yahoo, Amazon(AMZN), and the portal/commerce/mail cohort built the web-services ecosystem.

Then the device detonation. The internet’s real explosion came not from the PC but the smartphone. The iPhone introduced a mobile OS and the App Store as a platform; Apple ($AAPL) and the Android camp distributed pocket computers to the planet; the FANG platforms — Facebook/Meta, Amazon, Netflix, Google — seized the mobile/cloud service economy. And invisibly underneath, data center and cloud infrastructure investment exploded to carry it all.

Now overlay the AI era onto that template. The MS-DOS/Windows role belongs — not yet locked as a standard, but functionally — to OpenAI and the LLM developers: LLMs are the operating system of the AI era. Intel’s CPU seat belongs, so far, to $NVDA — supplying the GPUs that train LLMs and run AI services, sitting at the center of AI data center investment, growing exactly as Intel did by raising compute performance and shipping more chips.

As models scale, physical infrastructure investment follows: data centers, networks, power infrastructure, cooling, power semiconductors, grid. The Cisco seat of this cycle belongs to the companies building AI-dedicated data centers — power infrastructure/HVDC/ESS, high-performance networking, optical interconnect.

And after infrastructure, software. The sequence is already running in order of procurement sophistication: B2G first — national security, intelligence analysis, infrastructure management, public administration — where Palantir ($PLTR) is the canonical name winning contracts. B2B second — internal document search, workflow automation, accounting/legal/support, manufacturing optimization — where agentic AI (agents that execute tasks autonomously) is moving from pilot to production.

B2C last and widest — personal assistants, education, healthcare, personal finance, entertainment — riding smartphones and whatever device comes next. Which raises the device question: the smartphone era needed the iPhone to detonate the internet’s reach, and the AI era needs the device that best embodies AI.

A horizontal bar chart visualizing the scale of physical and economic integration as AI lands in devices. Colored bars scale upwards from PC Copilots and Phone Assistants at the top, to Wearables, Autonomous Cars, and Physical Robotics at the bottom with the largest reach.

Candidates abound, but many investors converge on $TSLA holding the physical device platform of the AI era through autonomous vehicles and humanoid robots. After device distribution comes service proliferation — the way delivery apps, the sharing economy, OTT, and social exploded once smartphones were ubiquitous, expect robots, autonomous driving, drones, and on-device AI services to spawn business models in volume.

The wave after that: quantum and space. As AI compounds, compute, power consumption, and data storage/transmission loads grow geometrically — and the next-stage technologies quietly preparing to relieve those bottlenecks are quantum computing and quantum networking, space-based data centers and satellite infrastructure.

The Bottleneck Principle: Where the Money Pools

Before the phase framework, one principle — arguably this entire series’ thesis compressed to a sentence: the largest capital pools form at the largest bottleneck. Think of a road system. With few cars, the traffic light is the bottleneck. Traffic grows, and lane count becomes the constraint. Widen the lanes, and the highway ramps and tollgates jam.

A structural flow diagram titled 'The Bottleneck Principle: Where Capital Pools' showing three sequential phases connected by down arrows: Phase 1 (Compute Bottleneck in red), Phase 2 (Data Flow Bottleneck in cyan), and Phase 3 (Real-World Bottleneck in green).

AI runs the identical sequence: first compute is scarce; add compute and network/storage can’t keep pace; clear that channel and data, services, and power become the new constraint. Every time the bottleneck moves, market leadership moves with it. This is the same logic that made $TSM’s packaging lines (Part 4) and HBM (Part 3) the chokepoints of the last two years — now generalized into a forward map.

The AI industry cycle divides into three phases:
Phase 1: Training EraPhase 2: Inference EraPhase 3: Service Era
Model layerLLM developmentLMM development, specialized DSLMs, MCP/A2ALAM development, world models
Binding bottleneckCompute, powerNetwork, storageData
Key silicon/hardwareGPU, AVP, HBMASIC, HBM4, network equipment, silicon photonics, CPO, HDD/SSDMulti-cloud, data platforms, digital provenance
Data center formHyperscale AI DC investment boomEdge data centersMicro data centers
Power themeGrid-scale buildoutGas turbines, nuclearOn-site generation, BESS, fuel cells
Additional themes“GPU = AI” narrativeMemory hierarchy expansion, “the era of substrates”Agentic / on-device / physical AI

Hold all three phases in view simultaneously — that is the map for reading where we stand and where capital flows next.

