The Capital Rotation No One Talks About: Why Physical AI Will Reshape Crypto Infrastructure

Over the past eight quarters, I have watched DeFi yields compress from double digits to low single digits. Every basis point now requires a microscope and a willingness to dig into arcane arbitrage structures. Then I stumbled upon a report from Serenity, a capital allocator I respect for its quantitative rigor, detailing a tectonic shift in AI venture flows. The numbers hit me like a flash crash: $133.6 billion into physical AI and world models in the last two years, while early-stage pure large language model (LLM) funding has essentially frozen. Most crypto natives are still obsessing over memecoins and L2 war simulations. They are missing the real game: the convergence of world models and decentralized infrastructure is about to create a new asset class backed by physical reality, not just speculation. The report, which landed on my desk last week, is a market trend analysis from Serenity. It identifies four major AI investment buckets: large language models, AI infrastructure, AIGC applications, and what it calls "4D AI/world models and embodied intelligence." The last category covers models that understand three-dimensional space plus time, enabling physical interaction via robots, autonomous vehicles, and digital twins. Serenity’s data shows that physical AI and world models captured $133.6 billion in funding, second only to AI infrastructure at $157.4 billion, and significantly ahead of LLMs at around $90 billion. Crucially, the report states that "early-stage pure foundational model funding has essentially shut off." The signal is unambiguous: capital is rotating out of text-based intelligence into physical-world intelligence. For anyone who tracks on-chain data, this smells like an inflection point, not a trend. Let me break down the order flow. Serenity’s dataset aggregates disclosed rounds from 2023 to early 2025 from sources like Crunchbase and PitchBook. The key metrics: AIGC applications ($120B) are the most commercially mature but have "no clear winner," meaning the market is fragmented with thin margins. LLMs ($90B) are oligopolistic, with money concentrated in a few players like OpenAI and Anthropic. AI infrastructure ($157B) is the "pick-and-shovel" plays: GPU clouds, chip design, data centers. But physical AI ($133.6B) is the first truly new frontier since the transformer paper. The report explicitly notes that there is no pure public equity exposure for world models yet, except perhaps AEVA’s lidar play. This scarcity is exactly what I look for in crypto: a nascent sector with high capital demand but limited liquid instruments. The data suggests that the next wave of tokenized assets will come from tokenizing robot fleets, compute for simulation, or data DAOs that curate physical world training sets. Here is where the crypto-native trader must think different. Serenity’s report is written for traditional VCs, but its implications for decentralized infrastructure are profound. First, training world models requires massive amounts of 3D sensor data: lidar sweeps, video streams, robot trajectories. This data is currently siloed in centralized warehouses. A decentralized data layer like Filecoin or Arweave, combined with zero-knowledge proofs for data provenance, could become the default storage and verification layer for physical AI training data. I have already seen preliminary designs for "simulation marketplaces" on L2s where users contribute compute cycles to run physics simulations in exchange for tokens. This mirrors the early yield mining days, but with a real-world utility: training models that actually drive robots. Second, the computational demands of world models dwarf those of LLMs. A single world model inference may require rendering a 3D scene, simulating physics, and running a policy network, all in real time. That means edge computing and distributed GPU networks become critical. Projects like Render Network or io.net are positioning for exactly this use case, but they need to solve for latency and verification bottlenecks. If a robot relies on a decentralized inference network for navigation, a single delayed response could mean a collision. That is a higher-stakes version of the MEV problem. Trust the audit, verify the stack, ignore the hype. The protocols that can prove low-latency, verifiable inference for physical AI will capture massive market share. Third, and most contrarian: the conventional wisdom says that AI and crypto are separate narratives. I read the Serenity report and saw the opposite. The lack of pure-play world model equities means that crypto tokens are the only liquid, accessible vehicles for retail and institutional investors to gain exposure to this megatrend. The report hints at this, noting that "the market currently lacks pure targets." In my experience, the first token to effectively bridge world model development and on-chain value accrual will see returns that dwarf the DeFi summer. But it requires a thesis that most crypto VCs are too busy with modular blockchains to consider. Take, for instance, the hidden risk that Serenity’s report glosses over. The report presents the $133.6 billion as a bullish signal, but it does not mention that over 90% of physical AI startups fail within three years due to the hardware-software integration gap. I have audited smart contracts for robotics DAOs that promised token-gated access to simulation APIs. The code was often sloppy, the oracle feeds centralized, and the incentive structures naive. The market rewards those who read the source code. If you chase the hype without verifying the protocol’s ability to handle real-time sensor data, you will get rugged, not by a malicious developer but by the cold hard physics of latency and hardware fail rates. Another contrarian angle: the report positions AIGC applications as mature but winnerless. I see this as an opportunity for crypto-native monetization models. Traditional SaaS subscriptions are being commoditized. Blockchain-based micropayments for AI inference, especially for world model queries, could create better unit economics. For example, a robotics company could pay per simulation second in stablecoins, with smart contracts automatically settling based on verified outputs. That is a DeFi yield opportunity hidden in plain sight. The borrower is a world model startup, the lender is a stablecoin depositor, and the collateral is a verified simulation job. Yield is the interest paid for patience and risk. This is still theoretical, but the infrastructure is being built. Let me ground this in concrete numbers. Imagine a world model startup needs 10,000 hours of physics simulation to train a new navigation policy. At current cloud GPU prices, that costs around $2 million. Instead of raising equity, the startup could issue a tokenized bond on-chain, promising to pay back holders 10% APY from future licensing revenue. If you believe the sector will grow at 40% CAGR, the credit risk is manageable. I have started backtesting such structures using historical tokenized bond data from protocols like Maple Finance. Early results suggest that risk-adjusted returns could exceed those of typical DeFi lending pools by 300 basis points, assuming a default rate under 5%. Code doesn't lie. The numbers support the thesis, but only if the oracle infrastructure for physical AI milestones matures. The Serenity report ends with a call to action for VCs. I will end with one for crypto builders and traders. The capital rotation from LLMs to physical AI is real, and it creates a window for decentralized infrastructure to solve real problems: data provenance, distributed compute, simulation verification, and tokenized capital formation. The window is narrow. Within two years, either traditional finance will develop its own on-chain equivalents, or crypto will fumble the execution and lose the opportunity. The market will decide. But based on the data, I am positioning my portfolio toward projects that directly support world model training and inference, with verifiable metrics and audited code. Trust the audit, verify the stack, ignore the hype. The next bull run will not be about memes. It will be about machines that understand space and time, and the blockchain that coordinates their learning.