AI Waves #8: SpaceX sold Anthropic 220,000 GPUs, three layers repriced in five days, and the agent stack consolidated at the hyperscalers

When the going gets tough, the tough sell GPUs to the competition

The Anthropic complex

Most of the news worth flagging this week is about the same company. In thirty days Anthropic has signed compute commitments across three silicon platforms, taken offers at a reported $900 billion valuation, launched a financial-services product line, formed a Wall Street consortium to deploy Claude inside private-equity portfolios, published the founding research agenda for The Anthropic Institute, and continued litigating the Pentagon’s supply-chain designation. Most of it landed in the last seven days. The composite is closer to a sovereign actor than a startup.

The instinct now is to extrapolate. The discipline is not to. What is true this week is that Anthropic has assembled a posture nobody else in the field can match. What will be true in August is not a question this issue tries to answer.

Compute from a rival

Anthropic and SpaceX announced on May 6 that Anthropic would use all of the compute capacity at Colossus 1, the Memphis data center originally built to train Grok, with more than 300 megawatts of capacity and roughly 220,000 Nvidia GPUs available within the month. Anthropic raised its five-hour rate caps for paid subscribers the same day.

The cluster is excess capacity, in xAI’s telling. The company has cut over to Colossus 2, a gigawatt-scale build that expanded to 1.5GW in April, where Grok 5 is now in training. Musk has called Grok 5 an AGI candidate. xAI has not yet shipped a model the AI community treats as a frontier peer to Claude Opus 4.7, GPT-5.5, or Gemini. Whether Colossus 1 was outgrown or surplus to a model line that did not arrive, the inference capacity is now Anthropic’s, and xAI is monetizing compute it once intended to train on. SpaceX, which acquired xAI in a $1.25 trillion all-stock deal in February, owns the Memphis site, and the deal is part of SpaceX’s pre-IPO neocloud recasting ahead of a reported mid-2026 listing. After years of calling Anthropic “Misanthropic,” Musk posted the same evening that in recent meetings with Anthropic’s senior team “no one set off my evil detector.”

The deal landed while Musk is in federal court in Oakland, two weeks into a trial against Sam Altman and Greg Brockman seeking $150 billion in damages. Mira Murati testified by video on May 6, the same day SpaceX and Anthropic put their compute deal into the public record. One Musk entity is suing OpenAI’s CEO. Another is selling 220,000 Nvidia GPUs of inference compute to OpenAI’s largest competitor.

The control plane lands

Microsoft Agent 365 reached general availability on May 1 at $15 per user per month, or $99 inside the new M365 E7 bundle. The product gives every AI agent an Entra identity and governs it through Defender, Purview, and Intune. At $15 per user, Microsoft has priced agent governance on the same per-seat schedule it uses for human user services.

ServiceNow’s answer arrived four days later. Action Fabric, announced May 5 at Knowledge 2026, opens ServiceNow’s “system of action” (flows, playbooks, approvals, catalogs) to any AI agent through a generally available MCP Server, with every action routed through the AI Control Tower for identity verification, permission scoping, and audit. Otto, ServiceNow’s unified agent layer combining Now Assist, Moveworks, and AI Experience, sits on top.

Microsoft bundles agent governance into Windows and the M365 stack and sells it per seat. ServiceNow’s bid is to sit underneath every vendor’s agent as the audit layer the enterprise actually trusts. The argument from #5, that agents are the new users and would need identity, billing, governance, and tooling rebuilt for them, keeps shipping infrastructure. Payments hardened in #7. Governance and identity hardened this week.

A third agent rail ships

AWS launched Amazon Bedrock AgentCore Payments in preview on May 7, integrating Coinbase’s x402 protocol and Stripe’s Privy wallet natively. Enterprise agents on Bedrock can authenticate, sign, and settle micropayments through a single API call, with USDC settling on Base and Solana in roughly 200 milliseconds, time-bound budget controls (the example: “$1.00, expires in 5 minutes”), CDP Facilitator handling sanctions and illicit-finance compliance, and full audit logs. Agents transact directly with thousands of x402 services exposed through AgentCore Gateway, including Exa, Messari, and Browserbase.

This is the third agent-native payment rail in five weeks. Stripe’s Machine Payments Protocol on Tempo and Mastercard’s Verifiable Intent on AP2 landed in #7. AWS now ships the same primitive on the largest cloud, with crypto-native settlement. The x402 Foundation has processed 169 million payments across 590,000 buyers and 100,000 sellers in its first year. The agent payment layer is now distributed by the largest payments network, the largest card network, and the largest cloud.

