The Era of Free Compute is Over, and That's Actually Fine
Published on 30.04.2026
TLDR
GitHub Copilot is ditching flat premium requests for token-based billing starting June 1. SpaceX is partnering with Cursor and deploying a million-H100-equivalent supercomputer behind it. OpenAI is now on AWS Bedrock. These three stories look separate but they're telling the same story: compute is expensive, the subsidies are drying up, and whoever controls the infrastructure controls the economics.
GitHub Copilot's Billing Shift Is a Signal, Not Just a Change
Starting June 1, GitHub Copilot moves to AI Credits calculated by token consumption — input, output, and cached. Each plan includes credits equal to its subscription cost: $10 for Pro, $39 for Pro+. Once those are gone, there's no fallback to a simpler free model. That last part is the actual news.
For a while, GitHub absorbed the difference between a quick chat message and a multi-hour autonomous coding session. Those two things cost the same to you, but very different amounts to GitHub. That math stopped working once agentic usage became the default rather than the exception. Running an agent through an entire repo for an hour is not a chat. The compute demands are genuinely different.
I think this is honest, even if it stings. Flat-rate pricing for unlimited AI inference was always a bet that usage patterns would be predictable. Agentic workflows broke that assumption completely. The question now is whether developers who built habits around unlimited access will absorb the change or push back hard enough to shift the market.
GitHub Copilot moves to usage-based billing
The SpaceX-Cursor Deal Is a Compute Story Wearing a Productivity Costume
Most coverage of the SpaceX-Cursor deal focused on SpaceX standardizing on a fast AI coding tool. That part is true but misses the point. SpaceX is partnering with Cursor to build a next-generation "coding and knowledge work AI," with an option to acquire the company for $60 billion. Behind that deal sits Colossus, SpaceX's supercomputer with the claimed equivalent of a million Nvidia H100 chips.
A rocketry company does not do a $60 billion software deal because they like the IDE. They do it because compute capacity is the scarce resource and demand-layer software — tools that route millions of developer requests — is what determines where that capacity flows. Cursor sitting at that demand layer, with tens of thousands of xAI chips already training its latest model, is infrastructure as much as it is product.
Here's what I keep thinking about: xAI is already claiming a 3 trillion parameter Grok release is coming as soon as May. If that lands, the bottleneck is not model quality. It's compute throughput. Every deal that secures guaranteed inference at scale, whether through ownership, partnership, or long-term contract, is really a hedge against that bottleneck.
OpenAI on AWS Bedrock Changes the Provider Calculus
OpenAI renegotiated its exclusivity with Microsoft and landed on AWS Bedrock. GPT-5.5 is now available alongside Anthropic, Meta, DeepSeek, and others. For enterprise customers, this means OpenAI models inherit all the AWS infrastructure controls they already depend on — IAM, CloudTrail, encryption — and usage likely counts toward existing AWS cloud commitments.
The Azure/OpenAI lock-in was always a bit artificial. Two genuinely separate companies sharing exclusive distribution made sense when OpenAI needed the capital and Azure needed the differentiation. What changed is that OpenAI is big enough now that exclusivity is a constraint, not a benefit. Microsoft got freed from revenue-sharing commitments in exchange. Both sides got something.
What this means practically: developers on AWS who wanted GPT models had to either build custom routing or use the Azure endpoint. That friction goes away. For teams already running Bedrock, adding OpenAI to the mix is now just another model choice rather than a separate integration. That's a genuine quality-of-life improvement, even if the underlying models are the same.
Model Freedom Is the Only Stable Foundation
The through-line across all of this: closed systems reprice when the economics shift. That's not a criticism, it's just how it works. GitHub absorbed inference costs until it couldn't. Flat-rate subscriptions made sense until agentic usage made them unsustainable. Exclusive partnerships made sense until they didn't.
Developers who locked into a single provider are now at that provider's mercy every time the model lineup changes, a price tier gets restructured, or a model gets deprecated without warning. BYOK coverage and transparent token pricing aren't premium features at this point. They're the baseline for any tool that expects to stay useful when the next pricing restructure lands.
Kilo's transparent pricing and BYOK coverage
Key Takeaways
- GitHub Copilot's move to token billing removes the safety net fallback model — agentic sessions now cost proportionally to actual compute consumed
- The SpaceX-Cursor deal is a compute alignment play, not a productivity tool procurement; owning the demand layer while securing H100-scale throughput is the real prize
- OpenAI landing on AWS Bedrock ends an artificial exclusivity; enterprise teams get OpenAI models with native AWS controls and no separate integration path
- Every major AI platform is converging on usage-based pricing because subsidized inference was always unsustainable at agentic scale
- Model portability and BYOK coverage matter more now — single-vendor lock-in means absorbing every future repricing decision that vendor makes