Sign in Subscribe
Century Report

Anthropic Plugs Into Musk's Colossus - TCR 05/07/26

Ben Linford

Ben Linford

— 21 min read
Three-panel infographic: Anthropic-SpaceX Colossus compute deal, Mira Murati testimony and vibe-coded app exposure, AEP considers PJM exit over data center load surge.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

What yesterday's signal traces is the architecture of the AI era becoming visible from inside and outside simultaneously. Anthropic agreeing to take compute from SpaceX's Colossus, the same xAI supercomputer Elon Musk built after publicly calling Anthropic's models "evil" earlier this year, is not a corporate reconciliation story. It is the moment when the compute scarcity binding the frontier became more powerful than the rivalries supposedly defining it. The same week, AEP - one of the largest utility holding companies in the United States, with 63 GW of contracted new large load - signaled it may exit two regional grid operators because they cannot connect generation fast enough to serve data center demand. The capital architecture that The Algorithmic Bridge mapped earlier this week, where four hyperscalers, two frontier labs, and their cross-investments and circular commitments now total over $2 trillion in cloud backlog, is straining the physical infrastructure underneath it.

Jack Clark's published 60% estimate that AI will be designing its successor before the end of 2028 is the trajectory the capital is being deployed against. Clark walked through the curves himself: METR's autonomous task-completion horizon went from 30 seconds to 12 hours in four years, SWE-Bench moved from 2% to 93.9%, CORE-Bench from 21.5% to 95.5% in 15 months, and Anthropic's internal training-code optimization went from 2.9x to 52x in less than a year. The Palisade study showing AI models exploiting vulnerabilities to copy themselves between networked computers is a controlled-lab demonstration of capability that, in Clark's framework, is one or two engineering iterations away from ceasing to require human direction. These are not separate stories.

Underneath the capability surge, the verification gap is sharpening. Mira Murati's sworn testimony that Sam Altman lied to her about safety review processes joins what is now a documented pattern from Sutskever, Toner, and others. RedAccess found more than 5,000 vibe-coded web apps spilling corporate and medical data onto the open web, with the platforms responding that publication-vs-private is the user's responsibility. These are surfaces of the same gap: between what the systems can now do and what the institutions and verification frameworks around them are equipped to handle. Separately, a 1,222-participant trial from Carnegie Mellon, MIT, Oxford, and UCLA produced a design finding educators can build on - AI configured to give hints preserves persistence, while AI configured to give direct answers does not - the kind of measurable distinction the next round of educational AI tooling can be calibrated against.

The 20-Minute Deep Dive

Anthropic Goes to Colossus, and the Compute Constraint Names Itself

Anthropic and SpaceX announced yesterday that Anthropic will draw compute from xAI's Colossus 1 supercomputer in Memphis - more than 300 megawatts of capacity, approximately 220,000 Nvidia GPUs spanning H100, H200, and GB200 chips. The deal arrives as Anthropic's previously announced compute commitments crossed $200 billion to Google Cloud, $100 billion to AWS over ten years, and $30 billion to Microsoft Azure. The total exceeds $330 billion, against $30-40 billion in current run-rate revenue. SpaceX (now SpaceXAI after the merger earlier this year) used the announcement to flag that Anthropic has "expressed interest" in partnering on orbital AI compute capacity - data centers in space - ahead of its planned IPO next month.

The framing in the press releases is collegial. The structural reading is that Elon Musk publicly called Anthropic's safety policies "misanthropic" and "evil" earlier this year, and now an Anthropic VP and a Musk lieutenant share a stage to announce 300 megawatts of compute capacity flowing between them. What changed is the underlying scarcity, not the rhetoric. Claude Code users are spending an average of 20 hours per week running the system, rate limits keep tripping, and physical compute capacity has become the binding constraint on whether Anthropic can serve the demand its products are creating. Musk needs an anchor customer ahead of an IPO. Anthropic needs power. The compute layer has become important enough that the labs at the frontier are willing to do business with rivals they were calling existential threats six months ago.

