Ask any restaurant owner what keeps them awake and it won’t be the menu. It’s the walk-in freezer contract, the produce supplier who might disappear, the lease that expires in eighteen months. The food is the fun part. The supply chain is the business.
That framing helps explain why Akamai announced on September 24, 2026 that Anthropic has committed $11.6 billion over seven years for its cloud infrastructure and software, with room to grow toward $20 billion. Anthropic also picked up a warrant to acquire up to 5% of Akamai’s stock. Akamai shares surged on the news. Reuters, the WSJ, and Akamai’s own release all landed on roughly the same shape of the story: this is a big expansion of a relationship that already existed, reportedly more than six times the size of a prior $1.8 billion arrangement between the two companies.
I build bots for a living. My monthly compute bill would not cover a rounding error in that contract. But the logic behind a deal like this scales all the way down to the side project running on your laptop, and that’s the part worth pulling apart.
Capacity is a product decision, not an ops decision
Every builder I know has hit the same wall at some point. You ship a bot, it works, someone shares it, and suddenly your rate limits are the feature that defines the user experience. Not your prompt engineering. Not your retrieval logic. Your ceiling.
Seven-year commitments exist because that ceiling is expensive to raise on short notice. When a company locks in infrastructure years ahead of demand, it’s making a bet that demand will show up and that the alternative — scrambling for capacity in a tight market — is worse than overpaying for certainty. Anyone who has watched a model provider throttle API access during a launch week understands the instinct.
The practical lesson for us: treat your compute budget as part of your architecture, not as a bill that arrives later. If your bot’s design assumes unlimited inference calls, you don’t have an architecture. You have a demo.
The warrant is the interesting bit
The stock option is what separates this from a standard vendor agreement. A customer with the right to buy up to 5% of its supplier isn’t just a customer. That structure aligns both parties on the same outcome, which is a very different relationship than the one you have with your hosting provider.
Most of us will never negotiate equity into a services contract. But the underlying idea — that dependency runs both directions — is something small builders consistently get wrong. We assume we’re the powerless side of every vendor relationship. Sometimes we are. Sometimes we’re the reference customer a platform desperately wants in its case study, and that’s worth something in support responsiveness, in early access, in pricing. Ask.
What this means for your stack
Concentration of AI spending into multi-year infrastructure deals has a downstream effect on the people building on top. A few things I’d actually change in response:
- Abstract your inference layer. One interface, swappable backends. If provider economics shift, you change a config value, not a codebase.
- Measure cost per task, not cost per token. Token pricing is noise. What matters is what it costs you to answer one user question end to end, including retries and failed tool calls.
- Cache aggressively. The cheapest inference call is the one you don’t make. Semantic caching on common queries is unglamorous and pays immediately.
- Know your degraded mode. When your primary model is unavailable or too expensive, what does your bot do? “Nothing” is a valid answer only if you chose it deliberately.
The quiet signal
Big infrastructure commitments are a forecast dressed up as an invoice. Nobody signs seven years of anything unless they expect to need it. Read the deal that way and it tells you something about where the people closest to model demand think that demand is heading.
For those of us further down the chain, the useful response isn’t awe at the number. It’s a slightly more serious relationship with our own capacity planning. The infrastructure layer of this industry is consolidating into long-term contracts between very large parties. Your bot sits on top of that. Build it so a shift in the foundation doesn’t take your project with it.
The food is still the fun part. But somebody has to think about the freezer.
🕒 Published: