Good morning. The day after I/O has a habit of producing the more interesting story — the announcement that didn't fit on the keynote stage. Today that story is SpaceX's IPO filing, which puts a $15 billion-a-year number on what Anthropic is paying to access Colossus. The compute race isn't a metaphor anymore: it has line items, depreciation schedules, and an S-1. The rest of today's brief reads as that bet's downstream — Ramp using Codex on GPT-5.5 to compress engineering cycles, Code with Claude framing software work as agent-collaborative by default, Google pushing Gemini Omni into the consumer-creator funnel, and an OpenAI model retiring a math problem that's outlived three generations of mathematicians. If you'd rather get this once a week, subscribe to the weekly brief.
- Anthropic is paying SpaceX $15B a year for Colossus — the compute race gets an SEC-filed price tag
- Ramp's engineering team productizes OpenAI Codex on GPT-5.5 as its code-review backbone
- Code with Claude in London reframes the developer conference as agent-collaborative by default
- YouTube Shorts Remix puts Gemini Omni generative video in front of a billion creators
- An OpenAI model closes the 80-year-old unit distance problem in discrete geometry
1. Anthropic is paying SpaceX $15B a year for Colossus — the compute race gets an SEC-filed price tag
The Verge reports that SpaceX's freshly filed IPO S-1 puts a number on the Anthropic compute partnership the two companies announced earlier this month: Anthropic is paying $15 billion a year for capacity at the Colossus data centers in Memphis, Tennessee. The headline is the dollar figure. The deeper point is the structure: a public-markets disclosure document is now the cleanest source of truth we have about the AI compute economy, because the buyers and sellers of training capacity at this scale are both compute-constrained, contract-laden, and reluctant to talk about either in detail.
The deal's shape is unusual on three axes. First, the buyer is a frontier AI lab, the seller is a rocket company; that's an industry combination that didn't exist as a category two years ago. Second, the underlying facility — Colossus, in Memphis — was originally built by xAI under Elon Musk's direction, and the supply-side story of how SpaceX ended up the entity collecting Anthropic's check is one of the more interesting subplots in the S-1. Third, the $15B/yr run-rate puts the deal in the same order of magnitude as Anthropic's reported annual revenue, which is the read that matters for the underwriting question: at what point does compute spend stop being a capex investment and start being the primary unit-economics constraint on a foundation model business? Anthropic's own announcement of the partnership earlier this month framed it as a long-term capacity arrangement; the S-1 confirms how long-term and how capacity.
Why it matters. If you're underwriting AI infrastructure as an investor, the SpaceX S-1 is the first time a hyperscale AI training contract has been documented with this level of clarity in a public filing — that's the comparable you've been missing. If you're a competing frontier lab, Anthropic locking in dedicated capacity at this scale signals that the answer to "where does the next 18 months of training compute come from?" is increasingly bilateral deals with non-traditional sellers, not the public hyperscaler menu. And if you're a security or sovereignty-minded enterprise customer, the geography matters: Memphis-based training compute, under a US-flagged rocket-company-turned-data-center-operator, is a different risk profile from the multi-region hyperscaler footprint you've been assuming. See our 2026 best AI coding assistants roundup and our ChatGPT vs Claude vs Gemini review for the model lineup this compute spend is meant to keep frontier.
2. Ramp's engineering team productizes OpenAI Codex on GPT-5.5 as its code-review backbone
OpenAI published a customer case study on how Ramp's engineering team is using Codex on GPT-5.5 to accelerate code review. The mechanics are the interesting part: rather than positioning Codex as a single reviewer-of-record, Ramp has integrated it into the PR workflow as a first-pass reviewer that surfaces substantive feedback — broken contracts, missed edge cases, regressions, suggested simplifications — within minutes of a PR being opened, so the human reviewer arrives at a triaged conversation instead of a cold diff. OpenAI's framing is that this shifts "wait-for-a-review-day" from hours to minutes, which is a believable claim because the bottleneck in most engineering teams isn't the actual reviewing — it's the asynchronous wait.
