Digital Colliers Daily Briefing — May 21, 2026

The S-1 era has come for SpaceX, and with it a forensic look at how AI compute money actually moves: Anthropic, the Claude developer, is wiring $1.25 billion a month to Elon Musk's xAI through 2029. Google used I/O to recast nearly every consumer surface around Gemini 3.5 Flash and a new agent stack centered on Antigravity and Gemini Spark. And OpenAI claims a general-purpose reasoning model has disproved one of Paul Erdős's longest-standing conjectures in discrete geometry — a result endorsed by Tim Gowers and other mathematicians. The throughline: capital, compute, and capability are consolidating around a small number of operators at extraordinary scale.

1. SpaceX's S-1 lays bare a $15B-a-year Anthropic–xAI compute pact

A mid-century businessman signing a contract inside an industrial control room.

What happened. SpaceX filed a nearly 400-page S-1 with the SEC on Wednesday, the first detailed look at its finances in 24 years, and disclosed Cloud Services Agreements signed in May with Anthropic. Under the contracts, Anthropic will pay $1.25 billion per month — roughly $15 billion a year — through May 2029 for access to compute at the Colossus and Colossus II data centers in Tennessee and Mississippi, with a reduced rate for May and June while capacity ramps. Either side can terminate on 90 days' notice. TechCrunch puts the headline total at over $40 billion in potential xAI revenue. The filing also confirmed $2.8 billion in recent commitments to gas turbines to power those facilities, and SpaceX targets a Nasdaq debut as soon as June 12 under the ticker SPCX at a $1.75 trillion valuation.

Why it matters. The deal formalizes the "neocloud" model TechCrunch describes: an AI lab simultaneously building for itself and reselling spare capacity to direct competitors. The subtext is sharp — xAI appears to have overbuilt for Grok, whose usage has dropped, and is monetizing the overhang via Anthropic ahead of its parent's IPO. For Anthropic, paying a rival roughly 1.5x its projected Q2 revenue ($10B+, per the Wall Street Journal) underscores how compute scarcity, not model architecture, now defines competitive position.

Who is affected. Anthropic gets a 300-megawatt runway through 2029 to feed Claude's coding products. xAI gets a multi-year, ~$40B revenue annuity that materially de-risks its IPO story. Hyperscalers — AWS, Microsoft, Google Cloud, CoreWeave — face a new merchant competitor with rocket-company economics. Pension funds, including those of California, New York, and New York City, have already called SpaceX's dual-class structure "novel and extreme," warning retail investors about a potential meme-stock dynamic if 30% of the offering goes to individuals.

What to watch next. Whether SpaceX signs additional "similar services contracts" Musk telegraphed on X; the SEC's response to governance provisions that, per Reuters, leave Musk firable only by himself; the Thursday Starship test and its bearing on the $15B Starship development overhang; and how environmental litigation around Colossus's turbines progresses as 46 portable units now operate at the Mississippi site.

Sources:

2. Google I/O reorients Search, Gemini, and developer tooling around agents

A vintage switchboard operator routing calls through a vast panel.

What happened. Google made Gemini 3.5 Flash generally available, claiming it outperforms Gemini 3.1 Pro on Terminal-Bench 2.1 (76.2%), GDPval-AA (1656 Elo) and MCP Atlas (83.6%) at Flash-tier latency, with Gemini 3.5 Pro slated for next month. The model is now the default for AI Mode in Search, which Google says has crossed one billion monthly users. The company also rolled out Gemini Omni, a video-first multimodal generation model; Gemini Spark, a personal agent built on Gemini 3.5 Flash and Antigravity that runs on phone or laptop and is initially limited to U.S. Google AI Ultra subscribers; Antigravity 2.0 as a standalone desktop app with subagents, hooks, and async task management; and a "Search agents" tier — information agents, generative UI built with Antigravity, and a Universal Cart spanning Search, Gemini, YouTube and Gmail. A new $100/month AI Ultra tier targets developers.

Why it matters. The Search box overhaul — Google's largest in 25 years, accepting text, images, files, videos, and Chrome tabs — combined with generative UI assembled by Antigravity in real time, effectively retires the ten-blue-links paradigm for a billion-user surface. The Verge separately reports that AI Search results now carry "Sponsored Product" placements with Gemini-generated explainers, signaling Google is rebuilding the ads stack around AI Mode rather than around it. On developer tooling, Google is consolidating: the open-source Apache-licensed Gemini CLI will stop working with paid subscriptions on June 18, replaced by the closed-source Antigravity CLI, a move Simon Willison flagged as a meaningful posture shift.

