Alphabet Just Raised $84.75 Billion in One Shot to Buy Compute. Your Cheap Inference Isn't Going Away — but Here's the Bet to Actually Make.
Alphabet Just Raised $84.75 Billion in One Shot to Buy Compute. Your Cheap Inference Isn't Going Away, but Here's the Bet to Actually Make.
On June 2, Alphabet priced an $84.75 billion equity raise. It was upsized from a previously announced $80 billion, it includes a $10 billion private placement from Berkshire Hathaway, and the company has described it as funding capital expenditure to scale AI infrastructure and global compute. Alphabet's 2026 capex guidance now sits at $180–190 billion. Several outlets called the raise the largest equity capital transaction in corporate history.
Strip away the superlatives and there's one question that touches your business: when a company spends that kind of money building data centers, does the price you pay per million tokens go up or down? For the next year or so, the answer is down. The longer answer is where the actual decision lives.
Why record capex pushes your token price down
This is supply. Alphabet is one of several hyperscalers (alongside Microsoft, Amazon, and Meta) pouring well over a hundred billion dollars each into compute capacity. The raise itself was structured in layers: roughly $45 billion from an immediate public offering, the $10 billion from Berkshire, and about $40 billion more through an at-the-market program starting in the third quarter.
When that much capacity comes online and the labs are in an open price war for your workload, inference gets cheaper. You can watch it happen in real time. OpenAI just previewed GPT-5.6, and its mid-tier model, Terra, is positioned at roughly half the price of GPT-5.5 for comparable work. Gemini Flash keeps undercutting on the cheap end. None of that is charity. It's overbuilt capacity meeting a land-grab for developers, and you're the developer.
For a solo operator, this is the tailwind of the decade. You are renting frontier-grade compute at prices that almost certainly sit below what it costs to serve you. A one-person business can run an AI feature today that would have needed a funded team and a six-figure infra budget three years ago. Take the cheap tokens. Ship the feature. The economics are genuinely on your side right now.
The part that should make you cautious
Here's the catch, and it's the whole reason I'm not just saying "party on." The price you're paying is a subsidized price, and the subsidy is funded by equity raises and venture money, not by the unit economics of inference. $84.75 billion of fresh capital is not a sign that this business prints cash at current prices. It's a sign that someone has to keep writing very large checks to keep the lights (and the GPUs) on.
Subsidized prices have a habit of normalizing once the land grab ends. We've watched it with cloud egress fees, with rideshare, with food delivery, with every platform that bought market share below cost and then spent years quietly clawing the margin back. The day a lab decides growth matters less than gross margin, the per-token price you've built your whole pricing model around can step up. It doesn't have to be dramatic. A 3x move on a workload that's central to your product is enough to turn a healthy margin into a loss.
I've been burned by the small version of this. I built a feature whose economics only worked at a promotional API rate, the rate ended, and I spent a weekend re-architecting to claw back margin that I should have designed for from the start. The cheap price was real. My assumption that it was permanent was the mistake.
The bet to make: take the cheap tokens, design for the expensive ones
So here's the actual move, and it's not "wait for prices to settle": that's how you ship nothing. Use today's cheap inference fully. But design your product as if your token cost were two to three times higher than it is, and make sure the business still works at that number.
Concretely, for a solo operator:
Know your cost per action. Not your monthly API total: the marginal token cost of the specific thing a user does that you charge for. If you can't state it in cents, you can't reason about a price increase. Price your product with margin to absorb a multiple, or build the levers to cut tokens fast: smaller models for the easy 80% of calls, caching, trimming context, and routing the expensive model only where it earns its keep. Keep your model choice behind a config flag so "use the cheaper tier" is a one-line change, not a refactor. And don't make a workload load-bearing if it only pencils out at a promotional rate: that's a feature built on someone else's balance sheet.
Do that, and a pricing reset is an afternoon of tuning instead of an existential event. The capex arms race hands you cheap compute today and the option to keep your margin tomorrow. Take both.
Where I might be wrong
The bear case on my own caution: maybe inference really is heading toward genuinely cheap-at-cost as chips improve and competition stays brutal, in which case designing for a 3x floor is over-engineering and you left growth on the table being conservative. That's possible: custom inference silicon and relentless competition could push the real cost down to where today's prices are sustainable. And reading a single equity raise as a signal about long-run pricing is exactly the kind of macro tea-leaf reading that's wrong as often as it's right; $84.75 billion funds a lot of things, and "your tokens will get more expensive" isn't a forecast anyone can make with confidence.
But the asymmetry is clear. Designing for a higher floor costs you a little conservatism. Assuming the subsidy is permanent and being wrong costs you your margin during the exact window when you can least afford the rewrite. I'll take the cheap tokens and keep my hand on the throttle.
Author
Lukas
@lukcombinator