Citrini’s 26 Trades for 2026

ELI5/TLDR

Citrini’s single highest-conviction trade for 2026 is that companies still paying Wharton grads $150,000 a year to realign logos on PowerPoint slides at 2am are about to figure out AI can do it for a fraction of a penny. The trade isn’t Nvidia — it’s the “AI losers”: bloated, employee-heavy firms like Accenture, Omnicom, SAP and insurance brokers that the market has given up on, which can quietly cut huge chunks of headcount and re-rate upward. The rest of the piece is a sprawling hunt for the un-obvious second-order beneficiaries: Japanese chemical companies with 90% market share in obscure AI-server films, gas turbine makers too traumatized to build new factories, and a Michael Sailor preferred stock designed to bleed.

The Full Story

The “bullshit jobs” trade: AI’s real ROI is firing people

Citrini puts this first in the piece for a reason. The usual AI trade is buying the capex — Nvidia, memory, power, data centers. That’s been great and probably continues. But trillions of dollars have been sucked into that side, and the second-order trade is just sitting there: companies that use AI to gut their own org charts.

The framing is important. The bear case on AI is that adoption is slow. Citrini flips it:

The tech adoption is fast, but the organizational adoption is glacial.

AI in 2025 got good enough to replace the bottom 20% of most white-collar organizations — the junior analysts, the logo-aligners, the insurance paper-pushers. The block isn’t capability; it’s that Fortune 500 bureaucracies don’t know how to rewire themselves. The moment one of them does it and reports better margins, the rest will panic-follow. He cites David Graeber’s Bullshit Jobs directly: Keynes predicted 15-hour work weeks by 2000 because of technology gains, and technology did deliver — we just invented new fake work to fill the hours.

The screen: take the S&P 500, find the bottom decile by net income per employee, then filter for high SG&A as a % of sales, then filter qualitatively for companies already talking about AI headcount reduction. That gets you roughly 30 names across consulting (Accenture, Booz Allen, Capgemini), ad agencies (Omnicom, WPP), insurance brokers, DocuSign, and enablers like SAP and IBM. The whole universe trades like “AI losers” — meaning the downside is already priced in and any margin improvement is pure upside.

He’s clear about the social dimension and doesn’t hide from it:

We spend a trillion dollars on a technology and the sole purpose of that technology is to replace people. As it gets better… we could see an economy in 2026 where the unemployment rate continues going up, but stocks also continue going up. That’s happened before plenty of times in history.

Advanced packaging: duct tape is the new silicon

For about 50 years, chips got better by making transistors smaller. That hit a wall — you physically can’t print a chip bigger than the lithography machine’s “reticle limit” without yields collapsing. So the new game is making a bunch of small chips (adorably called chiplets) and stitching them together so they behave like one big chip. That stitching is advanced packaging, and TSMC’s CoWoS process is the current global bottleneck for every AI chip on earth.

Citrini’s insight: stop trying to guess whether Nvidia, Google’s TPUs, Meta’s custom silicon, or Chinese ASICs win the chip war. They all need advanced packaging. Buy the duct tape.

The core longs: Intel (yes, really — their Foveros 3D packaging tech is becoming a relief valve for TSMC’s backlog, and Apple plus hyperscalers are reportedly sniffing around just for this layer), Amkor (takes TSMC’s overflow, trades at a boring valuation), Synopsys (owns the EDA software you literally cannot design a modern chip without — Cadence trades at 45x earnings, Synopsys at 30x, after a one-time contract dispute with Intel blew it up 40%), BESI, and Kulicke & Soffa (makes the specialized tools that physically bond chiplets together — ticker KLIC, pronounced “click,” and Citrini explicitly calls it “betting on the staplers”).

AI materials: the boring Japanese chemical companies nobody’s priced

This is the nerdiest section and arguably the best. The GPU is the brain. A ton of other stuff has to go into the skull and spine around it — resins, glass fibers, films, substrates. Most of it is made by obscure Japanese or Taiwanese chemical companies trading like chemical companies, not like AI plays.

