All of this was pulled from the Byblos live feed - 1,400+ VC firms, 60,000+ newsletters, one place.
See what else came in today at byblos.digital →
Among the founder calls and LinkedIn outreach, I decided to go and use my own Byblos App. Instead of reading thousands of VC newsletters, I skimmed through the intel section. The part that caught my attention was a piece by Edge & Node — a crypto fund, not even an AI one — wrote that good-enough AI is getting cheap faster than the data centers can pay for themselves.
That didn’t line up with anything else I’d read because everyone else was writing about the huge amounts of money going into those data centers.
So I went back and checked everything.
Nvidia announced a $500 billion financing program for AI compute, with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR.
Four separate funds flagged this too— Nazaré, TEN31, Haystack and Kucoin.
Among the deals, CoreWeave is leasing A100 GPUs through 2029. Riot Platforms is building a $9.1 billion data center with Anthropic as the tenant. Hut 8 signed a fifteen-year data center agreement in Texas.
Fifteen years.
That’s an office building contract. You sign that when you’re confident the thing you’re renting out is worth roughly the same in year twelve as in year one.
I found basically one person writing about what AI actually costs now.
Chris Zeoli:
AI token prices falling around 10x per year;
Alphabet cut the cost of serving Gemini by 78% in twelve months;
GPT-3-class pricing down 1000x;
DeepSeek published a 545% cost-profit ratio on serving.
Kucoin picked up the same shift in a different way as they wrote that the market has stopped asking how much companies spend on AI and started asking how efficiently they spend it.
That’s it. Ten funds on the financing. Basically one guy on the deflation.
I think that point is that these two things haven’t been put next to each other yet.
How to get started:
🟣 Sign up over here and get a free 7-day trial of all features
🟣 Attend a 1:1 setup call with us
🟣 Build your own watchlist from 1,400+ VC database
As I said before, bitcoin miners are the ones actually building these data centers.
Bitfarms renamed itself Keel Infrastructure and quit US mining entirely.
Bitdeer converted a mining site for a company called Volta.
Galaxy bought three data center sites in Texas.
Neoclouds are crypto mining infrastructure with new tenants.
— a16z
So the people who last built giant power-hungry facilities for an asset that crashed are building giant power-hungry facilities again. Except this time they’re landlords: they collect rent and someone else takes the price risk.
I think this is not just luck but that they’ve done this before and learned something.
I want to be fair, because the bull case is sitting in the same data.
Zeoli’s other number: Google’s token consumption grew 330x in two years, so price down 10x, volume up 330x. This basically means the revenue still goes way up.
But volume growth goes to whoever owns the customers. Price deflation hits whoever sells the compute. That’s two different groups.
I went looking for the bear case too, and found Michael Burry — short Micron, short Nebius, short Oracle, short the semiconductor ETF, warning that AI infrastructure demand isn’t sustainable.
I’m not saying Burry’s right. I’m saying if you only read VC newsletters, you would think there is no disagreement here at all if four funds on the financing, one on the short.
Figure out which side of the price curve you’re on.
If you resell compute, or your product is a thin layer on someone else’s model, the price you can charge falls as the model gets cheaper. You don’t capture the volume growth either, because the customer isn’t really yours.
If you own the customer, deflation helps you. Costs drop, you keep the margin.
And if you’re building physical infrastructure, Kucoin’s point is the one I’d underline. They wrote that power scarcity is now the main value driver for neoclouds. Chips get cheaper every year but a grid connection doesn’t.
One last thing I didn’t expect to find. Kucoin tracked DRAM prices surging, Apple testing Chinese CXMT memory for iPhones and MacBooks, and memory becoming a serious chunk of the iPhone 18 Pro’s bill of materials. The AI buildout has made ordinary memory expensive.
So if you build hardware, your costs went up last quarter because of AI. Even if your product has nothing to do with AI.
That’s the data. Now go build something.
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.