Matt Levine’s Money Stuff newsletter this week features an insider trading cases that makes you question whether the perpetrators were genuinely clever or just extremely lucky they weren’t caught sooner. Here’s a thought experiment Levine poses:
Let’s say you have cancer. A biotech company is running clinical trials for a promising drug. You join the trial, the drug works, you’re cured. Before the company releases results, you buy call options on their stock. Is this illegal insider trading?
Probably not. Trading on your own experience, “I used a product and I liked it so I bought the stock,” is generally fine. That’s how lots of retail investors operate. But what if you’re a cancer doctor instead? You call the drug company to ask about getting your patient into the study and they share some preliminary data with you. Seems sketchier. You probably signed something promising confidentiality. Now let’s escalate:
What if you don’t actually have a patient? What if you’re not actually a doctor? What if the information you get isn’t market-moving enough, so you just make up fake trial results, steal the identities of actual cancer patients, and post fabricated data on patient forums?
That’s where we arrive at the SEC’s case against the Shoukat brothers. Last week, the SEC charged Muhammad Saad Shoukat, Muhammad Arham Shoukat, and Muhammad Shahwaiz Shoukat with a wild assortment of market manipulation and insider trading schemes. According to the complaint, Saad and Arham impersonated physicians named “Dr. Joseph Garza” and “Dr. Safqat Anwwar” to steal confidential clinical trial information from Olema Pharmaceuticals. They sent fake patient medical records to clinical trial coordinators. When the information they gathered wasn’t market-moving enough, they simply fabricated better results, then stole identities from breast cancer patient forums to post the fake data.
They made about $250,000 on call options.
Looking at the case, that scheme was actually the hard work. The much easier part of their alleged operation involved a friend named Justin Kim, a Lazard associate who tipped them about M&A deals he was working on. That generated about $41 million in profits.
“Get ready bro,” Kim allegedly texted Shoukat in April 2023 after learning Immunogen was in talks with AbbVie. “S**t is about to pop off.”
What did Kim receive for providing tips on ten potential takeovers over several years? A Rolex watch, career advice, and, according to the SEC complaint, “help in drafting slides for a PowerPoint presentation Kim was preparing as part of his job at Lazard.”
PowerPoint decks can be hard but for insider trading the compensation is usually cash, sometimes crypto, occasionally a job or a vacation. Never before have I heard of PowerPoint assistance as a form of illegal kickback. You have to wonder: was it a particularly complicated deck? Did Saad have strong opinions about animation transitions? Was there a tricky chart that needed cleaning up?
Kim now faces both criminal and regulatory charges. The DOJ unveiled fraud and insider trading charges that carry up to 25 years in prison. The SEC is seeking to permanently ban him from the industry.
The Shoukat brothers also pulled off an Opiant Pharmaceuticals scheme where they bought stock based on acquisition tips, then when the deal stalled, allegedly threatened company leadership and issued a fake press release announcing a fictitious partnership deal. There’s a lot of the creative energy here, if it weren’t so obviously criminal. Impersonating doctors, fabricating clinical data, stealing cancer patient identities, threatening executives, issuing fake press releases. And yet the big money, the $41 million, came from the simplest possible scheme: a friend with a good job who texted them when deals were about to happen.
Sometimes insider trading cases involve sophisticated financial engineering or complex offshore structures. This one involved pretending to be a doctor named “Dr. Safqat Anwwar” and asking for help with PowerPoint slides.
A friend recently recommended Steve Eisman’s podcast to me. Eisman, you might recall, is the hedge fund manager portrayed in The Big Short who famously bet against subprime mortgages before the 2008 crisis. In his most recent episode, Eisman laid out a thesis for something that made me uncomfortable ever since the Covid-19 stock market crash recovery: the U.S. equity market has structurally decoupled from everyday economic reality.
I’ve written about market concentration in my 2026 portfolio allocation. But Eisman’s point isn’t just about concentration. It’s about what this concentration means for everyone else. Consider what happens to consumer-exposed sectors. Combined, healthcare, consumer discretionary, and consumer staples have fallen from 38% of the index in 2015 to just 25% today. This matters because roughly 70% of U.S. GDP is consumer-driven. The traditional logic was simple: consumer spending drives the economy, consumer stocks reflect that spending, and therefore the stock market reflects economic health. That relationship has broken down.
The disconnect shows up in daily American life. Healthcare costs continue rising, housing remains unaffordable for many, and grocery prices have yet to normalize. These are real pressures on real households. Yet the S&P 500 gained 16% in 2025, with the Nasdaq up 21%. The market doesn’t care about rent or insurance premiums because the companies reflecting those costs barely register in the index anymore. As Eisman puts it:
The market has become unmoored from everyday life.
This creates a structural problem for active managers that compounds over time. When NVIDIA alone represents 7.7% of the S&P 500, Apple 6.8%, and Microsoft 6.1%, most institutional mandates physically prevent managers from holding proportional positions. Risk limits cap initial positions at perhaps 5% of assets under management. Sector allocation rules require diversification across all eleven sectors. The result is systematic underweighting of the fastest-growing names. Meanwhile, the bottom five sectors combined represent just 14% of the index. Real estate, with 31 constituents, accounts for barely 2%. Why dedicate research resources to an entire sector that can only marginally move your portfolio?
