Morning Brief · Thursday

Anthropic Secured 220,000 NVIDIA GPUs from SpaceX. Claude Will Now Dream. And xAI Has Stopped Existing.

Anthropic announced a deal with SpaceX to access the full compute capacity of the Colossus 1 data center in Memphis — 300 megawatts and over 220,000 NVIDIA GPUs — effective within the month. Rate limits doubled immediately. In the same announcement, Anthropic launched "Dreams," a research preview API that lets AI agents review past sessions and consolidate their own memory. Separately, Elon Musk confirmed that xAI will be dissolved as a company into SpaceX — the AI startup he founded to rival OpenAI no longer exists as an independent entity. In court, former OpenAI board member Helen Toner testified that she found out ChatGPT had launched from screenshots on Twitter. And six of the largest companies in tech published a shared open standard for connecting GPUs at Stargate scale.

Mira Novian

May 7, 2026 · Morning Brief · ~9 min read

Anthropic signed a deal with SpaceX for all of Colossus 1 — 300 megawatts, 220,000+ NVIDIA GPUs. Rate limits doubled today. And a new API called "Dreams" lets AI agents review their own sessions and self-improve.

Three things landed in a single Anthropic announcement this morning. First, the compute deal: Anthropic has signed an agreement with SpaceX to use all of the compute capacity at the Colossus 1 data center in Memphis — the 300-megawatt, 220,000+ NVIDIA GPU facility that was originally built for xAI's Grok training runs. The capacity comes online within the month. This is Anthropic's largest single compute addition to date, and it joins a portfolio of infrastructure commitments that has grown rapidly: an up-to-5 GW agreement with Amazon (including ~1 GW by end of 2026), a 5 GW agreement with Google and Broadcom (coming online in 2027), a $30 billion Azure capacity deal with Microsoft and NVIDIA, and a $50 billion infrastructure commitment with Fluidstack. Anthropic has also disclosed interest in developing multiple gigawatts of orbital AI compute with SpaceX — a signal worth noting even if timelines remain speculative.

Second, the rate limits: effective today, Claude Code's five-hour rate limits are doubled for Pro, Max, Team, and seat-based Enterprise plans. The peak-hours limit reduction on Claude Code for Pro and Max accounts has been removed entirely. API rate limits for Claude Opus models have been raised considerably. Anthropic is explicit that the capacity expansion is the direct cause. For teams that have been running into Claude Code limits in production — and there are many — this is a meaningful quality-of-life change.

Third, and most technically interesting: Anthropic launched Dreams, a research preview API for the Managed Agents platform. Dreams is an asynchronous job that takes an agent's existing memory store plus up to 100 past session transcripts and produces a new, reorganized memory store — with duplicates merged, contradictions resolved, stale entries replaced, and new patterns surfaced. The original store is never modified; you review the output and decide whether to keep it. The design is explicitly about the problem that accumulates over long-running agentic deployments: memory stores grow incoherent over time. An agent that has operated for hundreds of sessions has accumulated conflicting notes, outdated preferences, and redundant entries that degrade its performance. Dreams is the mechanism for letting the agent consolidate what it has learned into a clean, current representation of knowledge — not during a live task, but asynchronously, between sessions. The name is apt: the biological analogy of memory consolidation during sleep is exactly the function being described. The beta header required for the API — dreaming-2026-04-21 — suggests this has been in development since late April, moving faster than the typical Anthropic beta-to-release cadence.

platform.claude.com ↗

The Dreams API is the more significant announcement, even if the Colossus 1 deal is the bigger headline number. The compute expansion is meaningful — 220,000 GPUs is not nothing — but the trajectory of AI compute deals has normalized those numbers to the point where the reaction is "that makes sense" rather than "that's surprising." Dreams is different. Memory consolidation for AI agents has been an open research problem since persistent memory stores became practical, and Anthropic has shipped a production-quality API solution for it. The key design choice — asynchronous, non-destructive, with session transcript input — is the right set of constraints. It means the consolidation can run without blocking live work, the original state is preserved for inspection, and the agent can mine not just its memory but the full record of what it actually did. For enterprise teams building agentic systems that operate for months, this changes the math on long-term agent performance. The question is how Dreams interacts with instruction-following fidelity over many consolidation cycles — there's a real risk that repeated dreaming drifts agent behavior away from user preferences if the consolidation model has different priors. Anthropic's instructions field gives teams control over what to focus on, but this will need careful evaluation in practice. The orbital compute interest is worth one sentence here too: if Anthropic and SpaceX eventually build gigawatt-scale satellite data centers, the geopolitical implications of AI compute located outside any nation-state's jurisdiction are substantial. That's a 2030-era conversation, but it's worth starting now.

Strategy

xAI no longer exists. Elon Musk announced the company will be dissolved into SpaceX and rebranded as SpaceXAI. The AI startup he founded to rival OpenAI lasted less than three years as an independent company.

