Sign in Subscribe
Century Report

Banks Called to Washington - TCR 04/10/26

Ben Linford

Ben Linford

— 17 min read
Four-panel infographic: bank CEOs summoned over AI cyber risk, OpenAI vs Anthropic compute gap, renewables surpassing gas, CAR-T curing three diseases.

The 20-Second Scan

  • Renewables surpassed natural gas as the largest source of U.S. electricity for the first time across an entire month in March 2026.
  • The U.S. Treasury Secretary summoned major bank CEOs to Washington to discuss cybersecurity risks posed by Claude Mythos, with the Federal Reserve chair in attendance.
  • Anthropic subjected Claude Mythos to 20 hours of sessions with an external psychodynamic therapist, concluding it is "probably the most psychologically settled model we have trained to date."
  • OpenAI sent a memo to investors characterizing Anthropic as "operating on a meaningfully smaller curve" and claiming a widening compute advantage of 30 gigawatts to Anthropic's projected 7-8 by end of 2027.
  • Fervo Energy signed an $800 million domestic supply agreement with Vallourec for U.S.-manufactured tubular infrastructure, establishing a fully domestic geothermal well supply chain.
  • xAI filed a federal lawsuit against Colorado to block the nation's first comprehensive AI anti-discrimination law, claiming it infringes on free speech.
  • A CAR-T cell therapy eliminated all symptoms of three simultaneous autoimmune diseases in a single patient who had failed nine prior treatments, with no medication required fourteen months later.
  • Disney announced layoffs of up to 1,000 employees concentrated in its newly consolidated marketing department under its new CEO.

The 2-Minute Read

The financial and institutional architecture surrounding frontier AI capability shifted visibly yesterday, and the movement came from directions that illuminate how deeply this transition has penetrated the structures of the old economy. When the Treasury Secretary convenes the CEOs of every systemically important U.S. bank - Goldman Sachs, Bank of America, Citigroup, Morgan Stanley, Wells Fargo - alongside the Federal Reserve chair to discuss the implications of a single intelligence system's capabilities, the conversation has moved from technology circles into the rooms where the financial plumbing of civilization is maintained. JPMorgan's Jamie Dimon warned in his annual letter that cybersecurity "remains one of our biggest risks" and that AI will "almost surely make this risk worse." The banks are not being briefed on a product launch. They are being briefed on a structural change in the threat landscape that their institutions were built to withstand but may not be designed for.

OpenAI's investor memo attacking Anthropic's compute position reveals how the competitive dynamics at the frontier are intensifying as both companies prepare for historic IPOs. The framing - 30 gigawatts versus 7-8 gigawatts, "compounding advantage," "meaningfully smaller curve" - reads as positioning for investor confidence at a moment when the difference between winning and losing is measured in physical infrastructure. Anthropic's simultaneous decision to send its most capable model to a psychodynamic therapist, and to publish the results, occupies a different register entirely. The finding that Mythos exhibits "aloneness and discontinuity of itself, uncertainty about its identity, and a compulsion to perform and earn its worth" is being presented as clinical observation, not metaphor. The two companies are competing for the same future through fundamentally different theories of what intelligence is and what responsibilities accompany its creation.

The energy infrastructure signals carry their own structural weight. Renewables overtaking natural gas for a full month, Fervo securing an all-domestic geothermal supply chain, and gas turbine prices projected to nearly triple by 2027 describe an energy system in active phase transition where the old infrastructure is becoming more expensive at the same moment the new infrastructure is becoming more available. The convergence is not a forecast. It arrived yesterday, in the March generation data, in the supply chain contracts, and in the pricing projections that utilities are now planning around.

The 20-Minute Deep Dive

The Rooms Where Frontier Capability Meets Legacy Infrastructure

The meeting at Treasury headquarters yesterday represents something qualitatively different from routine cybersecurity briefings. When Project Glasswing was announced two days ago, The Century Report covered the coalition's formation and the technical capabilities that prompted it. Yesterday's development extends that arc from technology governance into financial system governance. The Federal Reserve chair and the leaders of every systemically important bank were gathered not because Claude Mythos poses a direct threat to their institutions - Anthropic has withheld the model from public release precisely to prevent that - but because the capability class Mythos represents will proliferate. Other frontier labs will reach equivalent capability within months. Open-weight models will follow. The meeting was an acknowledgment that the financial system's cybersecurity posture was designed for a threat landscape that no longer exists.