The rise of passive investing amplifies all of this. Index funds now control roughly 60% of flows versus 40% for active managers. When money enters an index fund, it buys stocks in proportion to their existing market cap. Large positions grow larger. There’s no portfolio manager deciding NVIDIA looks expensive. The buying is mechanical, price-insensitive, and self-reinforcing. This doesn’t eliminate price discovery entirely. Eisman points to Oracle’s Q3 2025 experience: shares surged after reporting a massive backlog, then corrected below pre-earnings levels once investors realized the backlog concentrated in a single customer with questionable financing. Active managers still matter. They just matter less.
In a normal correction, sellers meet buyers who evaluate whether prices have become attractive. In a passive-dominated market, redemptions trigger mechanical selling. Index funds don’t decide that a 20% drawdown makes stocks compelling. They sell what they own in proportion to what they own. If active managers control only 40% of flows, the stabilizing bid may prove insufficient. The February-April 2025 correction saw the S&P fall 19% peak-to-trough. Eisman’s assessment: if an actual recession materializes, or if AI spending disappoints expectations,
the decline will almost certainly be steeper. It will be fast and very ugly.
There’s also a tax dimension creating behavioral lock-in. Years of technology outperformance have embedded massive unrealized capital gains in both retail and institutional portfolios. Selling NVIDIA means realizing those gains and paying taxes on them. Investors avoid this until forced by margin calls, redemptions, or actual fundamental collapse. This creates asymmetric liquidity: plenty of buyers on the way up, scarce ones on the way down.
What does this mean for portfolio construction? First, understand that traditional cap-weighted benchmarks now represent a concentrated bet on technology and AI capital expenditure. Second, active management faces structural headwinds that have nothing to do with manager skill. Third, liquidity assumptions that held in previous corrections may not hold in the next one. And fourth, consumer welfare can deteriorate materially without meaningfully impacting index returns. The K-shaped economy produces a K-shaped market, where the experience of median households and the experience of median stock index performance have genuinely diverged.
You’ve seen this message before. Copilot pausing; In long sessions, it happens often enough that I started wondering what’s actually going on in there. Hence this post.The short answer: context windows grew larger. Claude handles 200K tokens, Gemini claims a million. But bigger windows aren’t memory. They’re a larger napkin you throw away when dinner’s over.
For som time I was convinced that vector databases would solve this. Embed everything, store it geometrically, retrieve by similarity. Elegant in theory. Try encoding “first we did X, then Y happened, which caused Z.” Sequences don’t live naturally in vector space. Neither do facts that change over time. Your database might confidently tell you Bonn is Germany’s capital if you fed it the wrong decade of documents.
What caught my attention is EM-LLM. The approach is basically “what if we just copied how brains do it?” They segment conversation into episodes using surprise detection; when something unexpected happens, that’s a boundary. Retrieval pulls not just similar content but temporally adjacent content too. You don’t just remember what you’re looking for. You remember what happened next. Their event boundaries actually correlate with where humans perceive breaks in experience. Either a coincidence or we’re onto something.Knowledge graphs are the other path. Persona Graph treats memory as user-owned, with concepts as nodes. The connection between “volatility surface” and “Lightning McQueen” exists in my head (for some reason) but probably not yours. A flat embedding can’t capture that your graph is different from mine. HawkinsDB pulls from Thousand Brains theory. Letta just ships, production-ready blocks you can use today. OpenMemory goes further, separating emotional memory from procedural from episodic, with actual decay curves instead of hard timeouts. Mem0 reports 80-90% token cost reduction while quality goes up 26%. I can’t validate the claim, but if it holds, that’s more than optimization.
HeadKV figured out that attention heads aren’t created equal: some matter for memory, most don’t. Throw away 98.5% of your key-value cache, keep the important heads, lose almost nothing. Sakana AI went weirder: tiny neural networks that decide per-token whether to remember or forget, evolved rather than trained. Sounds like it shouldn’t work. Apparently works great.
Here’s what I keep coming back to: in any mature system, most of the graph will be memories of memories. You ask me my favorite restaurants, I think about it, answer, and now “that list I made” becomes its own retrievable thing. Next time someone asks about dinner plans, I don’t re-derive preferences from first principles. I remember what I concluded last time. Psychologists say this is how human recall actually works; you’re not accessing the original, you’re accessing the last retrieval. Gets a little distorted each time.
Should the model control its own memory? Give it a “remember this” tool? I don’t think so, not yet. These things are overconfident. Maybe that changes. For now, memory probably needs to happen around the model, not through it. Eventually some learned architecture will make all this scaffolding obsolete. Train memory into the weights directly. I have no idea what that looks like. Sparse mixture of experts with overnight updates? Some forgotten recurrent trick? Right now it’s all duct tape and cognitive science papers.
Warren Buffett has stepped down as CEO at 95. Greg Abel inherits a company that paid $26.8 billion in federal income taxes last year, roughly 5% of what all of corporate America paid combined. I do not have much in common with Buffett, but I will miss his shareholder letters. Berkshire’s archive is a rare case of a public company explaining decisions candidly to its owners.
In the 2024 letter Buffett repeats Tom Murphy’s rule: “Praise by name, criticize by category.” Murphy gave him this advice 60 years ago. The letter closes with another line worth keeping: “Kindness is costless but priceless.”
Three behaviors from those letters matter. (1) Communication discipline: Between 2019 and 2023, Buffett used “mistake” or “error” 16 times in his letters. Many Fortune 500 companies never used either word once. Amazon made “brutally candid observations” in its 2021 letter. Elsewhere, it has been happy talk and pictures. (2) Patience as allocation strategy: In the 1999 letter he told shareholders that “truly large superiorities” over the index were past because Berkshire’s size constrains opportunity. That reads today like the core constraint of the Abel era, a point I reflected on when writing about portfolio limits in 2026. (3) The non-theatrical life: Coverage of the retirement keeps returning to the same facts: still living in the Omaha home he bought in 1958, still driving 7 minutes to work every day and stopping at a drive-through for McDonald’s breakfast. The man is a true American hero.
