Frontier Labs Are Running Out of Substrate. Your Procurement Playbook Hasn't Noticed.

In April and early May, Anthropic developers watched their tooling get noticeably worse. Pro accounts at $20 a month lost Claude Code access after seven days of normal use. Some corporate accounts were banned without warning, with no clear path to appeal. Forums filled with the kind of complaints that signal a vendor losing its developer base. Then, almost overnight, the rate limits relaxed. Claude Code limits doubled. Peak-hour restrictions were removed. The fix did not arrive because Anthropic improved its software. The fix arrived because Anthropic, through a deal with SpaceX and xAI’s Colossus 1 cluster, found 220,000 NVIDIA GPUs to lease.

Dario Amodei, on the Big Technology Podcast, gave the explanation in one sentence. “We originally planned for 10x growth, and we’ve seen something like 80x growth in revenue and usage.” That is the operational story. The procurement story is the one no one is writing yet.

Two weeks before the SpaceX deal, Anthropic acquired Bun, the JavaScript runtime. In March, OpenAI announced plans to acquire Astral, the team behind uv, a Python package manager that already moves more than 150 million PyPI downloads per month. These are not coincidences. They are the same problem expressed at two different layers of the stack. Frontier labs need both compute and the runtime that agents live in. They are buying both. And procurement teams at the buyer end are still pricing these vendors as if the conversation were about a SaaS feature set.

This is a procurement problem before it is anything else.

What the 80x Number Actually Means

Plan for 10x. Get 80x. The two numbers diverge by an order of magnitude, and the difference is what gets billed to your team in degraded service.

When a vendor underprovisions by 8x, capacity rationing becomes the only lever they have. The first version of rationing is invisible: longer queue times, slower model responses, bursty errors. The second version is visible: rate limits that strangle paying customers. The third version is what Anthropic shipped in April: account bans without warning, plan tier downgrades, and developer-facing tools that stop working mid-task. By the time a vendor is doing the third thing, the second-order effects have already started. Teams build workarounds. Teams negotiate fallback providers. Teams write internal memos questioning the vendor relationship.

The Colossus 1 deal restored capacity. It did not restore the trust the developers spent April losing. And the underlying issue, that demand for frontier inference is growing faster than any single lab can provision against, has not gone away. It has been deferred to the next capacity wall.

If your contract with a frontier lab does not have language for what happens at the next capacity wall, you do not have a contract. You have a hope.

What Bun and uv Acquisitions Tell You

Anthropic bought Bun. OpenAI plans to buy Astral. The companies bought are not acquihires in the classic sense. Bun is a runtime that competes with Node.js. uv is a package manager that has, in eighteen months, displaced large portions of the pip ecosystem for Python developers who care about speed.

LogRocket’s analysis of the Astral deal puts the strategic logic plainly. Frontier labs need their agents to run somewhere. The “somewhere” is a runtime. If the runtime is owned by a competitor, or by a neutral foundation, the lab does not control its own substrate. By acquiring the runtime team, the lab controls the speed of release, the prioritization of features that benefit agents over humans, and the long-term roadmap of the tool every Python or JavaScript developer in the customer base uses.

The OSS license does not change. The maintainer’s incentives do.

A package manager that prioritizes agent workflows over human ergonomics will, over time, drift toward agent workflows. A runtime owned by Anthropic will optimize for Claude Code’s needs first. None of this is malicious. It is the predictable behavior of any team whose paycheck is signed by a single customer, even if that customer is the team’s own corporate parent.

The procurement implication: every OSS dev tool in your stack is now a candidate for absorption. Your fork-readiness for these tools is part of your vendor risk surface, whether you have written it down or not.

The Two Clauses Your Procurement Playbook Is Missing

If you procure from a frontier lab, your contract probably covers SLA, data residency, model behavior, and pricing. It probably does not cover compute supply contingency or OSS substrate continuity. Both belong in the next contract revision.

Compute supply contingency. What happens when the vendor hits a capacity wall and decides to ration your tier? You need explicit language on:

• Notice period before tier-level rate limits change. Forty-eight hours is the floor. Two weeks is fair.
• Compensation when announced limits are not honored. Service credits are weak; the right anchor is a refund proportional to the fraction of contracted capacity actually delivered.
• A fallback-provider clause that allows you to route a defined percentage of traffic to a competing model without breaching exclusivity. Your business continuity does not survive the next capacity crunch otherwise.
• Disclosure of any compute-supply deals that materially change the vendor’s ability to serve your tier. The SpaceX deal would have been disclosed in equity markets if Anthropic were public. Your contract should require it for you.

OSS substrate continuity. What happens when your stack depends on a tool whose maintainer team gets acquired? You need:

• An inventory, by name, of every OSS dev tool in your build, deploy, and runtime path that is now owned, sponsored, or majority-funded by a frontier lab. Bun. uv. The list is short today. It will not be short in twelve months.
• A fork-readiness assessment per tool. If the maintainer changes the license, deprioritizes a feature you depend on, or adds telemetry you cannot allow, can you fork? Who maintains the fork? What is the cost?
• A contractual obligation from any vendor whose product depends on these tools to disclose the dependency and the fork-readiness plan. If your CI/CD provider runs uv internally, you need to know.

These two clauses do not exist in any procurement template I have seen circulated. They need to start existing. The companies that write them first will set the negotiation floor for the rest of the market.

Do This Now

Pull your active contracts with any frontier lab and any tool whose maintainer team has been acquired or is rumored to be acquired in the next two quarters. Read the SLA section. Find the rate-limit language. If the contract gives the vendor unilateral right to change rate limits with no notice and no compensation, that is your top renegotiation item. If the contract has no language about OSS substrate dependencies, that is your second item. Bring both to your next vendor review.

Your CFO should ask one question at the next quarterly review. “What is our exposure if our top three AI vendors hit a capacity wall in the same week?” If the answer is a shrug, the procurement playbook is the work item, not the model evaluation.

The frontier-AI era is no longer a software era. It is a substrate-and-runtime era. The vendors who win will be the ones who control both. The buyers who survive will be the ones whose contracts assumed this from the beginning.

---

This analysis synthesizes The Pulse: Did Capacity Shortages Turn Anthropic Hostile to Devs? (Pragmatic Engineer, May 2026) and Why AI Companies Are Buying the Teams Behind Your Favorite Dev Tools (LogRocket, May 2026).

Victorino Group helps procurement and platform leaders write the vendor-risk clauses the frontier-AI era now demands. Let’s talk.