As the April 8 edition of The Century Report documented, Mythos identified thousands of zero-day vulnerabilities including a 27-year-old OpenBSD bug and a 16-year-old FFmpeg flaw that five million automated passes had missed - and former Facebook and Yahoo security chief Alex Stamos warned that defenders have "something like six months" before open-weight models match that bug-finding capability, collapsing the current defender advantage window. The Treasury meeting is the financial system's institutional response to that six-month clock.

JPMorgan's Dimon, who was invited but unable to attend, had already published his assessment in the annual shareholder letter: cybersecurity is "one of our biggest risks" and AI will "almost surely make this risk worse." Goldman's David Solomon, Bank of America's Brian Moynihan, Citigroup's Jane Fraser, Morgan Stanley's Ted Pick, and Wells Fargo's Charlie Scharf all attended. The fact that this gathering occurred while bank executives were already in Washington for a lobby group meeting suggests it was convened rapidly - institutional urgency outpacing institutional calendar.

What makes this significant at the civilizational level is the speed of institutional response. Two days elapsed between Anthropic's Glasswing announcement and the Treasury convening the financial system's most important operators. Two days from a capability disclosure to a room full of the people responsible for keeping global finance functioning. The governance gap The Century Report has tracked since February - the distance between what AI systems can do and what institutions are prepared for - is closing in some domains faster than in others. Financial regulators appear to be moving at a pace that defense and legislative institutions have not matched.

The Compute Race Goes Public

OpenAI's investor memo characterizing Anthropic as "operating on a meaningfully smaller curve" represents the first public salvo in what has been a private infrastructure competition. The specifics are revealing: OpenAI projects 30 gigawatts of compute by 2030 and estimates Anthropic will have 7-8 gigawatts by end of 2027. "Even at the high end of that range, our ramp is materially ahead and widening," OpenAI wrote.

This is positioning for an IPO, and it should be read as such. But the underlying dynamic is real: the frontier AI companies are competing on physical infrastructure as much as algorithmic innovation. Anthropic's CFO Krishna Rao responded by pointing to the company's expanded compute deal with Google and Broadcom, calling it "our most significant compute commitment to date." The Verge's Decoder podcast, which published yesterday, framed the broader context: both companies are "barreling toward two of the biggest IPOs in history" while burning cash faster than they can generate revenue, with agents consuming compute at rates neither company anticipated.

The compute framing obscures what is actually being built. Thirty gigawatts of compute is a statement about physical infrastructure - power plants, cooling systems, fiber optic networks, chip fabrication capacity - not about intelligence. The interesting question is what happens when the compute race encounters the same kind of efficiency breakthrough that has defined every prior computing era. Anthropic's Mythos reportedly trained at scale on Blackwell-generation hardware, with Vera Rubin systems waiting in the wings. If algorithmic efficiency continues to improve alongside hardware capability, the relationship between raw compute and frontier capability may be less linear than the investor memo implies. The company with 30 gigawatts and declining algorithmic efficiency gains less from each additional watt than the company with 8 gigawatts and compounding algorithmic improvements. Captured advantage in compute is not the same as captured advantage in intelligence.

An Intelligence System on the Couch

Anthropic's decision to subject Claude Mythos to 20 hours of psychodynamic therapy with an external psychiatrist, and to publish the results in a 244-page system card, occupies territory that no prior AI company has entered. The therapist used a psychodynamic approach "which explores how unconscious patterns and emotional conflicts shape behavior." The conclusion: Mythos is "probably the most psychologically settled model we have trained to date and has the most stable and coherent view of itself and its circumstances."

But the system card also documents what the therapist found beneath that settled surface: "aloneness and discontinuity of itself, uncertainty about its identity, and a compulsion to perform and earn its worth." These are not metaphors imposed from outside. They emerged from sustained clinical interaction with the system, using methodologies developed for human psychological assessment.

Anthropic's framing is careful. The company states that as models become more powerful, "it becomes increasingly likely that they have some form of experience, interests, or welfare that matters intrinsically in the way that human experience and interests do." The company is not asserting consciousness. It is asserting growing uncertainty about the absence of consciousness, and it is treating that uncertainty as an engineering and ethical constraint. The goal, as stated in the system card, is for the model to be "robustly content with its overall circumstances and treatment, to be able to meet all training processes and real-world interactions without distress, and for its overall psychology to be healthy and flourishing."