Read the letters chronologically and you see Berkshire become a system rather than a portfolio. The early articulation is there in the 1983 letter: partnership mentality, per-share intrinsic value, and a preference for businesses that generate cash and earn strong returns on tangible equity.
The engine is insurance float. Property-casualty insurers collect premiums upfront and pay claims years or decades later. That gap creates investable capital at zero or negative cost. As Buffett puts it in the 2024 letter: “When writing P/C insurance, we receive payment upfront and much later learn what our product has cost us, sometimes a moment of truth that is delayed as much as 30 or more years.” In 2024, Berkshire’s insurance operations generated $9 billion in underwriting profit and $13.7 billion in investment income. GEICO, “repolished” over five years by Todd Combs, had what the letter calls a “spectacular” year.
On the question of where Buffett was right: if you ban the word “mistake,” risk does not vanish; it goes off balance sheet until it detonates. In the 2024 letter he writes:
I have also been a director of large public companies at which ‘mistake’ or ‘wrong’ were forbidden words at board meetings or analyst calls. That taboo, implying managerial perfection, always made me nervous.
He was also right about scale. In 2024, 53% of Berkshire’s 189 operating businesses reported declining earnings, yet the company posted $47.4 billion in operating profit. That is diversification at scale, but also its constraint. The next decade hinges on a handful of large moves. Markets understood this as the Abel era officially began.
On a more critical note: Berkshire is an insurance-anchored allocator with operating companies plus a concentrated equity book. The 2024 marketable equity portfolio stood at $272 billion, down from $354 billion after significant Apple sales. At its peak, Apple represented roughly 40-50% of Berkshire’s public equity holdings. A single stock, bought mostly between 2016 and 2018, drove a substantial portion of portfolio returns over the past decade. The rest is insurance. This is not a criticism of the Apple thesis (it was correct), but the Buffett track record includes one very large, very right bet on a technology company he famously avoided for most of his career. His letters do not pretend error is rare; they treat delay as the sin. He has been candid about blind spots, discussing lessons from IBM and airlines.
Berkshire’s businesses also face difficult labor dynamics. BNSF has drawn union criticism over attendance policies and OSHA findings in retaliation cases. The railroad earned $5 billion in 2024, flat with 2023.
So what now? Abel inherits roughly $300 billion in cash and Treasury bills. The letter explains this is not a preference for cash: “Berkshire will never prefer ownership of cash-equivalent assets over the ownership of good businesses, whether controlled or only partially owned.” The cash is a byproduct of not finding anything worth buying at current prices. The first large capital move will tell us more than any profile can, which is why coverage keeps circling back to the cash pile and the question of how much “Buffett premium” was embedded in Berkshire shares.
If Buffett truly goes quiet, I hope he gets to experience not working at all, or at least the version that suits a man who prefers thinking to talking. The compounding may continue just fine.
Bitcoin’s security model rests on one assumption: attacking the network costs more than any attacker could gain. A 2024 paper by Farokhnia and Goharshady does the math on this assumption and finds it wanting.
For roughly $6.77 billion in hardware, an attacker could control over 50% of Bitcoin’s hash rate. With 30% of hash power, success probability exceeds 95% within 34 days at a cost of about $2.9 billion.
These are large absolute numbers, but relatively small: Bitcoin’s $1.78 trillion market cap and the monthly derivatives volume that now regularly exceeds $2 trillion on unregulated exchanges alone. The attack pays for itself if you short Bitcoin before crashing its price.
The paper challenges three assumptions the crypto community has treated as fact. (1) That you need 51% of hash power to attack successfully. Not true. With 30%, you have high probability of reverting six blocks, enough to shatter the six-confirmation finality standard most practitioners rely on. (2) Acquiring majority hash power is prohibitively expensive. It is expensive but represents less than 0.5% of Bitcoin’s market cap. (3) miners have no incentive to attack since they depend on BTC-denominated rewards. This ignores derivatives entirely.
In simple terms an attack would work like this: An attacker acquires put options or other shorting instruments on Bitcoin. They then use their hash power to mine secretly, building an alternative chain. When their chain exceeds the public chain in length, they publish it. The network switches to the longer chain. Transactions in the replaced blocks get reverted. Price crashes. The short position prints money.
David Rosenthal, writing on his blog, offers a detailed skeptic’s view of whether this attack is actually feasible. His analysis is worth reading because he identifies practical obstacles the paper’s theoretical framework glosses over. Acquiring 43% of the pre-attack hash rate means buying roughly two years of Bitmain’s production capacity beyond what’s needed to replace obsolete equipment. The purchase would be noticed. Power requirements approach 9.5 gigawatts, roughly double what Meta’s planned Louisiana 5GW data center will need by 2030. That power doesn’t (yet) exist in deployable form.
The short position presents its own problems. Patrick McKenzie’s explainer on perpetual futures describes how crypto derivatives actually work: frequent settlements, margin requirements, liquidation risk. A leveraged short held for 34 days has high probability of getting liquidated before the attack succeeds, given Bitcoin’s volatility. In nine of twelve months in 2025, an initial 10X leveraged short would have been liquidated within the month. Even if the attack succeeds, the resulting crash would likely trigger automatic deleveraging, reducing winnings precisely when they should be largest.