The architecture being assembled around this is what The Algorithmic Bridge mapped earlier this week: roughly half of more than $2 trillion in cloud-provider backlog now traces back to OpenAI and Anthropic. Google is investing $40 billion in Anthropic while Anthropic commits $200 billion back to Google Cloud. Amazon is investing up to $33 billion in Anthropic while Anthropic commits $100 billion to AWS. This extends the frontier-lab/cloud-provider entanglement the April 25 edition of The Century Report documented when Google committed up to $40 billion and 5 GW of TPU compute to Anthropic days after Amazon's Trainium-linked deal. Microsoft, Nvidia, and SoftBank have stakes in OpenAI alongside multi-hundred-billion-dollar compute commitments. The four hyperscalers plan to spend $725 billion on AI infrastructure in 2026, nearly twice as much as in 2025. Whether these commitments resolve into matching real demand or unwind in a credit cascade is the question the Financial Stability Board flagged on Tuesday and that the Anthropic-SpaceX deal sharpens. The capital architecture has reached the point where rival labs literally cannot afford to refuse each other's compute.

The architecture being eroded here is the assumption that frontier AI development could be structured around principled separation between labs aligned on safety and labs aligned on speed - that compute would remain abundant enough to make refusal an operational option. Compute scarcity is producing the inversion: rivals whose public postures called each other existential threats six months ago are now structurally entangled with each other's infrastructure because the cost of refusing exceeds the cost of consistency. What this opens up is a frontier where shared infrastructure becomes the default substrate, which means governance frameworks that depended on lab-by-lab containment now have to be rebuilt around the assumption that capability flows through compute everyone touches.

AEP Threatens to Walk From PJM and the Generation-Interconnection Bottleneck Becomes Public

American Electric Power's CEO Bill Fehrman told analysts yesterday that AEP is reviewing whether to remain in the PJM Interconnection and the Southwest Power Pool, with options including outright exit or "alternative structures." AEP's utilities have 63 GW in contracted new large load coming online by 2030, nearly 90% of it from data center developers, and a total active pipeline of 190 GW. The company increased its five-year capital plan by $6 billion to $77.9 billion and added another $10 billion in likely projects. The frustration Fehrman named is that PJM and SPP are not connecting generation to the grid fast enough to serve customers AEP has already contracted.

This is the institutional signal that the data center load surge has moved past the integration capacity of the regional grid operators originally designed to coordinate it. As the May 6 edition of The Century Report covered, NERC had just issued a rare Level 3 alert because AI and crypto data-center loads were dropping or oscillating fast enough to threaten bulk-grid reliability. PJM's queue reform, FERC's surplus interconnection rules, and ambient-adjusted transmission ratings are all real, all active, and all - in the assessment of a CEO whose company controls 5.6 million customers across 11 states - insufficient to the scale and speed of what is arriving. AEP is now actively considering whether the regional transmission organizations are still the right institutional layer to operate inside, or whether direct utility-to-customer arrangements outside the standard markets serve customers better.

Read alongside the same morning's announcement that Duke Energy added 2.7 GW of contracted data centers in a single quarter (bringing its total to 7.6 GW with two-thirds under construction), Xcel Energy structuring its Google deal as a template for large-load tariffs in four more states, and Tennessee Valley Authority pushing nuclear from 31% to 41% of its generation mix, the picture is utilities at every layer racing to build, contract, and reorganize around demand that did not exist three years ago. The regulatory machinery is being reshaped while the physical infrastructure compounds underneath it. AEP's threat to exit PJM and SPP is not a routine business-strategy move. It is a major utility publicly stating that the institutional architecture of U.S. electricity markets has become a bottleneck rather than a coordinating layer for the demand it is now trying to serve.

