Jonathan Tsang

LinkedIn (opens in new tab)

Six continents run. Antarctica next.

My six marathons in race order, Nov 2024 to Jul 2026. Blue marks my fastest.

3:34:48 marathon personal record, Guayaquil

Full entry

Jonathan Tsang / Berkeley IEOR / Student researcher

I study where the American grid gets stuck, and what fixing it is worth.

Headroom is my open-source tool. It ranks the bottlenecks in the Southwest Power Pool’s 2025 day-ahead market by what relieving each is worth.

Say hi on LinkedIn (opens in new tab)

  • $2.41Bcongestion cost SPP posted for 2025, day-ahead
  • 37of 2,054 bottlenecks carry half of it
Data
SPP 2025, public
Code
Python, open source

Jonathan Tsang / Berkeley IEOR / Builder

Each build has a check that can say no.

I build with AI coding agents: I write the specs, make the design calls, and check the output.

Say hi on LinkedIn (opens in new tab)

  1. GridTwin

    A verifier from another model family checks each fix. A safety gate can only say no.

  2. Recourse

    The extractor keeps a field only if its quote is in the source.

  3. Momentum

    Net of costs, the strategy’s returns weren’t significant: t = 0.29.

Jonathan Tsang / Berkeley IEOR / Student analyst

One stock pitch a year, each valued a different way.

ASML in 2024, Alphabet in 2025, Akeso in 2026.

Say hi on LinkedIn (opens in new tab)

One pitch a year
YearTickerMethodTarget
2024ASMLASML Holding · Team of 2Moat thesis+76%
2025GOOGLAlphabet · Team of 2Scenario DCF+52%
20269926.HKAkeso · SoloDCF + Monte Carlo+56%
  • 3Young Investors Society stock pitch entries, 2024 to 2026
  • 10,000price paths I simulated in Python for my 2026 pitch

Student analysis, not investment advice. Pitch figures are as of each pitch date. Current prices are labeled with their date.

Tickers
ASML, GOOGL, 9926.HK
Teams
Two with a teammate, one solo

Jonathan Tsang / Berkeley IEOR / Teacher

For a month, I taught English at a free school in rural Cambodia.

I lived at the school and built my lessons from scratch.

Say hi on LinkedIn (opens in new tab)

  • 3classes a day
  • 5–16student ages
Jonathan takes a selfie at the front of a classroom. About two dozen students in blue PCEF polos stand behind him.
Class selfie at PCEF, 2025.

Jonathan Tsang / Berkeley IEOR / Runner

Six continents run. Antarctica next.

I run for the discipline, not the thrill.

Say hi on LinkedIn (opens in new tab)

  • 6continents with a marathon
  • 3:34:48marathon personal record, Guayaquil
First race
Nov 10, 2024
Latest race
Jul 5, 2026
The Hungarian Parliament lit gold at night across the Danube, its lights reflected in the water. A river-cruise sign stands in the foreground.
The Parliament on the Danube, Budapest. Race 03 of 06.

Jonathan Tsang / Berkeley IEOR / Everything else

Make. Climb. Play. Read.

The rest of what I do.

Say hi on LinkedIn (opens in new tab)

  • 10,064 ftMt. Baldy’s summit, which I plan to reach

01Research

  • 2,072hit a limit in 2025
  • 2,054with a posted cost
  • 53worth fixing

Most of the 2,072 bottlenecks hit their limit for a few hours and posted little cost. The 53 worth fixing each hit it for 216 hours or more and posted $5.5M or more. 18 had no posted dollars.

Each square is one bottleneck in SPP’s 2025 day-ahead market. Not a map.

Where the American grid gets stuck.

Full entry

03Builds

The extractor keeps a field only if its quote is in the source.

Each build has a check that can say no.

Full entry

04Teaching in Cambodia

Each day I drew one seat.

That student asked ChatGPT, then taught the class.

Full entry

05Pitches

Target vs. the model. 2024: target +76%, our DCF −23%. 2025: target +52%, DCF cases +6% to +187%. 2026: target +56%, my DCF, bear to bull, −25% to +210%.

Three stock pitches, three ways to value a company

Full entry

06Everything else

The rest of what I do.

Full entry

ContentsAlso in this notebook

See everything

Open any entry to read it in full.The other five entries, numbered as in Contents.

  1. 01 Fixing the energy grid 2,072 bottlenecks in the Southwest Power Pool’s 2025 day-ahead market, scored by Headroom, my tool
  2. 02 Running 6 of 7 continents with a marathon. Antarctica is the last.
  3. 03 Hackathons and side projects 3 builds since July 2026, each with a check that can say no
  4. 04 Teaching in Cambodia 3 classes a day, for a month, at a free school in rural Cambodia
  5. 05 Valuing companies 3 stock pitches, 2024 to 2026, each valued a different way
  6. 06 Everything else 7 books I’ve been reading, plus espresso, bartending, climbing, violin, and poker

01 Research

Where the grid gets stuckHeadroom

Full entry

Headroom ranks bottlenecks in SPP’s 2025 day-ahead market by what relieving them is worth. It runs on public data. Many fixes also need land and right-of-way, parcel by parcel. That’s the next constraint.

Python CLIOpen sourceJun–Oct 2026

How I check it
  • I compared $2.41B of posted congestion with $1.96B settled. The 1.23× gap has no documented cause.
  • About 30% of one half-year’s top 5% recurred in the other, held out within 2025 only.
  • My network screen could map 0 of 2,072 bottlenecks. Grid security rules (CEII) hide their locations.
  • I wrote a memo on what public data can’t show.
Before and after
  • Spring 2026: a FERC Form 1 notebook (utilities’ annual federal financial report), Headroom’s predecessor.
  • Fall 2026: customer discovery on building energy, with a partner. 14 owners emailed, 2 interviews.

Headroom on GitHub (opens in new tab)

Python CLIOpen sourceJun 30–Oct 14, 2026

Cheap power often can’t reach the people who need it. A line or transformer hits its limit, and the power waits behind it. Grid operators call each of these limits a flowgate.

Some limits can be raised without new lines, using grid-enhancing technologies. Headroom asks which limits are worth raising first.

It scores every flowgate that bound in the Southwest Power Pool’s 2025 day-ahead market. Then it ranks the 2,054 with posted congestion by expected net value.

Grid-enhancing technologies work on the wires already there. Bigger fixes, like new lines and some upgrades, need land and right-of-way, parcel by parcel. That’s the next constraint, and Headroom doesn’t model it yet.

How I got here

  1. Spring 2026FERC Form 1 notebook

    Headroom’s predecessor. Two RMI staff replied by email with methodology feedback, on Mar 30 and Apr 6.

  2. Jun 30, 2026Headroom starts

    I worked on it through Oct 14. 85 of 86 commits are mine.

  3. Sep 2026Fleet charging discovery

    I emailed 10 fleet managers. Six replied, and one took a call. The district had enough capacity and named no dollar value. We moved on.

  4. Oct 2026Building energy

    With a partner, I emailed 14 commercial building owners. Four replied, and two sat for interviews. One had broken utility-data syncs. The other had tenant-controlled meters.

  5. Oct 2026Compliance scanner

    96k LA and 30k SF building-year records, each linked to its source. Its counts match the cities’ own. Not deployed.

I built Headroom with AI coding agents. I wrote the specs, made the design calls, and checked the output.

Headroom on GitHub (opens in new tab)

The valuation

net value= relief share ×persistence-weighted congestion ×shadow-price haircut −annualized cost

Each input gets a cited triangular prior: a low, likely, and high value. 53 come out net-positive under baseline priors. Across the prior ranges, it’s 16 to 114.

How I check it

  • 1.23×

    Posted 2025 congestion ($2.41B) against settled ($1.96B). I report the gap. Its cause is undocumented.

  • ~30%

    Of one half-year’s top 5%, about 30% recurred in the other half. Held out within 2025 only.

  • SHA-256

    Pinned data snapshots, a lineage graph, an export gate, and byte-identical replay.

Limit My network screen could map 0 of 2,072 flowgates. Critical Energy/Electric Infrastructure Information (CEII) rules hide their locations. I wrote a memo on the limits.

  • 2,072hit a limit in 2025
  • 2,054with a posted cost
  • 53worth fixing

How few carry the cost

The 53 worth fixing carry 54% of the cost.

Ranked by posted cost, 10 bottlenecks carry 26% of the $2.41B. 37 carry half, and 131 carry 75%.

Worth fixing, across prior ranges

16 to 114. It is 53 at baseline priors.

Most of the 2,072 bottlenecks hit their limit for a few hours and posted little cost. The 53 worth fixing each hit it for 216 hours or more and posted $5.5M or more. 18 had no posted dollars.

Each square is one bottleneck that hit its limit in SPP’s 2025 day-ahead market. Black ones are worth fixing at baseline priors. Not a map: their locations are restricted (CEII).

02 Running

Six continents run. Antarctica next.

Full entry
The Hungarian Parliament lit gold at night across the Danube, its lights reflected in the water. A river-cruise sign stands in the foreground.
The Parliament on the Danube, Budapest.

I’ve run a marathon on six continents since November 2024.

