AMD just launched its first rack-scale AI system and landed Microsoft as a customer. Netflix dropped $587 million on Ben Affleck’s AI filmmaking startup. And Google and Apple are fighting the EU over giving rival AI assistants access to your phone. Three very different stories, one theme: the infrastructure, the money, and the regulation of AI are all accelerating at once. Let’s dig in.

AMD Helios Takes Aim at Nvidia’s 95% Monopoly

AMD launched Helios, its first rack-scale AI system, on Monday and added Microsoft as a buyer alongside Meta, OpenAI, and Oracle. This is AMD’s most direct challenge yet to Nvidia’s stranglehold on data center GPUs. Each Helios rack packs 72 AMD Instinct MI400X accelerators connected by AMD’s Infinity Fabric, and Microsoft plans to deploy it on Azure for frontier model inference.

Shipments to all buyers start in the second half of 2026. The timing matters: Nvidia’s Vera Rubin platform was just announced for post-training workloads, and the two are now on a collision course. AMD’s pitch is simple: an alternative to Nvidia’s walled garden, with Microsoft as its biggest validation yet. This is infrastructure play, not feature play. AMD doesn’t need to beat Nvidia on every benchmark. It just needs to be good enough that hyperscalers have a second source.

Netflix’s $587 Million Bet on AI Filmmaking

Netflix paid $587 million in cash for InterPositive, the AI filmmaking startup co-founded by Ben Affleck, according to a newly disclosed regulatory filing. The acquisition closed in March but the price tag only surfaced now. InterPositive develops AI tools for filmmakers, and the deal signals that streaming giants see AI production tools as a strategic asset worth real money.

That’s not a vanity acquisition. Netflix is betting that AI-assisted production can lower costs and speed up content creation at a time when every streaming dollar is under scrutiny. Affleck’s startup was founded in 2022 and had been quietly building tools for script analysis, storyboarding, and post-production. The $587 million price tag tells you Netflix thinks this is the future of content creation, not a side experiment.

Google and Apple Push Back Against EU AI Assistant Rules

The European Commission wants Android users in the EU to be able to activate competing AI assistants by voice and let them operate apps with permissions similar to Google’s Gemini. Google must broaden Android access for third-party AI agents by July 2027. Apple has already delayed its new Siri assistant in the EU, citing privacy concerns.

Google argues that outside assistants would need sensitive device permissions that could expose personal and business data. Kent Walker, president of global affairs at Google, also said mandatory search-data sharing could create national security risks. Privacy specialists acknowledge the risks but point out that tech companies’ objections deserve scrutiny when privacy arguments also protect their market position. Omdia estimates about 427 million Android phones and iPhones are subject to EU rules, and ChatGPT is already installed on about 30% of EU smartphones. The Digital Markets Act is shifting from app distribution battles to control of AI services embedded in operating systems.

Kimi K3: 2.8 Trillion Parameters and a US-China Flashpoint

Moonshot AI’s Kimi K3 model has reopened the US fight over open AI. At 2.8 trillion parameters with a 1-million-token context window, it scores close to Claude Opus 4.8 and GPT-5.5 on general intelligence while leading on several coding and agentic benchmarks. The weights will be released under a Modified MIT license on July 27.

An OpenAI executive warned that Washington could target Chinese AI systems with regulatory uncertainty, and the benchmark results have revived concern that China is narrowing the US lead. This is the same pattern we’ve seen before: a Chinese lab releases competitive open weights, US policymakers panic, and the debate over open vs. closed AI gets another jolt. Kimi K3 uses Kimi Delta Attention and Attention Residuals to handle its massive scale efficiently. The question isn’t whether it’s good. It’s whether open-weight models from China change the geopolitical calculus.

NVIDIA Goes All-In on Edge AI and Post-Training

NVIDIA had a busy week on multiple fronts. It launched Cosmos 3 Edge, a 4-billion-parameter open-source world model for physical AI on edge devices, at SIGGRAPH 2026. It’s an omnimodel that handles text, image, video, ambient sound, and action for robotics, autonomous vehicles, and smart infrastructure. The Vera Rubin platform was positioned around a new metric: intelligence per dollar for post-training workloads, targeting the reinforcement learning and fine-tuning loops that power agentic AI. And the new Jetson Thor T2000 and T3000 modules bring up to 400 TFLOPS of AI performance to smaller, power-efficient robotics and edge systems.

Three launches, one strategy: NVIDIA is defending its data center dominance while expanding into the edge where the next wave of AI workloads will run. Cosmos 3 Edge being open source is notable. NVIDIA usually keeps its models close. Opening this one suggests they want to set the standard for physical AI before anyone else does.

Quick Hits

Anthropic opened applications for its AI for Science rare disease research grants, offering up to $50,000 in Claude credits to researchers studying rare genetic diseases. Also launched Claude for Teachers and committed $10 million to Canadian AI research.

Hugging Face disclosed a July 2026 security incident (details still emerging). AllenAI published lessons from building Shippy, their agent platform. IBM Research dropped a deep dive on model routing complexity. And the Real World VoiceEQ benchmark launched to measure human-quality voice AI.

Microsoft announced it’s expanding Azure AI and HPC infrastructure with AMD, though details were thin. Also published on new agentic capabilities in Microsoft 365 Copilot and Dynamics 365 for customer service, and how AI agents can train their own skills.

Meta AI released Muse Spark 1.1 (July 9) and Muse Image and Muse Video (July 7), continuing its push into generative media. Also published Brain2Qwerty research on non-surgical brain-to-text communication.

Google DeepMind continues its steady drumbeat of model releases and research partnerships, though no major new announcements this week beyond what was already covered.

Rundown for July 21, 2026. Sources: Yellow, Anthropic, Hugging Face, NVIDIA, Microsoft AI, Meta AI, Google DeepMind.