InnerPeace’s End of May 2026 Portfolio Update
AI Disclosure: This post was substantially generated with the assistance of LLMs, working from my research direction, my investment framework, and my personal transaction records. The analysis reflects my own views and decisions. I build AI systems for a living, and I hold these AI-assisted analyses to the same standard I’d apply to my own manual research. I hope the quality justifies your reading time, despite the board’s policy on AI-generated responses.
Portfolio Overview
2026 YTD Return: +72.3%
Drawdown from Nov 2025 Peak: Recoverd
Month-over-Month: Jan -5.9% → Feb -26.5% → Mar -10.2% → Apr +75.1% → May +54.9%
| Ticker | May 31 Weight | Q1 Weight | Change |
|---|---|---|---|
| IREN | 21.9% | 21.05% | +0.85pp |
| ALAB | 20.9% | 17.26% | +3.6pp |
| APP | 19.4% | 16.67% | +2.7pp |
| NBIS | 15.2% | 17.93% | -2.7pp |
| NVDA | 9.6% | 13.42% | -3.8pp |
| CRDO | 7.6% | — | NEW |
| NET | 3.7% | 6.52% | -2.8pp |
| HIVE | 1.7% | 0.96% | +0.7pp |
| RBRK | — | 6.20% | SOLD |
Transactions since Q1 (March 31):
- Sold RBRK, bought CRDO — The launch of Anthropic’s Mythos made me reassess small cloud security firms. I had thought cybersecurity SaaS was immune to LLM disruption. It isn’t. I exited RBRK and initiated CRDO as a new position in high-speed connectivity, which sits squarely in the AI infrastructure thesis.
- Added NVDA — Bought back some of what I trimmed in Q1. The valuation reset made it compelling.
- Added APP (twice) — Q1 earnings showed 59% YoY revenue growth re-accelerating from FY2025’s 16%. The Axon AI advertising platform and social media app launch are underappreciated catalysts.
- Trimmed NBIS, added IREN — NBIS had grown past 20% of the portfolio. IREN was lower. I cut NBIS and added to IREN to rebalance. IREN’s 8.8x upside to the $200B target (vs 3.4x for NBIS at current prices) makes it the better risk/reward at these weights.
- Trimmed ALAB — ALAB also crossed 20%. I trimmed back to my comfort zone and redirected some into APP.
The portfolio has recovered substantially from the Q1 lows. The AI infrastructure thesis — exponential token burn driving infrastructure demand — continues to play out. RBRK is out, CRDO is in, and the concentration in AI infrastructure has deepened.
Company Updates
IREN (21.9%) — The Asymmetric Bet
IREN remains my largest position for one reason: the math. My base-case model assumes 5 GW of active power by ~2030, $2-2.5B quarterly revenue per GW, 20% net margins, and a 25x P/E. Under those assumptions, IREN could be a $200-250B company. At current market cap, that’s 8.8x upside. Even if I’m wrong by 50% on every assumption, the return is still multiples of the current price. This is a model, not a forecast — the range of outcomes is wide.
The key driver is power. Here’s where the Big Three neoclouds stand:
| Active Now | YE 2026 Target | Contracted | Secured Pipeline | |
|---|---|---|---|---|
| CoreWeave | >1 GW | — | >3.5 GW | — |
| NBIS | ~170 MW | 800 MW–1 GW | >4 GW target | — |
| IREN | 810 MW | +480 MW (Childress Phases 1–4) | 810 MW fully contracted | 4.5 GW |
IREN has the deepest secured pipeline and needs only ~10% of it to hit its ARR target. But the pace is deliberately slower than NBIS or CoreWeave. This is both a strength — less execution risk concentrated in any single project — and a weakness. NBIS is signing the Meta and Microsoft deals now. CoreWeave already has >1 GW active. If the hyperscaler capacity shortage eases, the first movers will have locked in the best contracts.
The Q3 FY2026 results showed a 94.2% increase in AI cloud revenue. The transition from Bitcoin mining to AI infrastructure is showing in the numbers. But the legacy Bitcoin business creates noise in the financials — revenue from mining can decline even as AI cloud grows. I don’t care about the Bitcoin revenue. I care about the AI infrastructure buildout. The market sometimes confuses the two, which creates opportunities.