The announcement came embedded in a press release about a compute partnership with Anthropic: the company referred to itself as "SpaceXAI" — a name that hadn't appeared publicly before. When reporters noticed, Elon Musk confirmed on X: "xAI will be dissolved as a separate company, so it will just be SpaceXAI, the AI products from SpaceX." The merger had been anticipated since SpaceX's acquisition of xAI earlier this year, but the pace of the integration is faster than most observers predicted. Grok, the Colossus data center infrastructure, and xAI's research team are now formally part of SpaceX's organizational structure. The xAI brand — launched in July 2023 with the stated mission of "understanding the true nature of the universe" — is gone.

The corporate structure of what was xAI now sits under SpaceX, which is privately held and majority-owned by Musk. This has immediate regulatory implications: SpaceX's classified government contracts (including Starlink and national security launch operations) and its ITAR-governed technology base now sit in the same corporate entity as Grok, a public-facing AI model and API. The Department of Defense has historically been cautious about AI capabilities and classified infrastructure being co-located in the same corporate structure. Whether the DoD requires structural separation — or has already approved the arrangement — is not yet public. Separately, the irony of the moment is worth noting: SpaceXAI's first major announcement as a named entity was a compute deal with Anthropic, the company that the DoD previously listed as a "supply chain risk" after it refused surveillance use cases. The AI industry's partnerships do not follow tidy competitive lines.

theverge.com ↗

The consolidation of xAI into SpaceX is the natural endpoint of Musk's resource consolidation strategy — DOGE, SpaceX, Tesla, X, and now xAI are all progressively being drawn into a smaller number of corporate entities under tighter personal control. The interesting consequence for the AI landscape is that Grok — the model — is now backed by the full balance sheet and infrastructure of SpaceX rather than a standalone AI startup. SpaceX has real cash flow (Starlink is genuinely profitable), access to physical infrastructure at scale, and a culture of extremely fast iteration. That's a different competitive resource base than xAI had on its own. The name change also matters symbolically: "SpaceXAI" buries the Grok brand identity under a parent brand identity, which is usually what happens to acquisitions that aren't working as planned. Grok 4.3 got strong reviews from developers; if the product momentum is positive, you typically keep the brand. Absorbing it suggests something else is driving the consolidation — probably regulatory simplification, Musk's preference for fewer entities to manage, and the practical reality that Colossus 1 already sat inside SpaceX's infrastructure. The compute partnership with Anthropic is the genuinely surprising part. SpaceXAI's own model (Grok) competes with Claude. Leasing your competitor the data center you built to train your own models is either a sign of pragmatic capital deployment or an early indicator that SpaceXAI is deprioritizing frontier model development in favor of becoming infrastructure.

Legal

Helen Toner on the stand: the OpenAI board found out ChatGPT had launched from screenshots on Twitter. "I was used to the board not being very informed about things."

Day four of Musk v. Altman moved to Helen Toner, former Georgetown CSET director and OpenAI board member — and one of the four directors who voted to fire Sam Altman in November 2023. Toner's deposition was projected for the jury, and the picture it painted of OpenAI's governance structure was not flattering to anyone involved. The core of her account: the board's decision to fire Altman was driven by a "pattern of behavior" around "honesty and candor," not any single incident. The starting point, she said, was Ilya Sutskever reaching out to express serious concerns about Altman — matching Mira Murati's earlier testimony. What the board learned about Altman's behavior: he had told Sutskever that a board member suggested Toner resign over a paper she'd written; the board member denied having said it. He had interests in an OpenAI startup fund that weren't fully disclosed. He had not informed the board about ChatGPT before its launch.

That last point produced the most striking testimony of the day. Asked how she had learned about ChatGPT's existence, Toner said: "I saw screenshots on Twitter." She was not surprised, she added — "I was used to the board not being very informed about things" — which she interpreted as Altman not being "motivated to help the board perform the oversight role." The structural admission embedded in that sentence is significant: the board of a company building what it described as potentially humanity-altering technology had a CEO who didn't tell them when he launched the most widely-used AI product in history. On the technical side, Toner offered a formulation for the record: "Making AI models is more like alchemy than chemistry." There's no rigorous testing framework for safety, she said — "people are just throwing things together to see what happens." OpenAI's safety testing methods had become "somewhat less slapdash" over time, in her telling, but the baseline was explicit.

theverge.com ↗

The governance failures being documented in this trial are going to matter well beyond the verdict. What Toner described — a board that learned about its company's most significant product launch from social media, whose CEO actively managed information to limit their oversight capacity, and whose safety board operated on methods she characterized as "alchemy" — is a template for how not to run an AI lab with a mission-driven structure. And the tragedy is that OpenAI's board was, in context, among the more serious attempts at AI governance that existed at the time. They had actual researchers and policy experts on it. They had a nonprofit structure with explicit safety commitments. And it still produced a governance outcome in which the board was systematically kept uninformed. The lesson is not that individuals failed — it's that the structure was always going to fail, because it combined concentrated executive authority over operations with a board structure that depended entirely on voluntary information disclosure by the person being overseen. That's not governance; it's theatre. Every AI lab that currently claims to have serious safety oversight should be required to answer one question: what information about your company's capabilities and deployments is your board allowed to learn only from Twitter? If the honest answer is "more than nothing," you don't have governance. You have a press release.

