Remember when the shovel sellers were the safe bet? That was the whole pitch of the gold rush metaphor. Prospectors go broke, but the guy selling picks and pans walks away rich. For about two years, that was the consensus read on AI infrastructure: models come and go, but compute is compute, and whoever owns the racks collects rent forever.
Then the shovel sellers started needing loans to buy shovels.
Nscale, an AI compute provider, is in talks to raise $3.5 billion in pre-IPO financing, per TechCrunch. Part of that is reportedly up to $1.5 billion in convertible notes sold to a group of investors. The round comes right on the heels of a $45 billion deal with Anthropic. The stated purpose is bolstering infrastructure ahead of a public offering and strengthening the company’s financial position going into that IPO.
Read those two numbers next to each other. Forty-five billion in commitments. Three and a half billion in fundraising. That ratio is the entire story, and it’s worth sitting with before anyone starts calling this a vote of confidence.
The customer is also the reason you need money
Here’s what I find genuinely interesting about the sequencing. Nscale announced a deal with Anthropic, then went looking for billions. In most businesses, landing your biggest customer ever is the thing that reduces your need for outside capital. You have revenue now. You have contracted cash flow. Banks like you.
In AI compute, signing a giant customer is the thing that creates your capital problem. You can’t serve a $45 billion agreement with the data centers you already have. You have to go build them. GPUs, power contracts, cooling, real estate, all of it upfront, all of it before a single dollar of that contract converts into cash in your account.
So the contract is simultaneously the asset you show investors and the liability that sent you to investors in the first place. That’s not a scandal. It’s just how capital-intensive buildouts work. But it does mean the celebratory framing around these announcements deserves more skepticism than it usually gets.
Why convertible notes are the tell
The reported structure matters. Selling up to $1.5 billion in convertible notes ahead of an IPO is a specific choice with specific tradeoffs.
- You get cash now without setting a firm valuation today, which is useful if you think the public market will price you higher than private investors would.
- Investors get downside protection. They’re lending, not buying, until conversion. If things sour, they’re higher in the capital stack than equity holders.
- You defer dilution rather than avoiding it. The shares get issued eventually, usually at a discount to the IPO price.
That’s a deal structure for a company that wants a specific number in a specific window and is willing to give up future upside to get it. It’s not a distress signal. It’s also not the behavior of a company with a surplus of use in negotiations, no matter how large its contract book looks.
What this means if you actually use these tools
Most readers of this site aren’t buying pre-IPO convertible notes. You’re picking which API to build on, which agent platform to trust with your production workload, which vendor won’t disappear mid-quarter. So let me connect this to something practical.
The compute layer underneath your favorite AI tool is being financed with debt, notes, and IPO expectations tied to contracts that stretch years into the future. That’s a real dependency chain. When your agent framework calls a model, that model runs on hardware someone borrowed money to buy against revenue they haven’t collected yet.
This doesn’t mean anything breaks tomorrow. Large infrastructure has always been built this way, and the demand for AI compute is not imaginary. What it means is that pricing stability at the application layer is downstream of financing conditions at the infrastructure layer. If capital gets more expensive, that pressure travels. It shows up as rate limits, tier changes, and quiet repricing of the plans you built your product around.
The part nobody wants to say out loud
The AI infrastructure space has developed a habit of treating announced deal value as accomplished revenue. A $45 billion agreement is a promise about the future between two companies that both need the other to succeed. Anthropic needs compute. Nscale needs a marquee customer to justify its buildout and its IPO story. Both parties benefit from the number being enormous and public.
None of that makes it fake. It does make it a forecast rather than a fact. And the $3.5 billion raise is Nscale acknowledging, in the most concrete way available, that the gap between the promise and the capacity to deliver on it is measured in billions of dollars it does not currently have.
I’d watch what the IPO prices at, not what the contracts announce. Deal headlines are marketing. A public market valuation is a group of people with money on the line making a judgment. That’s the number that will tell you whether the shovel business is as good as everyone assumed.
🕒 Published: