Let Tim Cook

📈 AT&T and Apple keep climbing, flexing their fiber and AI plans. Meanwhile, SoFi got smacked despite its “best year ever,” and AI stocks are shaking as China’s DeepSeek disrupts the game.

January 28, 2025

Silver banana goes to…




In this issue of the peel:

  • 🏚️ Home sales hit rock bottom—lowest since ‘95, when Clinton was president, and America didn’t know Monica yet. Blame sky-high mortgage rates and a housing supply that moves slower than Biden standing up.

  • 📈 AT&T and Apple keep climbing and flexing their fiber and AI plans. Meanwhile, SoFi got smacked despite its “best year ever,” and AI stocks are shaking as China’s DeepSeek disrupts the game.

  • 🤯 AI drama as DeepSeek builds ChatGPT-level tech for a fraction of the cost of budget hardware. Nvidia is crying, but Zuckerberg is still spending billions of dollars like it’s monopoly money.

Market Snapshot

Banana Bits

Register Now: The Pitchbook Evolved

Ensure your pitchbooks win every time with Stifel, Leverest, and UpSlide.

UpSlide is hosting CIB experts from Stifel and Leverest to share their strategies and tips for deal teams to gain a competitive edge in 2025.

They’ll break down some of the biggest obstacles faced by pitching teams, as well as how to overcome them. Plus, they’ll share insights on:

  • Why pitchbooks are shrinking while deal competition is rising

  • What boutique teams can do to stay ahead of their bulge bracket counterparts

  • How to automate your work in PowerPoint and Excel to impress your MD

Save the date:
📆 January 30th
⌚ 10:30 am EST
⌛ 45 minutes

Macro Monkey Says

What Happened?

Honestly, we could’ve written this story months ago. The writing’s been on the wall for a while, but as of last Friday, the data was officially confirmed.

2024 saw the fewest home sales in the United States since 1995. For context, Bill Clinton was President in 1995, but the rest of the country still had 3-years to wait before meeting Monica Lewinsky.

Anyway, high mortgage rates, stubbornly high prices, and the cause of it all—an anemically low supply of housing—made sure 2025 was absolutely not the year for housing to come back.

The Numbers

4.06mn existing homes were sold from Sea to Shining Sea last year, just below the 4.09mn sold in 2023. That means home sales are down by a third from 2021.

Every cost you could think of related to not being homeless is at or near all-time highs.

Since late 2022, getting a mortgage under 6% has been like finding a college-age male who doesn’t bet on sports—physically impossible.

In fact, over the last 2 years, there was really only a period of about 4-months where the average 30-year rate on new mortgages sat below 6.5%. 

Prior to the pandemic, the last time average 30-year rates were that high was in 2002, a full year before 50 Cent dropped Get Rich or Die Tryin’.

Usually, when mortgage rates rise—or the interest rate linked to any asset—the principal value of that asset falls. However, that has rarely been the case in the post-pandemic housing market.

Three big forces are at play here. First, homeowners who bought houses during the ZIRP era don’t want to double their mortgage rate, causing them to stay in their current homes when many may have otherwise moved or downsized.

Second, the Millennial and Gen Z generations are two of the largest in American history. Both generations are at or entering their home-buying/household formation years, and combined with WFH, it’s easy to see why demand is so high, keeping prices elevated.

Third—our favorite thing to complain about—is the housing supply. Inventory levels have been subpar since the GFC traumatized homebuilders and, with permitting speeds slower than Joe Biden standing up, don’t expect this to change anytime soon.

America’s housing problem summed up in one sentence: We built half as many homes during the 2010s as we did in the 1970s.

Completions have since been in decline since August, only exacerbating future issues.

The Takeaway?

That’s not even to mention the additional cost of skyrocketing home insurance prices, increasing property tax rates across much of the country, and the general increase in the cost of living over the past 5-years.

Time to drop the “American” from “The American Dream.” For most young Americans, owning a home looks about as realistic as just another dream.

Pending rate cuts could help to spur demand, but unless supply is increased dramatically, record-high demand on the sidelines will keep prices elevated.

