Essay № 001

Where Is the Next Wave of AI Wealth?

A 100-year pattern that keeps repeating — and what it tells us about the next three to five years in AI.

~2,500 words 12 min read 3 case studies

Over the past two years, names across the AI infrastructure stack have posted returns of 3×, 5×, 10×. The first wave of the AI boom was captured by the picks-and-shovels players upstream. That wave isn't over. But the companies most likely to generate the next 10× are probably not the ones that already ran.

To understand where the next wave accrues, you need to understand a pattern that has governed every major information industry for over a century. It has repeated three times. AI is about to run it a fourth time.

The most profitable company at the end of every information cycle is never the one that built the pipe. It's the one that knew what was flowing through it.

Three Waves, Every Time

A token is, at its core, the same thing as a telegram word, a phone minute, or a gigabyte of mobile data — the smallest billable unit of information transfer. And every time humanity has defined a new billable unit, wealth has followed the same three-wave sequence.

Wave 01
Infrastructure

The cable layers, the copper runners, the tower builders, the GPU makers. This wave always breaks first and biggest. The people who build the pipes capture the earliest and largest share of value.

RUNS FIRST
Wave 02
The Compressors

The price per unit always collapses — driven by a specific class of company that makes the infrastructure more efficient. They profit by making information units cheaper to produce and capturing the efficiency margin.

FOLLOWS QUICKLY
Wave 03
The Gatekeepers

Doesn't build infrastructure. Doesn't compress it. Inserts itself between the infrastructure and the end user — owning access, shaping habits, controlling where people begin. Owns the front door.

ARRIVES LAST · LARGEST
Wealth capture sequence — each information era
HIGH MED LOW TIME → VALUE CAPTURED WAVE 1 — Infrastructure WAVE 2 — Compressors WAVE 3 — Gatekeepers AI TODAY

The Telegraph Era

1866 – 1910 $10 / word

When the first transatlantic cable opened in 1866, sending a single word cost $10 — roughly ten weeks' wages for a skilled laborer. The minimum message was ten words. A single telegram could cost a month's salary.

Wave 1 Eastern Telegraph and the other companies laying undersea cable. By 1900, whoever controlled the cables controlled the flow of global financial information. Eastern Telegraph became one of the largest multinational corporations of its era — not because it was clever, but because it owned the only pipe.

Wave 2 Arrived almost immediately, because $10 a word is an extraordinary incentive to compress. Merchants invented telegraphic codes — entire sentences collapsed into single words. When technology arrived, Western Union deployed multiplexing: a single cable carrying multiple simultaneous signals, tripling effective throughput without laying a new inch of wire. By the late nineteenth century, Western Union controlled over 80% of American telegraph traffic.

Wave 3 Julius Reuter never laid a single mile of cable. In 1850, he discovered a gap in the European telegraph network near Aachen, Germany, and filled it with 45 carrier pigeons — delivering European stock prices six hours faster than the mail trains. His business model: buy telegraph capacity wholesale from the cable companies, repackage it as a news service, sell it to newspapers, banks, and merchants. Reuters sat between the infrastructure and the people who needed information. It owned the front door. It made more money than the cable companies did.

The Telephone Era

1915 – 2000 $7 → $0.06 / min

In January 1915, AT&T completed the first transcontinental telephone call. A three-minute conversation cost $20 — roughly $600 in today's money.

Wave 1 AT&T. A telephone network is a physical monopoly: any competitor wanting to enter would have to re-lay the entire network. AT&T controlled over 80% of American telephone lines and remained the largest private company in the United States for more than half a century.

Wave 2 Two technologies did the compression work. First, the automatic switchboard eliminated human operators — every call had previously required a manual patch by a human being. Second, long-distance signal amplifiers allowed voice to travel thousands of miles without relay stations. Together, these technologies compressed the per-minute transcontinental rate from $7 in 1915 to six cents by 2006. The companies that captured this value weren't AT&T — they were AT&T's equipment suppliers. The most important was Western Electric. When its assets were spun out and rebranded as Lucent Technologies, the 1996 IPO was the largest in American history at the time.

Wave 3 The telephone Yellow Pages charged merchants for listings rather than charging users for calls — pure gatekeeper economics. Long-distance resellers built internal software tools to manage thousands of enterprise accounts. Those tools eventually spun out and became Salesforce and the CRM software category. The Yellow Pages became Google Maps and Yelp. Wave 3 from the telephone era is still compounding today.

The Mobile Data Era

2007 – present Pipe vs. Platform

This case study is best told through what happened in China, because nowhere else was the dynamic more extreme or more visible.

In 2007, China Mobile had a market cap of $410 billion — larger than Microsoft, the largest technology company in the world at the time. Its business model was simple and seemingly unassailable: charge per SMS, per minute, per gigabyte. No one could route around the pipe.