Phase 1 — The Training Era: When GPU Meant AI

Phase 1 was the period of training frontier language models. GPT-3 and GPT-4-class models required learning hundreds of billions to trillions of parameters, and the first bottleneck was obvious to everyone: the GPU. No GPUs, no training run; and one training run occupies an entire data center for weeks to months. Big Tech fought open procurement wars to secure allocation first.

A macro photograph of an advanced AI hardware accelerator chip and HBM memory on a circuit board glowing with complex orange, gold, and blue data pathways, representing Phase 1 Training Era infrastructure.

Solving the compute bottleneck elevated three axes in tandem: the GPU/AI accelerator as compute engine, HBM as the ultra-fast memory feeding it, and advanced packaging physically bonding the two. The market narrative was brutally simple — GPU is AI — and the spotlight concentrated almost entirely on $NVDA, the memory producers who could execute HBM, and the foundry supplying CoWoS. Readers of Part 3 and Part 4 will recognize this as precisely the period those installments documented in real time.

Phase 2 — The Inference Era: From Computing Chips to Moving Data

A trained model goes into service — chatbots, search, translation, image/video generation, workflow automation, coding assistants — and AI converts from a research project into infrastructure carrying live traffic. The physics change with it. Training runs one model for weeks; inference serves enormous user counts making short, frequent, simultaneous calls. The bottleneck stops being explainable by GPUs alone: the GPU sits ready to compute while data fails to arrive on time. Network and storage move to the center of the industry.

A cinematic close-up of computer processors running on a motherboard illuminated by high-speed, vibrant magenta and electric pink laser lines, representing Phase 2 Inference and silicon photonics optics.

Inside an AI data center, tens to hundreds of GPUs operate as one cluster — $NVDA’s NVL72 links 72 GPUs into what functions as a single giant GPU. Sustaining that architecture requires switch silicon, network equipment, high-speed optical modules, silicon photonics, and CPO (co-packaged optics).

As line rates climb through 400G, 800G, and 1.6T, purely electrical transmission stops coping with the power, heat, and signal loss — converting electricity to light becomes core infrastructure. Phase 2’s contest is how well you connect GPUs, not just how many you own. (The terminology matters less than the flow: data movement is the new scarcity.)

Storage restructures under the same pressure. Inference is not reading one large file; it is reading and writing small fragments at extreme frequency and minimum latency. Core data center storage re-forms around SSD + NVMe-oF, while bulk low-cost capacity stays on HDD — a two-tier structure hardening into standard: the SSD as the librarian handing you the book instantly, the HDD as the vast archive deep in the warehouse. Phase 2, compressed: the industry expands from the era of computing chips into the era of flowing data.

Leadership broadens accordingly — GPUs and HBM stay essential, but network silicon and equipment, optical communication, SSD/HDD, storage controllers, and cluster/scale-out systems all enter the investment radar. Add the source framework’s parallel themes: edge data centers, gas turbine and nuclear power procurement, memory hierarchy expansion, and — connecting straight back to Part 4 — the era of substrates.

Phase 3 — The Service Era: Agents, Devices, and Physical AI

Once Phase 2 infrastructure reaches sufficient scale, AI stops being any single company’s experimental service and seeps into daily life as a default function. The service era spreads along three branches:

A futuristic editorial design featuring a sleek, white concept sports car glowing with cyan accents parked next to a high-tech white humanoid robot, representing Phase 3 Physical AI.

Agentic AI — agents that complete work, not just answer questions: scheduling meetings, sending mail, drafting reports, chaining approval/booking/settlement flows across systems automatically. The stack above the model becomes investable: data platforms, agent orchestration (MCP/A2A/RAG), logging/security/billing systems — plus the source framework’s adjacent themes of digital provenance, ontology, blockchain, and stablecoins. Hardware still matters underneath: large-scale data centers, HBM4, high-performance networking, and storage carry the load.

On-device AI — AI running directly on personal hardware: AI laptops, AI smartphones, AR glasses, analyzing photos and translating in real time without a connection. The critical components shift to ultra-low-power NPUs and AI SoCs, LPDDR-class memory, UFS/NVMe storage, cooling and packaging, high-efficiency batteries — the era of silicon that is small, smart, and survives being always on. The inference ASIC ↔ NPU/AI SoC axis, low-power displays, low-latency networking/memory, and power management ICs define the bill of materials.

A conceptual digital artwork of a human hand holding a smartphone, with a glowing, intricate 3D holographic neural network and human brain emitting blue light above the screen, representing On-device AI.