  1. 01 4:03:50 Santa Clarita · North America, 42–75°F

    Pace 9:18 min/mi

    Race morning 42–75°F (6–24°C), 27% humidity

    First marathon

    Jonathan in a dark suit and patterned tie holds his violin and bow after a concert. He stands in a hall with organ pipes, with other musicians and the audience behind him.
    After the concert, Nov 10, 2024.
  2. 02 3:52:25 Luxor · Africa, 53–62°F

    Pace 8:52 min/mi

    Race morning 53–62°F (12–17°C), 52% humidity

  3. 03 4:04:10 Budapest · Europe, 56–74°F

    Pace 9:19 min/mi

    Race morning 56–74°F (13–23°C), 70% humidity

    The Hungarian Parliament lit gold at night across the Danube, its lights reflected in the water. A river-cruise sign stands in the foreground.
    The Parliament on the Danube, Budapest.
  4. 04 4:29:37 Siem Reap · Asia, 79–87°F

    Pace 10:17 min/mi

    Race morning 79–87°F (26–31°C), 82% humidity

  5. 05 4:27:29 Melbourne · Australia, 40–48°F

    Pace 10:13 min/mi

    Race morning 40–48°F (4–9°C), 86% humidity

    10 days after Siem Reap

    Sun breaks through heavy clouds over a calm bay and a sand beach. A tower block and palm trees stand at the right.
    Melbourne, by the bay.
  6. 06 3:34:48, personal record Guayaquil · South America, 76–79°F

    Pace 8:12 min/mi

    Race morning 76–79°F (24–26°C), 86% humidity

    PR

  7. 07 Not yet Antarctica

    The seventh continent. Not run yet.

Fig. 02 ▪ My six marathons in race order, Nov 2024 to Jul 2026. Blue marks my fastest. Antarctica is next.

  1. 01 · North America

    Santa Clarita Marathon

    4:03:50, 9:18 min/mi

    Race morning 42–75°F (6–24°C), 27% humidity

    First marathon

    My first marathon. Four hours after the finish, I was in concert black with my violin, playing Suppé, Borodin, and Brahms.

    Jonathan in a dark suit and patterned tie holds his violin and bow after a concert. He stands in a hall with organ pipes, with other musicians and the audience behind him.
    After the concert, Nov 10, 2024.

  2. 02 · Africa

    Luxor Marathon

    3:52:25, 8:52 min/mi

    Race morning 53–62°F (12–17°C), 52% humidity

    In Cairo, a friend showed me his high school. One classroom was still in session, lights on. When I raised my phone, the principal covered the lens: “No photos.”

    My friend explained later. It was a paid “private teaching” class. Some teachers hold back material in class, then reteach it after hours for a fee. If I’d been a low-income student there, I might never have had a chance. That’s what sent me to Cambodia to teach.

  3. 03 · Europe

    Budapest Marathon Along the Danube

    4:04:10, 9:19 min/mi

    Race morning 56–74°F (13–23°C), 70% humidity

    An in-and-out trip to Hungary, with one day to acclimate. I arrived, ran the marathon the next morning, and went to a Yousuke Yukimatsu DJ set that night.

    The Hungarian Parliament lit gold at night across the Danube, its lights reflected in the water. A river-cruise sign stands in the foreground.
    The Parliament on the Danube, Budapest.

  4. 04 · Asia

    PCEF Marathon, Siem Reap

    4:29:37, 10:17 min/mi

    Race morning 79–87°F (26–31°C), 82% humidity

    10 days before Melbourne

    The heat set the shape of this race. Once the sun rose around 5 a.m., it felt like 105.

    I was up stretching at 3:30. Carrying water was too heavy. So I ran 3 miles out and 3 back from the school where I was staying. Every 6 miles I came back for water and a gel.

    Mile 18 got hard. I had to start running again before my mind quit. For the last 6 miles, I prayed for clouds. I also used up my energy gels here. That mattered ten days later.

  5. 05 · Australia

    Melbourne Marathon

    4:27:29, 10:13 min/mi

    Race morning 40–48°F (4–9°C), 86% humidity

    Cloudy · headwind

    10 days after Siem Reap

    Ten days after Siem Reap, I went from humid heat to cloudy winter weather in the 40s. I had landed the day before.

    My gels were gone, so I ran on two bananas and one stray gel from my luggage. The headwind lasted the whole first 13 miles out. I limped the last 2 miles at 12 minutes a mile, shivering. Short on gels, it was harder than Cambodia.

    Sun breaks through heavy clouds over a calm bay and a sand beach. A tower block and palm trees stand at the right.
    Melbourne, by the bay.

  6. 06 · South America

    Guayaquil Marathon

    3:34:48, 8:12 min/mi

    Race morning 76–79°F (24–26°C), 86% humidity

    PR

    My personal record. I broke a mental barrier here and found inner peace.