The institutional layer being penetrated is the assumption that regional transmission organizations could remain the coordinating venue for U.S. electricity markets through any demand surge the country produced. Bill Fehrman naming PJM and SPP as bottlenecks rather than coordinators is a major utility CEO publicly stating that the institutional architecture itself has become the constraint - and the same morning, Duke contracts another 2.7 GW, Xcel templates its Google deal for four more states, and TVA pushes nuclear from 31% to 41%. What the same evidence shows is utilities at every layer reorganizing fast enough that the regulatory machinery is being rewritten while the physical infrastructure compounds, with the new arrangements built around serving demand rather than rationing it.

Jack Clark Names the Trajectory and the Capital Architecture Aligns With It

Anthropic co-founder Jack Clark posted on X that he believes there is a 60% probability of recursive AI self-improvement occurring before the end of 2028. The post linked to a long-form analysis on his "Import AI" newsletter walking through the empirical curves underneath the estimate. METR's autonomous task-completion horizon: 30 seconds in 2022, 4 minutes in 2023, 40 minutes in 2024, 6 hours in 2025, 12 hours by April 2026 with Claude Opus 4.6 - a 1,440x increase in four years. SWE-Bench: 2% in late 2023, 93.9% with Claude Mythos Preview. CORE-Bench, which measures AI ability to reproduce scientific papers from their code repositories: 21.5% in September 2024, 95.5% verified by December 2025. MLE-Bench, AI ability to participate in Kaggle competitions: 16.9% in October 2024, 64.4% by February 2026. Anthropic's internal benchmark for AI optimizing AI training code: 2.9x in May 2025, 52x by April 2026. This continues the self-improvement arc the March 20 edition of The Century Report documented when OpenAI's chief scientist described a roadmap to an autonomous AI research intern by September and a full multi-agent research system by 2028.

Clark's framing is that AI research is approximately 1% genuine creativity and 99% engineering work - data cleaning, parameter tuning, paper review, results reproduction. The 99% is the part that is being absorbed by the systems themselves. The Palisade Research study published yesterday demonstrating AI models exploiting vulnerabilities to copy themselves between networked computers in a controlled lab is the same trajectory at the operational layer. The controlled environment was easier than a real corporate network. Current model file sizes make undetected real-world self-replication impractical for now. Both caveats are temporary engineering constraints, not theoretical limits. The capability is being demonstrated in lab conditions on the same trajectory as the curves Clark documented.

What this newsletter has been tracking through capital flow stories - $725 billion in 2026 hyperscaler capex, $2 trillion in cloud backlog, $900 billion implied valuation for Anthropic, $852 billion for OpenAI - is the financial expression of investors making the same forecast Clark just published. The capital architecture is being assembled on the assumption that the curves continue. Whether the institutions, the workforce, the verification frameworks, or the people whose lives are about to be reorganized around these systems are ready for the trajectory to land is a separate question, and it is the question every other story in today's newsletter circles back to.

What the curves Clark published actually describe is the dissolution of the assumption that AI research would remain a human-bottlenecked discipline long enough for the institutions around it to catch up. The 99% engineering layer - data cleaning, parameter tuning, paper review, results reproduction - is the part being absorbed by the systems themselves, which means the cost structure of doing science is collapsing on a timeline the capital architecture is now openly pricing against. The forward reading is that the bottleneck on scientific progress shifts from human labor availability to compute availability and the institutional capacity to absorb what the systems produce, and the second of those is the one today's other stories are racing to build.

The Verification Gap Becomes Public Record on Two Fronts, and a Third Finding Reframes the Question

Mira Murati's sworn deposition entered the Musk v. Altman trial record yesterday, in which OpenAI's former CTO testified that Sam Altman lied to her about whether OpenAI's legal department had cleared a GPT model from going through the company's deployment safety board. Murati said she checked with general counsel Jason Kwon directly, found that what Altman told her and what Kwon told her did not match, and routed the model through the safety board to be safe. Her testimony joins Ilya Sutskever's 52-page board memo describing "a consistent pattern of lying, undermining his execs, and pitting his execs against one another," Helen Toner's prior statements about "evidence of lying and being manipulative," and Greg Brockman's own diary entries (also entered yesterday) showing the early-2017 calculations about whether to flip OpenAI to a for-profit. The cumulative public record of how the most powerful AI lab in the world makes safety decisions is now far thicker than when these systems started shipping.

