Higgsfield just closed a $400 million round at a $5.4 billion valuation. The AI video startup claims $700 million in annualized revenue as of August, 35x growth in a year, and 30 million users across 238 countries. The numbers are loud. But the silence in the fine print is louder.
Context: The Sora graveyard
OpenAI shut down Sora in early 2026. The reason? Daily inference costs reportedly hit $15 million — against a lifetime revenue of $2.1 million. Sora was a consumer product. Higgsfield is enterprise. But the underlying physics of video generation doesn’t care about your business model. The compute cost per frame is the same. The only difference is who pays.
Higgsfield’s CEO told the FT that the new capital is partly to “reserve compute capacity” — pre-paying for GPU services months in advance. This is a signal that the company’s growth is already constrained by hardware supply. They are buying compute futures, not building moats.
Core: The on-chain evidence chain
Let’s follow the compute. Nansen’s Smart Money flows show that decentralized GPU networks like Render Network and Akash Network have seen a 200% increase in token velocity over the past two quarters. Why? Because AI video startups are quietly hedging against centralized cloud lock-in. I’ve been tracking this since my 2024 report on ETF flows — capital moves where the bottleneck is. The bottleneck for AI video is not model quality. It’s the cost per frame.
Code does not lie. Check the contract: Render’s token burn rate spiked 40% in the same week Higgsfield’s funding was announced. That’s not a coincidence. Smart money is positioning for the compute cost crisis to shift demand to decentralized alternatives.
Higgsfield’s $700 million ARR — if real — means they are generating millions of videos per month. Each video, depending on length and resolution, costs anywhere from $0.50 to $5 in inference compute. At scale, that’s hundreds of millions in annual infrastructure spend. The company hasn’t disclosed its gross margin. When a growth-stage company hides its margin, it’s usually because the number is ugly.
Contrarian: The Intel trap
Intel invested in this round. That’s not just capital — it’s a strategic play to lock Higgsfield into Intel’s Gaudi chips. Gaudi is cheaper than NVIDIA, but the software ecosystem is years behind. If Higgsfield optimizes for Intel’s hardware, they risk being stuck on a slower iteration cycle. The moment Google Veo or Meta’s video model ships a better quality-to-cost ratio, Higgsfield’s customers will have low switching costs. Their current moat is product integration, not technology.
Liquidity leaves before the crash hits. The hype around Higgsfield’s revenue growth is a narrative built on self-reported data. The actual cash flow could be negative after compute costs. If the $700 million is inflated by multi-year commitments or non-recurring project fees, the real P/S ratio could be 15x or higher. That’s expensive for a company with a high burn rate and no margin visibility.
Takeaway: The next signal
Watch the GPU utilization rates on decentralized networks. If Higgsfield’s compute cost per video exceeds $1, they will be forced to seek cheaper alternatives. That’s when Render and Akash see a demand spike. The smart money is already front-running that shift. The question is not whether Higgsfield can sustain its growth — it’s whether the cost of compute will break the centralized model before the decentralized model scales.
Code does not lie. Follow the compute.