// AI Deep Dive
The AIE Index at Two Years — Up 96%, and Everything I Got Wrong
An analyst report on the AIEI — where the AI trade has been, where it's going, and the rebalanced index for the next two years
EXECUTIVE SUMMARY
Back in August 2024, when I built the AIE Index, the market was having one of its periodic AI panic attacks. Everyone was suddenly worried the whole thing was a bubble. My argument at the time was that we'd seen this movie before — Isaac Newton lost his shirt in the South Sea Bubble, the dot-com crash wiped out a generation of paper wealth, and yet the internet still ate the world. I said AI was a strong long-term growth driver and that the infrastructure being built would generate substantial value, the same way all that dark fiber from 1999 eventually carried Netflix.
Classic Amara's Law: we overestimate technology in the short term and underestimate it in the long term. So I picked 22 stocks across six categories — pure AI plays, hardware and infrastructure, tech giants, industrial applications, supporting software, and data center real estate — weighted them roughly equally, and let it ride.
Two years later, here's the scorecard.
// The Number
+96% — AIEI return since inception (Aug 30, 2024), vs. +53% for the large-cap tech benchmark. An equal-weight buy-and-hold of the original 22 returned +85% with dividends, vs. +37% for the S&P 500. Roughly 39% annualized.
The index nearly doubled the benchmark and ran almost triple the S&P. I'd love to tell you that was all skill. It wasn't — and understanding which parts were skill and which were luck is the whole point of this report.
// The Two-Year Scorecard
Fourteen of twenty-two picks made money. Eight lost money or went nowhere. But the distribution is the story: just four names — Intel, Palantir, Vertiv, and AMD — delivered about two-thirds of the entire gain.
The winners:
Stock | 2-Yr Return | What Happened |
|---|---|---|
Intel (INTC) | +398% | Bought at $22, days from a multi-decade low. Lip-Bu Tan took over, 18A shipped, and the foundry turnaround became real. |
Palantir (PLTR) | +303% | Up 140% and 135% in back-to-back periods, then gave back 29% in 2026 as the multiple came back to Earth. |
Vertiv (VRT) | +284% | Power and cooling for data centers. The purest "electricity is the bottleneck" trade, and it compounded every single period. |
AMD | +276% | The MI series finally landed. Up 160% in 2026 alone — outrunning NVIDIA itself. |
TSMC (TSM) | +153% | Everybody's chips come out of the same fabs. |
The losers:
Stock | 2-Yr Return | What Happened |
|---|---|---|
C3.ai (AI) | −62% | Tom Siebel's health forced him out, revenue fell 35%, and the company flirted with a sale. |
Adobe (ADBE) | −61% | Down every period. The market decided generative AI disrupts Adobe rather than feeding it. |
Salesforce (CRM) | −34% | The poster child of the SaaSpocalypse — down 38% in 2026 alone. |
Super Micro (SMCI) | −35% | The short report hit weeks after I bought it. Sometimes the market just punches you in the mouth. |
Microsoft (MSFT) | −6% | The genuine shocker. Negative over two years while funding the biggest capex boom in history. |
// The Callout: The Stack Ate the Software
Average two-year return by layer: semiconductors +142%, infrastructure +138%, physical AI +109%, hyperscalers +35%, software and applications +30% — and negative 25% if you take Palantir out.
In 2024 the reasonable thesis was "AI will supercharge software companies." What actually happened is AI supercharged the companies selling shovels to AI and started eating the software incumbents. Microsoft and Meta shareholders effectively wrote checks that Intel, AMD, TSMC, Vertiv, and Arista shareholders cashed.
// What I Got Right, and What I Got Wrong
Right: The blend. Owning every layer of the stack meant I didn't have to predict which layer would win — and the layer rotation was violent. 2025's leaders (Palantir +135%, Symbotic +151%) became 2026's laggards, while 2026 belongs to semis and power (Intel +198% YTD, AMD +160%, Vertiv +97%). Equal weighting capped every mistake at about a 2.8-point loss while letting winners run uncapped. Intel alone added 18 points.
Wrong: The application layer. Five of my six software-and-applications picks lost money. I said at the time I couldn't personally vouch for the pure plays — I just knew what they were building. Turns out that hedge was doing a lot of work. C3.ai and UiPath weren't pioneering digital transformation; they were standing between the frontier labs and the enterprise customer, which in 2026 is the single worst place on the org chart to be standing.
Luck, acknowledged: August 30, 2024 was a historically great day to buy Intel and a historically terrible day to buy Super Micro and Adobe. Net-net, the timing luck broke my way. I grew up on a farm — sometimes it just rains when you happen to have planted.
// Where the Market Is Right Now
Let me double click on the current setup, because the chart tells you more than the headline number.
The index took two gut-check drawdowns — the April 2025 tariff shock that took everything briefly back to the flatline, and an April 2026 air pocket — and made new highs after both. That's what a secular trend punctuated by cyclical panic looks like. It's the 90s internet chart with better unit economics.
Four things define mid-2026:
The trade rotated from GPUs to everything around GPUs. Custom silicon (Broadcom's AI revenue +143%), high-bandwidth memory (Micron guiding to ~100% growth), networking, and neoclouds (CoreWeave's backlog: $99 billion). NVIDIA is still the king, but it was only my tenth-best pick. The monopoly profits are diffusing through the supply chain.
The market now separates capex takers from capex spenders. Microsoft and Oracle fell 20–28% in 2026 not because AI demand slowed but because investors want to see returns on the hundreds of billions being spent. The companies receiving those dollars are the ones being rewarded.