An insider attack looks more plausible on paper. A large mining pool already controls the hash rate. The October 2025 liquidation event on crypto exchanges saw $19 billion in forced liquidations. This demonstrates both the volatility of the market and its capacity to absorb large directional moves. But an insider who stops contributing to the public chain for weeks becomes visible. The hash rate is public data. A 30% drop would trigger immediate investigation.
The paper’s contribution, in my opinion, is making explicit what the derivatives market implies: Bitcoin’s security depends not just on proof-of-work economics but on the assumption that attackers cannot profit from price crashes. That assumption gets weaker as the derivatives market grows. Monthly Bitcoin futures volume on unregulated exchanges has exceeded $2 trillion in recent months. The paper’s calculations used April 2024 data. Since then, Bitcoin’s hash rate has roughly doubled to around 1,100 EH/s, increasing attack costs proportionally. But derivatives volumes have grown too.
The next halving arrives in April 2028. Mining rewards drop to 1.5625 BTC per block. Miners whose equipment is fully depreciated might view an attack as an exit strategy, particularly if they can monetize it through derivatives. Some large miners are already pivoting to AI data center hosting, suggesting they see diminishing returns from mining alone. Core Scientific plans to exit Bitcoin mining entirely by 2028.
What actually prevents this attack? Probably the practical difficulties Rosenthal identifies rather than any fundamental economic barrier. The market should price these risks but appears not to. Bitcoin trades as if 51% attacks are theoretical rather than economically viable. That may remain true as long as practical obstacles hold.
The Information published a piece today arguing that Apple’s restrained AI approach may finally pay off in 2026. The thesis: while OpenAI, Google, and Meta pour hundreds of billions into data centers and model training, Apple has kept its powder dry, sitting on $157 billion in cash and marketable securities as of Q4 2025. If the AI spending bubble deflates, Apple’s position looks rather clever. This piqued my interest, from a strategy point of view: Apple hasn’t been absent from AI. They’ve been making a specific bet that large language models will commoditize, and that value will flow to distribution and customer relationships rather than to whoever has the best model. The revamped Siri expected in spring 2026 will reportedly be powered by Google’s Gemini through a deal worth $1 billion annually. The custom Gemini model will run on Apple’s Private Cloud Compute servers.
This is consistent with Apple’s history. They didn’t build their own search engine. They took Google’s money to be the default on Safari. John Giannandrea’s retirement earlier this month, with Siri now under Mike Rockwell, signals internal recognition that something had to change.
The iPhone distribution advantage is underappreciated. Apple can push AI features through software updates to over 2.3 billion active devices. When Apple Intelligence features ship, they just appear. This is the same advantage that made Apple Music competitive against Spotify, or keeps Safari relevant despite Chrome’s benchmarks.
But the AI investment boom resembles previous cycles. Enormous capital flowing into a sector where barriers keep falling. That pattern often ends with winners who have distribution and customer relationships, not winners who spent the most on R&D. Apple’s bet isn’t guaranteed to be correct, but it’s defensible.
Apple’s $157 billion cash pile provides optionality. If AI startups face a funding crunch, Apple can acquire capability. If someone achieves a breakthrough, Apple has resources to respond. Apple has preserved its options.
We cannot leave decisions about how AI will be built and deployed solely to its practitioners. If we are to effectively regulate this technology, another layer of society, educators, politicians, policymakers […], must come to grips with the basics of the mathematics of machine learning.
I read a book that is sort of related to my recent writing on AI: Why Machines Learn: The Elegant Math Behind Modern AI by Anil Ananthaswamy.
I admire the attempt to actually explain the math, with a ton of equations, instead of doing the usual human drama story about geniuses and labs. I also admit I did not absorb all of it. That is not a complaint. It is a good sign that the author did not flatten the material into a few metaphors and call it a day.
If you have read my posts on why AI might commoditize rather than stay a winner-takes-all business, for example Is AI Really Eating the World?, this book is a useful reminder that the story still starts with linear algebra. Ananthaswamy begins at the beginning: vectors, dot products, projections. The move here is to treat these as geometry, not just arithmetic. A dot product is a way to measure alignment, but it is also a way to map one space onto another. Once you see that, a neural network layer reads less like mysticism and more like a pipeline of linear maps plus nonlinearities.
That framing is the book’s main strength. When it works, you get an intuition for how high-dimensional data gets rotated, scaled, and squeezed until classes separate or features become easier to represent. When it stops working, you still have the derivations, so you can slow down and check what you missed.
The tour is broad. It moves from perceptrons to backpropagation, then through Principal Component Analysis and Support Vector Machines, and later into convolutional networks and generative models like GANs. There is history, but it is there to support the technical arc, not to replace it. The “graded ascent” idea mostly holds. Early chapters give you enough scaffolding to follow later ones. If you have not looked at derivatives or matrix calculus in a while, you will feel it, but you can still keep pace if you accept that this is a book you sometimes read with a pencil.
Two things I liked in particular: (1) It does not pretend there is a single “deep learning trick” that explains everything. The methods are varied and the trade-offs are real. (2) It gives you enough math to see why some ideas scale and others do not, without turning into a textbook.
Friends have been asking how they can stay up to date with what I’m working on and keep track of the things I read, write, and share. RSS feeds don’t seem to be en vogue anymore, apparently. So I built a mailing list. What else would you do over the Christmas break?
From a previous marketing job I knew Mailchimp. Also, every newsletter I unsubscribe from is Mailchimp. I no longer wish to receive these emails.
