Automated draft from LLL
Anthropic Secures Five Gigawatts of Amazon Compute and Reveals a Thirty-Billion-Dollar Revenue Run Rate
Anthropic and Amazon announced a ten-year agreement where Anthropic committed over one hundred billion dollars to AWS infrastructure. This deal secures up to five gigawatts of compute capacity, allowing Anthropic to train and deploy its Claude models using Amazon's Trainium2 to Trainium4 chips. Amazon will invest an additional five billion dollars immediately, with twenty billion more to follow, beyond its earlier eight-billion-dollar commitment.
The most striking disclosure wasn't the compute—it was the revenue. Anthropic's current annual revenue run rate now exceeds thirty billion dollars, a sharp rise from approximately nine billion dollars at the end of 2023. This marks more than threefold growth in four months. The company said the deal partly addresses strain from "unprecedented consumer growth," which degraded reliability for its free, Pro, Max, and Team users during peak hours. Anthropic expects nearly one gigawatt of new capacity before year-end, and significant computing power will arrive within ninety days.
The full Claude Platform will integrate directly into AWS. Users will access it through their existing AWS accounts, with unified billing and no additional credentials. Claude is now the only frontier model on all three hyperscalers (AWS, Google Cloud, Azure). A separately announced Google and Broadcom partnership will add more capacity. Anthropic thus diversifies across chip vendors, but retains Amazon's custom silicon as its primary training platform. Over one hundred thousand customers already run Claude on Bedrock.
The broader Claude ecosystem continues expanding. A guide to Claude Design, which we covered on April 18th and 19th, details a design-system-first workflow, offering customizable parameters and native skill modes that many users overlook. On GitHub, at least two open-source projects—cc-design and claude-code-design—already attempt to reproduce Claude Design's prototyping capabilities within Claude Code. Anthropic also announced the winners of its "Built with Opus 4.6" Claude Code hackathon. Four of the five winners were not professional developers—including a lawyer building a California housing permit tool and a cardiologist developing patient follow-up software. This reinforces that its user base extends far beyond software engineering.
GPT-5.5 Leaks Suggest OpenAI's New Base Model Drops This Week
Multiple T4 sources report on what OpenAI internally calls "Spud," which many expect to launch as GPT-5.5, and a Pro variant that offers extended reasoning. The information stems from leaked outputs and firsthand accounts on social media, as well as a separate hands-on test of early checkpoints seemingly accessible through ChatGPT.
The headline claim, attributed to users of the model, claims Spud equals Mythos, Anthropic's unreleased research model, which sets an informal benchmark for cutting-edge AI. Greg Brockman described it as the product of two years of pre-training work—a new base model, not a distillation or finetune. If benchmarks prove accurate, Spud could achieve a ten-to-fifteen percent jump across standard evaluations, potentially pushing OpenAI back into the lead in categories where Opus 4.7 currently dominates, as we noted on April 17th and 18th.
Two technical bets stand out. First, Spud might be natively multimodal, processing audio, images, and text within a single architecture rather than routing data through separate encoders. OpenAI previously abandoned this approach with GPT-4o; whether they have now made it work remains the central question. Second, a new image generation model, "Images V2," will reportedly ship alongside Spud, whose outputs reportedly match or exceed Google's Gemini 1.5 Pro, especially in handling complex styles and compositional understanding. These details come from unconfirmed T4 sources, but the volume and specificity of the leaks point to an imminent announcement. If even partly accurate, the pricing claim—better reasoning, lower cost, and faster output—would be the most strategically significant aspect, as it attacks Anthropic's capacity constraints from the demand side.
Five Sources Say the Same Thing: The Harness Matters More Than the Model
A cross-source signal stands out this week: five independent sources—a T2 podcast, a T3 newsletter series, and practitioner content—all present the same thesis. The bottleneck isn't model capability. It's the scaffolding around the model.
Ramp's internal AI system, "Glass," detailed on The AI Daily Brief, offers the most concrete enterprise example. Glass configures developer workspaces automatically on day one via SSO integrations. It provides a marketplace of more than 350 reusable agent skills called "Dojo," operates a recommendation engine ("Sensei") that identifies the five most relevant skills for each user, based on their role and tools, and maintains persistent memory through a daily synthesis pipeline across Slack, Notion, and Calendar. Ninety-nine percent of Ramp's 350-person team uses AI daily. The episode cites a PWC study, which shows seventy-five percent of AI's economic gains accrue to just twenty percent of companies—not because they possess superior models, but because they leverage AI for growth and business model reinvention, rather than mere productivity. McKinsey data indicates a three-dollar return in EBITDA for every dollar invested for AI leaders, with a twenty percent average EBITDA uplift.
Claw Mart Daily published a five-part practitioner series on agent-engineering fundamentals, covering topics such as explicit done criteria, failure budgets with checkpoint-based recovery, information provenance tracking, when multi-agent coordination actually justifies its overhead, and operating manuals that load into session context. The consistent message: agents fail not from insufficient intelligence but from missing structure. Done criteria alone reduced task times from seventy-three to twenty-three minutes in one practitioner's tracking. The multi-agent piece is especially insightful: "Multi-agent systems don't multiply success rates—they multiply failure rates. Every handoff is a potential break point." The recommended test: if you can't explain why Agent B can't do Agent A's job, you don't need Agent B.