Two proof-of-concepts worth internalizing:

Nitto Boseki (Nitobo) makes “T-glass” — a special glass fiber that doesn’t expand when heated. 70% global market share. You cannot build a high-end AI server board without it. The stock has gone up every single trading day for eight months, and the first headline about a T-glass shortage was eight months ago. By the time you read about the shortage, the trade has already happened if you weren’t on the watchlist.

Ajinomoto — yes, the MSG company — has 90% global market share on “ABF” buildup film, the insulation layer in every high-end chip. It trades like it makes MSG. It makes the glue holding AI together.

The top conviction pick here is Resonac, which has “the most shots on goal” — it doesn’t dominate any single shortage yet, but it has meaningful market share across many potential bottlenecks: BT resin, non-conductive film (used to glue HBM chips together — it has to conduct heat but not electricity), and more.

The operational approach Citrini recommends is clever: don’t try to be a semiconductor engineer. Keep a watchlist of 20 obscure materials companies cross-referenced to the bottleneck each one supplies. Set a news alert for “Taiwan shortage” or “Japan shortage.” When one hits, Ctrl-F the watchlist, find the 90%-market-share company, and buy. Trades that can run for a year.

Post-Traumatic Supply Disorder

The best-named thesis in the piece. Look at the best-performing stocks of 2025 — gas turbine makers (GE Vernova, Siemens Energy) and memory (Micron, SK Hynix, Kioxia, Sandisk, Western Digital). Everyone explains this with “AI exposure.” That’s incomplete. Lots of companies have AI exposure and got crushed.

The common thread: every one of these industries got burned during the past five to ten years. They built factories into a “secular” demand forecast, demand fell off a cliff, shareholders crucified them. Now, even with demand genuinely ramping, they refuse to build new capacity. They’re hoarding backlogs and raising their long-term margin targets instead.

They’re treating debt like an STD they got in college.

When a traumatized oligopoly meets real demand, you get pricing power and multi-year earnings surprises. The screen: (1) trauma exists, (2) demand is inflecting, (3) management shows capital discipline, (4) the industry is concentrated enough that one cowboy competitor can’t flood the market. Gold miners fail #4. Solar, analog semis, offshore drillers, lithium, wind turbines, and probe-card makers (Technoprobe) all qualify. Special mention for First Solar — Chinese solar is getting locked out for national security reasons, which de facto makes US solar an oligopoly for the first time.

Natural gas as the AI power trade

Nuclear is great. Solar is great. But you can’t build machine god on a nuclear plant that takes 10 years to permit, and most of the data centers actually being built right now run on natural gas. The thesis isn’t “natural gas is going up because winters are cold” — it’s that the market is priced as if the Permian can infinity-ramp supply forever, and that’s a complacent assumption given that US LNG export capacity is about to come online and start competing with hyperscalers for the same molecules.

The catalyst to watch: fixed-price long-term contracts between hyperscalers and gas producers, analogous to what Vistra and Constellation did for power. Once EQT or Comstock signs one, growth investors can start pricing the equity on hyperscaler demand rather than front-month weather, and the whole sector re-rates. EQT has hinted at this but hasn’t pulled the trigger yet.

The Michael Sailor short — or, “strictly for nerds”

Citrini is long Bitcoin but short STRD, one of MicroStrategy’s preferred securities. The construction is genuinely cursed:

• It’s non-cumulative, meaning if Sailor skips a dividend there’s no penalty and no back-pay.
• It gives you capped upside on an asset (Bitcoin) whose entire value proposition is uncapped upside.
• If Bitcoin rips, best case you earn ~10%. If Bitcoin tanks, they can legally just stop paying you and the security tanks 60% with no “pull to par” to rescue you.
• It still trades near par. It shouldn’t.

This is a watchlist trade, not a pound-the-table trade — you can bleed on the coupon for a long time in a flat market — but the asymmetry is real.

Inference moves to the edge

The bottleneck for running AI locally on your phone isn’t the chip, it’s the RAM. Current iPhones would need twice as much, and RAM has gotten prohibitively expensive. But there are five separate algorithmic and hardware paths to improving memory efficiency. Break one, and inference moves on-device because the latency math is too good to ignore (800ms round-trip to a Texas datacenter vs. 200ms locally — irrelevant for chatbots, fatal for an agent booking you an Uber mid-conversation).