Pricing the layers

Sierra, Bret Taylor’s enterprise customer-facing agent platform, raised $950 million on May 4 at $15.8 billion, led by Tiger Global and Google’s GV, with Benchmark, Sequoia, and Greenoaks participating. Up from $10 billion last fall, on $150 million of run-rate revenue.

Cerebras filed its amended S-1 the same day, pricing its IPO at $115 to $125 a share. Roughly $26.6 billion at the top of the band on $510 million of trailing revenue. UAE-linked customers, principally G42, account for the majority of that revenue. The float is the first AI compute pure-play to test the public market since GTC 2026.

Application cleared at $15.8 billion private. Silicon cleared at roughly $26.6 billion public. Model cleared at a reported $900 billion private. All inside the same five-day window.

What We’re Watching

OpenAI ships memory sources. GPT-5.5 Instant rolled out as ChatGPT’s new default on May 5. The notable feature is “memory sources,” a view of which saved memories or past chats shaped any given response, with delete and correct controls. First time a frontier lab has made AI context selection auditable at the consumer surface. Whether the view shows everything the model actually relied on is a separate question.

MCP, still. The MCP design-flaw RCE we flagged in #6 (200,000 exposed servers, Anthropic’s “expected behavior” response) remains the canonical example of agent plumbing outpacing security. Provably’s AgentKit, below, is one of the first production answers.

Portfolio

Provably: the verification SDK ships. Provably released the Verifiable Data AgentKit, an open-source Python SDK for detecting hallucinations, verifying agent answers, and stopping bad data from spreading across workflows. Released on May 7, it works with major agent frameworks including OpenAI, Anthropic, LangChain, LangGraph, LlamaIndex, AutoGen, CrewAI, and Google GenAI. The SDK stops agents from fetching data from unapproved sources, records every response they do fetch, and cryptographically verifies that what one agent claims it found matches what the source actually returned when work is handed to another agent. That removes the need to re-run source queries just to check an answer, saving time and speeding up detection and healing in both single-agent and multi-agent systems.

Prime Intellect: Lab opens to all. Prime Intellect’s Lab moved from beta to general availability on May 7. The platform consolidates the full self-improvement loop (environments, evaluations, RL training, adapter deployments, inference, sandboxes) into one managed surface, multi-tenant across training and inference so customers pay per token actually moved through the model rather than per cluster-hour reserved. GA lands with 14 base models from Nvidia, OpenAI, Meta, and Qwen between 1B and 70B parameters. The beta drew hundreds of researchers running 10,000+ training jobs across math, code, browser tasks, games, customer support, and enterprise workflows. Issue #6 framed self-improvement as the closed labs’ defensive fight. Lab is the open-side equivalent: any team can now run that loop on hosted compute, per-token billed. The latest brick in what Prime Intellect calls the Open Superintelligence Stack.

Dimensional: SpatialMemory2 ships open-source robot memory. Dimensional released SpatialMemory2 on April 29, an open-source multimodal memory system for robots. Production robots generate thousands of hours of camera, lidar, and odometry data per deployment, far beyond any agent context window. SpatialMemory2 stores every frame as a spatial vector embedding with map metadata, then exposes algorithmic and cosine-similarity search at runtime so agents can query weeks or months of sensor data on demand (every place a robot saw a plant, every door it could not open, every transient obstruction across a month). Composable streams render spatially on the map. Robotics is hitting the context-window wall language agents hit two years ago. Dimensional shipped one of the first open-source answers.

Intelligent Internet: Factory ships agentic creative production. II launched Factory on May 6, an infinite-canvas creative environment where one sentence becomes a complete multi-medium production. Agent Director plans the project, picks models from a roster of nine frontier image and video options (NanoBanana Pro, Seedream 4.5, FLUX 2 Max, Seedance 2.0, Kling O3 Pro, Veo 3.1, others), generates assets, and wires them into an editable node graph. Two modes (Supervisor for step-by-step approval, Automation for end-to-end first draft) trade speed for control. The category claim: no single model wins every job, so the value sits in orchestration. Issue #6 made that argument at the evaluation layer with LayerLens. Factory makes it on the creative surface.

Closing

This is the week Anthropic ran a playbook that did not exist last year, and the layers underneath repriced in the same window. The discipline opened earlier still holds: what is true this week is not what will be true in August. The labs run their plays in months. The infrastructure builders run theirs in weeks. The distance between those clocks is the trade.