The case study is worth reading carefully because it's one of the few publicly documented examples of an engineering organization actually adopting an AI coding agent into a load-bearing part of the workflow, rather than as an opt-in tool individual engineers use in their editors. The pattern Ramp is describing — agent-as-first-pass-reviewer, human-as-final-call — is the shape that scales. It also accidentally answers a question OpenAI's GPT-5.5 launch earlier this quarter didn't quite get to land: which deployment shape does the new model unlock that prior models couldn't? Reviewing-at-PR-time, at production reliability, is one credible answer. The story is also a strategic counter to yesterday's CodeMender opening: OpenAI's reply isn't a new product, it's a named customer running the integration in production.
Why it matters. If you're an engineering manager, the Ramp pattern — Codex as a first-pass reviewer integrated into the PR workflow — is the most concrete blueprint you have for how to put an AI coding agent into the engineering process without breaking the human review culture that holds quality together. If you're choosing between coding agents, the case study is a useful data point for OpenAI Codex specifically, and a reminder that the question "which model is best in benchmarks" is less load-bearing than the question "which agent integrates cleanly into the workflow we already have." If you're at Anthropic or Google, the gap to close is the named-customer-in-production part — you have the product, you need the cited case study. See our review of OpenAI Codex vs Anthropic Claude Code 2026 for how the two agents stack up.
3. Code with Claude in London reframes the developer conference as agent-collaborative by default
MIT Tech Review's on-the-ground report from Anthropic's two-day Code with Claude event in London — which opened May 19, the same day as Google I/O in Palo Alto — captures the kind of moment that's almost impossible to fake. The MC asks the audience who has shipped a pull request in the last week that was completely written by an AI agent, and a meaningful fraction of the room raises a hand. That's the news. Not the product announcements (the event had several), not the keynote (Dario Amodei spoke), but the audience composition: a developer conference where a sizable share of attendees have already crossed the threshold from "I use AI in my editor" to "I shipped agent-authored code into production this week."
The framing matters because conferences are, among other things, leading indicators of which working norms have stabilized. Google I/O — happening across the Atlantic on the same dates — pitched the developer story as one of new tools and new model capabilities. Code with Claude pitched it as one of new working norms: pair-programming is being replaced not by solo coding but by agent-orchestrating, and the people doing the orchestrating are no longer fringe. The MIT Tech Review piece is honest about the discomfort that creates ("coding's future, whether you like it or not"), and it captures a sentiment that came up multiple times: even attendees who are excited about the productivity gains have not fully figured out what their job becomes in 18 months. That's a tension worth tracking, because it's the cultural variable that determines whether the productivity gains stick.
Why it matters. If you're a developer who hasn't yet shipped agent-authored code, the Code with Claude crowd is a real signal that this is becoming the median experience, not the early-adopter experience. If you're an engineering leader, the working-norms question is the one to put on the next quarterly retrospective: who orchestrates the agents, who reviews their output, who owns the bug at 2am when something they wrote misfires. If you're Anthropic, the event's success is the kind of qualitative validation that the company's developer-first positioning is landing — and the Tech Review piece's "whether you like it or not" framing is exactly the kind of cultural marker that helps a category cross the chasm. Pair this with the Ramp case study above for what production agent-collaboration actually looks like.
4. YouTube Shorts Remix puts Gemini Omni generative video in front of a billion creators
The Verge covers a YouTube I/O follow-on: a new Remix button at the bottom of every YouTube Short that lets a viewer prompt Gemini Omni — Google's multimodal video model — to "reimagine" the clip, restyle it, or insert themselves into it. Google's announcement frames it as a creator-distribution feature; the practical read is that it's the largest-scale consumer deployment of an end-to-end generative video model the industry has seen. YouTube Shorts is reported to have more than a billion monthly users; the Remix button puts a generative-video UX one tap away for every one of them.