Physical AI — AI that moves the real world: autonomous vehicles and EVs, logistics and manufacturing robots, drones, humanoids, medical robotics. The keywords here are not peak performance but safety, durability, and real-time guarantees: automotive-qualified SoCs, memory that survives heat and vibration, SiC/GaN power semiconductors, sensors and actuators, batteries and BMS, next-generation batteries, motors, and rare earths — sold as one integrated set.

Spanning all three branches, one axis covers the entire base infrastructure: cybersecurity. By Phase 3, semiconductor leadership is broader and more varied than at any prior point — data center silicon, on-device AI chips, vehicle and robot SoCs and power semiconductors, sensors, substrates, and packaging all carry distinct investment maps.

The Phases Overlap — They Do Not Hand Off

One correction to the clean three-act structure: reality does not run the phases sequentially to completion. Training infrastructure investment continues through Phases 2 and 3. Inference infrastructure — network, storage, substrates — grows larger as services proliferate. The service layer stacks on top of both. The framework tells you where the marginal bottleneck — and therefore marginal capital — is moving, not which spending stops.

A stacked area chart titled 'Capital Flow: Phases Overlap, They Do Not Hand Off' projecting from 2022 to 2034. It illustrates the compounding growth of Phase 1 Training in red, Phase 2 Inference in cyan, and Phase 3 Services in bright green.

One correction to the clean three-act structure: reality does not run the phases sequentially to completion. Training infrastructure investment continues through Phases 2 and 3. Inference infrastructure — network, storage, substrates — grows larger as services proliferate. The service layer stacks on top of both. The framework tells you where the marginal bottleneck — and therefore marginal capital — is moving, not which spending stops.

The Endgame: Three Layers of the AI Value Chain

Compress the whole cycle into its terminal structure:
LayerPhaseWinning LogicRepresentative Names
Infrastructure controlPhase 1–2Capital intensity; solving hardware bottlenecks$NVDA, $TSM, SK hynix, $ASML, Big Tech data centers
Software & agent diffusionPhase 2–3Productivity transformation; B2B/B2G solutions$PLTR, $MSFT, the open-source camp, specialized LLM developers
Service & device platformsPhase 3Daily-life penetration; physical expansion; mass adoption$TSLA (FSD/humanoid), $AAPL (on-device ecosystem), robot/AV/drone manufacturers

Note that the infrastructure layer is populated by exactly the companies this series spent five installments mapping — the foundry and memory economics of Part 3, the packaging and equipment monopolies of Part 4, the fabless leaders of Part 2. The AI Supercycle does not replace that map. It re-weights it, phase by phase.

A text-based matrix table titled 'The Endgame: 3 Layers of the AI Value Chain' divided into three rows from top to bottom: Layer 1 Infrastructure Control (NVDA, TSM, Hyperscalers), Layer 2 Software & Agent Diffusion (PLTR, MSFT), and Layer 3 Service & Device Platforms (TSLA, AAPL).

The conclusion follows the same discipline as Part 5’s allocation playbook: AI investment is not a theme that burns once and ends. Infrastructure gets laid, software runs on the infrastructure, and services plus new devices reach mass adoption — a long-breath megatrend with partial opportunities embedded at every stage: the model/GPU opening phase, the data center and power/network infrastructure phase, the B2G/B2B/B2C software diffusion phase, the physical and on-device AI phase, and the quantum/space infrastructure quietly preparing behind it.

Some investors got rich catching exactly one wave of the PC → internet → smartphone → cloud sequence; others compounded by catching several. The investor’s job now is identical to then: identify coldly which bottleneck the market is currently passing through, and position at the channel that opens next. Semiconductors are the one value chain that runs beneath every phase of it — the most reliable physical claim on the entire cycle.

Part 1 : .Semiconductor Value Chain Map: 6 Districts Shaping the Future of Tech

part 2 : .Top 30 Best Fabless Stocks: Inside Semiconductor Design Monopolies

part 3 : 3 Powerful Models of Semiconductor Factory Economics:
Foundries, IDMs,and the Memory Supercycle

Part 4 : Top 7 Advanced Packaging & Semiconductor Equipment Monopolies

Part 5 : Semiconductor Asset Allocation:
The Subscription Economics of Silicon and the Big Tech CapEx Engine

Posted in ,

Leave a Reply

Discover more from GearFlow Lab

Subscribe now to keep reading and get access to the full archive.

Continue reading