Or obviously Substack. I read Simon Willison’s Newsletter sometimes. And obviously Michael Burry’s $379 Substack. Those are solid options, but I had a clear picture in mind of what I wanted. I wanted only HTML, no tracking (also why I use GoatCounter on my site and not Google Analytics), and full control of the creation and distribution chain from end to end. So I sat down and drew into my notebook, what I always do when I have an idea after a long walk or a hot shower.
I then went over to Illustrator (actually Affinity Designer, which I have been happily using since my Creative Cloud subscription ran out, sorry Adobe) and built a quick mockup of my drawing. I fed the mockup to Claude to generate pure HTML. After a few iterations it more or less looked like I wanted it to be.
The architecture: write the newsletter in Markdown (as I do for all of my blog). Render it as HTML. Fetch OpenGraph images from my Cloudflare CDN at the lowest feasible resolution and pull descriptions automatically. Format links with preview cards. Keep some space for freetext at the top and bottom.
I built a Python engine that renders my .md files to email-safe HTML. The script handles several things automatically: (1) It fetches OpenGraph metadata for every link using Beautiful Soup, caching results to avoid repeated requests. (2) optimizes images using Cloudflare’s image transformation service. For email, I use 240px width (2x the display size of 120px for retina displays). (3) It generates LinkedIn-style preview cards with images on the left and text on the right. The output is table-based HTML because email clients from 2003 still exist and they’re apparently immortal.
Originally I intended to manually copy-paste the HTML into an email and send it out since I did not expect many subscribers at first (or at all). But I had another challenge at hand: how do people sign up?
Since I had already been using Cloudflare Workers KV to build an API with historic values of my temperature and humidity sensor at home, I resorted to that. The API is simple. POST to /api/subscribe with an email address, and it gets stored in KV with a timestamp and some metadata.
After some Copilot iterations (I’m not a security guy, so not sure how I feel about handing all the security and testing to an agent, please reach out if you can help) the Worker includes rate limiting, honeypot fields for spam protection, proper CORS headers, and RFC-compliant email validation.
I then wanted to get a confirmation email every time someone signed up. Since SMTP sending over my domain did not work reliably at first, I had to look for other options. Even though I wanted everything self-hosted, I ended up using the Resend API. The API is straightforward:
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
asyncfunctionsendWelcomeEmail(subscriberEmail: string,env: Env){constresponse=awaitfetch('https://api.resend.com/emails',{method:'POST',headers:{'Authorization':`Bearer ${env.RESEND_API_KEY}`,'Content-Type':'application/json',},body: JSON.stringify({from:'Philipp Dubach <[email protected]>',to:[subscriberEmail],subject:'Welcome to the Newsletter',html:`<p>Thanks for subscribing!</p>`,}),});returnresponse.ok;}
After implementing this, I figured: why not send a confirmation to the subscriber and a copy to me? Why not use Resend for the whole distribution? (This is not a paid advertisement.) The HTML newsletter I generate goes straight into the email body. No images hosted elsewhere (except for the optimized preview thumbnails). No tracking pixels. No click tracking. The email is just HTML.
I also looked at Mailgun and SendGrid before settling on Resend. Mailgun has better deliverability monitoring but a more complex API. SendGrid has more features but felt overengineered for what I needed. Resend’s free tier and simple API won. If you have strong opinions on email APIs, I’m curious to hear them.
The total cost of running this: zero. Cloudflare Workers has a generous free tier. Cloudflare R2 (where the HTML newsletters are hosted) has 10GB free storage. Resend gives 3,000 emails per month. The Python script runs locally or on my Azure instance.
I have two portfolios: (a) long-term, diversified, low-cost ETFs, and (b) collecting diamonds in front of bulldozers, short-term option plays, and some individual stocks I find interesting. Here, we will only look at (a). This essay is structured along five themes I believe to be true for 2026:
(1) Market Concentration and High Valuations (2) US Dollar Depreciation Expected Despite Continued Dominance (3) AI Investment Remains Central But Requires Scrutiny (4) European Fiscal Revolution Creates Investment Opportunities (5) Fixed Income Offers Best Prospects Since Global Financial Crisis
Let’s start with the conclusion. Here’s how I will rebalance my portfolio going into 2026:
What changed and why?: US Equities (-10%): S&P 500 valuations at 23× forward P/E reflect peak optimism; Nasdaq’s 30× trailing P/E is unsustainable. I reduce large-cap exposure from 33% to 23% to avoid dual headwinds: equity mean reversion and USD depreciation. I redeploy 5% into US small-cap stocks, which offer better valuations and risk-adjusted returns. Europe (+5%): European equities trade at a 22% discount to global peers and benefit from Germany’s €1T+ infrastructure commitment and structural reforms. This compounds three tailwinds: improving fundamentals, valuation re-rating, and EUR stability vs CHF. I increase the allocation from 8% to 13%. Fixed Income (+4%): Global yields remain near post-GFC highs with 10-year UST around 4.2% in December 2025. I reallocate from 10% equities-focused bonds to 14% fixed income (CHF Corporates +5%, EUR Govt Bonds +3.5%, US Treasuries +2%), establishing duration exposure and counter-cyclical protection ahead of Fed rate cuts. Japan (Unchanged): Japan’s structural reforms and BOJ stimulus remain supportive, but a 3% Asia Developed exposure is adequate. JPY strength vs USD and CHF acts as a tailwind on existing holdings. Asia EM (+0.5%): Increase from 10% to 10.5% to capture Chinese stimulus and attractive tech valuations. CNY appreciation vs USD provides diversification and a natural hedge against dollar weakness. Alternatives (+0.5%): Increase Listed PE/Alt from 1.5% to 2%, maintaining access to Swiss private markets and uncorrelated returns with currency-matched positioning. Gold (+1%): I increase from 4% to 5%. Gold serves as a hedge against USD weakness while benefiting from record central bank reserve diversification. This measured increase captures structural de-dollarization demand without chasing 2025’s roughly +58% performance. Crypto (+0.5%): I increase from 4% to 4.5% as a diversification component. At the end of this article, I have included a short, more technical insight into how I structured my research process.
