US export bans trigger an open-source surge, DeepSeek builds its own chips, and Grok 4.5 changes the coding game.

Last week, we traced how the US government started deciding who gets access to the next flagship model, with Anthropic and OpenAI both bending to Washington's demands. This week the story moved downstream, from the models to the machines underneath them. DeepSeek began building its own inference chip to cut its dependence on Nvidia and Huawei. SpaceXAI dropped Grok 4.5, a coding-first model trained to beat Claude Opus at a fraction of the token cost. Anthropic pushed Claude Cowork to mobile, turning its agent from a desktop tool into something that works while you sleep. And OpenRouter data showed the Big Three's combined share of AI traffic collapsing from 55% to 33% in six months as open-weight models filled the gap. Follow the leverage, and it is not moving toward whoever has the most restricted model. It is moving toward whoever can run, own, and ship AI regardless of what Washington decides next.

The US government's recent moves to restrict access to top-tier artificial intelligence systems from Anthropic and OpenAI have inadvertently fueled a massive surge in open-source adoption. Following the Trump administration's order for Anthropic to block non-Americans from its Mythos 5 and Fable 5 models, and OpenAI's agreement to let the government vet customers for GPT-5.6, developers and enterprises are seeking alternatives they can control.

This shift is starkly visible in usage metrics. On OpenRouter, the combined market share of Google, Anthropic, and OpenAI dropped from 55% in January to just 33% by June. Meanwhile, Chinese open-weight models like DeepSeek and Zhipu AI's GLM-5.2 are rapidly gaining ground. GLM-5.2, which performs competitively against top US models on several benchmarks, is free to download and run locally, offering a compelling, sanction-proof alternative for businesses wary of sudden access revocations.

The irony is palpable: efforts to maintain a strategic advantage by locking down proprietary US models may be accelerating the global adoption and maturation of foreign open-source ecosystems. As companies prioritize flexibility and cost-control—often finding Chinese open models to be 60% to 90% cheaper—the debate over open versus closed AI has shifted from philosophical to strictly operational.

Read the full story →

That's the combined estimated value of the pending Anthropic and OpenAI IPOs, alongside SpaceX's recent public offering, which together are set to generate more value than all US VC-backed exits since 2000.

Source: Pitchbook/NVCA Venture Monitor

Chinese AI champion DeepSeek is reportedly developing its own AI inference chip to reduce its dependence on Nvidia and Huawei hardware. The move, which began about a year ago, involves hiring chip-design engineers and partnering with external foundries. For global enterprises, this signals a major strategic shift toward vertical integration amid tightening US export controls.

SpaceXAI released Grok 4.5, its smartest model yet, specifically trained alongside Cursor to excel at software engineering and agentic tasks. Priced competitively at $2 per million input tokens, Grok 4.5 claims to beat Claude Opus at coding benchmarks while operating at a fast 80 tokens per second. For engineering leaders, this introduces a highly capable and cost-effective new option for AI-assisted development.

OpenAI unveiled GPT-Live, a new generation of full-duplex voice models that can listen and speak simultaneously. By decoupling continuous interaction from deeper reasoning (handled by GPT-5.5 in the background), the new architecture eliminates awkward pauses and allows for natural interruptions. For customer service and interface design, this sets a new standard for natural human-AI interaction.

Anthropic expanded its Claude Cowork agent to web and mobile platforms for Max subscribers. The update allows users to initiate complex administrative tasks on their desktop, monitor progress via their phone, and have the agent continue working even when the laptop is closed. This marks a critical push to embed agentic AI into broader office workflows beyond just software development.

Former Federal Reserve Chairman Ben Bernanke has joined Anthropic's Long-Term Benefit Trust, an independent body designed to ensure the company adheres to its public mission. The Nobel laureate brings significant gravitas to the trust, which has the power to appoint and remove a majority of the corporate board members. For enterprise buyers, this reinforces Anthropic's positioning as the safety-conscious and governance-focused frontier lab.

ChatGPT for PowerPoint — Now generally available for Enterprise and Edu workspaces, allowing users to build, update, and polish presentations directly within PowerPoint.

Copilot Cowork — Microsoft's agentic system is now generally available worldwide, enabling users to delegate complex, multi-step tasks across Microsoft 365 apps.

Meta Muse Image — Meta's new image generation model, built into Instagram and WhatsApp, allowing users to create and edit images natively within their social feeds.

TechCrunch explores the staggering scale of the impending AI IPOs. With SpaceX already public and Anthropic and OpenAI waiting in the wings, these three companies alone are poised to eclipse the value of every US VC-backed exit since the dot-com bubble. Essential context for understanding the financial gravity of the current AI boom.

If you only do one thing this week: Re-evaluate your model dependencies. With export controls causing sudden outages for top-tier proprietary models, testing an open-weight alternative like DeepSeek or Llama for your core workflows is no longer just an optimization exercise—it's basic business continuity.