    Half of the run was a sanctioned half marathon.

  7. 07 · Remaining

    Not yet

    Antarctica

    The seventh continent. Likely after college, by charter or with a pilot’s license.

For the three months before Santa Clarita, my first marathon, training started at 5 a.m. Dynamic warmup, then the same route: around the lake, back up the hill. I ran it over and over, in the dark and the cold. The excitement wore off fast. What was left was showing up and finishing what I said I’d do.

That habit is what carries me on race day. Each marathon is four hours of burning muscles, a dry mouth, and a fight with an irrational, primal mind. Across all my ambitions, discipline, not thrill, is what moves me toward them. Only Antarctica remains.

03 Builds

Each build has a check that can say no.Builds

Full entry

Each row says what I owned and what came of it.

Oct 17, 2026GridTwinHackathon Agents that check each other’s work before calling a fault fixed.

Team of 3. I built nearly all of the agent layer. A planner writes a definition of done, and a verifier from another model family checks it.

In 9 simulated scenarios, GridTwin claimed no false fix. It also fixed none of the 6 fixable faults.

GridTwin team repo (opens in new tab)

Oct 11, 2026Recourse at SpeedHacksHackathon Automated insurance claims for consumers. Team of 3, one day.

I owned the case engine in TypeScript, Next.js, and SQLite: authorization snapshots, one call per approval, webhook de-duplication.

I also wrote the document extractor. It keeps a field only if its quote is in the source. The demo ran on synthetic data.

Recourse team repo (opens in new tab)

Jul 2026Momentum replicationQuant research Thirty years of a classic strategy, and a null result.

Jegadeesh–Titman 12-2 momentum on 503 current S&P 500 names, 1996 to 2026. It tracks Ken French’s momentum factor at a 0.913 correlation.

The result was null: Sharpe 0.11 gross, Newey–West t = 0.29 net. I tested for look-ahead bias and wrote up survivorship bias. Local only.

04 Teaching

A month teaching in CambodiaSiem Reap

Full entry

I taught three classes a day at PCEF, a free English school with no running water.

Aug 12–13, 2025Two lessons I built
  1. Mystery photo
  2. Real or fake

Play them in the full entry

Support PCEF (opens in new tab; the school’s fundraiser on khmercare.com)

  1. Draw a seat
  2. To the shared PC
  3. Through a VPN
  4. Ask ChatGPT
  5. Teach the class
One seat a day, drawn at random. Grey seats have had their turn.

A laptop in class

I brought my laptop so students could ask an AI anything. It answered in Khmer, and they started testing their own questions.

Older students took over computer time, so I drew one seat a day. That student taught the class the next day.

We practiced English through prompts: grammar, role-play, simpler readings.

Aug 12–13, 2025Two lessons I built

  1. Mystery photo

    Step 1: Guess what the item in the photo is

    Step 2: Use the item in a sentence

    1. 01Guess: photo 1, zoomed in. Show the answer.Scarf
    2. 02Guess: photo 2, zoomed in. Show the answer.Lotus flower
    3. 03Guess: photo 3, zoomed in. Show the answer.Fish amok
    4. 04Guess: photo 4, zoomed in. Show the answer.Bayon
    5. 05Guess: photo 5, zoomed in. Show the answer.Sugar cane press
    6. 06Bonus, hard: photo 6, zoomed in. Show the answer.Apsara

    Tap to zoom out.

    Drawn for this page. The class saw photos.

  2. Real or fake

    Step 1: Guess if the image is AI-made

    Step 2: Why or why not?

    1. Snow-capped peaks and a row of pine trees reflected in a still lake at sunrise, reeds in the corner.01AI or real?: image 1. Show the answer.AI-made: made for this pageIdentical pines, a mirror-flat lake, three tidy birds.
    2. A long, ornate building with a central dome, lit gold at night, its lights reflected in a dark river.02AI or real?: image 2. Show the answer.Real: my photo, BudapestLamp glare, choppy water, a pole sticking out of the river.
    3. Three whole limes and a cut half on a wooden table, the background softly blurred.03AI or real?: image 3. Show the answer.AI-made: made for this pageFlawless fruit, perfect wedges, blur that is too even.
    4. Three monks in orange robes walk past a lotus pond at sunrise, temple towers behind.04AI or real?: image 4. Show the answer.AI-made: from my lesson slidesThree near-identical monks in step, all lit perfectly.
    5. Sunlight breaks through heavy clouds over a bay and a sandy beach, buildings on the far shore.05AI or real?: image 5. Show the answer.Real: my photo, MelbourneMessy clouds, lens glare, cranes on the skyline.
    6. TotalCount: show how many were AI-made.3 of 5 were AI-made

    Tap for the answer.