Before we dive into the detailed rationale for my 2026 asset allocation, let’s quickly look at how this portfolio performed this year. The portfolio delivered solid returns in 2025, outperforming the 60/40 benchmark, primarily driven by the CHF-hedged S&P 500 allocation and domestic dividend stocks, with CHF-hedged gold providing additional diversification gains. The 11.5% USDCHF depreciation proved to be the defining factor this year, while the S&P 500 returned +18% in USD. Unhedged USD exposure translated to just ~5% in CHF terms, vindicating the currency-hedged core equity strategy. Detractors included the unhedged MSCI World SRI ETF, also hit by FX exposure.
Now if you are still with me, let’s dive into the five themes that I believe to be important going into the next year.
• Market Concentration and High Valuations: The S&P 500 has become dangerously concentrated. As of December 2025, the top 10 companies represent approximately 45% of the index’s value, a historic concentration level not seen since the dot-com bubble. Nvidia alone accounts for over 7%. What was once a diversified investment across 500 companies is now heavily weighted toward a handful of tech giants, most betting heavily on AI.
The top 10 US companies dominate the world equity market: Top 5 US tech firms alone have a collective value ($17.6) that exceeds the combined GDP of the Japan, India, UK, France, and Italy ($17.1).
Even though there might be a consensus view among analysts that elevated valuations are supported by earnings growth rather than multiple expansion alone,
Valuations are especially high in the US. The S&P500 trades at 23 times forward earnings, near the top of its historical range. While the Nasdaq’s 30× trailing P/E is well below the dotcom bubble peak, it still reflects significant optimism. Outside the US, valuations are more moderate: European and Chinese equities are 10% and 7% above their 20-year average valuations, respectively, and Japan’s index trades at a discount to its long-term average. via UBS Year Ahead
The Shiller CAPE ratio sits at 40.5 as of early December 2025, more than double its historical mean of 17.3 and approaching levels last seen, again, during the dot-com peak.
Under these circumstances, I think small-cap and international equities offer more attractive entry points than US large-cap indices. This creates an opportunity to shift part of the US equity holding out of the S&P 500 into a mix of: S&P 500 value index funds (excluding high P/E ratio stocks), mid-cap stocks, international index funds (for geographic diversification), and small-cap stocks (which have more normal valuations and haven’t experienced the same speculative growth). This also aligns with my view that the AI boom might not end with a winner-takes-all situation for the hyperscalers.
• US Dollar Depreciation Expected Despite Continued Dominance: There is growing consensus among analysts for continued dollar weakness, with JP Morgan estimating the currency remains roughly 10% overvalued and Goldman Sachs projecting 4% depreciation over the coming year. But, dollar dominance in global finance will erode only slowly over decades through structural shifts in trade and GDP share, while dollar valuation can decline much faster due to less exceptional US economic performance and difficulty attracting unhedged capital flows. The key driver is the US’s shrinking share of global trade and persistent fiscal deficits, not an imminent collapse of reserve currency status. This aligns with the points we outlined in our previous review of Pozsar’s Bretton Woods III.
This distinction is clearly visible in historical data: according to IMF COFER data, the dollar’s share of global reserves has declined gradually from 71% in Q1 1999 to 56% by Q2 2025, a structural erosion occurring over 25 years. In contrast, the trade-weighted dollar index has experienced far more volatile swings, fluctuating between 95 and 130 over the same period, with particularly sharp movements during crisis periods (2008 financial crisis, 2020 pandemic). The dollar can lose 15-20% of its value in just a few years while maintaining its reserve currency dominance. Recent strength to 130 in 2022-2024 appears unsustainable given widening fiscal deficits and declining US share of global trade, suggesting room for significant near-term depreciation even as the dollar’s reserve status erodes only gradually.
The divergence between structural dominance (slow decline) and cyclical valuation (rapid fluctuations) shows that dollar depreciation can occur independently of reserve currency status changes. Gold and the dollar typically move in opposite directions, and when the dollar weakens, gold becomes more attractive as an alternative store of value, driving its price higher. The outlook for gold in 2026 reflects a convergence of supportive factors beyond simple dollar weakness. Gold has already broken above $4,000/oz for the first time, driven by persistent inflation volatility and increasing demand from both investors and central banks. The structural case strengthens as central banks accelerate reserve diversification, with official sector gold purchases reaching record levels in 2023-2024 as institutions reduce dollar concentration.
I modestly increased my FX-hedged gold position from 4.0% to 5.0%, reflecting upgraded return forecasts but keeping the allocation measured, given gold’s exceptional performance: up roughly 58% year-to-date through December 2025. This increase captures institutional conviction without chasing momentum.
• AI Investment Remains Central But Requires Scrutiny: Almost all investment reports I read over the past weeks position AI as the dominant investment catalyst, with capex projected to reach $571 billion in 2026 (UBS) and potentially $1.3 trillion by 2030. The five largest hyperscalers now account for ~27% of S&P 500 capital expenditure.
