Anthropic is spending $3.4 billion on Caterpillar natural-gas turbines capable of producing up to 1.7 gigawatts, a sharp break from its preference to rent infrastructure.
The purchase comes as turbine backlogs stretch into the early 2030s and grid-connection queues are similarly long. Anthropic disclosed in its May S-1 that it pays SpaceX subsidiary xAI $1.25 billion a month for computing power through May 2029, but rising demand appears to have pushed it toward owning generation.
Google, Amazon, Microsoft, and Meta are also pursuing on-site natural-gas plants. That strategy carries exposure to fuel prices: one analyst expects some markets to see prices more than triple by decade’s end.
Why it matters: Anthropic is locking in generation while turbine and grid queues extend into the early 2030s, giving the AI company an alternative to relying solely on rented computing capacity from xAI.
Meta launched Muse Ember, a lightweight model priced for the repetitive work AI agents perform at scale. Ember handles tool calls, coding subtasks, routing, extraction, and long-document parsing, with pricing of $0.10 per million input tokens and $0.50 per million output tokens through the Meta Model API and Muse Code.
Meta says Ember delivers more intelligence per dollar than Anthropic’s Claude Haiku 5.5 and OpenAI’s GPT-6 Luna. Artificial Analysis measured about $0.14 per task for Ember, versus $2.39 for Meta’s larger Muse Spark 1.3 at nearly the same score.
The trade-off is coding: Haiku scored 34.9% versus Ember’s 27.2% on FrontierCode 2.0-AA. Ember supports text, images, audio, video, and up to one million tokens.
Why it matters: Meta is targeting the expensive, repetitive layer of agent deployments rather than trying to replace frontier models outright. If the $0.14-per-task result holds beyond Meta’s preferred tests, Ember gives companies a cheaper execution engine while leaving harder coding work to models such as Claude Haiku.
AWS has added NVIDIA Nemotron 3.5 Super VL to SageMaker JumpStart, giving customers a deployable vision-language model for images, video, and text. The model combines a 120B-parameter hybrid Mixture-of-Experts backbone with 12B active parameters and supports context lengths up to 1 million tokens.
JumpStart offers an NVFP4 checkpoint for optimized inference and a 249GB BF16 reference version for customization. The NVFP4 model is about 75GB and runs on Blackwell GPUs, while the BF16 version supports SFT, reinforcement learning, distillation, and domain adaptation.
NVIDIA's model-card tests show NVFP4 broadly tracking BF16 across document, chart, video, and computer-use benchmarks. But JumpStart configurations serve only 520 to 15,600 tokens, not the headline 1-million-token maximum.
Why it matters: SageMaker customers can now test NVIDIA's multimodal stack without assembling a separate serving path, but the practical deployment envelope is narrower than the advertised 1-million-token context. For AWS teams building agents that read documents or screens, the 75GB NVFP4 checkpoint is the operationally relevant option.
Anthropic is moving ahead with plans to go public, with a valuation range expected this week and trading targeted for Thursday or Friday of next week. Institutional investors are pressing for a valuation below $1.5 trillion, down from a target discussed as recently as last week and well below the $2 trillion figure floated months ago. That puts the listing under pressure before shares trade, making the debut an immediate test of public-market appetite for AI.
The comparison point is OpenAI, Anthropic’s main rival: it was reportedly valued at $1.25 trillion last week after raising $11 billion, short of its initial $20 billion goal.
Why it matters: Anthropic’s pricing will give institutional investors a public-market yardstick for OpenAI’s $1.25 trillion valuation—and test whether they will still fund trillion-dollar AI companies after Nasdaq’s sell-off.
Amazon Bedrock Managed Knowledge Base now lets applications choose retrieval effort per request through AgenticRetrieveStream. Developers can set maxAgentIteration from 0 to 5 for retrieval-only, single-shot RAG, or agentic retrieval.
At 1, the service retrieves once and uses one foundation-model call to generate a grounded answer; at 0, it returns passages without an FM call. Higher values support planning, query rewriting, iterative searches, and evidence evaluation for complex questions. The API supports multiple knowledge bases across all three modes.
That gives banking operations assistants a single path for a checklist lookup and a cross-policy comparison. AWS says the simpler modes can reduce latency and model usage, but provides no independent performance or cost figures.
Why it matters: For teams building internal assistants, Bedrock turns retrieval tuning into a per-request control rather than an architecture decision: maxAgentIteration=0 avoids a foundation-model call, while complex cases can use up to five iterations. That makes the latency-and-quality tradeoff explicit, although AWS has not supplied independent benchmarks.
Google DeepMind used its AlphaProof Code framework to formally verify portions of the Signal protocol in a four-week proof of concept. The analysis found no new vulnerabilities in the code it examined, adding independent validation for cryptography used by Signal, Google Messages, and WhatsApp.
The test also correctly identified known, fixed vulnerabilities in older Signal versions and showed those implementations could not be formally verified. But it did not cover the entire Signal mobile app. DeepMind released its specification and proof, while researchers stressed that human review remains essential. The bigger question is economics: Google did the work free as research, declined to disclose its cost, and has not said whether it will sell a verification product.
Why it matters: A four-week DeepMind project makes AI-assisted verification look viable for security-critical code, but the free research demonstration leaves banks and other high-stakes operators without a known commercial cost or scaling path. Signal's results are a milestone, not proof that AlphaProof Code can secure entire software supply chains.
Data-center opposition is increasingly a fight over electricity, not just technology. In Virginia’s Loudoun County, the average facility already has 54 megawatts of capacity—more than Denver International Airport’s power draw. Activist Elena Schlossberg says the issue was “always about power,” while OpenAI global affairs chief Chris Lehane has framed the backlash as a proxy for affordability and corporate control.
Virginia now requires data centers to pay the full cost of new transmission lines, and developers anticipate possible grid cutoffs during peak demand. Elsewhere, AI campuses are bypassing weak grids with behind-the-meter natural-gas turbines and diesel backup. The smallest sites attractive for AI training are 300 megawatts; ideal facilities could reach 6,000 megawatts—roughly Chicago’s draw.
Why it matters: Anthropic’s gas-turbine plans show why avoiding a weak grid does not end the fight: nearby communities still face air pollutants from private generation and expect a say comparable to the one traditionally given for public-utility projects.