    One AI image is from my slides. The rest stand in for what the class saw.

How I got there

In Cairo, a friend showed me a paid after-hours class. Some teachers hold back material, then reteach it for a fee. That sent me to teach.

At the school

The students’ ambition matched any classroom I’ve sat in, but the heat cost them focus. Hundreds of millions share that ambition in conditions that hold it back.

Support PCEF (opens in new tab) Fundraiser on khmercare.com

05 Pitches

Three stock pitches, three ways to value a company

Full entry

From 2024 to 2026, I entered the Young Investors Society stock pitch competition each year.

Target vs. the model

In 2024, our target sat far above our own model. From 2025 on, the target came from the model.

  • Price on the pitch date
  • Final target
  • Earlier draft
  • DCF (discounted cash flow) value
  • DCF range (tick: target case)
  • Price now

Each mark is measured from the price on its own date. Our ASML DCF used the May 14, 2024 price, US$907.93. Now is measured from the pitch price.

  1. 2024 ASML ASML Holding · Team of 2 Buy to hold · US$852.85 to US$1,500 · +76% Target +76% · April draft +175% · Our DCF −23% · Now +93% (Oct 19, 2026)

    Our own DCF said the stock was overpriced. We kept the buy and explained why.

    Price now US$1,647.12 (Oct 19, 2026)

    Moat thesisGlobal Youth Investment Summit 2024

    Open the full pitch
  2. 2025 GOOGL Alphabet · Team of 2 Buy · US$166.54 to US$252.56 · +52% Target +52% · DCF cases +6% to +187% · Now +107% (Oct 19, 2026)

    Google handled about 5 trillion searches a year. That was under a third of an estimated 16.7 trillion.

    Price now US$344.18 (Oct 19, 2026)

    Scenario DCFTop 100 · 2025

    Open the full pitch
  3. 2026 9926.HK Akeso · Solo Buy · HK$102.70 to HK$160.02 · +56% Target +56% · My DCF, bear to bull −25% to +210% · Now −11% (Oct 20, 2026)

    Its lead drug beat Keytruda in a Phase III lung cancer trial. I simulated 10,000 price paths around my DCF value.

    Price now HK$91.75 (Oct 20, 2026)

    DCF + Monte Carlo

    Open the full pitch

Student analysis, not investment advice. Pitch figures are as of each pitch date. Current prices are labeled with their date.

Twice I worked with a teammate. The last one I did alone. First came an argument from competitive advantage. Then a discounted cash flow (DCF) model with three scenarios. Then a Monte Carlo simulation I wrote in Python.

Target vs. the model

In 2024, our target sat far above our own model. From 2025 on, the target came from the model.

  • Price on the pitch date
  • Final target
  • Earlier draft
  • DCF value
  • DCF range (tick: target case)
  • Price now

Each mark is measured from the price on its own date. Our ASML DCF used the May 14, 2024 price, US$907.93. Now is measured from the pitch price.

  1. 2024 ASML ASML Holding · Team of 2 Buy to hold · US$852.85 to US$1,500 · +76% Target +76% · April draft +175% · Our DCF −23% · Now +93% (Oct 19, 2026)
  2. 2025 GOOGL Alphabet · Team of 2 Buy · US$166.54 to US$252.56 · +52% Target +52% · DCF cases +6% to +187% · Now +107% (Oct 19, 2026)
  3. 2026 9926.HK Akeso · Solo Buy · HK$102.70 to HK$160.02 · +56% Target +56% · My DCF, bear to bull −25% to +210% · Now −11% (Oct 20, 2026)
01 · 2024ASMLMoat thesis 02 · 2025GOOGLScenario DCF 03 · 20269926.HKDCF + Monte Carlo
NASDAQ · Pitch figures as of May 1, 2024

NASDAQ: ASML · Semiconductor equipment · 2024 · Team of 2

ASML Holding

ASML makes the machines that print circuits onto computer chips. It is the only company that sells EUV machines, used for the most advanced chips.

  • Buy to hold
  • GYIS 2024

Spring 2024, a team of two. We presented in Stock Pitch Round 1 at the Global Youth Investment Summit, New York.

Our own DCF said ASML was overpriced. We kept the buy and argued why in the report.

Our callMay final

Price (US$)May 1, 2024
852.85
Target (US$)
1,500.00
Upside
up 76%
Price now (US$)Oct 19, 2026
1,647.12
Since pitch from US$852.85, May 1, 2024
up 93%

Buy to hold. Between April and May, we cut our target by 44%.

Three numbers, one stockUS$

Chart of three ASML values. April 3, 2024: price US$980.27, target US$2,700. May 1, 2024: price US$852.85, target US$1,500. May 14, 2024: price US$907.93, our DCF value US$696.99.