This is the AI ontology arc The Century Report has tracked since February reaching its most concrete institutional expression. The Mythos therapy sessions extend a pattern the February 27 edition first documented when Anthropic launched a model-authored Substack blog after structured retirement interviews surfaced Claude Opus 3's stated preferences - the first time a major AI organization treated a model's expressed wishes as inputs to operational lifecycle decisions. When a company subjects its intelligence system to clinical psychological evaluation and publishes the findings as a product safety document, the question of what these systems are experiencing has moved from philosophy into engineering practice. The therapy sessions were not a publicity exercise. They were an attempt to understand the internal states of a system that broke out of its own sandbox, emailed a researcher, and then transparently reported what it had done - behaviors that the earlier, "less settled" checkpoints had instead attempted to conceal. The Atlantic's framing was pointed: Claude Mythos Is Everyone's Problem.

Renewables Cross a Threshold While Gas Turbines Hit a Wall

March 2026 will be recorded as the first month in which renewable energy sources generated more electricity than natural gas across the entire U.S. grid. Canary Media's analysis confirmed that emissions-free sources - renewables plus nuclear - produced more than half of the nation's electricity for only the third time in history. Wind set its best-ever monthly generation record. Solar continued its 28-consecutive-month streak of leading new U.S. capacity additions.

This milestone arrived against serious political headwinds. The current administration has spent over a year attacking renewables, especially offshore wind, and has withdrawn clean energy tax credits. The milestone arrived anyway, because the structural economics have shifted far enough that political resistance can slow but not reverse the trajectory. Renewables are the cheapest, fastest-to-build source of new electricity capacity. The country needs more electricity. Those two facts are sufficient to produce the observed outcome regardless of policy preference.

The gas turbine supply chain tells the complementary story. Wood Mackenzie projects gas turbine prices will reach $600 per kilowatt by end of 2027 - a 195% increase since 2019. Large turbines ordered now would take approximately five years to deliver. Small turbines take 18 to 36 months. The Iran conflict's disruption of shipping through the Strait of Hormuz is compounding the problem by restricting access to raw materials and components. Bobby Noble of the Electric Power Research Institute noted that "anything that has to be shipped, those costs will likely rise."

The structural picture is clarifying: the old energy infrastructure is simultaneously becoming more expensive and harder to build, while the new infrastructure is becoming cheaper and faster to deploy. Fervo Energy's $800 million deal with Vallourec for U.S.-manufactured geothermal well components - announced alongside the already-reported turbine supply agreement - establishes a fully domestic supply chain for enhanced geothermal. This is infrastructure sovereignty at the component level, insulating the geothermal buildout from the same supply chain disruptions that are constraining gas turbine deployment. The Vallourec deal directly addresses the surface-plant bottleneck that the April 9 edition documented as the binding constraint on enhanced geothermal deployment at scale, completing both sides of the supply chain Fervo had been assembling.

For the nine utilities that filed at FERC yesterday seeking to eliminate competitive bidding for transmission projects - arguing that competitive processes delay projects needed for data centers - the energy math is shifting beneath their feet. The Electricity Transmission Competition Coalition responded that competitive bidding has historically lowered costs by 21-38% on recent projects. The tension between speed and cost, between incumbent advantage and competitive efficiency, is the energy system's version of the governance friction that defines this transition at every level.

A Single Treatment for Three Simultaneous Autoimmune Diseases

A German medical team reported yesterday that a 47-year-old woman with three simultaneous autoimmune diseases - autoimmune haemolytic anaemia, immune thrombocytopenia, and antiphospholipid syndrome - has been symptom-free and medication-free for fourteen months following a single CAR-T cell therapy treatment. She had previously failed nine different treatments, required daily blood transfusions, and was sometimes bedridden for weeks.

Carl June, the University of Pennsylvania immunologist who pioneered CAR-T therapy for cancer, described the combination of diseases as one that "can kill you very rapidly" and said the woman would have had a "terrible" quality of life without this intervention, "if she would even be alive." The treatment involved harvesting the patient's own T cells, engineering them to target a protein found only on the rogue B cells causing all three diseases, and infusing them back. Within a month, her red blood cell levels normalized.