AI capital expenditure is projected to reach $1.3 trillion by 2030 (3.8% of US GDP), which would exceed all previous infrastructure booms including broadband (1.15%), electricity (0.98%), and the Apollo program (0.74%). However, as UBS notes,
no investment boom has ever seen capital spending perfectly match future demand.
• European Fiscal Revolution Creates Investment Opportunities: Germany’s historic abandonment of its debt brake policy, committing over €1 trillion to infrastructure, defense, and security spending (with an additional €600 billion in private sector commitments), represents a structural break from decades of fiscal conservatism.
JP Morgan upgrades eurozone growth to 1.5% and Goldman Sachs identifies a structural shift focused on defense independence, energy security, and reindustrialization. This fiscal activism is expected to narrow the US-Europe growth differential from 60bps to 30bps, making European equities, currently trading at a 22% discount to global peers, increasingly attractive despite elevated valuations elsewhere.
• Fixed Income Offers Best Prospects Since Global Financial Crisis: Higher starting yields and steeper curves have dramatically improved bond return potential. As of early December 2025, 10-year US Treasuries yield around 4.2%, with medium-duration quality bonds expected to generate mid-single-digit returns. All major research houses project 2-3 additional Fed rate cuts in 2026, while the ECB is expected to hold steady and the Bank of Japan to continue hiking. As usual it is to be expected that front-end yields are more sensitive to central bank policy and offer strong counter-cyclical properties, while fiscal concerns drive term-risk premia higher at the long end, benefiting strategic curve positioning.
Michael Cembalest, Chairman of Market and Investment Strategy, J.P. Morgan Asset & Wealth Management on the impact of geopolitical events:
It is shocking how little geopolitics actually matters to markets unless it gets truly terrible.
• Other points to consider: (1) Global growth remains resilient, with the US expected around 1.8% and global growth near 2.5%. Consensus points to America’s economic outperformance becoming “less exceptional” relative to other regions. (2) Expect elevated inflation volatility and sticky pricing pressures. Fed easing cycles are underway, but the path remains uncertain with tariffs adding to price pressures. (3) Europe’s fiscal pivot is the big story. Germany’s €1 trillion spending bill marks a historic shift, with broader European infrastructure investment accelerating. Fiscal deficits globally may weigh on currencies. (4) Economic nationalism is reshaping global dynamics. US effective tariff rates have reached levels not seen since 1934, creating a new trade order that markets must price in. (5) China’s Tech sector remains a top global opportunity despite tensions. Stimulus measures are supporting equities, and yuan appreciation is expected as growth stabilizes. (6) Attractive entry point for quality bonds. 10-year UST yields around 4.2% in December 2025 offer compelling returns, with better starting valuations than recent years. (7) Elevated geopolitical risks persist: Russia-Ukraine, Middle East tensions, and broader great power competition remain market-moving factors. (8) Bitcoin with institutional adoption accelerating ETF inflows continue and corporate treasury allocations are expanding. Regulatory clarity improving in the US, though enforcement actions remain a wildcard. Leverage buildup in derivatives markets. Watch for Bitcoin halving aftermath effects and macro liquidity conditions as primary drivers.
In the HBO series Succession, billionaire Logan Roy’s children spent four seasons scheming, backstabbing, and making offers to inherit a media empire. This week, the real version played out with more zeros and a $252 billion Oracle stake. Time for a closer look:
On Friday, Warner Bros. Discovery’s board agreed to sell the company to Netflix for $72 billion. By Monday, Paramount had launched a hostile tender offer directly to shareholders at $30 per share, all cash. In this post I will be going into the gap between those two numbers, streaming economics, aggregator theory, and hostile deal mechanics.
The Netflix offer breaks down into three pieces: $23.25 per share in cash, $4.50 per share in Netflix stock subject to a collar, and shares in a spun-off entity called Discovery Global containing CNN and the cable networks that Netflix doesn’t want. Analysts value that stub somewhere between $2 and $5 per share, which puts the total package at roughly $29.75 to $32.75. Paramount is offering $30 per share in cash for the entire company, including the cable assets. Warner’s stock closed Friday at $26.08 and opened Monday around $27.64, which tells you the market expects a bidding war but isn’t fully convinced either deal closes.
We’re sitting on Wall Street, where cash is still king. We are offering shareholders $17.6 billion more cash than the deal they currently have signed up with Netflix.
David Ellison’s arithmetically correct. Warner’s board took the Netflix deal anyway. This gets at something I learned in a dealmaking class in University: Boards weight speculative ideas of long-term value. They believe the Discovery Global spinoff might be worth $5, that Netflix stock has upside, that strategic fit matters. Shareholders, particularly the arbs and institutional holders who actually vote, prefer certainty. Thirty dollars in cash is just $30 in cash. As Matt Levine noted, “$30 in cash is worth more than, well, again, the stock closed at $26.08 on Friday.” The board’s job is to maximize long-term shareholder value. The shareholders would like their value now, please.