The board discussion has focused extensively on comparing IREN to NBIS and CoreWeave. The consensus is that IREN is the tortoise, not the hare, but the secured power pipeline is the deepest moat. Smorgasbord1 raised the right question: “We keep getting talk of ‘watts,’ but that’s like saying ‘miles’ without saying whether those miles are planned, mapped, or actually driven.” For IREN, the 4.5 GW pipeline needs to convert to active power. The 810 MW operational today is the baseline. The 480 MW expansion is the near-term test.
My conviction in IREN is not that it will execute perfectly. It’s that the power assets are real, the demand for AI compute is real, and the current valuation prices in far less success than NBIS or CoreWeave. The asymmetry is what makes it the largest position.
ALAB (20.9%) — The Standard-Bearer
Astera Labs makes PCIe retimers (Aries product family) and CXL fabric switches (Scorpio platform) — industry-standard connectivity chips that sit between the GPU and everything else in an AI server. The numbers: 76.3% gross margins, 93.4% YoY revenue growth in Q1 2026, $10B TAM target for the Scorpio X-Series alone. Non-GAAP operating margin: 36.2%. The business is real, profitable, and growing fast.
In my hardware cycle framework, ALAB sits between NVDA and CRDO on the safety spectrum. PCIe is a 20-year industry standard. Every GPU needs PCIe connectivity. You can’t build an AI server without it. The switching costs are embedded in system architecture — you don’t rip out PCIe for a cheaper alternative the way you might delay a KLIC bonding tool purchase. This is not a commodity connectivity play; it’s a standards-based moat with a 5-10 year durability horizon.
The product diversification within the connectivity niche helps too. Aries covers PCIe retimers across multiple generations. Scorpio covers CXL fabric switches from 32 to 320 lane configurations. If one product line faces pricing pressure, the portfolio provides some insulation. That’s a meaningful advantage over CRDO’s more concentrated AEC cable business.
Compared to CRDO, ALAB has lower customer concentration (not flagged as extreme in any of the research), a more diversified product line built on industry standards rather than a single new technology, and gross margins that have been stable to improving. In a hardware downturn, ALAB’s revenue would decline — it’s not immune to the cycle — but the essential nature of PCIe connectivity means it declines less than discretionary equipment purchases. The PCIe standard has survived decades of technology transitions. That’s the kind of moat I want in a 20%+ position.
APP (19.4%) — The Platform Play
AppLovin delivered 59% YoY revenue growth in Q1 2026, re-accelerating from 16% in FY2025. The annual growth decelerated from 43% in FY2024 to 16% in FY2025 — a 27 percentage point drop — but Q1 2026’s 59% suggests the business has re-accelerated. The FY2025 annual figure may have been depressed by comparisons or one-time factors. Guidance for Q2 2026 implies continued strength, and management raised FY26 revenue guidance.
The business fundamentals are exceptional. 88.9% gross margins. 64.3% net profit margins. TTM free cash flow of $3.18B (Q1 2026 quarterly FCF was $1.3B). SBC is low. The Axon AI advertising platform — alongside MAX in-app bidding — is the growth engine. Revenue exceeded analyst estimates by 3.8% in the most recent quarter, continuing a consistent beat pattern that suggests management guides conservatively. Earnings growth of 109.2% YoY reinforces the margin story.
The board has been intensely focused on two catalysts: the social media app launch and the automated video creation tool. Whether the social media app itself succeeds is secondary. The real story is Axon’s AI advertising platform becoming the operating system for mobile ad spend. If Axon captures even a fraction of the $300B+ global digital advertising market, the growth runway is long. The CloudX/Axon ecosystem — automated video creation, AI-driven ad optimization — is where the competitive moat gets built.
The risks are real. An active SEC probe into data-collection practices creates uncertainty. The P/S ratio is elevated. The stock is volatile. But at 19.4% of the portfolio, I’m comfortable with the risk/reward given the margin structure and growth trajectory.
NBIS (15.2%) — Best Execution, Smaller Upside
Nebius has delivered the strongest execution in the neocloud space. The Meta deal over 5 years provides a contracted revenue backbone that neither IREN nor CoreWeave can match. The Microsoft deal is on track for 2026 delivery. Revenue grew 621.5% YoY in Q1 2026. Gross margins are 74%.