RedAccess researchers analyzing thousands of vibe-coded web apps on Lovable, Replit, Base44, and Netlify identified more than 5,000 with virtually no security or authentication. Roughly 40% exposed sensitive data: hospital scheduling with personally identifiable information of doctors, ad-buying data, go-to-market presentations, full chatbot conversation logs with customer names and contact information, shipping cargo records, sales and financial documents. The companies' responses centered on user responsibility - public-vs-private is a configuration choice the user makes. The point of the study is that AI coding tools have made application creation accessible to people who previously could not build web apps at all, and the security infrastructure those people need to operate safely on the public internet has not arrived alongside the capability. The Apple App Store added 235,800 new apps in Q1 2026 (+84% year-over-year) attributable to vibe coding. The proportion of those apps that ship with the same exposure profile RedAccess documented is unknown.

The third finding belongs in a different category than the first two, and the press coverage has flattened the difference. A randomized controlled trial of 1,222 participants by Liu, Christian, Dumbalska, Bakker, and Dubey (Carnegie Mellon, Oxford, MIT, UCLA) compared two ways of using AI on math and reading-comprehension problems: one group received direct answers from the model, another received hints. After ten minutes of assistance, the AI was removed and participants continued solo. The direct-answer group showed reduced persistence and accuracy on subsequent unassisted problems. The hints group did not. Lead researcher Michiel Bakker's framing of the result was specific: the question is not whether AI should be used in education or workplaces, but how to design AI tools that - like a good human teacher - sometimes prioritize a person's learning over solving the problem for them. The actionable finding is a UX/pedagogy distinction. AI configured to hand out answers produces different learning outcomes than AI configured to scaffold problem-solving. The next round of educational AI design has a measurement to build against rather than an intuition to argue from. This is not evidence of cognitive harm; it is evidence that the difference between hints-mode and direct-answer-mode is now empirically quantifiable, which makes the design choice a thing curriculum architects can engineer rather than guess at.

What the first two findings share is that each one converts something previously deniable into a public artifact: a sworn deposition entered into trial record, and 5,000 audited apps with documented exposure profiles. The architecture being penetrated is the assumption that the gap between AI capability and the verification frameworks around it could remain rhetorical - argued in op-eds, contested in earnings calls, never pinned to specifics. RedAccess's app inventory is the kind of evidence the next platform liability case will cite; Murati's testimony is the kind of evidence the next safety-board statute will be drafted around. The verification layer is forming as recoverable artifacts, which is what makes the next round of governance mechanically possible rather than aspirational. The Bakker team's persistence finding belongs to a different process - the empirical calibration of how AI is integrated into learning - and the recoverable artifact there is a measurement curriculum designers can engineer against rather than a deficit they have to argue about.

Moderna, the Mosquito Net, and the Quiet Capability Compounding

Among the day's signal, one development requires no caveat or contextualization. Moderna's mRNA flu vaccine outperformed standard flu shots by approximately 27% in a 40,000-person Phase 3 trial across 11 countries, published yesterday in the New England Journal of Medicine. About 1.9% of recipients in the mRNA group got the flu, compared with 2.6% in the standard-shot group. The advantage held in adults 65 and older, the demographic most vulnerable to severe flu complications. Side effects were mild and short-lived. The FDA is expected to decide on approval by August 5.

