The data is precise. On August 15, 2025, Soros Fund Management filed its 13F for Q2. Five new positions appear: Nebius Group (NBIS), DigitalBridge (DBRG), Apogee Therapeutics (APGE), Taylor Morrison Home (TMHC), and American Electric Power (AEP). Five are gone: Salesforce (CRM), GlobalFoundries (GFS), Intuitive Surgical (ISRG), and two others. The filing is a snapshot — June 30, 2025, reported 45 days late. Code does not lie, but it rarely speaks plainly. The surface story is a simple sector rotation. The deeper truth is a bet on infrastructure that mirrors the scaling challenges crypto faces today.
Context: The 13F Window 13F filings are required for institutions managing over $100M in US equities. They show only long positions, no derivatives, no shorts. Soros Fund Management, now run by Alex Soros, holds roughly $6.5B in US equities. The Q2 filing is the only public window into its thinking. The five new picks are not random. They cluster around two themes: AI infrastructure and rate-sensitive real assets. Nebius is a GPU cloud provider — think AWS for AI inference. DigitalBridge is a digital infrastructure REIT — data centers, fiber, towers. Apogee is a biotech, but the rest are pure plays on physical and digital infrastructure. This is not a macro hedge fund portfolio. It is a sector rotation from 2023’s winners (enterprise software, chip manufacturing) to 2025’s emerging backbone.
Core: The Code-Level Rotation Let me break down each position. I have audited similar infrastructure projects in the crypto space — zkSync’s sequencer, Arbitrum’s fraud proofs. The same logic applies: the value chain is shifting from application layer to infrastructure layer.
Nebius (NBIS): This is a direct bet on AI compute demand. Nebius operates GPU clusters for LLM training and inference. The company was relisted in October 2024 after a restructuring. In Q2 2025, its revenue was growing at 200%+ YoY, but still unprofitable. Soros bought in at a time when the market was skeptical of AI capex ROI. The key metric: GPU utilization rates. If Nebius can maintain 80%+ utilization, the unit economics work. If not, it’s a cash incinerator. In my own audit of decentralized GPU networks (like io.net), I found that utilization is the single biggest friction point. Beneath the friction lies the integration protocol — the software that matches compute supply with demand. Nebius’s proprietary scheduler is that protocol. It’s not open source, but the principle is the same.
DigitalBridge (DBRG): This is a REIT that owns data centers and digital infrastructure. It’s a proxy for the entire AI buildout. Soros likely saw that the hyperscalers (AWS, Azure, GCP) are building their own data centers, but the REIT structure provides a yield on top of the growth. The valuation is not cheap — P/AFFO of 25x. But the thesis is that data center rents will rise as power constraints bite. In the crypto world, we see the same dynamic: Bitcoin miners are pivoting to AI hosting, and Layer2 sequencers are competing for blockspace. The scarcity is not blockspace; it’s power and cooling.
American Electric Power (AEP): This is the most contrarian pick. AEP is a regulated utility. Its margins are capped by regulators. Why would a growth-oriented fund buy a utility? Because AI data centers are projected to add 10% to US electricity demand by 2030. AEP operates in the Midwest, where data centers are clustering. The real value is in the transmission upgrades. Soros is not buying AEP for the 3% yield; he is buying the option on grid expansion. In crypto, we talk about scalability of Layer2s, but the real bottleneck is the physical infrastructure underneath. The same principle applies: the base layer (Ethereum) is congested, so we build L2s. The grid is congested, so we build transmission lines. The investment thesis is identical.
Taylor Morrison Home (TMHC): This is a homebuilder. It seems unrelated, but it’s a macro hedge. If the Fed cuts rates in H2 2025, mortgage rates fall, and homebuilders benefit. Soros is positioning for a soft landing. He is not betting on a recession. The correlation with crypto is indirect: if rates fall, risk assets (including crypto) rally. But the direct signal is that Soros expects the US economy to avoid a hard landing. That is bullish for all growth assets.
Apogee Therapeutics (APGE): A biotech play on inflammation. This is a classic Soros move — a high-risk, high-reward binary bet. It’s a small position, likely a hedge against the rest of the portfolio.
The Cleared Positions: The sales are equally telling. Salesforce (CRM) is a legacy enterprise software play. AI-native tools are eating its lunch. GlobalFoundries (GFS) is a mature-node chip manufacturer. The CHIPS Act subsidies are priced in, but the company lacks the leading-edge technology of TSMC or Intel. Soros sold the winners of the previous cycle and bought the infrastructure of the next cycle.
Contrarian: The Blind Spots The 13F does not show shorts. Soros could have hedged these positions with put options or short ETFs. The filing also does not reveal the size of each position. If the Nebius stake is only $10M, it’s a token gesture. If it’s $200M, it’s a conviction bet. We don’t know. The timing is also critical: Q2 ended June 30. The filing is August 15. By now, Soros may have already sold some of these positions. The lag is the biggest risk for followers.
Another blind spot: the macro environment. Soros built this portfolio assuming the Fed cuts rates. If inflation re-accelerates (due to AI capex demand), the Fed will not cut, and both TMHC and AEP will suffer. The AI infrastructure bet is also crowded. Every fund is buying data center REITs and GPU cloud stocks. The entry price matters. I have seen similar patterns in crypto — the L2 narrative was crowded by mid-2023, and many projects that raised at high valuations are now trading below token price. The same pattern may repeat in AI infrastructure.
Takeaway: The Infrastructure Layer Is the New Scaling Layer Soros’s Q2 move is a microcosm of a larger shift: value is migrating from application to infrastructure. In crypto, we see this with L2s — the scaling infrastructure is where the value accrues (ETH, OP, ARB). In AI, the same: compute, data centers, and energy are the bottlenecks. The contrarian angle is that the most obvious infrastructure plays (like Nebius) may already be priced in. The real opportunity might be in the less obvious components — like the transmission grid (AEP) or the physical housing supply (TMHC). For crypto readers, the lesson is clear: the next bull run will be driven by infrastructure, not just speculation. Layer2s that solve real scalability (not just hype) will survive. The code does not lie — but it takes time to read the full stack. Beneath the friction lies the integration protocol, and Soros just placed his bets.