The power trajectory tells the story: 170 MW active today → 800 MW–1 GW connected by end 2026 → >4 GW contracted. That’s a 5-6x scaling in 12 months. If they hit it, it’s extraordinary execution. If they don’t, the stock reprices violently. NBIS is coming from a unique origin — restructured from Yandex’s international assets after the Russia exit. 2023 revenue was near zero. The growth from that base has been explosive, but it also means the track record is short. The bull case requires flawless execution through an unprecedented scaling challenge.
At current market cap, the upside to $200B is 3.4x — excellent, but less than half of IREN’s 8.8x. My trim in May was not a statement about NBIS’s quality. It was purely a risk/reward rebalancing: IREN offers more upside per dollar of current valuation, driven by the deeper secured power pipeline. The board’s comparison thread from February reinforced this — NBIS has more signed contracts, but IREN has more secured power, and power is the ultimate bottleneck in this space.
The risks: customer concentration (Meta and Microsoft dominate revenue), the scaling challenge, and deep negative FCF during buildout. But these are the right risks to take — they’re execution risks, not thesis risks. If NBIS executes, the returns are exceptional. If it doesn’t, the position size (15.2%) limits the damage.
NVDA (9.6%) — The Moat Is Real, But Watch the Edges
NVDA controls ~81% of the AI data center chip market. Its moat is three-layered: best-in-class hardware (H200, B200, Vera Rubin), the CUDA software ecosystem (decades of developer lock-in), and financial innovation — equity stakes in customers like CoreWeave that align incentives beyond simple chip sales.
The Q1 FY2027 numbers were extraordinary. Revenue grew 85.2% YoY to $81.61B. Net income was up 210.6% YoY (calculated from quarterly net income — the yfinance “earnings growth” field is unreliable). Gross margin: 74.9%. Operating margin: 65.6%. Net cash: over $40B. Forward P/E and P/S are current market data — check at time of reading. Beta: 2.24.
But the moat is asymmetric. It’s strongest in training — where CUDA optimization, CoWoS packaging access, and the software ecosystem create genuine lock-in. It’s weakest in inference — where workloads are simpler, AMD’s ROCm has “practically closed the software gap,” and open-source stacks (JAX, PyTorch direct) reduce CUDA dependency. Inference has already surpassed training as the dominant AI workload. Every ChatGPT query, every Claude response, every Gemini search — these are inference calls. As the workload mix shifts toward inference, NVDA’s moat erodes at the margin.
The other threat is Google’s Trillium TPU, currently exclusive to GCP. TPU doesn’t have CUDA’s ecosystem, but if Google ever makes TPU available outside GCP, the competitive dynamic shifts meaningfully.
I trimmed NVDA in Q1 and added back in April. At ~16x forward P/E on 85% revenue growth, the valuation is reasonable. I’m watching gross margins for any sign of inference commoditization. If margins dip below 65% without a clear temporary explanation, that’s a sell signal. The 81% market share means NVDA has only one direction to go in market share — the question is how fast, not whether.
CRDO (7.6%) — High Risk, Tactical Size
Credo makes Active Electrical Cables (AECs) — high-speed copper interconnects for AI clusters. As GPU clusters scale to 100,000+ units, the interconnect market explodes. AECs are replacing active optical cables because they’re lower power, lower cost, and more reliable. The Q3 FY2026 numbers were strong: revenue grew 201.5% YoY, gross margins expanded to 68.5%. Beta: 1.6. The DustPhotonics acquisition adds optical capabilities to the portfolio.
Guidance tells an interesting story. Q4 FY2026 guidance of $425M-$435M implies only ~5% sequential growth — a dramatic deceleration from Q3’s 51.9% sequential jump. Management guided FY2027 to over 50% revenue growth, which is strong, but the near-term deceleration raises the question: temporary digestion or genuine slowdown? The board has been closely analyzing whether this is customer absorption after a massive buildout quarter, or an early warning.
The bear case is severe: 88% of revenue comes from three customers (per the company’s 10-K filing). If one hyperscaler develops an internal AEC solution — and Google, Amazon, and Microsoft are all building custom silicon — CRDO could lose a major revenue source overnight. The other risk is standard risk. AEC is a relatively new category. If the industry converges on a different interconnect standard, CRDO’s product becomes obsolete. Cables in data centers get replaced every 2-4 years as speeds upgrade — there’s no long-term lock-in.
