AI Agents Broke GitHub and Gave Elon Musk a Shot at Owning the Code You Write

How Agents Turned Code Into a Firehose

Something changed in how code gets written, and the numbers are not subtle. In its 2025 Octoverse report, GitHub said developers pushed nearly 1 billion commits in 2025, up 25.1 percent year over year, with a record of nearly 100 million in a single month. It also logged a first wave of autonomous tooling, more than 1 million pull requests opened by its coding agent between May and September 2025, which GitHub itself framed as observational signals rather than causal claims.

Into 2026 the framing got louder. GitHub's own COO has publicly cited roughly 275 million commits per week in 2026, on track for billions for the year. The yearly projection is an extrapolation, but that weekly pace comes from GitHub itself, not just aggregators. The most attributable data point came from Nvidia CEO Jensen Huang, who said at GTC Taipei on June 1, 2026 that AI-generated commits on GitHub had tripled to about 1.4 billion in 2026. Different sources mix AI-generated commits, total commits, and annualized projections, so the exact figure is fuzzy. The direction is not. AI agents are now a large and growing share of everything committed.

GitHub Buckled, Then Changed the Rules

Human-era infrastructure and pricing were not built for machine-paced contribution, and it showed. Through 2026 GitHub saw repeated outages tied to the surge, and Microsoft began tapping AWS capacity to keep it running.

Then the bill arrived. On June 1, 2026, GitHub Copilot moved to usage-based billing, retiring premium request units for token-metered AI Credits. Code completions stay free, while chat, agent mode, code review, and the cloud agent consume credits, with overage charged at each model's API token rates. GitHub's stated reason, in its own words, was that the old model was no longer sustainable because a quick chat question and a multi-hour autonomous coding session can cost the user the same amount. It also paused new Copilot Individual sign-ups and tightened limits, and earlier, in March 2026, a rate limiter that had been undercounting tokens from newer models got fixed, which by GitHub's own admission mistakenly impacted many users with normal usage and triggered a revolt.

The practical takeaway is that heavy agentic users can exhaust a plan's included credits in days, not a month, then pay per token. Some reports describe bills jumping from tens of dollars to many hundreds, but the widely shared 39-to-800-dollar figure is illustrative of a risk category, not a documented case. Either way, flat-rate AI is over.

The Opening for Cursor to Build Its Own Forge

If GitHub's foundations strain under agent load, the obvious move is to build foundations made for agents. On June 16, 2026, Cursor announced Origin, a Git-compatible forge for the agentic era, built on Graphite, the code-review platform Cursor acquired in December 2025. Cursor's launch demo claimed throughput a human-era host was never designed for, around 22.6 commits per second in one repository, sub-400-millisecond global sync, and S3-backed failover. Those are vendor claims, not independent benchmarks. Origin completes a three-layer stack where Cursor and Composer write code, Graphite reviews it, and Origin hosts it. It is waitlist-only now, with general availability targeted for fall 2026. The pitch underneath the specs is simple, whoever owns the forge owns the workflow data and the trust layer for agentic development.

SpaceX Buys the Whole Thing

The same day Origin launched, SpaceX announced it is acquiring Cursor's parent, Anysphere, for 60 billion dollars in an all-stock deal expected to close in the third quarter of 2026, exercising an option that was structured back in April 2026. That matters because Grok and xAI sit under the same Musk umbrella, already training on SpaceX's Colossus supercluster, with Cursor's Composer being folded into a Grok Build CLI. In one stroke, one operator lines up the model in Grok, the editor and agents in Cursor and Composer, review in Graphite, and hosting in Origin.