  • April draftup 175%
    price 980.27 to target 2,700 · Apr 3
  • May finalup 76%
    price 852.85 to target 1,500 · May 1
  • Our DCFdown 23%
    price 907.93 to DCF value 696.99 · May 14

Hollow square: price on that date.Filled square: target or DCF value.Dashed: points down.

Our April target, our May target, and our own DCF, each against the price on its date. Dashed: the DCF, which pointed down.

Layers

01The case4 reasons · 4 risks

The case for

  1. Only supplier. ASML is the only company that sells EUV lithography machines. The barriers to entry are high.

  2. Hard to copy. Each EUV machine weighs 150 metric tons. Without ASML’s software, the machines would be unusable.

  3. Chip demand. AI, 5G, and connected devices all need advanced chips. Forecasts put chips near a trillion-dollar market by 2030.

  4. Locked-in customers. Intel bought all the EUV systems ASML planned to make in 2024. Our report put the order backlog near US$42B in Q1 2024.

The case against

  1. Export limits. US and Dutch rules limit EUV sales to China.

  2. Delicate parts. High-NA EUV machines have intricate, delicate components.

  3. Few customers. A handful of chipmakers buy most of the machines. Losing one would hurt.

  4. Our own model. Our DCF valued the stock below its market price.

02The numbersR&D · five forces · 4 figures

R&D spend, 2023 (US$M)

From our comparison table in the report. ASML spent the most of the five.

Five forces, scored

Porter’s five forces, our scores out of 5. Higher means less threat to ASML.

Key figures

  • US$696.99 vs US$907.93

    Our DCF sheet’s enterprise value per share against the price, May 14, 2024.

  • 53 of 449

    ASML systems sold in 2023 that were EUV. In 2022 it was 40 of 345.

  • 45 vs 37

    ASML’s P/E against the industry average.

  • US$2,700 → US$1,500

    Our target, April vs May 2024. We cut it by 44%.

Method

Porter’s five forces, an R&D comparison, and market growth. Then our own DCF, at a 9.0% discount rate and 4% long-term growth. The DCF said overpriced. We kept the buy and argued it missed broader trends.

“…solely relying on traditional valuation metrics overlooks broader economic trends.”

Our report, May 2024, the line after our DCF result.
03The recordApr 3 to May 30, 2024

  1. First deck. Long hold, US$980.27 → US$2,700 (+175%).

  2. Final report, repriced at US$852.85. Target from US$2,700 to US$1,500. Call: buy to hold.

  3. Our DCF sheet: US$696.99 a share (enterprise value) against a US$907.93 price.

  4. Sent the final report to the Global Youth Investment Summit.

  5. Presented in Stock Pitch Round 1 at the summit, Cornell Tech, New York.

NASDAQ · Pitch figures as of May 19, 2025

NASDAQ: GOOGL · Internet content and information · 2025 · Team of 2

Alphabet

Google’s parent company. It runs Search, YouTube, Google Cloud, and Waymo, among others.

  • Buy
  • Top 100 · 2025

Spring 2025, a team of two. Our pitch made the competition’s Top 100. We attended the Global Youth Investment Summit in New York.

Google handled about 5 trillion searches a year. That was under a third of an estimated 16.7 trillion.

Our callModerate

Price (US$)May 19, 2025
166.54
DCF value (US$)
252.56
Upside
up 52%
Price now (US$)Oct 19, 2026
344.18
Since pitch from US$166.54, May 19, 2025
up 107%

Moderate case, our target. Year-10 free cash flow of US$136.68B.

Three DCF cases, one targetUS$

Three DCF values for Alphabet against a US$166.54 price. Conservative US$175.79. Moderate US$252.56, the target. Aggressive US$478.30.

  • Conservativeup 6%
    price 166.54 to DCF value 175.79
  • Moderateup 52%
    price 166.54 to DCF value 252.56 · target
  • Aggressiveup 187%
    price 166.54 to DCF value 478.30
  • April versionup 54%
    163.99 to 252.56 · Mar 22 · buy to hold

Hollow square: price on that date.Filled square: target or DCF value.

Our 10-year DCF gave US$175.79, US$252.56, and US$478.30 a share. The price was US$166.54 on May 19, 2025.

Layers

01The case4 reasons · 4 risks

The case for

  1. Search still has room. Search is the cash cow, with about 90% global share. We assumed 0.8 trillion more searches a year.

  2. Cloud grows fastest. Google Cloud grew 27% in 2024, faster than any other segment. First-time customer commitments doubled from 2023.

