Jim Liu@jiahanjimliuMacro Thesis of the AI BuildoutCommon…

Hosted by @i · 2026-08-04 · 19 min · Tags: MACRO

TLDR

Jim Liu presents a macro framework linking AI infrastructure investments to historical patterns of liquidity expansion, productivity gains, and controlled inflation. He argues AI represents an extreme productivity revolution that concentrates gains among few players while enabling sustained liquidity without broad consumer inflation.

Historical parallels from industrialization to digitalization show liquidity flowing into assets rather than goods
AI accelerates diffusion on existing digital infrastructure unlike the dot-com era
Profits concentrate in frontier labs, hyperscalers, neoclouds, power, and silicon layers
Open source captures limited revenue; closed models and orchestration drive real value
Risks include overcapacity, accelerated depreciation, commoditization, and regulatory capture
Jevons paradox implies efficiency gains fuel higher intelligence utilization
Key investments tracked in inference, power, photonics, and robotics applications

Speakers

Jim Liu — Delivers comprehensive macro thesis tying AI CapEx to liquidity, production, and inflation dynamics; analyzes buildout layers from inference to physical power; highlights investment targets and risks

Notable quotes

“AI is a more extreme revolution compared to industrialization, financialization and digitalization” — Jim Liu
“This concentration results in less inflation” — Jim Liu
“AI will be working Twenty four seven three hundred sixty five with p h d plus i q, this will result in deflation” — Jim Liu
“The bust was a limitation on how fast humans could reorient the physical world” — Jim Liu
“This higher rate of diffusion will allow the AI build-out to significantly longer than the dot com boom” — Jim Liu
“Vertical integration on the physical layer will be an significant advantage as the build out scales” — Jim Liu
“Jevons' paradox on what is potentially the most fungible resource ever, intelligence” — Jim Liu