At 7.6%, the position size correctly reflects the asymmetry. Big upside if AEC becomes the standard for AI cluster interconnects. Existential risk if it doesn’t. I bought CRDO after exiting RBRK — the AI infrastructure thesis is where I want to be concentrated, and CRDO is a pure play on that theme, just with higher risk than the other names. This is not a “buy and forget” position. It requires active monitoring of customer concentration and standard evolution.
Special Topics: A Research Framework for AI Infrastructure Investing
Over the past month, I’ve developed a structured research framework to test the assumptions underlying my portfolio. Here are the five key findings.
1. From Cyclical to Growth? A Hardware Cycle Framework
Traditional semiconductor investing follows a bullwhip pattern: under-supply → surging profits → massive CapEx → over-supply → earnings crash. This cycle has repeated for 40 years. Every memory upturn has ended in a downturn. Every equipment boom has ended in a bust.
The AI era may be different — or it may just be a longer cycle. Long-term HBM contracts (Micron’s 2026 capacity is already sold out with multi-year agreements), TSMC’s persistent CoWoS packaging bottleneck (sold out through 2026), and NVDA’s CUDA ecosystem are stretching the up-cycle in unprecedented ways. But different companies benefit differently:
←─── Secular Growth ─────────────────────── Cyclical ───→
NVDA ALAB CRDO MU KLIC
GPU+CUDA PCIe+CXL AEC cables HBM memory bonding tools
10yr+ 5-10yr 2-5yr TBD <2yr
The key determinant is how close the product is to the end consumer of compute, and how deep the switching-cost moat is. Ecosystem lock-in (CUDA) > industry standards (PCIe) > technology lead (AEC) > supply bottleneck (HBM undersupply) > CapEx cycle position (equipment).
The framework validates my portfolio construction. The heaviest weights are on the growth side of the spectrum. The tactical positions are on the moat-dependent side. And I’ve avoided the purely cyclical names despite their apparent cheapness — a low P/E on cyclical earnings is not a bargain, it’s a trap.
The key question for every position: when the supply constraints ease and GPUs are no longer scarce, who still has pricing power?
2. The Inference Shift and Open Source Disruption
Inference has already surpassed training as the dominant AI workload. This matters because inference economics are fundamentally different from training economics. Training requires cutting-edge GPUs with CUDA optimization. Inference can run on older GPUs, consumer cards, or even non-NVDA hardware. AMD’s ROCm has “practically closed the software gap with CUDA” for inference workloads. The moat that protects NVDA in training doesn’t fully extend to inference.
The open source earthquake — DeepSeek, Llama, Mistral — compounds this. When anyone can run a competitive model, proprietary API pricing power erodes. Token prices fall. But Jevons’ Paradox kicks in: “We made every token a hundred times cheaper, and then we built products that consume ten thousand times more tokens.” Total inference demand explodes, not contracts.
This is the core bull case for neoclouds. The infrastructure layer — GPUs, data centers, power — captures the volume growth regardless of which model wins or what the token price is. When the LLM companies compete themselves to death on pricing, the picks-and-shovels providers still get paid per GPU-hour. If total GPU-hours consumed goes up 10,000x while per-hour pricing stays stable, neocloud revenue grows massively.
The risk scenario: if open source models become so good that proprietary AI labs face severe financial pressure, near-term GPU procurement could slow. This wouldn’t mean bankruptcy, but belt-tightening at OpenAI or Anthropic would create a demand air pocket. Not the base case, but worth monitoring.
The board discussed this extensively during the January 2025 DeepSeek panic. WillO2028 posted the Jevons’ Paradox argument within hours, and it proved prescient — GPU demand did not collapse. The same logic holds today.
3. Macro Risk and the High-Beta Reality
Saul-style portfolios fall 50%+ roughly every 2-3 years, while the SOX index falls that much every 5-10 years. The same characteristics that drive outsized returns — hypergrowth, high multiples, concentrated positions, emerging tech exposure — amplify drawdowns. It’s not a bug; it’s the mathematical cost of the strategy.