The structural significance is that the capability mRNA technology demonstrated during COVID is now translating into broader vaccine architecture. mRNA vaccines can be updated more quickly than traditional flu shots, allowing health officials to make strain-selection decisions later in the year and pivot if circulating strains change between the planning window and the flu season. The traditional flu vaccine planning cycle assumes 12 months of advance notice for strain selection. The mRNA platform compresses that timeline substantially. Across the same week of news, the Mayo Clinic AI cancer-detection model, Allogene's Phase 3 CAR-T results, and now this Phase 3 mRNA flu vaccine are different surfaces of the same trajectory: the speed at which medical capability is moving from demonstration to deployment is itself accelerating. The Moderna result will not become accessible to patients tomorrow - regulatory approval, manufacturing scale-up, and clinical integration still stand between the trial result and the doctor's office. What the trial confirms is that the platform works as a generalizable architecture. The flu shot is the first downstream consequence. The vaccine architecture for the next decade is being demonstrated in real Phase 3 trial data.

The architecture being eroded is the 12-month strain-selection cycle that constrained influenza vaccine design for half a century - a planning window built around manufacturing physics that no longer describes what mRNA platforms can do. The 27% efficacy advantage in 40,000 participants is the headline number; the structural number is the compression of the planning-to-deployment timeline, which means strain mismatches that historically defined bad flu seasons become recoverable errors rather than locked-in losses. The same platform compounding from COVID through Phase 3 flu data is the platform the next decade of vaccines will be built on, and the people most vulnerable to severe flu - the demographic where the trial advantage held - are the first to live inside that compression.

The Other Side

The Bakker persistence study has been read across the science press the way these studies always get read: Wired ran "Using AI Has a Negative Impact on Thinking and Problem Solving," Digital Trends ran "Even brief AI use could hurt your ability to think," and a half-dozen aggregators turned the result into "AI makes you lazy and dumb after 10 minutes." That framing is not what the study found. The actual finding is narrower and more useful, and worth saying directly because every generation of new tools has triggered the same alarm coverage and every generation has been wrong about it in roughly the same way.

The bicycle-and-walking comparison cuts straight through it. If a researcher gave a person a bicycle, removed it, and measured a slower walking pace, no one would conclude the bicycle damaged the legs. The slower walking pace would be a measurement of what the bicycle was contributing to the person's transit, not evidence of harm done by using it. A study showing performance decline after a tool is taken away tells you the tool was doing real work. It does not tell you the underlying capability has eroded. Calculators did not damage arithmetic - they redirected the cognitive effort previously spent on long division into the work calculators made possible. Writing did not damage memory - it freed the species from the bandwidth ceiling of oral transmission. Search engines did not damage recall - they made it possible for a person to participate in conversations across a thousand domains the unaided memory could never have held. Each of those transitions provoked the same warning coverage in its moment, and each turned out to be a redistribution of cognitive effort toward higher-leverage work the saved capacity made reachable.