Transcript

Jim Liu: Jim Liu at Gihann jim liu macro thesis of the AI buildout common justifications of the AI buildout circle around CapEx commitments, productivity boom, and either ASI or AGI. However, I want to tie together a broader macro picture. After all, hyperscalers, frontier labs, neoclouds derive profits from their end customers, which sit in a larger economic system: liquidity, production, and inflation. First, I need to cover history. Historical precedents to illustrate dynamics in liquidity, production, and inflation. Without a loss of generality, I will oversimplify to not stretch into tangents. Prior to industrialization, investment into small items such as farming tools to large items such as canals led to an increase in production. These investments were made using both credit on a personal level and bonds on a government level. If credit was too loose or currency expanded too fast, society would face inflation. During the Industrial Revolution, the steam engine and coal powered factories allowed for much larger investments without inflation. First, the multiplier in production was so much greater. Second, the investment was split into labor and machines. Most of currency that went towards machines was concentrated into capitalists and hoarded rather than spent. During age of financialization and digitalization, outsourcing and computers introduced productivity improvements through cheaper labor. Automation, these investments flowed into large enterprises and into finance and tech workers. Both financialization and digitalization required much fewer workers than industrial labor, thus the liquidity increase was concentrated in a finance and tech upper middle class that stored much of their value into stocks, bonds and the real estate of n y c, London, Hong Kong, Tokyo, Bay Area, etcetera. Inflation concentrated in real estate and stocks does not drive Drive up inflation that much in daily goods, AI, liquidity, production and inflation, AI is a more extreme revolution compared to industrialization, financialization and digitalization. The investment cash flow goes to a very concentrated segment of the population, namely the companies driving the AI build out. Within these companies, there is less human headcount than ever before. In a society addicted to liquidity expansion, this concentration results in less inflation. The liquidity be heavily stored in assets and those who invest in AI will benefit disproportionately more from this liquidity expansion. Moreover, the productivity increase is compounded on top of industrialization, financialization, and digitalization. AI will be operating factories through sensors and robots, access everything through an efficient financial system, and work directly on top of the digital world that humans have pioneered for decades. AI will be working Twenty four seven three hundred sixty five with p h d plus i q, this will result in deflation, which allows the government to expand liquidity. Some people understand this concept a different way: the value of the dollar hasn't gone up despite massive productivity increases. AI provides the environment that stretches the limits of this dynamic, overbuilds, and speed of technological diffusion. We want to look at the fundamentals underlying the booms and bust due to euphoria and overbuild. Without getting too bogged down in history this time, I'll focus on the dot com bust. All the promises of the nineteen ninety-nine dot com bubble came fifteen years late in the form of Google Search, Amazon Web Services, Facebook social media, Microsoft enterprise software, and Apple iPhones. The bust was a limitation on how fast humans could reorient the physical world and society to fit the digital revolution. AI has a higher speed of technological diffusion because it skates on- On top of the digital world, and the digital world reorients much faster than the physical world, general software engineering has grinded to almost a standstill in face of more software getting written than ever before. Digitalization brought shopping online, but ad targeting pioneered by Google Ads, Meta AI, and Amazon has completed the revolution. Data analysts now work with the output of AIs to optimize supply chains. Self-driving is on the cusp of full city limit rollouts; we are seeing the first of AI-generated images and videos for ads and entertainment. AI adoption in drones will change the military complex. This higher rate of diffusion will allow the AI build-out to significantly longer than the dot com boom. Additional, there are race effects between enterprises and sovereigns to make sustained investments before seeing concrete returns. Layers of the AI build-out: I will give intros to major layers of the AI build-out and include some investments I'm monitoring as high potential. I will link longer form research posts in the comments as I write them. Inference layer every on X is hyped about open- Open source and open source will generate a lot of revenue, but there's no way to argue around the revenue numbers that Anthropic and OpenAI are putting out. Anthropic is at eighty B A R R, ya gan, and OpenAI is at forty-two point six B one two. There's a misconception that open source is cheaper because the model itself is free, however, that isn't necessarily true as OpenAI has higher efficiencies in inference and compute which allow them to serve ChatGPT. Five point six Luna cheaper and more intelligent than open source three, where open source models achieve significant market share as enterprises incorporating proprietary data, while many AI platforms announce day zero Kimmy K three inference four at Fireworks AI underscore h q is the only AI platform which has specialized post-training for Kimmy K three