AI Weekly Briefing: 18-24 July 2026

AI infrastructure, health products and model competition led the week. AMD secured a large Anthropic deployment, OpenAI connected ChatGPT to medical records, and new Chinese models continued to narrow the gap with leading US systems.

AMD commits up to $5 billion to Anthropic

AMD and Anthropic announced a strategic partnership covering up to two gigawatts of AMD Instinct MI450 GPUs in Helios rack-scale systems. The first gigawatt is due to begin deployment in the first half of 2027, while AMD has committed to a future equity investment of up to $5 billion. Anthropic already uses AMD’s MI355X hardware.

The agreement gives AMD a major frontier-model customer and expands Anthropic’s supply beyond Nvidia, Amazon’s Trainium and Google’s TPUs. At this scale, access to power, racks and several chip ecosystems has become part of the competitive position of an AI laboratory, rather than a procurement detail.

OpenAI expands ChatGPT Health across the US

OpenAI began rolling out Health in ChatGPT to logged-in US users aged 18 and over on web and iOS. Eligible users can connect supported medical records and Apple Health data, then ask ChatGPT to compare results, summarise changes and prepare questions for appointments. The feature is available across Free, Go, Plus and Pro plans.

OpenAI says connected records, Apple Health information and conversations using that data are not used to train its foundation models or target advertising. The company still warns that ChatGPT can make mistakes and does not replace qualified medical care. That distinction carries more weight now that the product can reason over longitudinal records rather than information pasted into a single chat.

Chinese models keep pressure on the frontier laboratories

Moonshot AI’s Kimi K3 and Alibaba’s preview of Qwen3.8 prompted another round of comparisons with OpenAI and Anthropic’s leading systems. The releases drew scrutiny after claims that Kimi’s performance might have come from distilling Anthropic’s Fable model, although researchers quoted by TechCrunch said distillation alone was unlikely to explain a system of its reported strength.

The broader development is the pace and consistency of Chinese model releases, particularly systems distributed with open weights or developer-friendly access. Engineering teams now have credible alternatives arriving from several laboratories, which makes evaluation across coding, tool use, languages, cost and deployment constraints more useful than treating a single benchmark as a league table.

Etched reaches a $10.3 billion valuation with an inference-first chip

AI chip startup Etched reached a reported $10.3 billion valuation after attracting new investment from prominent backers. The company is developing its Sohu hardware and accompanying memory components specifically for transformer inference, claiming it can run current AI models without conventional GPUs.

Etched is taking a narrower architectural bet than general-purpose accelerator vendors. Specialisation can produce substantial inference gains if transformer workloads remain dominant, but it leaves less room to absorb a major shift in model architecture. The valuation shows how much capital is chasing alternatives to Nvidia as inference demand becomes a larger share of AI spending.

MCP prepares its largest specification revision

The Model Context Protocol project published the release candidate for its 28 July specification. The revision introduces a stateless core designed for ordinary HTTP infrastructure, an extensions framework for capabilities such as long-running Tasks and MCP Apps, tighter alignment with OAuth 2.0 and OpenID Connect, and a formal deprecation policy.

Remote MCP servers should become easier to operate behind standard load balancers because deployments no longer need the same sticky-session and shared-state arrangements. The release contains breaking changes, including a new lifecycle for teams that adopted the experimental Tasks API, so implementers have a migration job alongside the simpler production architecture.

Monday.com cuts 20 per cent of its workforce around an AI strategy

Monday.com announced plans to remove about 20 per cent of its global workforce, affecting roughly 630 employees, as it reorganises around its AI Work Platform. Its product direction now combines work management with no-code application building, workflow automation, chat and configurable agents that can update dashboards and generate reports.

The company expects continued revenue growth, so the restructuring is a product and operating-model decision rather than an emergency response to falling sales. It is a stark example of established SaaS companies trying to reduce costs while repositioning products for a market where agents can perform parts of the work their software previously helped people organise.

Runway adds automatic model routing for generative media

Runway launched Media Router through Runway Dev, its recently introduced developer platform for image, video and audio generation. Developers can specify whether quality, speed or cost matters most, and the router selects a suitable model instead of requiring an application to encode those choices directly.

Model routing is becoming a product layer in its own right as providers add competing generators with different prices and performance profiles. Runway’s approach reduces integration work for teams that care about an output target more than the name of the underlying model, while giving the platform control over how requests move across its catalogue.

Substack lets readers estimate AI use in newsletters

Substack added reader-facing AI detection through a partnership with Pangram. Readers can scan eligible posts, notes, comments and replies published from 21 July and receive an estimate of how much text appears human-written or AI-assisted. Publishers can disable detection for individual posts.

The feature turns disclosure about AI-assisted writing into a platform mechanism, but the percentage remains a classifier’s estimate rather than proof of authorship. Substack is testing whether readers value that signal enough to accept false positives, disputes and the incentive for writers to alter their prose to satisfy a detector.

Also worth noting

Anthropic upgraded Claude’s voice mode with more capable models and actions that can help draft emails or reschedule meetings, extending the competition to conversational assistants that can complete routine work rather than only answer spoken questions.

OpenAI introduced Presence, an enterprise platform for deploying voice and chat agents across customer and internal workflows. The company says its own deployment now resolves 75 per cent of inbound support issues without human assistance.

NTT DATA reported that a Codex-assisted incident analysis took 30 minutes after previously requiring five experienced engineers for three days. The case study covers one organisation’s deployment, but it gives engineering leaders a concrete workflow and time comparison to examine.