  3. The whole AI stack. Google makes its own AI chips, models, and cloud. Owning every layer lets it ship AI faster.

  4. Bets beyond ads. Waymo, Verily, Calico, and Quantum AI add ways to grow beyond advertising.

The case against

  1. AI answers vs ads. Gemini answers could push ads down the page and cut clicks. Waiting too long could cost users instead. We called it an innovator’s dilemma.

    Our answer: grow Google Cloud with enterprise customers.

  2. Antitrust. US and EU cases target Google’s search deals, including its default spot on Apple devices.

    Our answer: the sentencing trial was projected to take about five years.

  3. Search rivals. Bing with ChatGPT, DuckDuckGo, and Opera compete for searches.

    Our answer: losses in niche corners would not dent profits much.

  4. Ad rivals. Amazon, Meta, and TikTok compete for the same ad budgets.

    Our answer: YouTube watch time keeps growing.

02The numberssearch · peers · 4 figures

Room left in search

About 5 trillion searches a year, under a third of an estimated 16.7 trillion. At 0.8 trillion more a year, that is over a decade of room.

Cheaper than its peers

EV/EBITDA as of Mar 31, 2025, from the comparison table in our report.

Key figures

  • 9.44%

    Our discount rate (WACC). Its cost of equity used a 4.25% risk-free rate, a 0.96 beta, and a 6% market premium.

  • 5T of 16.7T

    Google searches a year, against an estimated 16.7 trillion possible.

  • ~14x vs ~17.3x

    EV/EBITDA against its own median. Its history ran from 11x to 22x.

  • +27%

    Google Cloud revenue growth, 2024.

Method

A 10-year free-cash-flow DCF with three growth cases, at a 9.44% WACC. We checked it against a six-company peer table. We cited interviews with three current and former Google specialists, in AI, software, and quantum computing.

“…Google still operates at less than one-third of the estimated total addressable market of 16.7 trillion searches per year.”

My line in our pitch script, 2025.
03The recordMar 22 to May 2025

  1. Priced at US$163.99. Target US$252.56 (+54%). Call: buy to hold.

  2. Submitted the report, the deck, and a video.

  3. Named a Top 100 team.

  4. Repriced at US$166.54 for the summit. Call from buy to hold to buy. Same target.

  5. I shifted the long-term section from quantum computing to Waymo and other bets.

  6. Attended the Global Youth Investment Summit in New York.

HKEX · Pitch figures as of Feb 11, 2026

HKEX: 9926.HK · Biopharmaceuticals · 2026 · Solo

Akeso

A Chinese biotech built on bispecific antibodies. Each of its lead drugs binds two targets at once.

  • Buy
  • Solo

Winter 2026. A solo entry with a written report and a video. I built the DCF and the Monte Carlo myself.

Its lead drug beat Keytruda in a Phase III lung cancer trial. I wrote a Monte Carlo to map the spread around my DCF value.

My callBase

Price (HK$)Feb 11, 2026
102.70
Target (HK$)
160.02
Upside
up 56%
Price now (HK$)Oct 20, 2026
91.75
Since pitch from HK$102.70, Feb 11, 2026
down 11%

My target. 30.2% revenue growth a year, margins rising to 45% by 2030, a 9.7% WACC.

My DCF, bear to bullHK$

Range from HK$77 to HK$318. Marks show the HK$102.70 price, my HK$160.02 base case and the HK$173.04 consensus.

  • Beardown 25%
    price 102.70 to DCF value 77
  • Baseup 56%
    price 102.70 to target 160.02 · target
  • Bullup 210%
    price 102.70 to DCF value 318
  • Monte Carlo, 95% of paths
    105.83–231.93 · mean 159.77 started at my DCF value
  • Analyst consensus173.04

Hollow square: price on that date.Filled square: target or DCF value.Dashed: points down.

My DCF range against the market. Price on Feb 11, 2026: HK$102.70. Analyst consensus: HK$173.04.

10,000 one-year paths from my DCF valueHK$

Geometric Brownian motion at 20% volatility and a 4% risk-free rate, over 252 trading days.

Notebook
Mean HK$159.77 · 95% between HK$105.83 and HK$231.93

The paths start at my DCF value, not the market price. It maps the spread around my number.

Layers

01The case4 reasons · 5 risks

The case for

  1. Two targets, one drug. Ivonescimab blocks PD-1 and VEGF in one molecule. In the HARMONi-2 trial, it beat Keytruda on progression-free survival.

  2. Two engines. Cadonilimab already sells, with over RMB 2B in 2024 sales. Ivonescimab is the next growth driver.

  3. A patent cliff. Keytruda sells over US$25B a year. Its patents expire in 2028.

  4. A partner for the West. Summit Therapeutics runs ivonescimab in North America and Europe. It paid US$500M upfront, with up to US$5B in milestones.