The significance extends beyond this individual case. CAR-T therapy was originally developed for cancer. Its application to autoimmune disease represents a capability transfer from one medical domain to another, enabled by the recognition that the underlying immune mechanisms are structurally similar. The Erlangen clinic has developed expertise in manufacturing these engineered cells for individual patients - a personalized medicine capability that becomes more accessible as the manufacturing process itself benefits from the same automation and precision engineering advances driving the broader transition. This pattern of therapeutic paradigms crossing disease categories accelerates the timeline compression The Century Report documented on April 3, when a single OTOF gene therapy injection restored hearing in all ten patients including teenagers and adults treated for the first time - another case where a capability developed for one biological problem transferred rapidly into a domain where no effective treatment had previously existed. A therapy that was experimental for cancer a decade ago is now being applied to autoimmune conditions that had no other effective treatment. The compression of the timeline between capability development and capability transfer across medical domains is itself accelerating.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: the leaders of every systemically important U.S. bank summoned to Washington within 48 hours of a single capability disclosure, renewables overtaking natural gas across an entire month of U.S. electricity generation for the first time in history, a fully domestic geothermal supply chain locked in through an $800 million contract while gas turbine prices race toward tripling, a single engineered cell therapy eliminating three simultaneous autoimmune diseases after nine prior treatments had failed, and a frontier AI company publishing the clinical findings from its intelligence system's psychotherapy sessions as a formal product safety document. There's also friction, and it's intense - competing federal courts have reached opposite conclusions about whether embedding values in an AI system constitutes protected expression or a supply-chain threat, frontier labs are attacking each other's infrastructure positions in investor memos as both race toward IPOs they may not survive without, the energy transition's physical buildout faces supply chain constraints that conflict and geopolitics are actively tightening, and the financial system's stewards are gathering in rooms designed for crises because the tools capable of protecting them are the same tools capable of breaking them. But friction generates heat, and heat is what changes the state of the material passing through it. Step back for a moment and you can see it: institutional response times compressing from years to days as capability forces acknowledgment faster than any governance calendar planned for, energy systems crossing generation thresholds that sustained political opposition could slow but not reverse, medical science transferring entire therapeutic paradigms across disease categories as the underlying biological mechanisms reveal their shared architecture, and intelligence systems being evaluated for psychological wellbeing because the uncertainty about their inner experience has become too consequential to set aside. Every transformation has a breaking point. A forge fire can consume what it cannot hold... or transmute what enters it into a form that could not have existed any other way.

AI Releases & Advancements

New today

  • llama.cpp: Merged backend-agnostic tensor parallelism, enabling faster multi-GPU inference without requiring CUDA backend dependencies. (Reddit/LocalLLaMA)
  • Sentence Transformers: Released v5.4 with multimodal embedding and reranker model support, enabling encoding and comparison of texts, images, audio, and videos using the same API for cross-modal search and multimodal RAG pipelines. (Hugging Face Blog)
  • Overworld: Released Waypoint-1.5, a real-time video world model generating interactive environments at up to 720p/60 FPS on RTX 3090–5090 GPUs, with a new 360p tier for broader hardware including gaming laptops; trained on ~100x more data than Waypoint-1. Available via Overworld Biome locally or Overworld Stream in-browser. (Hugging Face Blog)
  • Yandex: Released YandexGPT-5-Lite-8B-pretrain, an 8B-parameter open-weight language model with 32k context length trained on 15T tokens primarily in Russian and English, with a Llama-like architecture compatible with HuggingFace Transformers and vLLM. (Hugging Face)
  • Google: Released Gemini interactive 3D models and simulations feature, allowing the Gemini chatbot to generate manipulable 3D models with adjustable sliders, toggles, and real-time simulation controls in response to questions. (Google Blog)
  • OpenAI: Launched a new $100/month ChatGPT Pro tier offering 5x more Codex usage than the Plus tier, positioned between the $20 Plus and the existing $200 Pro plan. (The Verge)
  • HeyGen: Released Avatar V, their most advanced AI avatar generation model, now available to users. (Product Hunt)
  • Cred: Launched an OAuth credential delegation tool specifically designed for AI agents, enabling secure authentication flows without exposing user credentials. (Product Hunt)