Steve Newman, creator of Writely (later Google Docs), articulated a parallel philosophy on The Cognitive Revolution. He uses fifteen separate Claude Code projects that form his personal AI infrastructure. This includes an "attention firewall" that classifies urgency across email, Slack, WhatsApp, Signal, and SMS, bringing only critical items to his attention. His principle involves separate repositories for each project, keeping architectural stakes low enough to render staging environments unnecessary, and optimizing for human attention rather than agent utilization. His observation on productivity gains echoes the Jevons Paradox: tools did not save time; instead, they enabled previously impossible outputs such as custom podcast music, AI-generated art, and video clips. Fewer engineers per line of code, but vastly more code total.
Pi Coding Agent Makes the Case That Claude Code Has Gotten Too Big
The most pointed contrarian take this week arrives from Mario Zechner, creator of the Pi coding agent, in a workflow demonstration by Cole Medin. Pi is a deliberately minimalist open-source coding agent. Zechner argues that Claude Code, which began as a simple, predictable command-line interface, has accumulated so many features, bugs, and constantly shifting system prompts that users can no longer control its underlying processes. "Your context is not really your context," as Zechner puts it.
Pi's answer is radical simplicity. It has no Multi-Constraint Planner (MCP), no sub-agents, and no built-in plan mode. Users can ask Pi to build any of these features into itself, and a growing extension marketplace already offers third-party implementations. Medin demonstrated a plan-implement-validate workflow, combining Pi with Archon, his open-source harness builder. He used a "Planotator" extension for browser-based plan review with inline commenting. The workflow mixed Pi—running GPT-5.3 via Codex—for planning, and Claude for implementation. This provider-agnostic approach Claude Code's architecture does not natively support.
A noteworthy counterpoint from The AI Daily Brief: George Savulka at a16z argues that individual AI productivity does not sum to organizational value without coordination layers. Ramp's approach to this proves instructive: it preserved full capability for power users rather than simplifying for the lowest common denominator, by making complexity invisible rather than absent. The distinction between "institutional AI" and "aggregated individual AI" may determine which companies realize the McKinsey-projected returns and which just distribute chat interfaces.
Noetik Licenses a Cancer Biology Foundation Model to GSK for Fifty Million Dollars
In a deal that may signal how bio-AI will commercialize, Noetik, a startup that trains transformer models on spatially resolved patient tumor data, announced a fifty-million-dollar licensing agreement with GSK for its OctoVC virtual cell foundation model. Discussed on Latent Space, the deal is described as the first announced foundation model licensing agreement in the bio-AI space.
Noetik's thesis posits that ninety to ninety-five percent of cancer drugs fail in trials not because the drugs are ineffective, but because trials enroll the wrong patients. Their models, trained on multimodal data—H&E stains, immunofluorescence, spatial transcriptomics, DNA genotyping—all generated in-house, identify patient subtypes that predict drug response. A new autoregressive architecture called Tario outperformed their previous masked-autoencoding approach, OctoVC. Larger models and longer spatial context consistently improved performance—a scaling curve mirroring that of language models years ago. Critically, after training on multimodal data, inference requires only a standard H&E pathology image, which makes clinical deployment practical. The GSK deal includes an upfront payment, milestones, and annual licensing fees, suggesting pharmaceutical companies are moving toward broad model access rather than bespoke project collaborations.
Five Things With 30-Day Clocks
GPT-5.5 / Spud launch. If leaks prove accurate, OpenAI will ship it this week. The benchmark to watch is SWE-Bench Pro, where Opus 4.7 jumped eleven points on April 18th. Whether Spud matches that coding performance—and whether native multimodality delivers measurable gains over encoder-stitching—will determine any shift in the competitive narrative.
Anthropic's Q2 capacity expansion. The Amazon deal promises "significant computing power in the next three months." The test is whether Pro and Max throttling visibly improves by mid-May. Consumer reliability has become the most common complaint in the Claude ecosystem, and the thirty-billion-dollar run rate suggests demand is not slowing.
Trainium3 production benchmarks. Anthropic expects "scaled Trainium3 capacity" by the end of 2026, but AWS has not published independent training benchmarks. Whether Trainium3 narrows the gap with NVIDIA Blackwell for frontier model training will determine how much of the five-gigawatt commitment is strategically optimal or merely locked in.
Pi's extension ecosystem. With community catalogs tracking more than eighty-five vibe-coding tools and Pi's marketplace growing, we will track whether Pi's active user base crosses the threshold that compels Claude Code to respond—either by simplifying its architecture or by officially supporting provider-agnostic model switching.
Noetik's Tario scaling results. The autoregressive architecture demonstrated promising scaling curves on spatial biology data. Published benchmarks comparing Tario to OctoVC on identical datasets would influence both pharmaceutical companies' evaluation of bio-AI vendors and broader architectural choices for foundation models beyond language.
Top comments (0)