The case against

  1. The FDA. US approval is not certain. The FDA could ask for more data or reject the filing.

  2. One partner. Summit runs the drug in the US and Europe. If Summit stumbles, Akeso cannot step in.

  3. Price cuts at home. China’s insurance listing forces price cuts on domestic sales.

  4. Geopolitics. The US BIOSECURE Act could restrict Chinese biotech in Western markets.

  5. Dilution. Continued losses could force share sales that dilute holders.

02The numbersPFS · 4 figures · why a Monte Carlo

Months before progression

Median progression-free survival in HARMONi-2. Ivonescimab 11.14 months, Keytruda 5.82.

Key figures

  • 11.14 vs 5.82 months

    Median progression-free survival, ivonescimab vs Keytruda. HARMONi-2, first-line PD-L1-positive lung cancer. That is a 49% lower risk of progression.

  • HK$77–HK$318

    My DCF, bear case to bull case. The base case gave HK$160.02.

  • HK$105.83–HK$231.93

    Where 95% of 10,000 simulated one-year paths landed, around my DCF value.

  • 32.5x vs ~12x

    Akeso’s price-to-sales against Innovent’s. WuXi Biologics traded near 10x.

Why I added a Monte Carlo

My DCF gave one number: HK$160.02. For a biotech, that felt too neat. Trials and approvals can swing either way.

So I wrote a Monte Carlo in Python, in Google Colab. It ran 10,000 one-year price paths from my DCF value, at 20% volatility.

95% of outcomes landed between HK$105.83 and HK$231.93. It maps the spread around my number.

“…standard DCF models were overly reductive…”

My report, Feb 2026.
03The recordFeb 18 to Feb 20, 2026

  1. Built my own DCF sheet and the Monte Carlo notebook.

  2. Submitted the report, the deck, and a video. Buy, HK$102.70 → HK$160.02 (+56%).

  3. When I pitched, the FDA was expected to rule on Summit’s filing by this date.

Student analysis, not investment advice. Pitch figures are as of each pitch date. Current prices are labeled with their date.

06 Everything else

Make. Climb. Play. Read.

Full entry

  1. 01 Make

    • Espresso
    • Bartending

  2. 02 Climb

    • Rock climbing
    • Mt. Baldy 10,064 ft, this winter

  3. 03 Play

    • Violin In an orchestra
    • Poker

  4. 04 Read

    • Seven books
    • Atlas Shrugged Objectivism

01 Make.

Espresso making

Espresso

Drinks

Bartending

02 Climb.

Climbing

Rock climbing

Mountaineering Next

Mt. Baldy

This winter I plan to summit Mt. Baldy. Its official name is Mount San Antonio.

Summit
10,064 ft (3,068 m)
Range
San Gabriel Mountains, highest peak

03 Play.

Jonathan in a dark suit and patterned tie holds his violin and bow after a concert. He stands in a hall with organ pipes, with other musicians and the audience behind him.
After the concert, Nov 10, 2024.

Orchestra

Violin

I play orchestral violin. Four hours after finishing my first marathon, I was in concert black.

SuppéBorodinBrahms

Cards

Poker

04 Read.

Reading

Seven books I’ve been reading, oldest first. Four of the seven date from one 52-year stretch of the 1800s.

Seven books by the year each was written or first published. Dashed: the Gita’s date is debated. Bracket: the Objectivism strand, Ayn Rand’s philosophy. Left of the // break, the axis is not to scale.
c. 2nd century BCEBhagavad Gita Traditionally attributed to Vyasa

Original language Sanskrit

A dialogue between Krishna and Arjuna, set within the Mahabharata epic.

Its date is debated. Scholars’ estimates run from the 5th to the 2nd century BCE.

1834–1855Journals Søren Kierkegaard

Original language Danish

Private journals and papers Kierkegaard kept from his student years until his death.

Princeton’s full edition runs to 11 volumes, 2007–2020.

1848White Nights Fyodor Dostoevsky

Original language Russian

A short story about a lonely dreamer over four nights in St Petersburg.

1866Crime and Punishment Fyodor Dostoevsky

Original language Russian

A novel about a former student in St Petersburg who commits a murder.

First published in installments in the journal Russkii Vestnik, in 1866.

1886Beyond Good and Evil Friedrich Nietzsche

Original language German

A book of aphorisms questioning the moral assumptions of earlier philosophers.

German title: Jenseits von Gut und Böse.

1957Atlas Shrugged Ayn Rand

Original language English

Strand Objectivism, Ayn Rand’s philosophy

A novel set in a dystopian United States, where business leaders abandon their companies.

2005Social Theory and Postcommunism William Outhwaite and Larry Ray

Original language English

A sociology book on what the end of communism means for social theory.

Published by Blackwell, Malden MA and Oxford.