Feel free to blame my dad again. He’s a carpenter, so everyone can call him and tell him to get back to work building our houses. His number is 248-434-5508.

Career Corner

Question

I was wondering if, in interviews for IB internships, the interviewers ask questions about Excel. Do we need to know Excel well and how to build models or DCFs in Excel for interviews?

Answer

Unlikely, you'll get a question about Excel specifically. I would focus more on understanding the concepts behind modeling, DCFs, LBOs, etc.

You may get a case study, in which case Excel work and speed will matter, but unlikely in a face-to-face interview.

Head Mentor, WSO Academy

What's Ripe

AT&T (T) 6.3%

  • Wondering what the hot gift of the year was for Holiday szn 2024? Wonder no more—clearly, the top gift of the year was a bundled 5G and fiber plan from AT&T.

  • Actually, it doesn’t sound like a bad gift at all, and neither was Ma Bell’s fourth quarter. AT&T beat estimates on sales and earnings, reporting $0.54/sh on $32.3bn in revenue vs expectations for $0.50sh on $32.04bn.

  • The wireless provider’s 5G and fiber packaged deals led to subscriber additions of 482k, beating analyst estimates for just 425k. Alone, AT&T’s fiber business added 307k subs, its strongest growth ever.

Apple (AAPL) 3.3%

  • This is your daily reminder to work smarter, not harder, or maybe even not at all. After doing checks notes f*ck all with AI so far, Apple was the lone tech stock to survive Monday’s Massacre.

  • We give you the rundown below, but Apple defied the gravity of its sector on Monday as markets are seemingly starting to appreciate Tim Cook’s AI strategy.

  • Letting the minions do the work for them, Apple has no plans to build its own LLM but to use its distribution to monetize the best model(s) on iPhones and other devices going forward.

  • In that context, Deepseek’s R1 breakthrough is a huge boon to Apple. The firm just wants the best model it can get to maximize distribution and monetize users at the application layer. See what happens when we let Tim Cook? 

What's Rotten

SoFi Technologies (SOFI) 10.3%

  • Markets wasted no time telling SoFi CEO Anthony Noto where he could shove his “2024 was SoFi’s best year ever” idea. Yes, it’s true, but his mistake was expecting the market to care.

  • Q4 GAAP EPS exploded to $0.29/sh, up from $0.05/sh in Q3 and $0.02/sh in Q4’23. Net revenue of $739.1mn beat estimates for $682.2mn and grew 24.4% from last year.

  • Member growth crossed the 10mn user line, helping boost SoFi’s strategy of cross-selling banking and wealth management products to users. Loan originations grew 66% from Q4’23.

  • Despite delivering their first-ever full-year profit, markets focused on lackluster guidance, projecting Q1 EPS of 0.03/sh vs the $0.06/sh analysts were pricing in.

AI Stocks (ARTY) 8.4%

  • 5th-grade history class teaches us about the “shot heard ‘round the world” that started the American Revolutionary War. Last week, we got the AI heard ‘round the world…

  • We dive into this below, but for now, just know a Chinese startup built an LLM on par with the ChatGPTs of the world for less than 1/10th the cost on lower-quality hardware.

  • This shines a bright light on how quickly best-in-class technology can become your grandpa’s tech, especially in the AI market.

  • Nvidia lost 16.9%, the biggest market cap loss in history at almost $600bn. Broadcom fell 17.40%, Vistra Energy lost 28.28%, and Oracle shed 13.79%. 

Thought Banana

The Fan’s Been Hit

You would’ve thought Silicon Valley had their own version of the 2020 Beirut explosion the way these stocks opened yesterday.

Nvidia must’ve replaced CEO Jensen Huang with Garfield Monday morning as shares opened down 12.5% from Friday’s close.

Let’s dive in.

What Happened?

Sh*t hit the fan in the world of AI last week when everyone learned about a little-known Chinese startup.

DeepSeek—founded in May 2023 by Liang Wenfeng and backed by hedge fund High-Flyer, which Wenfeng also founded—is a Chinese AI startup that, until last week, existed in relative obscurity.