Who is affected. Publishers and SEO-dependent businesses face a Search experience that increasingly answers without referrals. Developers gain a tightly integrated agent harness, AI Studio-to-Antigravity export, and Managed Agents via a single API call — but lose an open CLI option. Anthropic, OpenAI, and Cursor face direct competition from Spark and Antigravity 2.0 across personal-agent and coding-agent categories. Consumers in nearly 200 countries get Personal Intelligence connections to Gmail, Photos, and (soon) Calendar without a subscription.

What to watch next. Whether Gemini Spark's Agent Gateway, DLP enforcement, and ephemeral VM isolation hold up against prompt-injection attacks once it reaches Ultra subscribers next week; Gemini 3.5 Pro's June rollout; the summer launch of Universal Cart on YouTube and Gmail and how merchants integrate the Universal Commerce Protocol; and whether AI Mode's monetization preserves Search's margins as queries shift to longer, agentic sessions.

Sources:

3. OpenAI model disproves Erdős's planar unit distance conjecture

A mid-century mathematician writing equations on a chalkboard.

What happened. OpenAI announced that an internal general-purpose reasoning model disproved a central conjecture in discrete geometry: Paul Erdős's 1946 planar unit distance problem, which asks how many pairs of points in the plane can be exactly one unit apart. The prevailing belief since Erdős's original work was that rescaled square-grid constructions were essentially optimal. The model produced an infinite family of configurations achieving a polynomial improvement, using tools from algebraic number theory — including infinite class field towers and Golod–Shafarevich theory — to construct number fields with the required symmetries. A companion paper by external mathematicians, including Fields medalist Tim Gowers, verified the proof; Princeton's Will Sawin produced a refinement establishing an explicit exponent. Latent Space reports the model — speculated to be a GPT-5.6 variant — ran under 32 hours for under $1,000 and produced roughly 125 pages of chain-of-thought.

Why it matters. Gowers called the result "a milestone in AI mathematics," and number theorist Arul Shankar said current AI models are "capable of having original ingenious ideas, and then carrying them out to fruition." Crucially, the model was not a specialized math system like AlphaProof and was not scaffolded for the problem — the same architecture that handles general reasoning produced a publishable mathematical discovery. As TechCrunch notes, OpenAI's track record here is uneven: former VP Kevin Weil's October claim that GPT-5 had solved ten Erdős problems collapsed when it turned out the solutions already existed in the literature. This time, OpenAI shipped with endorsements from Noga Alon, Melanie Wood, and Thomas Bloom — the same mathematicians who previously criticized the earlier overreach.

Who is affected. Research mathematicians now have working evidence that frontier reasoning models can contribute novel constructions in active subfields, not just reproduce known results. For AI labs, the result strengthens the inference-time-scaling thesis underpinning current product roadmaps. For adjacent sciences — biology, physics, materials — OpenAI explicitly frames the same long-horizon reasoning capability as transferable, though that claim remains untested.

What to watch next. Whether algebraic number theorists revisit other open problems in discrete geometry along the bridge this proof opened, as Bloom predicts; whether the model — or its successor — is released publicly with comparable capability; how peer review handles the increasing volume of AI-assisted proofs; and whether competing labs (DeepMind, Anthropic) produce comparable autonomous research results outside Olympiad-style benchmarks.

Sources:

Today's three stories trace a single arc. The compute that Anthropic is buying from xAI is the same kind of compute that produced OpenAI's Erdős proof in under 32 hours, and the same kind that Google is now packaging into Antigravity-orchestrated agents on a billion-user Search box. The capital intensity revealed in SpaceX's S-1 — $15B in annual contracts, $2.8B in turbines, gigawatt-scale facilities — is the infrastructure layer; agentic Search and personal AI are the consumer monetization; and autonomous mathematical discovery is the upper bound on what that stack can now do unsupervised. The frontier is no longer about which model scores highest on a benchmark, but about who controls the data centers, the agent harnesses, and the long-horizon reasoning loops on top of them.

Digital Colliers Daily Briefing — May 21, 2026

1. SpaceX's S-1 lays bare a $15B-a-year Anthropic–xAI compute pact

2. Google I/O reorients Search, Gemini, and developer tooling around agents

3. OpenAI model disproves Erdős's planar unit distance conjecture

Digital Colliers Daily Briefing — June 12, 2026

Digital Colliers Daily Briefing — June 11, 2026

Digital Colliers Daily Briefing — June 10, 2026