The technically interesting question is what Gemini Omni actually does in the Remix loop. Restyling an existing short — applying a visual filter, swapping a background, lifting the subject and re-rendering them in a different setting — is a different generation problem from synthesizing video from scratch, and the public-facing demos so far suggest Google is starting with the conservative end of the menu (re-stylize and re-cast) rather than the radical end (synthesize entirely new scenes). That choice matters for the downstream content-moderation and rights-management story, both of which are quietly enormous: a remix that inserts you into someone else's Short is a rights conversation; a remix that restyles a Short is much less of one. Expect YouTube to be careful about which Remix options are available on which Shorts, and expect the creator-side controls — turn off remix, turn on attribution, royalty share — to evolve quickly.
Why it matters. If you're a Shorts creator, the practical question is not whether viewers will remix your content but how the Remix UI surfaces attribution and reach — both of which are unannounced and which we'll be watching as the feature rolls out. If you're a brand or marketing team running paid creator partnerships, the Remix surface adds a new variable to the contract: who owns derivative remixes, and how is performance measured across the remix tree. And if you're tracking the broader trajectory of generative video, this is the first time the technology has had a billion-user front door — the data Google collects from the rollout will reshape the next generation of Omni and its competitors. See our best AI image generators roundup for the upstream creative tools many of these remixes will start in.
5. An OpenAI model closes the 80-year-old unit distance problem in discrete geometry
OpenAI announced that one of its models has disproved a central conjecture in discrete geometry — specifically, an open problem related to the unit distance problem, which has been open for roughly 80 years. The result is one of a small number of cases where an AI model has produced a counterexample or proof that closes a long-standing open question in mathematics, rather than reproducing a known result or improving an existing bound. Both the math and the way the result was found will get scrutinized by the mathematics community over the coming weeks, which is the appropriate next step.
The substance to focus on is whether the proof is independently verifiable by working mathematicians using only standard tools, and whether the model's approach generalizes beyond the specific problem. OpenAI's post includes detail on the method; the discrete-geometry community will respond with replication attempts and a careful read of the argument over the coming weeks. The arc to compare against is AlphaProof and AlphaGeometry, where the meaningful claim wasn't a single solved problem but a generalizable method that other researchers can use. The interesting question for the field is whether OpenAI's geometry result is a one-off or the leading edge of a method that mathematicians can plug into their own work. This story also pairs with yesterday's Soohak benchmark: as AI starts producing actual mathematical results, the evaluation methodology shifts from "can it solve olympiad problems" to "can it close open questions," which is exactly the bar Soohak is trying to set.
Why it matters. If you're a mathematician, the practical question is whether the proof reads cleanly enough that your students could verify it as an exercise — that's the bar that determines whether the result enters the canon or stays an asterisk. If you're a research lab competing in AI-for-math, OpenAI putting a named open-problem closure on the board is a positioning event analogous to DeepMind's AlphaProof IMO result — the next 12 months will be measured in problems closed, not benchmarks topped. And if you're a research funder, the AI-for-math thread is the part of the AI-for-science story that is closest to producing legible, citable, mathematician-verified outputs, and the funding flow will start to reflect that.
What to take from today
Three threads. First, the AI compute race now has a SEC-filed price tag — Anthropic's $15B/yr Colossus arrangement is the cleanest documented data point we've had on what frontier-lab training capacity actually costs, and the buyer-seller structure (foundation lab buys from rocket company) is the first time a deal of that scale has crossed industry boundaries that cleanly. Second, the agent-collaborative engineering story moved from possible to mainstream this week: Ramp's Codex case study and the Code with Claude audience that's already shipping agent-authored PRs are two sides of the same shift, and the working-norms question — who orchestrates, who reviews, who owns — is the variable to track. Third, generative video is now a billion-user surface — YouTube Shorts Remix puts Gemini Omni in front of every Shorts viewer with a single tap, which makes the next data-collection cycle the largest the technology has ever seen.
Tomorrow's brief lands at 08:00 UTC. If you'd rather read this in your inbox once a week — just the five stories that actually matter — subscribe here.