We’ve had five macro shocks in more than twelve months: DeepSeek R1 panic (Jan 2025, CRDO -25%), Trump Liberation Day tariff (Apr 2025), AI Capex “Show Me the ROI” rotation (Dec-Mar), Iran War with Strait of Hormuz closure (Mar 2026, oil past $120), and Fed markets betting on rate hikes (May 2026). Each one recovered. But collectively they demonstrate the environment. A 50% drawdown over the next 12 months is historically normal.
What’s different from 2022: the companies are profitable. NVDA has $40B in net cash. ALAB has 75% gross margins. But many 2021 SaaS companies also had 70%+ margins and were profitable before the crash. The difference is not margins — it’s that today’s AI infrastructure demand is backed by hyperscaler CapEx measured in hundreds of billions, not speculative software TAMs. A -90% drawdown is less likely. A -50% is not.
The practical implication: size positions for the drawdown I can hold through. Keep margin low. Have conviction in the businesses, not the stock prices. As DoctorRob posted during the February selloff: “As Saulites, we know our companies. When the market goes down, instead of a dwindling fortune we see a fire sale.”
4. Our Deliberate Deviation from Saul
My two largest positions — IREN and NBIS — were bought before revenue hypergrowth was evident. IREN was primarily a Bitcoin miner when I built the position. NBIS was a restructured shell of Yandex with near-zero revenue. This is a deliberate deviation from Saul’s method, which buys after 30%+ revenue growth is demonstrated.
What we’re doing sits in a gray zone between VC and Saul-stage investing: pre-inflection, thesis-driven, higher risk/reward. The upside is captured in the returns in NBIS and IREN. If we had waited for Saul-style confirmation (30%+ YoY demonstrated over multiple quarters), we would have entered at much higher prices and captured far less of the gains.
The risk is the failure rate. Saul’s 30+ years give him pattern recognition — he’s seen thousands of “transitioning” companies. Most don’t successfully transition. I don’t have that database. Every pre-inflection bet is a higher-variance decision.
The Nov 2025 portfolio post documented the specific deviations that caused damage: margin (forced selling during drawdowns), failure to trim oversized positions (IREN exceeded 20%), too many positions to follow closely (11 names). The lesson is clear: pre-inflection investing can work, but only if the rest of Saul’s framework — no margin, disciplined trimming, concentrated focus — is followed with even more rigor than Saul himself applies to later-stage positions. I’m not deviating from the philosophy. I’m applying it to an earlier stage of the company lifecycle. Whether that’s wisdom or hubris will be clear in a few years.
5. The Neocloud Race: IREN, NBIS, and the 5 GW Prize
IREN, NBIS, and CoreWeave have emerged as the Big Three independent neocloud operators. Each started from a different origin but is converging on the same prize: providing GPU capacity for the AI era.
My base-case model: 5 GW of active power by ~2030 × $2-2.5B quarterly revenue per GW × 20% net margins × 25x P/E = $200-250B enterprise value. Under these assumptions, the upside is 8.8x for IREN, 3.4x for NBIS, and 3.3x for CoreWeave. This is a model, not a prediction — all three companies could massively under-execute or exceed these numbers. The asymmetry — IREN offering nearly 3x the upside of its peers — is why it’s the largest position.
Power is the real moat. Not revenue. Not earnings. Power. The distinction between active (GPUs running), contracted/connected (binding agreements, utility connections built), and secured pipeline (agreements in place, not yet built) is the difference between a promise and a revenue-generating asset. CoreWeave leads on active power (>1 GW). NBIS is the most aggressive in growth rate. IREN has the deepest pipeline (4.5 GW secured) and needs only ~10% of it for its target.
IREN doesn’t need to win the whole race. It can under-execute relative to its potential and still deliver extraordinary returns. The lower valuation prices in less success. If IREN converts even half its pipeline, the re-rating will be substantial.
The risk for all three is execution. The key metric to watch is active power growth quarter over quarter. Revenue and earnings will follow. If active power stalls, the story breaks — regardless of what the contracted or pipeline numbers say. HIVE remains a small position as a potential consolidation play; its power assets could be valuable to an acquirer as the industry consolidates.
Thank you for reading. I welcome pushback on any of these positions or frameworks — the best ideas on this board come from being challenged.
— InnerPeace123