The SaaSpocalypse is real. Adobe, Salesforce, and UiPath are being repriced as AI's victims, not its beneficiaries. Agentic AI — and yes, I use that word a lot, because it's the right one — turns out to be the competition for seat-based software, not the upsell.
The froth is showing. SK Hynix just pulled off the largest US listing by a foreign company ever ($26.5 billion, popped 13% on day one — two days ago). SpaceX priced at $135, spiked to $225, and gave a third of it back. Anthropic is marked at $965 billion privately. None of this means the top is in. All of it means the easy money has been made.
// AIEI 2.0: The Rebalance
I'm making the biggest changes since inception: eight names out, eight names in, and moving from equal weight to three tiers. The two-year data earned these decisions — I'm cutting the disrupted, not the merely disappointing.
Out (8): C3.ai (−62%), Adobe (−61%), Super Micro (−35%), Salesforce (−34%), Synopsys (−14% — losing money in the greatest semi bull market in history is disqualifying), UiPath (−9%), Tesla (+90%, but nearly all of it in one post-election spike; ~59 robotaxis on the road doesn't cover a $1.5T valuation), and Digital Realty (+25% — fine, but it's a bond in a rocket portfolio).
In (8): Broadcom, Micron, SK Hynix, CoreWeave, ServiceNow, Dell, Astera Labs, GE Vernova.
Tier | Weight | Holdings |
|---|---|---|
Anchors (8) | 5.5% each | NVDA, AVGO, TSM, GOOGL, AMZN, META, AMD, VRT |
Core (8) | 4.5% each | MSFT, MU, ANET, ORCL, PLTR, NOW, IBM, DELL |
Satellites (6) | ~3.3% each | INTC, SKHY, CRWV, ALAB, GEV, SYM |
The logic, layer by layer:
Memory is the new bottleneck. SK Hynix owns 56% of HBM and roughly 70% of NVIDIA's HBM4 orders for the Rubin platform — and as of Friday you can finally buy it on Nasdaq (SKHY). Paired with Micron, I own two-thirds of the HBM oligopoly. On Samsung: I looked hard at it — it leads HBM4 validation and it's the value entry. But there's still no US listing, only a thin OTC pink sheet, and Micron has actually overtaken it in HBM share. Two memory names is a thesis; three is a leveraged bet on a famously cyclical industry. If Samsung lists an ADR the way SK Hynix just did, I'll revisit.
Power is the layer I didn't own. Vertiv's +284% proved electricity is the constraint, yet I owned zero generation. GE Vernova fixes that. Watch this space — the grid, not the GPU, is the binding constraint through 2028.
Connectivity is where the next NVIDIA-sized margin pools hide. Astera Labs is the pure play on the guts of the AI rack, and a named beneficiary of Anthropic's buildout.
Intel drops to a satellite. Up 398%, now a $550 billion company. I'm not a genius for buying at $22 and I won't pretend the next $88 is as easy as the last. Take the win, keep a position.
Broadcom was my biggest omission, full stop. Custom silicon for Google, Meta, Anthropic, and OpenAI — the pick-and-shovel play on every private lab. It should have been in the index a year ago. It's an anchor now.
// The Rule: Anthropic and OpenAI
Anthropic filed its S-1 on June 1 and is targeting an October Nasdaq listing near its $965B mark. OpenAI is leaning toward 2027. An index that claims to be the best blended AI portfolio cannot exclude the model layer once it's investable — these two companies are the sun this entire solar system orbits.
But I lived through enough IPO pops in 1999 to know better than to buy the first print. So the rule, committed in advance: initiate each lab at a capped 3% starter weight after 90 trading days or the first public earnings report, whichever comes first. Scale to full weight after lock-up expiry. For an October Anthropic listing, that means starting around January 2027 and sizing up around April. Funded by trimming the hyperscalers — Google and Amazon already own big Anthropic stakes and Microsoft owns ~27% of OpenAI, so holding all of them at full weight after the IPOs is double-counting the same exposure.
// Where It Might Move
Three scenarios I'm watching, in order of importance:
1. The Anthropic IPO is the event of the fall. If a $1.1 trillion listing holds its price, it validates the entire private AI complex and probably kicks off the next leg. If it does the SpaceX pop-and-fade, it marks the near-term top for AI sentiment. Either way, it's the single best price signal we'll have gotten in three years of this boom.
2. The 2027 memory glut question. All three HBM makers are adding capacity at machine speed. Memory has been a boom-bust business since before I had gray hair. If Rubin-generation demand absorbs the supply, SKHY and MU keep running; if not, 2027 looks like every other memory cycle top. I'm sized for the thesis, capped for the cycle.
3. Enterprise ROI has to show up in the numbers. The SaaSpocalypse repriced the losers, but the capex spenders' stocks are now demanding proof that agents generate revenue, not just demos. If enterprise AI spend shows up in 2027 earnings, the hyperscalers re-rate higher and this keeps going. If it doesn't, capex estimates get cut and the whole supply chain — my whole index — corrects together. That's the honest risk of a thematic portfolio: it's diversified across layers of exactly one idea.
The tail risks haven't changed: Taiwan, rates, and the increasingly circular financing where the labs' biggest costs are their investors' biggest revenues. I said in 2024 that the infrastructure would create value even if individual companies flamed out, and two years and 96 points later I believe it more, not less. We're still early. We're just not cheap anymore.
The AIE Index is an educational exercise, not an investment vehicle of any kind. This doesn't constitute financial advice — I'm a publisher with a spreadsheet, not your fiduciary. Do your own research, and remember: if your AI is hallucinating, that's a problem. If your portfolio is, that's a margin call.