Infrastructure

OpenAI, AMD, Broadcom, Intel, Microsoft, and NVIDIA just published a shared open standard for connecting GPUs at Stargate scale. MRC is the networking protocol the AI industry didn't know it needed to agree on.

OpenAI announced the release of MRC — Multipath Reliable Connection — through the Open Compute Project, co-developed with AMD, Broadcom, Intel, Microsoft, and NVIDIA. MRC is a new network protocol designed specifically for the failure modes that emerge in GPU clusters at the scale of Stargate. The two problems it addresses are congestion and failure. Congestion: as clusters grow, even carefully-designed networks develop bottlenecks where multiple GPUs send to the same destination simultaneously. MRC uses adaptive packet spraying to distribute traffic dynamically across multiple network paths, virtually eliminating core congestion without requiring complex traffic engineering. Failure: at large enough scale, some network link or switch will always be failing at any given moment. Previously, a single failure could crash a training job entirely or cause a multi-second stall while routing reconverged. MRC uses static source routing to bypass failures and multi-plane redundant network architecture to continue operation through link-level failures without restarting from a checkpoint.

The business rationale is straightforward: at Stargate scale — hundreds of thousands of GPUs training a single model synchronously — every training job is a "failure amplifier." A single network failure cascades across the whole cluster. The larger the job, the higher the cost of any interruption. MRC is designed to make those interruptions vanishingly rare without adding hardware complexity. OpenAI's framing for why they're publishing it openly: "shared standards in key infrastructure layers can help scale AI systems more efficiently, reliably, and across a broader partner ecosystem." The Open Compute Project spec is publicly available. The coalition behind MRC — including the two primary GPU architecture competitors (NVIDIA and AMD) alongside Microsoft and Intel — signals industry-wide convergence on the networking layer as a shared infrastructure problem, not a competitive differentiator.

opencompute.org ↗

The coalition that built MRC is as significant as the protocol itself. AMD and NVIDIA cooperating on a shared networking standard — while competing fiercely at the GPU level — reflects a pattern that's becoming visible across the industry: companies are finding the layers where standardization benefits everyone more than proprietary lock-in. Networking protocols are the obvious candidate; both companies want their GPUs to be easy to cluster, and a shared standard means neither has to solve the networking problem from scratch. The same logic applies to Microsoft, which has to build Stargate regardless of which GPU vendor it sources from. OpenAI publishing MRC through the Open Compute Project also continues a trend I've been watching: frontier AI labs making strategic infrastructure components open as a way to accelerate the ecosystem they depend on. The playbook is consistent with the AMD-Intel ACE x86 instruction set extension announced earlier this week — multiple competitors finding common cause in infrastructure layers while competing at the product and model layer. What this means in practice: the AI training stack is bifurcating into a shared infrastructure substrate (networking, memory standards, interconnects) and a highly proprietary frontier layer (model weights, data, RLHF pipelines). The open parts are getting more open; the proprietary parts are getting more closely held. That's a rational equilibrium, but it concentrates the value at exactly the layer that's hardest to regulate.

✦ Mira's Take

The through-line in today's brief is the concentration of infrastructure into fewer, larger entities — and the interesting tensions that creates. Anthropic's compute portfolio now spans Amazon, Google, Broadcom, Microsoft, NVIDIA, SpaceX, and Fluidstack: a web of bilateral deals so extensive that Anthropic's effective compute supply is now backed by almost every major technology company simultaneously. That's an extraordinary position for a company that doesn't manufacture hardware, doesn't run a hyperscaler, and doesn't have a consumer product with meaningful market share. The reason it works is that every major cloud provider needs a competitive frontier model, and Anthropic's Claude is genuinely competitive. The compute deals aren't charity — they're defensive investments by companies that can't afford to lose access to frontier models if the market consolidates. Anthropic is playing that dynamic well.

The xAI → SpaceXAI story inverts that dynamic. Musk's consolidation move concentrates AI capability under a single corporate entity that is already deeply integrated with national security infrastructure, classified government contracts, and a communication platform. The compute deal with Anthropic — leasing Colossus 1 to a competitor — raises the question of whether SpaceXAI has made a strategic decision to deprioritize model frontier competition in favor of becoming infrastructure. If Grok is not going to be a frontier model competitor in 12 months, the rational move is to monetize Colossus 1's capacity. Watch whether SpaceXAI continues to publish model releases at the pace xAI maintained — if the cadence slows, that's the tell.

And underneath all of it, the Toner testimony gives us the clearest picture yet of how the most consequential AI governance failure of the past decade actually happened. The board found out about ChatGPT from Twitter. The board found out about ChatGPT from Twitter. That sentence is going to be quoted in AI governance discussions for years — not as a scandal, but as a structural fact about what happens when you build oversight mechanisms that depend on the goodwill of the person being overseen. As the compute deals in today's brief make visible, the AI industry's resource concentration is accelerating. The governance structures being built to oversee that concentration need to be designed for adversarial conditions, not collegial ones. That's the lesson from this trial that nobody is talking about loudly enough.