five for tracking open source provider market share, the closest proxy tracker is OpenRouter. Yes, OpenRouter is only 1% of global tokens, but once you account for the fact that the vast majority of global token is OpenAI, Anthropic, hyperscalers, and on-prem, you will see that OpenRouter is a significant share of non-free open-source tokens. Most importantly, OpenRouter is at non-time's trolley, whose inference service gets chosen because switching between providers is seamlessly on OpenRouter because there are so many semi-commoditized providers of O. Open source models, profit will flow to the rest of the layers. For example, for DeepSeek v4 Pro, there are eighteen providers on OpenRouter, eighteen, and that's not even including the Azure, Oracle, and GCP who serve it directly. Eight, nine, ten. Closed source Frontier Lab revenue will also drive profits to most of the bottom layers, except for orchestration layer which is in-house and potentially some AIs will be in-house. Orchestration layer obviously hyperscalers and Frontier Frontier Labs excel here, but at Fireworks AI underscore H Q and C R W V have the best orchestration relative to their market cap. Fireworks AI leads in cache hit rate among Kimmy K three providers outside of Moonshot AI six. Cache hit rate is subject to variance in customer requests but is a proxy measurement of intelligent request routing, prefix cache efficiency, and KV cache management. Outside of the hyperscalers, CoreWeave has the best Quality and diversification of enterprise customers, with customers such as Anthropic, OpenAI, Jane Street, Cloudflare, Mistral, Cohere, Perplexity, Cursor, Cognition, Midjourney, Runway AI, CrowdStrike, Mercado Libre, OneOne, Nvidia has been growing up from CUDA, Network c c l into this layer, and there is a risk of commoditization within this layer. Nvidia Run dot AI, which is part of d s x o s, utilizes advanced orchestration. Invidia run dot AI significantly enhances GPU efficiency and workload capacity twelve. Power Datacenter bare metal i r h has the strongest power portfolio out of the neoclounds with five point eight gigawatts of power that is grid firm and one hundred percent powered by renewable energy thirteen. NAS and Spain not updated on website are eight hundred megawatts and four hundred ninety megawatts. At this physical layer, iREN is the only neoclound That is vertically integrated on all its sites and doesn't use any colocation providers. Vertical integration on the physical layer will be an significant advantage as the build out scales and iREN's workforce builds institutional knowledge through each site. Vertical integration in this layer is investment in a team that learns and improves time to compute. For neoclouds that depend on colocation, this institutional knowledge builds in each supplier. Behind the meter will also see large growth. N u a i is a binary bet on a team with expertise in natural gas in the Permian Basin and a site in t c d c with triple gas pipelines for redundancy. One four. B e is the default option for clean behind the meter power generation. Silicon design n v d a is the king of not only the AI compute engine g p u's and edx six via Grok but has a moat in AI networking, data center design. Software layers like CUDA underscore n c c l and d s x o s and leader position in the being able to secure the semiconductor supply chain, h b m has risen from commodities off the back of two developments: one, yield challenges in stacking have turned the h b m fabrication process into a moat much like how yield for t s m c is a core moat; two, h b m is now co-designed with g p u's and a s i c's. Skate oberwy, gullbot at, and samson have a triopoly on hbm c x m t makes d ram but not hbm. Many people talk about photonics because the number of connections from photons are replace copper. Fiber has long dominated long distances and moved to smaller and smaller distances, and no, it's not the cable where the money is made, but rather it's the transceivers. At the smaller distance, exponentially more connections are made, which means exponentially more transceivers. The transceiver is composed of silicon components or co-packaged components. The photonic integrated circuit PIC is silicon that manipulates light. N P is the laser, while the photodetector is often germanium. Light is the leader here. Silicon fabrication: There are only three fabs in the world capable of the most advanced nodes: TSMC, INTC, Samsung. TSMC is dominant, but INTC is the most interesting play here. Intel was dominant in silicon fabs for many years and only fell behind under Brian Krzanich, who was a bean counting quarterly performance optimizer who tried to chase every buzzword and diversified Intel into IoT processors for wearables. Fabs, 5G modems, fpga's, storage, memory, cameras, self-driving drones, fifteen, yes, Intel drones, fifteen, Intel actually has some of the best IP in both silicon etching, packaging, and backside power delivery. Now with refocusing by Gelsinger and Lip, Bhutani, Intel is set to potentially surprise in eighteen A and fourteen A. Oh, and Intel still makes the best CPUs. Among fabs is also Dalla T. Huett, which is the leader in silicon. The Tonix Fabs application, I separate application from inference as it's the harness and business application that utilizes the tokens. There will be many new companies that will be born in this layer as AI matures. A current winner is Meta. Meta is worth like we one point seven Terillions compared to Google's four point three five Terillisteds, but is similar in many aspects: one, mastered integrating AI into ad targeting; two, MuseSpark one point one greater Gemini; three, TPU greater m t i a; four, dominant network effects in social media versus