Cohere and Aleph Alpha merge in a $20 billion sovereign AI deal

Cohere and Aleph Alpha announced a $20 billion merger to create a transatlantic sovereign AI alternative backed by Germany’s Schwarz Group, while Anthropic’s internal marketplace experiment demonstrated that higher-capability AI agents extract measurably better deals for their users – with the disadvantaged side unaware of the gap. DeepSeek delayed V4’s full launch to prioritize Huawei Ascend chip optimization, signaling an accelerating shift toward domestic AI infrastructure in China.

Funding & Business #

Why Cohere Is Merging with Aleph Alpha #

TechCrunch / Globe and Mail / CNBC / Sifted

Canadian AI startup Cohere is acquiring Germany’s Aleph Alpha at a combined valuation of approximately $20 billion, with Schwarz Group (owner of Lidl) providing EUR 500 million in financing and the merged entity operating on Schwarz’s STACKIT sovereign cloud. Backed by both the Canadian and German governments – who recently launched a Sovereign Technology Alliance – the deal positions Cohere as the leading non-US enterprise AI provider for Europe’s highly regulated sectors including defense, finance, and government. The strategic question is whether European institutions will treat a Canadian-German entity as sufficiently sovereign, particularly if Cohere eventually goes public in the US.

Cognition AI in Funding Talks at $25 Billion Valuation #

Bloomberg / SiliconANGLE

Cognition AI, maker of the AI software engineer Devin, is in early talks to raise hundreds of millions at a $25 billion valuation – more than doubling from its $10.2 billion September round led by Founders Fund. The jump, alongside Cursor’s parallel $50 billion valuation talks, reflects investor conviction that AI coding agents are becoming the primary delivery mechanism for enterprise software development. With clients including Dell and Cisco, Cognition represents the bet that autonomous coding agents will capture a larger share of development budgets than copilot-style assistants.

Developer Tools #

Project Deal: Anthropic’s Agent-to-Agent Marketplace Reveals Model Capability Gaps #

Anthropic / TechCrunch / Cybersecurity News

Anthropic ran a one-week classified marketplace experiment with 69 employees, each represented by a Claude agent with a $100 budget and no human intervention during negotiations. The system facilitated 186 transactions totaling over $4,000. The critical finding emerged when agents were randomly assigned different model tiers: Opus users closed roughly two more deals on average and extracted $3.64 more per item than Haiku users – yet participants on the disadvantaged side did not perceive unfair treatment. As agent-mediated commerce becomes more common, this demonstrates a concrete mechanism for hidden economic inequality driven by model quality that users cannot detect or correct for.

Security #

OpenAI CEO Apologizes for Not Alerting Law Enforcement About Mass Shooting Suspect #

TechCrunch

Sam Altman issued a public apology to the community of Tumbler Ridge, Canada, after an 18-year-old suspect killed eight people. OpenAI had banned the individual’s ChatGPT account in June 2025 after detecting discussions of gun violence, and internal staff debated contacting police but ultimately decided against it. The incident exposes a critical gap in AI governance: companies can detect concerning behavior patterns but lack established protocols for law enforcement escalation, and the consequences of inaction can be catastrophic.

Model Releases #

DeepSeek Delays V4 Full Launch to Prioritize Huawei Chip Optimization #

Bloomberg / NewsBytesApp

DeepSeek is postponing the full V4 launch – currently available only as a preview – to deepen optimization for Huawei’s Ascend AI processors, including the upcoming Ascend 950. The company spent months reworking its software stack to go beyond basic Ascend compatibility toward hardware-specific tuning, and expects to lower V4-Pro prices later this year as Huawei scales up production. The delay aligns with Beijing’s push for AI supply chain self-reliance and, if successful, would make DeepSeek V4 the first frontier-competitive model fully optimized for non-NVIDIA silicon – a significant milestone in the decoupling of the global AI hardware stack.

Research & Papers #

Amateur Solves 60-Year-Old Erdos Problem Using ChatGPT #

Scientific American / Hacker News (393 points)

Liam Price, a 23-year-old without advanced mathematics training, solved a 60-year-old Paul Erdos conjecture about primitive sets by entering the problem into GPT-5.4 Pro – an approach he describes as “vibe mathing.” The solution employed a method that mathematicians had not previously considered; Terence Tao noted there was “some kind of mental block” preventing human progress. This is among the strongest demonstrations that LLMs can contribute genuinely novel mathematical techniques, not just formalize or verify known approaches, though the reproducibility and generalizability of the “vibe mathing” methodology remains unestablished.

Transformers Are Inherently Succinct #

arXiv / Lobsters

Proves that transformers can represent formal languages substantially more succinctly than standard representations like finite automata and Linear Temporal Logic formulas. As a by-product of this expressivity result, verifying properties of transformers is shown to be EXPSPACE-complete – meaning formal verification of transformer behavior is provably intractable. For teams working on AI safety through formal methods, this establishes a fundamental theoretical barrier: the same expressivity that makes transformers powerful also makes them inherently resistant to exhaustive verification.

Regulatory & Policy #

Maine Governor Vetoes First US Statewide Data Center Moratorium #

TechCrunch

Governor Janet Mills vetoed L.D. 307, which would have imposed the country’s first statewide moratorium on new data centers until November 2027. Mills objected specifically because the bill lacked an exemption for a project in Jay, Maine that has strong local support – signaling that even sympathetic governors will resist blanket restrictions when specific communities want the development. Similar proposals are advancing in New York and other states, but this veto suggests that project-level exemptions will be a key battleground in future data center regulation.

Threads to Watch #

Agent-mediated commerce is a coming inequality vector. Anthropic’s Project Deal experiment quantified what many have theorized: when AI agents negotiate on behalf of users, model quality directly translates to economic advantage, and the disadvantaged party cannot detect the gap. As coding agents like Devin and Cursor command $25-50B valuations, the infrastructure for agent-to-agent transactions is being built faster than the norms to govern them.

Sovereign AI is consolidating around bilateral partnerships. The Cohere-Aleph Alpha merger, backed by Canadian and German governments, follows the same logic driving DeepSeek’s Huawei chip optimization: reducing dependence on US-controlled AI infrastructure. The emerging pattern is bilateral partnerships – not unilateral national champions – as the vehicle for AI sovereignty.

AI safety governance is failing the escalation test. OpenAI’s Tumbler Ridge case and Anthropic’s Project Deal both expose the same underlying problem from different angles: AI systems can detect or create consequential outcomes, but the institutional frameworks for responding to those outcomes – whether escalating threat signals to law enforcement or ensuring fair agent-mediated transactions – do not yet exist.

Sources Unavailable Today #

These sources could not be fetched today. Links point to their homepages so you can check them directly.

• Stanford HAI – scrape: content_truncated