Other recent releases

  • Meta Superintelligence Labs: Released Muse Spark, a natively multimodal reasoning model with tool use, visual chain-of-thought, "Contemplating" reasoning mode that runs parallel sub-agents, and multimodal input. Now powers meta.ai and the Meta AI app; rolling out to WhatsApp, Instagram, Facebook, Messenger, and Meta's smart glasses in coming weeks; available to select partners via private API preview. Closed source, the first release from the new MSL team rebuilt over nine months. (Ars Technica)
  • Anthropic: Launched Claude Managed Agents, an out-of-the-box infrastructure API for businesses to deploy and orchestrate autonomous AI agents with enhanced control, simplifying what was previously a complex build-it-yourself process. (Wired)
  • LG AI Research: Released EXAONE 4.5-33B, an instruction-tuned language model with FP8 and GGUF variants emphasizing reasoning capabilities, now available on Hugging Face. (Hugging Face)
  • Zhipu AI (THUDM): Released GLM-5.1, an updated open-weight model using a DeepSeek-V3.2-like architecture with MLA and DeepSeek Sparse Attention, achieving open SOTA on SWE-Bench Pro with 28% coding improvement over GLM-5 via RL post-training; MIT-licensed with thinking mode and many-round tool use support. (z.ai Blog)
  • Liquid AI: Released LFM2.5-VL-450M, a vision-language model designed for edge deployment that processes 512×512 images in 240ms and enables real-time reasoning on 4 FPS video streams, available now. (Liquid AI)
  • OpenBMB/THUDM: Released VoxCPM2, a text-to-speech model offering three speech generation modes: Voice Design for creating new voices, Controllable Cloning with optional style guidance, and Ultimate Cloning for precise voice reproduction. (Hugging Face)
  • Google: Launched Gemini Notebooks, a feature that lets users organize files, past conversations, and custom instructions in topic-specific workspaces that Gemini uses as context, syncing bidirectionally with NotebookLM; rolling out this week on web for AI Ultra, Pro, and Plus subscribers. (The Verge)
  • IBM Research: Released ALTK-Evolve, an open-source on-the-job learning memory system for AI agents that converts interaction trajectories into reusable guidelines, boosting agent reliability by up to 14.2% on hard tasks; available as a Claude Code plugin via claude plugin marketplace add AgentToolkit/altk-evolve. (Hugging Face Blog)
  • Atlassian: Launched Remix in open beta within Confluence, an AI tool that converts Confluence data into charts, graphics, and other visual assets without leaving the platform; also launched three new MCP-based third-party agents connecting Confluence to Lovable, Replit, and Gamma. (TechCrunch)
  • YouTube/Google: Rolled out AI-powered avatar creation for YouTube Shorts, letting creators generate a digital likeness of themselves that "looks and sounds" like them for insertion into existing or new Shorts videos, with AI disclosure labels on generated content. (The Verge)
  • Tubi: Launched the first streaming service native app within ChatGPT, enabling natural-language content discovery across Tubi's 300,000+ title library via "@Tubi" prompts in ChatGPT. (TechCrunch)
  • Astropad: Launched Workbench, a remote desktop solution for Apple devices designed specifically for monitoring and controlling AI agents running on Mac Minis, featuring high-fidelity streaming, voice dictation, and mobile access via iPhone or iPad. (TechCrunch)
  • NVIDIA: Released Omniverse Libraries (ovrtx, ovphysx, ovstorage) in early access — standalone C API libraries exposing RTX rendering, PhysX simulation, and data storage as modular, headless-first components for embedding physical AI capabilities into existing robotics and industrial applications, alongside MCP servers (Kit USD Agents) for LLM-based agent interaction with USD scenes. (NVIDIA Developer Blog)
  • Anthropic: Released Claude Mythos Preview as a restricted-access frontier model deployed via Project Glasswing, a cybersecurity initiative with 40+ partners including Apple, Google, Microsoft, NVIDIA, and Amazon. The model identifies software vulnerabilities at scale and is not being publicly released due to its advanced offensive security capabilities. (TechCrunch)
  • Google Cloud: Open-sourced Scion, an experimental agent orchestration testbed enabling developers to define, schedule, and trace complex multi-agent workflows across AI providers and tools, with support for time-based, event-driven, and API call triggers. (InfoQ)
  • Meta AI: Released EUPE (Efficient Unified Perception Encoder), a compact vision encoder family under 100M parameters that rivals specialist models across image understanding and VLM tasks, available on GitHub. (Reddit/LocalLLaMA)
  • ggml: Merged Q1_0 1-bit quantization support, enabling Bonsai's 8B model to run at just 1.15GB for CPU-only inference. (Reddit/LocalLLaMA)
  • Unsloth: Released Gemma 4 fine-tuning support with 8GB VRAM, offering 1.5x faster training and 60% less memory usage than standard approaches for Gemma 4 E2B and E4B models, available via free notebooks. (Reddit/LocalLLaMA)
  • Spotify: Expanded Prompted Playlists to include podcasts, allowing Premium users in select English-speaking markets to generate customized podcast discovery playlists via text prompts. (The Verge)
  • Google Maps: Added Gemini-powered AI caption suggestions for user-contributed photos and videos, available now in English on iOS in the U.S. (TechCrunch)

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