The Flywheel That Turns Your Code Into Fuel

This is the part worth slowing down on. Cursor holds what analysts call the richest dataset of how developers actually write code outside of GitHub, the prompts, edits, accept-and-reject signals, and step-by-step agent actions. Pair that high-frequency coding signal with xAI's compute and you get a feedback loop, where more Cursor and Origin usage produces better training data for Grok, a better Grok makes Cursor more useful, and that drives more usage. The concern, flagged across the coverage, is twofold. First is model lock-in, since Cursor's appeal was being model-agnostic and routing to Anthropic, OpenAI, and others, and the incentive now is to favor Grok. Second is data governance, whether your code and coding telemetry get used to train the owner's model. The honest line matters here. The defensible, less legally fraught asset is the workflow telemetry and public code, because training on customers' private repositories would be an intellectual-property and trust minefield, so that part is a raised concern, not an announced policy.

The Moat Under It All

Here is the piece most people miss. While competitors wait years in grid-interconnection queues, Musk's empire moves at the speed of its own power infrastructure. xAI's Colossus was brought online by bypassing the grid, with 35 on-site gas turbines producing about 420 megawatts, a large bank of Tesla Megapacks to smooth the draw, and a 150-megawatt substation built in 97 days where a utility would normally take 2.5 years. xAI is now reportedly importing an entire overseas power plant for a one-million-GPU, two-gigawatt site.

The distributed angle the rumor mill loves is real, if early. On June 18, 2026, Tesla filed a trademark for MEGAPOD, modular AI inference hardware meant to run at Supercharger sites, where Musk says roughly 7 gigawatts of power sits available across more than 65,000 stalls when cars are not charging. He has also floated idle Teslas as a 100-gigawatt distributed inference network, with owners paid to opt in. The caveats matter, because this is a trademark filing plus statements, not a deployed product, and the Supercharger and fleet play is for inference, which distributes well, not training, which needs concentration. Still, the structural point holds. Musk controls power generation, storage, grid-connection speed, and a distributed compute substrate that no rival has.

Was It All Carefully Planned

It is tempting to read this as a master plan, but the honest answer is more useful, a consistent strategic direction executed opportunistically rather than a chart drawn years ago. The deliberate signals are there, the April option on Cursor, the December 2025 Graphite buy, and a years-old Musk thesis of owning scarce inputs and building flywheels. So is the opportunism, since the AI wave is industry-wide and not Musk-engineered, these are separate companies with separate shareholders, and we are partly connecting dots in hindsight. For your purposes, intent matters less than outcome. A vertically integrated, model-to-metal AI software stack under one operator is taking shape, by design or by accumulation.

What This Means for Your Team

You do not need to pick a side in the Musk story to act on it. The shifts are concrete.

• Cost control is now an engineering problem. Usage-based AI billing means an unsupervised agent can run up real money, so put guardrails on agent runs, watch token burn, right-size plans, and consider self-hosted CI runners to keep metered work off someone else's meter.
• Treat forge and model choice as a governance decision, not a tooling preference. Ask where your code lives, who can learn from it, and what the training and usage terms say, especially for private and enterprise repositories.
• Design for lock-in resistance. Keep repositories Git-standard and portable, keep your CI and review policy in your own config, and avoid wiring critical workflow into a single vendor's substrate you cannot leave.
• Plan for agent-scale load. Branch hygiene, merge policy, and review automation break first when agents open branches faster than people can read them, and that groundwork pays off on any forge.

If you want help putting cost guardrails on AI-driven pipelines, hardening review and merge gates, or keeping your stack portable, that is exactly what our CI/CD pipeline and DevOps as a service work covers. For the launch that kicked this off, see our note on Cursor Origin.

Sources

GitHub commit and AI figures come from GitHub Octoverse 2025, Jensen Huang's GTC figure via CryptoBriefing, and the weekly-pace projections via Quasa. Copilot pricing and limits come from GitHub on usage-based billing, Changes to Copilot Individual plans, The Register on the rate-limit revolt, and UsageBox on overage mechanics. Cursor Origin is covered by cursor.com/origin and AlphaSignal. The SpaceX acquisition and Grok ties are reported by CBS News, TechTimes, and HyperFRAME Research. Energy and compute figures come from DataCenterDynamics on Colossus Megapacks, Tom's Hardware on the imported power plant, the Tesla MEGAPOD trademark, and Tom's Hardware on idle-fleet compute.