🥛 Where crypto fits in AI's $7.6T tab 💰

GM. This is Milk Road, the crypto newsletter that's like having a Bloomberg terminal, but with personality.

Here’s what we’ve got for you today:

✍️ Where crypto fits in AI’s $7.6T tab.
🎙️ The Milk Road Show: The Bitcoin Setup Right Now Looks NOTHING Like Previous Cycles w/ Matt Crosby.
🍪 Bitmine is six weeks from owning 5% of all ETH.

Squid is the cross-chain infrastructure that makes crypto actually work. Something big is coming: Learn more here!

Prices as of 2:00 p.m. ET. Powered by Coingecko.

WHERE CRYPTO FITS IN AI'S $7.6T TAB 💰

Goldman dropped a report this week putting AI infrastructure spend at $7.6T between 2026 and 2031.

That's chips, data centers, cooling, and power - totaling roughly $765B in 2026 alone, and scaling to $1.6T per year by 2031.

Source: Goldman Sachs

Most of the noise around this number is about demand:

Will companies and consumers use AI enough to pay for all this?

But Goldman's flipping the question to the supply side - i.e. the actual cost of building the thing.

(How much will chips cost? How often will they need replacing? How much will new data centers cost to construct?)

The point they’re getting at: those costs are way more uncertain than people think - and a few boring assumptions can swing the total by hundreds of billions.

The four assumptions doing all the work:

1. Chip lifespan.

Accountants depreciate AI chips over 4-6 years. But NVIDIA ships a new architecture every single year with massive performance jumps. If chips hit economic obsolescence in 3 years instead of 6, you're buying replacements twice as often.

2. Data center costs.

Old cloud data centers ran ~$10M per megawatt. New AI ones cost $15-20M. Even facilities built 2 years ago may not handle next-gen chips because power density and cooling have moved that fast.

Source: Goldman Sachs

3. Chip mix (GPUs vs ASICs).

NVIDIA earns ~75% gross margins on its data center GPUs. Custom chips (ASICs) are cheaper. But Goldman believes cheaper chips just lead to more compute usage, so the total bill stays roughly the same.

4. Bottlenecks.

Transformer lead times, power interconnect queues, specialized labor… these don't change unit costs, but they stretch timelines.

That last one is a sneaky little bugger. ☝️

If projects keep slipping on their timelines, that supply friction feeds demand doubt, and that's where capital pulls back.

(Constricted compute access = higher costs for AI services = a smaller amount of customers that can actually afford it.)

The optimistic flip side of it all, goes like this…

IF YOU BELIEVE IN CRYPTO, YOU BELIEVE IN SQUID

Do you believe in crypto? In DeFi? In cross-chain? In RWAs?

Then you already believe in Squid.

Since 2023, Squid has been seamlessly connecting it all - moving $6B+ across 100+ chains and powering 1,000+ apps you probably already use.

Here's what Squid lets you do:

Buy, swap, bridge & send 20k+ tokens across 100+ chains
Tap into 130+ liquidity sources in a single transaction
Skip the chain-hopping, wallet-juggling headaches

Across chains, assets, and apps - wherever you go in crypto, Squid already is.

Something significant is coming from Squid. Learn more here.

WHERE CRYPTO FITS IN AI'S $7.6T TAB (P2) 💰

If the industry pulls off the buildout, history says we can expect cheaper compute to create demand that didn't exist at higher prices.

Meaning today's $7.6T estimated build-out cost might not be enough for whatever AI looks like by 2031.

Sweet, so where does crypto fit in?

Three places:

1. Decentralized GPU networks absorb the trailing edge.

NVIDIA's older chips (A100s and H100s) still rent at prices that suggest 5-6+ year useful lives.

Frontier AI labs constantly upgrade to the newest hardware, but the older chips don't get thrown out.

They get repurposed for lighter work like inference (inference = an AI running your prompt) instead of training (teaching the model in the first place).

That's where DePIN comes in.

Networks like Render and Akash plug older GPU capacity into a global rental market. Bittensor takes it further with subnets that run decentralized AI services on top.

Translation: as frontier labs rotate to new chips, decentralized networks become the natural home for those second-life GPUs.

Source: Render Network

2. Onchain credit for the buildout.

$7.6T is a lot of money. Traditional project finance can absorb some of it, but data center and energy operators are hunting for new capital fast.

That's where rails like USDAI come in. Tokenized debt and onchain credit let infrastructure projects raise from a global pool without waiting on slow private credit syndication.

It's still early, but the build-out will require new financing channels to open up - and onchain credit is one of those channels.