The elegant trade expression: long MediaTek (20x earnings, designing next-gen TPUs, big upside to on-device inference) and Qualcomm; short Lenovo, Dell, and similar hardware makers whose bills of materials are getting hammered by RAM costs. The Nintendo Switch is singled out as especially exposed — 41% of its BOM is RAM, which is why the price already jumped from $300 to $450.

Quick hits

• World Cup trade: US budget hotels like Choice Hotels (CHH) have been destroyed. When a billion people descend on North America and every luxury room is booked, beggars can’t be choosers. Base effects will be enormous.
• Q1 tax refund trade: refunds will be 30-50% larger than last year. That money flows to deferred consumer durables — mattresses, appliances — not houses or cars. Added optionality if Trump pushes a “tariff rebate” before midterms.
• Robotics in weird places: Sweetgreen, Cava, Chipotle. Assembly-line food, everything in the same location every time, trivially replaceable by a Fanuc arm. Sweetgreen already sold its robotics division on a cost-plus agreement.
• Don’t fight the CCP: China is going to spend whatever it takes to build domestic AI silicon. ACM Research (ACMR) is a US-listed ADR whose six-share equivalent trades in China at 5x the US price. You can’t arbitrage it directly, but the valuation gap eventually closes.

The meta-lesson

Every trade here starts from the same move: find something the market is pricing as an “AI loser” or a boring commodity, then ask what has to be true for it to become an AI winner. Citrini calls this thematic factor investing — the idea that “AI” as a factor now pulls in companies that were never in your growth screen, like the day Siemens Energy suddenly started trading with a 1:1 beta to Nvidia because a utility is now an AI stock.

It’s the one time a year where we’re not just neurotic about being right.

Claude’s Take

Citrini is on the strongest ground when he’s doing supply-chain detective work — the Nitto Boseki / Ajinomoto / Resonac stuff is genuinely differentiated research, and the “90% market share in an obscure input to every AI server” pattern is a real investable edge that most generalists will never find. The advanced packaging framing as “bet on the duct tape, not the chip war winner” is a clean insight that sidesteps a debate nobody actually knows the answer to.

The “bullshit jobs” trade is provocative and directionally right, but it’s squishier than he acknowledges. The empirical case that AI is already replacing labor at scale is thinner than the vibes suggest — 2025 saw lots of corporate AI announcements but very few clean examples of a public company posting quarter-on-quarter margin expansion specifically attributable to AI headcount reduction. The thesis requires a belief that 2026 is the year qualitative AI adoption finally shows up in hard numbers, which is plausible but not proven, and he admits the 12-month shot clock himself near the end. If nothing concrete materializes by early 2027, the bears’ “parlor trick” framing gets a lot more oxygen.

On the depreciation debate (Chanos/Burry vs. bulls), his rebuttal is clever — “capex is replacing future operating expenses” — but it’s weaker than he presents. That argument works for the end customer (Meta saving on labor) but doesn’t actually address whether hyperscalers are accounting honestly for chip useful life. He more or less concedes this by saying “I’m not as good at accounting as Jim Chanos,” which is fair but also the whole ballgame. The cleaner point, which he stumbles into, is that accelerated depreciation would be bearish for OpenAI/Microsoft margins and bullish for Nvidia — those are the same trade.

PTSD is a great name for a real pattern but it’s less novel than framed. “Capital-disciplined oligopolies in a demand upcycle” is the oldest cyclicals playbook in the book; he’s re-marketing Jeremy Grantham circa 2010. Still useful as a screen.

One blind spot worth flagging: Citrini runs a newsletter that sells conviction. The 132-basket format is impressive but also guarantees that something outperforms every year and lets the narrative follow the winners. The Citrindex is auditable and the returns are genuinely good, which is a much higher bar than most newsletter writers clear — credit where due. But “we publish 26 trades and half of them outperform” is mathematically not far from chance when your universe is thematic baskets in a bull market.

The STRD short is the one trade you should probably not do unless you really understand preferred stock mechanics. He’s right that it’s mispriced, but “bleeds slowly in a flat market” is a description of almost every professional short that blew up 2020-2023.

Word count

~2,100