Good morning. Today is a five-story brief: OpenAI and Microsoft formally retire the partnership terms that defined the GPT era, China hands Meta its first major AI-deal defeat, Google DeepMind plants a sovereign-AI flag in Seoul, OpenAI's hardware ambitions get a date, and Meta signs the most exotic energy contract in the AI buildout so far. If you'd rather get the highlights by email once a week, subscribe to the weekly brief.
1. OpenAI and Microsoft sign an amended agreement that resets the partnership for the AGI era
OpenAI and Microsoft jointly announced the next phase of their partnership, an amended agreement that "simplifies the partnership, adds long-term clarity, and supports continued AI innovation at scale." The headline change: the original 2019 exclusivity structure — under which Microsoft had first-refusal rights on OpenAI compute and on commercializing OpenAI's IP — has been retired in favor of a cleaner contractual framework that contemplates an eventual AGI declaration without triggering a chaotic unwind.
The redo was not a surprise; it had been telegraphed for months as OpenAI's structure shifted toward the public-benefit corporation it spent late 2025 negotiating with Delaware regulators. What's new is that both companies now have a written settlement of how revenue share, model access, and the AGI clause work in practice — rather than hoping a 2019 letter still fits a 2026 reality.
Why it matters. The 2019 deal was the single most consequential commercial agreement in modern AI: it underwrote the Azure-hosted GPT family that defined the last cycle. Resetting it cleanly removes the largest unresolved governance risk hanging over both companies — and over the broader enterprise AI stack that runs on Azure OpenAI Service. Expect Microsoft customers to get faster confirmation that their long-running deployments are not at legal risk from any future AGI declaration.
What to do. If your AI roadmap depends on Azure OpenAI Service contracts, this is the trigger to ask your account team for the updated commercial language — particularly around exclusivity carve-outs, custom-model access, and whether the new agreement changes the IP terms on fine-tunes you've shipped.
2. China orders Meta to unwind its $2 billion Manus acquisition
China's State Administration for Market Regulation has ordered Meta to unwind its multibillion-dollar acquisition of AI agent startup Manus, ending a months-long antitrust review with an outright veto, according to TechCrunch. Meta announced the roughly $2 billion deal earlier this year as a centerpiece of its push into agentic AI; Manus had built one of the most-discussed Chinese-origin agent products before relocating its headquarters to Singapore.
The block is unusual in scope. Beijing has repeatedly used merger-review delays as a foreign-policy lever, but a flat unwind order against a US tech giant — for a target that had already restructured outside mainland China — signals that the regulator views agent-platform consolidation as strategically sensitive in its own right, not just as a generic foreign-investment concern.
Why it matters. Every US lab and hyperscaler with a 2026 agent roadmap has been quietly looking at the Chinese agent ecosystem as an acquisition surface — both for talent and for the consumer-facing agent UX patterns China is iterating on faster than the West. If Manus can be vetoed after relocating to Singapore, the assumption that "we'll just buy them after they leave the mainland" is dead. Expect more US-side investment in domestic and European agent startups as a result, and expect any China-rooted agent company entertaining a US offer to receive harder questions about jurisdictional risk during diligence.
3. Google DeepMind signs a sovereign-AI partnership with the Republic of Korea
Google DeepMind announced a partnership with the Republic of Korea to "accelerate scientific breakthroughs using frontier AI models." The framing is squarely sovereign-AI: government-backed access to frontier capability for use in Korean scientific research, with the implication that DeepMind's models will be available to Korean institutions in a configuration distinct from the standard commercial Google Cloud offering.
The deal lands in the same week as the Cohere/Aleph Alpha sovereign-AI merger we covered yesterday, and it points the same direction: governments outside the US are buying preferred access to frontier model capacity rather than waiting for it to arrive through ordinary cloud channels. Korea — home to Samsung, SK Hynix, and a substantial public R&D budget — is a particularly strategic anchor customer for DeepMind given the country's centrality to the global semiconductor supply chain.
Why it matters. The map of who-runs-which-frontier-model-where is being redrawn in 2026 along government lines, not just commercial ones. If you sell into regulated Korean industries — finance, healthcare, defense, public sector — expect Gemini-family models with Korea-specific deployment guarantees to become a more visible procurement option over the next two quarters. The mirror question for US-headquartered enterprises is whether the same models will continue to be uniformly available everywhere, or whether deployment regions will start to fragment.
4. OpenAI's agent-first phone targets 2028 mass production
An analyst note circulated this morning pegs OpenAI's Jony Ive-collaborated hardware project as a phone-class device that could enter mass production in 2028, with AI agents replacing the conventional app launcher as the primary user-interaction surface, per TechCrunch. OpenAI has not commented publicly on the form factor, but the company's 2024 acquisition of Ive's design firm io and a steady drumbeat of supplier reporting through 2025 had already pointed at a phone-class product as the most likely first shipment.
The strategic read here is interface, not silicon. OpenAI's bet is that the home screen — the grid of apps that has organized mobile computing since 2007 — gets replaced by a single conversational agent that fans out to apps as services rather than destinations. That's the same thesis Apple's incoming CEO John Ternus is reportedly counter-positioning against (per yesterday's brief), and it's why a 2028 launch matters now: it sets the timeline both Apple and Google have to beat with their own agent-first hardware refreshes.
Why it matters. If you build a consumer app, the hardware surface you'll be designing for in 2028 may not have an icon. Start asking now what "your product" looks like when its only user-facing artifact is a tool description an agent reads and a pricing page a buyer never visits. The MCP-style commerce specs we covered earlier this week are the early version of that contract.
5. Meta inks the first hyperscaler contract for solar power beamed from space
Meta signed a power purchase agreement with Overview Energy for solar power generated in low-Earth orbit and beamed to a ground receiver — the startup's first commercial contract. The volume is modest by data center standards, but the structural pitch is the part to watch: orbital solar arrays generate around the clock, sidestepping the duck-curve and battery-storage costs that have made terrestrial solar a partial answer to AI's 24/7 power profile.
Space-based solar has been a 50-year "five years away" technology. The reason it's plausibly real now is the same reason everything else in orbit is real now: Falcon-9 and Starship have collapsed the dollar-per-kg-to-LEO number to the point where the construction logistics finally close.
Why it matters. The data center capacity question is the AI buildout's hardest physical bottleneck — yesterday's Maine veto was the political face of the same problem. Hyperscaler willingness to sign the first contracts for genuinely exotic round-the-clock power tells you how much they expect demand to outrun the grid over the next decade. Overview Energy, like every space-power outfit, is a long-shot startup; the signal is that Meta now thinks even long-shot supply is worth locking up early.
What to take from today
Two threads. First, the partnership architecture of the AI industry is being rewritten in plain sight: OpenAI/Microsoft retiring 2019 terms, Cohere absorbing Aleph Alpha under Schwarz Group, DeepMind signing Korea, and China vetoing Meta's Manus deal are all part of the same restructuring — frontier AI is being repackaged as a strategic asset that nation-states want under their own commercial logic. Second, the 2028 hardware pin OpenAI just put down on its agent phone forces every other consumer-tech roadmap to pick a side: app-grid or agent. Three years is a short runway for that pivot.
Tomorrow's brief lands at 08:00 UTC. If you'd rather get this in your inbox once a week — just the stories that actually matter — subscribe here.