Wave 2 dismantled China Mobile's pricing power systematically. Video compression improved until an HD film that once required several gigabytes compressed to a few hundred megabytes. Per-gigabyte pricing collapsed from tens of renminbi to fractions of a cent.

Wave 3 was the finishing blow — and it came from WeChat, Alibaba, and ByteDance. Before WeChat, sending a message meant paying China Mobile. After WeChat, those messages rode data instead. But the money users saved didn't stay in their pockets. It flowed into the platforms: into ads, purchases, food delivery. Every byte flowing through China Mobile's pipes represented real commercial value — but China Mobile could collect none of it. It could only charge for the pipe.

Tencent's market cap today exceeds HK$4 trillion — more than twice China Mobile's. China Mobile's stock has never recovered its 2007 peak. That gap is the distance between owning the pipe and owning the front door.

Wave Telegraph Era Telephone Era Mobile Era AI (Now)
Wave 1 Eastern Telegraph AT&T China Mobile Nvidia · CoreWeave
Wave 2 Western Union Lucent Technologies Compression algorithms DeepSeek · Inference infra
Wave 3 Reuters Yellow Pages → Salesforce WeChat · Tencent TBD — forming now

AI Is Running the Same Script

Three eras. One pattern, repeated exactly. Wave 1 captures value first. Wave 2 compresses the cost of the unit and captures the efficiency margin. Wave 3 arrives last, owns the relationship with the end user, and ultimately takes the largest share.

Where does AI sit today? The data is unambiguous.

AI Gross Profit Distribution — Today vs. Mature Cloud Stack
AI Stack Today
Semiconductors 79%
~$225B gross profit · Nvidia dominates
Infrastructure 14%
~$40B · Clouds + inference providers
Applications 7%
~$20B · OpenAI, Anthropic + others
Mature Cloud Stack (Target State)
Semiconductors 6%
Commoditized · margin compressed
Infrastructure 24%
Scaled · stable margins
Applications 70%
Dominant · owns the customer
Source: Apoorv Agrawal, "The Economics of Generative AI: Two Years Later" (2026). Cloud stack comparison based on mature SaaS-era benchmarks. At the current pace of 4 percentage points per year of shift, full inversion could take a decade or more.

Wave 1 is at its zenith. That's not a warning — Nvidia at $5 trillion is Wave 1 doing exactly what Wave 1 always does. But the pattern tells us what comes next, and the data tells us we are earlier in the transition than most investors assume.

Wave 2 is opening now. When compute costs this much, every point of efficiency is worth real money. The companies building inference acceleration, quantization, and architectural efficiency are the Western Unions and Lucents of this cycle. DeepSeek demonstrated what this looks like at the model level. The hardware and software companies enabling that efficiency at scale are Wave 2.

Wave 3 is real but patient. The gatekeeper layer in AI — whoever ends up owning where users and developers begin their AI interactions — will likely capture more value than the infrastructure beneath it. That's what the pattern says, and the pattern has never been wrong. But the data above is sobering: at the current pace, it could take a decade or more for the application and gatekeeper layer to reach the kind of profit share that applications enjoy in cloud. Wave 3 is the right long-term bet. It is not the near-term trade.

Three Positions, Three Time Horizons

1
Wave 1 · Still Running
Upstream infrastructure — be selective

Infrastructure names have run, but the buildout continues. The key discipline is separating the structural winners from the names that are simply in the vicinity of the spend. Not every company near Nvidia benefits the way Nvidia does. Core picks-and-shovels still have room, but the easy money has been made.

Horizon: 1 – 3 years
2
Wave 2 · Opening Now
The compression layer — near-term opportunity

Companies making AI cheaper to run — inference efficiency, quantization, architectural innovation, serving infrastructure — are in the position of Western Union and Lucent. The more expensive compute remains, the more valuable this layer becomes. This is where the next 3–5 year opportunity concentrates.

Horizon: 3 – 5 years
3
Wave 3 · Decade-Long Bet
The gatekeepers — position early, harvest late

Whoever owns the front door in AI — the place where users and developers habitually begin — will eventually capture the largest share of value. Reuters didn't lay cable. The Yellow Pages didn't run telephone lines. WeChat didn't build cell towers. They owned where people started. That's worth more than the infrastructure underneath. But it takes time.

Horizon: 7 – 10+ years

The semiconductor layer capturing 79% of AI gross profit today is not a permanent state. It's Wave 1 at its peak — exactly where Eastern Telegraph was in 1900, exactly where AT&T was in 1950, exactly where China Mobile was in 2007. The compression wave is already underway. The gatekeeper wave is forming.

The script is the same. The only question is where you want to be positioned when the next act begins.

Data references: Apoorv Agrawal, "The Economics of Generative AI: Two Years Later," Tailwinds (2026). Historical case studies sourced from public records. This essay represents personal investment views only and does not constitute financial advice.