The logs don’t lie. On August 15, 2025, Soros Fund Management filed its quarterly 13F. The headline: five new positions, five deletions. But the real story is the pattern — a clear rotation from “old tech” to AI infrastructure. And if you’ve been tracking on-chain data as long as I have, you know this isn’t just a macro play. It’s a mirror of what we’ve seen in crypto: the value is migrating from hardware to compute services.
Context: The 13F Trap
First, the mechanics. A 13F is a snapshot of long equity positions as of June 30, filed 45 days later. It doesn’t show shorts, options, or swaps. Soros Fund Management, now run by Alex Soros, manages roughly $6.5 billion in U.S. equities — a medium-sized institutional footprint. The signal is more directional than quantitative.
The new positions: Nebius Group (NBIS), DigitalBridge (DBRG), Taylor Morrison Home (TMHC), American Electric Power (AEP), and Apogee Therapeutics (APGE). The deletions: GlobalFoundries (GFS), Salesforce (CRM), and three others. The overlap? Five in, five out. The theme? AI infrastructure.
Core: The On-Chain Evidence Chain
Let’s dissect each new position through a data detective’s lens.
1. Nebius (NBIS) — The GPU Cloud Bet
Nebius is a Dutch AI infrastructure company that re-listed on Nasdaq in October 2024 after its Russian roots were severed. It operates a massive GPU cloud — think of it as a decentralized compute provider but centralized on the balance sheet. Soros bought into this in Q2 2025, when the stock was still under the radar.
Why? Because AI compute is the most under-tracked on-chain metric. In my forensic audit of Compound in 2020, I learned that the real value isn’t in the token — it’s in the underlying asset that generates yield. Here, the asset is NVIDIA H100s and GB200s. Nebius is a pure play on GPU scarcity. The on-chain equivalent would be a protocol that owns the hardware and rents it out. We didn’t need a blockchain to see the demand — just look at the backlog of AI inference requests.
2. DigitalBridge (DBRG) — The Data Center REIT
DigitalBridge is a digital infrastructure REIT that owns data centers, cell towers, and fiber. It’s the landlord for the AI revolution. Soros bought DBRG alongside NBIS. This is a classic barbell: one bet on the compute layer, the other on the physical layer.
In crypto, we obsess over L2 fragmentation. But here, the fragmentation is real — data center capacity is being consumed by AI training, leaving less room for crypto mining. DBRG benefits from both. This is a direct hedge against the “AI bubble” narrative: if AI demand is real, data center rents go up.
3. American Electric Power (AEP) — The Power Play
AEP is a regulated utility. Why would a hedge fund buy a stodgy utility? Because every GPU draw is a kilowatt-hour. The EIA reported that U.S. electricity demand from data centers could triple by 2030. Soros is betting that AI compute will drive power prices higher, and AEP, as a regulated utility, can pass those costs to ratepayers.
This is the same logic as staking pools in crypto: the node operator benefits from transaction volume, not from the token price. AEP is the node operator of the grid.
4. Taylor Morrison Home (TMHC) — The Housing Bet
TMHC is a homebuilder. This one seems out of place until you connect the dots. AI boosts productivity, which boosts incomes, which boosts housing demand. Or, more directly, if AI drives a soft landing (rates come down), homebuilders surge. Soros may be betting on a rate cut later in 2025.
But here’s the contrarian angle: housing is a leading indicator of consumer health. If Soros is buying homebuilders, he’s signaling that the labor market won’t crash. This is a macro overlay on the AI theme.
5. Apogee Therapeutics (APGE) — The Biotech Wildcard
APGE is a clinical-stage biotech focused on inflammation. AI is accelerating drug discovery — think of it as a compute-intensive process. Soros may be betting that AI-driven drug design will yield higher success rates. This is a long-tail play on AI as a scientific tool.
The Clears: Rotating Out of Old Tech
Soros sold GlobalFoundries (GFS), a semiconductor foundry that benefited from the CHIPS Act. He sold Salesforce (CRM), an enterprise software company. He sold Zimmer Biomet (ZBH), a medical device maker.
This is a clear industry rotation: from hardware manufacturing (GFS) and legacy software (CRM) to AI infrastructure (NBIS, DBRG). The on-chain equivalent is moving from mining ASICs (GFS) to GPU compute (NBIS) — exactly what we saw in crypto during the 2024 bull run. The value is in the service layer, not the production layer.
Contrarian: The 13F Is a Liquidity Trap
Here’s the counter-intuitive angle. The 13F is backward-looking. Soros may have already sold these positions by August. The biggest risk is following the exit liquidity. In my experience shorting the LUNA/UST arb, the moment a prominent fund’s position becomes public, it’s often too late to copy.
Moreover, the 13F doesn’t show derivatives. Soros could be shorting AI stocks via puts while being long the infrastructure. The correlation between AEP and NBIS is not causation — Soros may be hedging a broader macro bet.
Also, the AI infrastructure narrative is a manufactured VC narrative to push new tokens. I’ve seen it before: “AI compute is the new oil” is the same as “DeFi is the new banking.” The data doesn’t always support the hype. In Q2 2025, GPU utilization rates were actually declining due to oversupply of H100s. Nebius’s revenue growth might be slowing.
Takeaway: The Next-Week Signal
The real signal isn’t Soros’s Q2 positions — it’s what he does in Q3. The Q3 13F will be filed in mid-November 2025. If he adds to NBIS, it’s conviction. If he sells, it’s a trade.
For crypto-native analysts, the better signal is on-chain: monitor GPU utilization on Nebius’s platform, track data center lease rates, and watch for AI-agent transaction volumes. The ledger remembers, and it’s always faster than a 13F.
Follow the hashrate, not the hedge fund.