The closure of OpenAI’s Sora was a whisper that turned into a roar. A product that could generate any video from text—a dream of the creative class—was shut down. The official reason: overwhelming inference costs. Then, in the same quarter, Higgsfield, a lesser-known AI video platform, announced a $400 million raise at a $5.4 billion valuation. Its revenue? A claimed $700 million annualized run rate, up from $20 million a year ago.
On the surface, this is a story of two destinies: one that ran out of cash, and one that found a market. But the ledger tells a deeper truth. Both Sora and Higgsfield are wrestling with the same invisible antagonist: the cost of compute. And in that struggle, I see the contours of a new economic layer—one that crypto is uniquely positioned to own.
Let me be clear: Higgsfield is not a blockchain company. Its CEO, Alex Mashrabov, talks about enterprise marketing videos, not smart contracts. But the infrastructure it depends on—the GPUs, the data centers, the energy contracts—is a system of centralized scarcity. That scarcity is the crack where the light of decentralized compute enters.
The Context: A Market of Two Halves
Higgsfield’s pivot from consumer to enterprise is the textbook example of finding product-market fit. The company started as a consumer app, amassing 30 million users across 238 countries. But the real money came from brands like Dollar Shave Club, which now produces multiple videos daily using Higgsfield’s platform. In eight months, enterprise revenue went from a quarter of total to the majority. That’s a velocity that growth equity firms like Goldman Sachs (now an investor) dream of.
But the cost of serving those enterprise clients is hidden. Sora’s daily inference cost was reported at $15 million—a figure that, even if inflated, signals the terrifying unit economics of video generation. Higgsfield’s COO, Michael Mignano, admitted that the company’s $700 million run rate is partly driven by the need to “pre-reserve compute capacity.” In other words, the company is spending heavily on GPU futures, locking in capital to ensure supply.
This is where the crypto narrative begins. The centralized cloud model of AWS, Azure, and GCP is a form of rent extraction. Every video generated is a transaction that pays a tax to the cloud provider. For a company scaling at 35x, that tax becomes a existential question: can the margins survive the growth?
The Core Insight: Compute as a Variable Cost, and the DePIN Solution
In my years of auditing tokenomics—from the 2017 ICO boom to the 2024 DeFi shakeout—I’ve learned that any system where the cost of production scales linearly with usage is a system that will eventually hit a liquidity crisis. Video generation is the ultimate example. Each frame requires a forward pass through a massive diffusion transformer. The more popular the platform, the more GPUs it needs. In a centralized cloud, that means you’re at the mercy of a single provider’s pricing.
Decentralized physical infrastructure networks (DePIN) offer a counterpoint. Networks like Render, Akash, and io.net aggregate idle GPU capacity from around the world, creating a spot market for compute. The cost is often 30–60% lower than centralized cloud, and the supply is elastic. For a company like Higgsfield, which generates millions of video frames per day, even a 40% reduction in compute cost could mean the difference between a 10% margin and a 50% margin.
But there’s a deeper advantage: the liquidity of compute. In a centralized system, you pay for reserved instances whether you use them or not. In a decentralized system, you pay per task. That’s a variable cost model that matches the unpredictable demand of a viral video campaign. It’s the same logic that made file storage cheaper on IPFS versus AWS—but for the most expensive resource in AI.
A transaction is just a promise frozen in time. The promise of a GPU is that it will compute for a certain number of cycles. That promise, when tokenized, becomes a tradable asset. This is the thesis behind projects like Golem’s new GPU market or the compute derivatives on L2s. I believe the next bull run will be driven not by NFTs or memecoins, but by the tokenization of compute hours.
The Contrarian Angle: Higgsfield’s Valuation Is a Bet on Centralization
Here’s the contrarian view that most analysts miss: the $5.4 billion valuation is not a bet on AI video. It’s a bet on the continued dominance of centralized cloud providers. Intel’s investment in Higgsfield is a hedge—Intel wants to ensure its Gaudi chips have a marquee customer. Goldman Sachs is betting on the growth of a SaaS model that, historically, fails when the underlying resource gets commoditized.
But the history of tech is the history of commoditization. Mainframes were once the only way to compute; then PCs came, then the cloud, and now we’re moving to decentralized fog. The cost of a GPU is dropping, but the cost of a centralized GPU is still high because of the monopoly of providers. The moment a decentralized compute network reaches a critical mass of supply, the pricing power of AWS will collapse.
Higgsfield’s window is real. It has 30 million users, a strong enterprise onboarding flow, and a brand that brands trust. But its moat is not the model—it’s the customer relationship. The model itself is a diffusion transformer, similar to what Google Veo or Meta’s moviegen can do. The true differentiator is the data flywheel: every video generated by a brand teaches the model about that brand’s style. That data is valuable, but it’s not locked to Higgsfield. If a competitor offers a better price for the same quality, the brand will switch. The switching cost is low.
Silence is the loudest market signal. The silence here is about gross margins. Higgsfield has not disclosed its compute cost per video. If it’s 50% of revenue, the company is barely breakeven at scale. If it’s 80%, the company is losing money on every video and hoping to make it up on volume. That’s a Ponzi of compute.
The Takeaway: The Next Unicorn Will Be a Compute Layer
I’ve spent the last 17 years watching markets form and dissolve. The 2017 ICO boom was about the tokenization of value. The 2021 NFT boom was about the tokenization of culture. The 2025–2026 cycle is about the tokenization of physical resources—bandwidth, storage, and now compute.
Higgsfield’s raise is a signal that the demand for AI video is real. But the infrastructure to serve that demand is brittle. The real opportunity is not in building another video generation platform; it’s in building the decentralized compute network that all platforms will rely on. The companies that will win the next decade are not the ones that own the best AI model, but the ones that own the cheapest, most reliable, and most liquid compute market.
When the GPU costs more than the video, the market will find a way to share the load. That’s the crypto answer to the AI cost crisis. And it’s an answer that doesn’t require a $5.4 billion valuation to start. It just requires a ledger that can record the promise of a frame, and a community that trusts that promise to be executed.
Trust is a luxury good in a digital world. But compute, when tokenized, becomes a commodity. And commodities are the foundation of every economic cycle. The question is: who will own the commodity of the next cycle?