Then, on January 20th, DeepSeek released DeepSeek R1, an LLM that made ChatGPT’s blow-up look slower than the adoption of the printing press by the Catholic Church.

R1 shot to the top 10 across almost every LLM performance leaderboard instantly. Over the weekend, DeepSeek R1 became the #1 most downloaded app on Apple’s App Store and is still #1 in Productivity apps, above ChatGPT at #2.

But popular apps emerge out of nowhere all the time. That’s not why the firm was scared enough to send Nvidia 16.86% lower on the day.

The remarkable aspect of DeepSeek R1 is how dirt cheap it was to make. 

DeepSeek was able to make R1, a model currently competitive with ChatGPT-o1, with total training costs of $5.576mn. Training for ChatGPT-4, by comparison, ran OpenAI $78mn while Google’s Gemini Ultra cost $191mn.

Moreover, using ChatGPT’s API has a cost of $15 per million input tokens and an output cost of $60. While delivering the same quality, DeepSeek R1’s API cost per million input tokens is just $0.55, while output tokens are just $2.19.

So, training for this model was ~15x cheaper than ChatGPT, and per-unit API costs are ~27.25x cheaper.

To make matters even more insane, DeepSeek did this on junk hardware. 

Export restrictions have left Chinese AI firms hanging, unable to get their hands on Nvidia’s latest chips, which set the standard for best-in-world. DeepSeek can’t access Blackwells, H100s, or A100s—Nvidia’s latest best-of-the-best AI data center chips.

But, before export restrictions took place, DeepSeek got its hands on a few thousand A800 and H800 chips, older and less advanced versions of the above chips.

So, really pouring salt in the wound, DeepSeek’s model achieved performance on par with U.S. LLM leaders at a way cheaper cost on worse hardware.

For comparison, Meta AI’s Llama 3.1 405b model used 30,840,000 H100 GPU hours for training. DeepSeek used 2,788,000 H800 GPU hours to get the same if not a better, model quality.

We can keep going, too—it’s just crazier. 

DeepSeek was able to do this by tweaking a few aspects of the model creation, training, and inference process.

We don’t have time to go into the technical side, but DeepSeek’s 1) Chain-of-Thought inference methodology, 2) Reinforcement Learning, and 3) Distillation were the fundamental game changers making every AI engineer in the world right now think, “Why didn’t I think of that??!”

The Takeaway?

This is one of those “better, faster, cheaper” moments where a new product disrupts the existing industry by displaying all three of those characteristics against competitors. 

DeepSeek was able to create a model on par with slightly older versions of the U.S.’s current best-of-the-best AI models at less than 1/10th the cost and training time while on worse hardware. Nvidia is in shambles.

That means companies might not need to buy as many chips and on-chip training hours from Nvidia. 

Arguably the best part is that while all this was going on, Mark Zuckerberg announced that Meta still plans to spend $65bn on AI model training and improvements in 2025 rather than figuring out how tf DeepSeek just did that at 1/1,000ths the cost. 

Probably best to just pull a Microsoft and invest it all in DeepSeek.  

Here’s a great, quantitatively-based article that goes deep on DeepSeek if anyone wants to know more.

The Big Question: How will U.S. companies respond, especially given that DeepSeek is open source? What other questions should we be asking?

Banana Brain Teaser

Previous

A dance troupe has a total of 50 dancers split into 2 groups. The costumes worn by Group A cost $80 each, and those worn by Group B cost $90 each. If the total cost of all the costumes if $4,270, what is the total cost of the costumes worn by Group B?

Answer: $2,430

Today

A doctor prescribed 18 cubic centimeters of a certain drug to a patient whose body weight was 120 pounds. If the typical dosage is 2 cubic centimeters per 15 pounds of body weight, by what percent was the prescribed dosage greater than the typical dosage?

Send your guesses to [email protected]

❝

A 10% decline in the market is fairly common—it happens about once a year. Investors who realize this are less likely to sell in a panic, and more likely to remain invested, benefitting from the wealthbuilding power of stocks.

Christopher Davis

How Would You Rate Today's Peel?


Happy Investing,
David, Vyom, Ankit & Patrick