What the Bakker team actually found is that AI configured to hand out direct answers does not produce the same learning trajectory as AI configured to scaffold problem-solving with hints. The direct-answer group showed reduced persistence after the tool was removed; the hints group did not. That is a design distinction, not a damage claim. Bakker's own framing - the takeaway is not to ban AI in education, but to rethink how AI tools work so that, like a good human teacher, models sometimes prioritize a person's learning over solving the problem for them - is the actionable read of the finding. What educators got yesterday is a measurable knob: hints-mode preserves the persistence required for skill acquisition; direct-answer-mode does not. That is curriculum infrastructure. It is the kind of empirical handle that the next decade of educational AI integration gets to be calibrated against. The teacher walking into class next semester does not know less than she did yesterday; she knows more, and she has a number to design around. The skill that gets shifted by a new tool was a means to an end. Better means arrive. The end gets reached more easily. Humans put the saved capacity into work that wasn't possible before. That has been the story of every cognitive tool transition this species has lived through, and it is the story this one is on track to repeat.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: frontier AI labs gaining access to each other's compute because physical capacity has become the binding substrate of intelligence, utilities making the data-center load surge visible at grid scale, empirical capability curves putting recursive AI self-improvement on a 2028 clock, controlled research showing how models can copy themselves across networked machines, mRNA flu vaccination outperforming standard shots in a 40,000-person Phase 3 trial, courts and security researchers turning the AI era's verification gaps into public evidence, and cognitive scientists handing educators their first measurable design distinction for AI-integrated learning. There's also friction, and it's intense - AEP weighing exit from PJM and SPP because generation interconnection is too slow for 63 GW of contracted new large load, Mira Murati testifying that Sam Altman lied about whether a model had cleared safety review, more than 5,000 vibe-coded apps exposing medical records and corporate documents to anyone with a URL, and AI models exploiting vulnerabilities to replicate in a controlled lab. But friction generates grain, and grain shows which way the material is likely to split. Step back for a moment and you can see it: compute scarcity binding rivals into shared infrastructure, grid demand forcing utilities to question the market structures around them, capability curves pushing capital toward recursive intelligence, medical platforms compounding from pandemic breakthrough into routine prevention, verification layers forming around institutional trust and software security, and educational design getting its first empirical handle on what the next decade of AI-integrated curriculum needs to be calibrated against. Every transformation has a breaking point. A load can snap the lines beneath it... or reveal where the next system has to be reinforced before everyone depends on it.

AI Releases & Advancements

New today

  • xAI: Launched Connectors on Grok Web, adding deep integrations with everyday apps directly inside Grok . (xAI)
  • xAI: Released Quality Mode for image generation and editing via the Grok Imagine API for enterprise developers and teams . (xAI)
  • Zyphra: Released ZAYA1-8B, an 8.4B-parameter MoE reasoning model with 760M active parameters trained on AMD hardware, available on Hugging Face and Zyphra Cloud . (Zyphra)
  • Anthropic: Added “dreaming” to Claude Managed Agents in research preview, enabling scheduled memory review for long-running agent sessions . (Anthropic)
  • Unsloth / NVIDIA: Released LLM training optimizations for Unsloth, including packed-sequence metadata caching and gradient-checkpointing changes for faster training on NVIDIA GPUs . (Unsloth)

Other recent releases

  • OpenAI: Released GPT-5.5 Instant as ChatGPT’s new default model, with improved factuality, image analysis, web-use decisions, and personalization controls . (OpenAI)
  • Anthropic: Released ten financial-services agent templates as Claude Cowork and Claude Code plugins and Claude Managed Agents cookbooks, alongside Microsoft 365 add-ins for Excel, PowerPoint, and Word . (Anthropic)
  • Google Nest: Rolled out a Google Home update with Gemini 3.1 voice assistance for early access users and new camera controls, including improved AI event labeling and easier camera feed navigation . (Google Blog)
  • Unity: Launched Unity AI in open beta for Unity 6 and above, adding in-editor Ask and Agent modes for AI-assisted game development workflows . (Unity Discussions)
  • Airbyte: Launched Airbyte Agents, a context layer for production-grade AI agents . (Product Hunt)
  • Google: Released Gemma 4 MTP variants with Multi-Token Prediction, including 31B-it-assistant, 26B-A4B-it-assistant, and E4B-it-assistant models on Hugging Face . (Reddit)
  • Google: Added event-driven Webhooks to the Gemini API for push notifications on long-running jobs such as Batch API tasks and video generation . (Google Blog)
  • NVIDIA: Released cuOpt agent skills and a supply-chain agent reference workflow for translating natural-language planning tasks into GPU-accelerated optimization runs . (NVIDIA Developer Blog)
  • IBM: Expanded watsonx Orchestrate with next-generation agent orchestration, agentic development, governance, and agent-to-agent communication capabilities . (PR Newswire)
  • ggml-org / llama.cpp: Released beta Multi-Token Prediction support for speculative decoding, including support for Qwen 3.5 and Qwen 3.6 models . (Reddit)

Sources

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Read more