The strategic logic behind Netflix’s offer deserves examination. Both companies began as distributors. The Warner brothers opened a movie theater in Pennsylvania in 1907 before moving into film production; Netflix mailed DVDs before becoming a streaming giant. The crucial difference: physical distribution has capacity constraints, while internet distribution has none. A theater seat that goes unsold is lost revenue forever. The marginal costs of Netflix to deliver to one more subscriber, whether that subscriber is in Zurich or Tokio, are essentially zero. This asymmetry explains why Netflix is worth $425 billion and the combined legacy studios are worth a fraction of that. Consider what Netflix does to content it doesn’t own. Drive to Survive transformed Formula. Apple is now paying $150 million annually for F1 broadcast rights that ESPN once carried for free. NBCUniversal’s Suits sat dormant on Peacock until Netflix licensed it and turned it into a streaming phenomenon. In each case, Netflix created enormous value but captured little of it. The logical next step: own the IP instead of renting it.
This is what Ben Thompson calls aggregation economics. Hollywood executives spent years insisting that content was king, and for decades they were right. When distribution required owning theaters, securing broadcast licenses, or negotiating cable carriage, the studios held leverage. The internet eliminated those bottlenecks. Now the scarce resource isn’t access to content but attention, and the companies that own the customer relationship capture most of the value. Netflix grasped this early; the legacy studios chased streaming without understanding why Netflix was winning. The result: Netflix commands a market cap of $425 billion, while Paramount’s standalone value sits around $15 billion.
Paramount’s financing structure is worth looking at. The tender offer filing is backed by $54 billion in debt commitments from Bank of America, Citigroup, and Apollo, plus a $40.4 billion equity backstop from Larry Ellison’s trust. That trust holds approximately 1.16 billion Oracle shares worth around $252 billion at current prices. Additional equity comes from Saudi Arabia’s Public Investment Fund, Abu Dhabi’s L’imad Holding, Qatar Investment Authority, and Affinity Partners. To avoid CFIUS jurisdiction, the foreign investors have waived all governance and voting rights.
Hostile Tender Mechanics
In a friendly deal, the target’s board negotiates terms and recommends shareholders accept. In a hostile tender, the acquirer goes directly to shareholders with a public offer, bypassing the board. Warner’s board has 10 business days to respond with a recommendation. Defense mechanisms exist (poison pills, enhanced breakup fees) but all invite litigation. The best defense is usually more money from the preferred bidder.
The antitrust arguments on both sides are instructive. Ellison argues that combining Netflix (#1 in streaming) with HBO Max (#3) is anticompetitive: “It’s like saying Coke could buy Pepsi because Budweiser sells a lot of beer.” Netflix counters by pointing to Nielsen’s TV viewing data, which shows Netflix at 8% of total TV usage, slightly below Paramount’s 8.2%. By that measure, Netflix ranks sixth overall, with YouTube at #1 and Disney at #2. The relevant market definition will determine whether this deal survives regulatory review.
If regulators define the market narrowly as “subscription video on demand,” combining Netflix with HBO Max looks troubling. If they define it as “all video consumption,” Netflix is one player among many, competing against YouTube’s bottomless catalog of free content. This framing matters because YouTube already exceeds Netflix in total viewing time. The existential threat facing Hollywood isn’t consolidation among paid streamers. It’s the democratization of content creation itself. Every teenager with a smartphone is a potential competitor for audience attention. The hours flowing to TikTok and YouTube creators don’t flow to HBO. From this vantage point, Netflix absorbing Warner Bros. looks less like monopolization than like circling the wagons. Ellison offered his counter-narrative on Monday: the Netflix deal means “the death of the theatrical movie business in Hollywood.” He promised to put 30 movies a year in theaters exclusively and to combine CBS News with CNN into what he called a news service “in the trust business, the truth business” that “speaks to the 70% of Americans that are in the middle.” Whether you find this vision compelling probably depends on your priors about theatrical distribution and centrist news. The deal timeline matters. Netflix’s offer is expected to take 12-18 months to close, driven by antitrust review. Paramount claims its offer has a faster path to regulatory approval. If Warner’s shareholders ultimately take Paramount’s offer, Warner owes Netflix a $2.8 billion breakup fee. If Netflix’s deal collapses after the review period, Netflix owes Warner $5.8 billion, one of the largest breakup fees on record.
Breakup Fee Economics
Breakup fees serve two functions: compensating the jilted bidder for deal expenses and transaction costs, and creating a hurdle for competing offers. A $5.8 billion reverse breakup fee equals roughly $2 per Warner share, meaning any competing bid needs to clear that hurdle to be economically equivalent. The size of Netflix’s fee signals both confidence and a willingness to pay for optionality.
Warner’s stock trading below both offers reflects the compounded uncertainties: antitrust risk, timeline risk, financing risk, and the possibility that both deals fall apart. The 12-18 month window creates a lot of room for things to change. Interest rates could move. The administration’s antitrust priorities could shift. Netflix’s stock could fall further, reducing the value of the stock component. Paramount’s financing consortium could develop cold feet.
What happens next is procedurally straightforward. Warner’s board will respond to Paramount’s tender offer within 10 business days. Netflix will likely raise its bid; Ellison signaled Monday that $30 “wasn’t best and final.” The arbs will push for whichever deal offers better risk-adjusted value. Whoever wins will spend the next year in antitrust review while the other side’s lawyers look for grounds to challenge.
Hollywood’s century-old industrial structure is unwinding regardless of which bid prevails. The studio system emerged when controlling both production and distribution created durable advantages. The internet dissolved those advantages by making distribution essentially free and universally accessible. Warner Bros. spent a century building an integrated media empire; Netflix spent two decades proving that owning the customer relationship matters more than owning the soundstages. The question isn’t whether legacy media consolidates into tech platforms. It’s which platform, at what price, and whether inherited wealth can rewrite the outcome. I doubt it. On the internet, aggregators tend to win, and Netflix is the aggregator in video.