search YouTube. However, AI has some cannibalization risk to Google's core search business while AI cinnabar Energizes well with social media. Yes, Google's search business is still growing because AI increases digital use in general, but you can see the difference in that Meta's social media numbers are growing much faster than Google's search numbers. The big delta in market cap is explained by GCP. However, with Meta looking to enter as a cloud provider, if Meta's model improves faster than Google's and MetaCompute has a decent chance of growing faster than GCP, Robotics dollar TSL- A is in a unique position because its manufacturing advantage will convert into a data advantage, which will allow it to lead in FSD and generalized applications of Optimus. Yes, dollar T S L A stock been inside a volatile consolidation range for five years and opportunity cost has been high, but when T S L A breaks out, it's significant. Although D T S L A has a very large market cap already, which all else equal means smaller percentage gain, it's important to track D T S L A as it will have an impact to the development of the robotics industry. If T S L A is for the lidar-less applications, then dollar U S T Is for the lidar applications. The team at Dallase Luvituk is phenomenal and many robotics applications which rely on proprietary data but don't have sheer volume of data for an image based neural net will use lidar. In drones, Anderil dominates the high tech, expensive military drones while Dody ID's does very well with tailored deployments for airports, stadiums, etc. Drones will all be autonomous and behind it will be potentially massive military spending. For counter drones, EOS dot A X is interesting as a laser based defense to drones. Risks: Every good investment analysis includes assessment of risks. The first risk is accelerated depreciation, which is closely tied to overcapacity. As long as compute is under capacity, accelerated depreciation will be unlikely because in a scarce regime, companies have to use what's available. If that's A one hundredths S, that's A one hundredths, which is exactly what we are. Seeing today, A100s came out May twenty twenty and are six years old now, and with A100s being renewed this year, they will go into year seven. Overcapacity is something we have to monitor closely as it's a complex function of model and efficiency improvements, but from the supply side, the current dominating force is limitation in silicon fabs, memory fabs, potentially photonics materials, firm power, blue collar labor to build the data centers, on the demand side. There is Jevons' paradox on what is potentially the most fungible resource ever, intelligence. Commoditization can be a serious risk to some layers. For example, AI itself is eating away at the software moats with CUDA no longer being a moat and AI being increasingly being able to optimize much of the orchestration umbrella of request routing, prefix cache hits, KV cache management, batching, model routing, GPU scheduling, auto-scaling, load balancing. For every strong ROI story like Meta ad targeting, there will be many weak ROI as enterprises like Klarna have failed to replace customer service with AI. Many of these ROI failures will be solved with intelligence in models. For example, although AI has been surprisingly underwhelming in customer service, ChatGPT voice models are becoming increasingly better. Voice models will require more compute, and enterprises failing to integrate AI into customer service are doomed. Afterthought: Those using the old transcript into small model approach now that only addresses quality, not ROI. However, AI compute drastically increasing inefficiency and Jevons' paradox will make using expensive voice models in lower value tasks like customer service become positive ROI. The same risk that drives commoditization helps improve ROI. Scientific breakthrough is complex and not fully within the scope of this post, however, when you take a look of the entire- History of AI, you'll see that compute and algorithm efficiency have continuously been made, and that's what has enabled the level of intelligence we have today. Efficiency breakthroughs are swallowed up utilization of compute for higher intelligence. This is the intelligence corollary of Jevons' paradox: the more efficient AI compute and algorithms are, more of the resource will be used to unlock the next level of intelligence. Space DCs are a risk to the power layer and neocloud specifically as it potentially allows X company SpaceX to have an innate advantage. However, anyone who has seen data center operations knows that continuous maintenance is crucial and the energy and cost consumption isn't neither green nor cheap. Customer concentration is another risk and real. AI will generate wealth for a more concentrated set of companies than ever. That's why I invest in leaders or refundable layer. Like grid connected power, no whether open source or open AI, Anthropic become dominant, AI doesn't print powered data centers and GPUs and Iridion will be able to sell to the winner. Regulation and politics are a risk. If this happens, certain companies may be able to do regulatory capture of profits, and it will be prudent to invest along those company's supply chain. Security is another risk, but the way to battle AI enabled hacks is to use AI. This drives up the demand. Demand for AI: Political backlash is an often mentioned risk in the news these days, but fear of losing China will be a more sensible fear. We've already lost much of manufacturing to China. If the West loses intelligence via AI to China, it's over. six forty seven a m org fourth twenty twenty six three point one kill views eleven one twenty seven twenty three.