7 April 2026
Who builds gigawatt AI clusters when Iranian drones are already hitting petrochemical plants next door?
Anthropic partnered with Google and Broadcom to plan gigawatt-scale compute infrastructure, signaling a new phase in the raw power requirements of frontier model training.
Trump set a midnight deadline threatening devastating action against Iran as Israeli strikes hit Tehran, Ahvaz, and Shiraz, while Iranian drones struck Saudi Arabia's SABIC petrochemical complex.
Markets plunged into extreme fear territory as oil prices surged ahead of the midnight deadline, with investors pricing in the possibility of a wider Gulf conflict.
UBTech's $18 million salary offer for a single AI scientist reveals that the talent war now rivals sovereign defense budgets, even as the physical world those models depend on burns.
Click to expand
OpenAI announced the OpenAI Safety Fellowship, a new program to support researchers and practitioners working on AI safety. The initiative will foster knowledge exchange and collaborative projects aimed at improving the safety landscape across the AI community. This marks a concrete institutional investment in safety research talent beyond OpenAI's internal teams. Source →
Anthropic announced a major expansion of its partnership with Google and Broadcom targeting multiple gigawatts of next-generation compute capacity. The deal focuses on large-scale AI processing infrastructure, signaling Anthropic's preparation for training runs far beyond current scale. The partnership aligns with Anthropic's rumored Claude Mythos 5 development, which prediction markets expect before April 30. Source →
Microsoft released its MAI model family on April 6, including MAI-Transcribe-1, which the company claims is the most accurate speech-to-text system available at $0.36 per hour. The launch also includes MAI-Voice-1 for speech synthesis and MAI-Image-2 for image generation, all optimized for speed, quality, and safety in human-centric applications. The release further distances Microsoft's in-house AI efforts from its OpenAI dependency. Source →
Chinese robotics company UBTech is offering an annual salary of $approximately 18 million to recruit a top AI scientist, according to Bloomberg. The figure represents an extraordinary escalation in the global talent war for AI researchers, particularly those working at the intersection of embodied intelligence and robotics. UBTech is a leading humanoid robotics firm based in Shenzhen. Source →
Meta AI released TRIBE v2, a predictive foundation model trained on human brain responses to visual, auditory, and linguistic stimuli. The model builds on Meta's Algonauts 2025 work and represents a step toward AI systems that can predict and simulate neural activity. The research has implications for brain-computer interfaces and neuroscience-informed model design. Source →
Harvard researchers have created microscopic 'neurobots,' living robots equipped with functional nervous systems. Published in IEEE Spectrum, the work represents a convergence of synthetic biology and robotics, producing biological machines capable of sensing and responding to their environment through neural signaling rather than electronic circuits. Source →
Meta published a detailed technical blog on its MTIA custom silicon program, revealing it has shipped four generations of its in-house AI inference chips in just two years. The chips power AI experiences for billions of users across Meta's platforms, from recommendation systems to generative AI features. The pace underscores Meta's commitment to reducing reliance on NVIDIA for inference workloads. Source →
Hugging Face published its State of Open Source report for spring 2026, examining shifts in competition, geography, technical trends, and emerging communities across the open-source AI ecosystem. The report arrives as open-weight model releases have accelerated dramatically, with March 2026 alone logging over 30 distinct launches across multiple labs. The analysis provides a benchmark for how open models are closing the gap with proprietary frontier systems. Source →
OpenAI's GPT-5.4 Thinking variant, which uses test-time compute to 'ponder' problems, scored 75.0% on OSWorld-Verified, surpassing human-level performance and improving 27.7 percentage points over GPT-5.2. The model also reduced claim errors by 33% compared to its predecessor. These results position GPT-5.4 Thinking as the current leader in agentic desktop task completion. Source →
Venture funding for AI companies in Q1 2026 reached a record $267.2 billion, dominated by mega-rounds from OpenAI and Anthropic alongside the landmark SpaceX acquisition of xAI. The figure reflects a systemic industrialization of the AI sector, with capital flowing toward agentic systems and inference infrastructure. The pace of investment has fundamentally altered the competitive dynamics of the industry. Source →
The AI model release cycle has compressed to an average of one major launch every 72 hours, according to multiple tracking sources. March 2026 saw 12 models from six labs in a single week, while April has already delivered releases from Microsoft, Mistral, Anthropic, and Google in its first seven days. Developers now face a monthly, not annual, model selection problem, with specialized models outperforming generalists on narrow tasks by 8 to 14 percentage points. Source →
President Trump threatened to destroy every bridge and power plant in Iran if Tehran does not reopen the Strait of Hormuz and accept a Pakistan-brokered ceasefire by midnight EDT on April 7. At a White House press conference, Trump declared "the entire country can be taken out in one night, and that night might be tomorrow night," dismissing accusations that such strikes would constitute war crimes. Defense Secretary Pete Hegseth and Joint Chiefs Chairman Gen. Dan Caine flanked the president during the briefing. Source →
Iran formally rejected the 45-day truce proposal and countered with 10 clauses demanding sanctions relief, an end to regional conflicts, safe strait passage, and reconstruction commitments. Tehran vowed to ramp up attacks on Persian Gulf energy sites. Over 5,200 people have been killed across the Middle East since the war began on February 28, with Pakistani-mediated talks described as reaching a "critical stage" by Iran's ambassador. Source →
Iranian drones hit a SABIC petrochemical facility in Jubail, Saudi Arabia, igniting a fire at one of the kingdom's largest industrial complexes. The attack marked an expansion of Iranian strikes against Arab states, with Saudi Arabia, the UAE, and Bahrain all issuing safety alerts to citizens. The strike underscores Tehran's strategy of targeting Gulf energy infrastructure as leverage against the US-led coalition. Source →
Israeli airstrikes hit Iranian infrastructure across Tehran, Ahvaz, Shiraz, and other cities, with combined US-Israeli operations reportedly killing over 25 people. A synagogue in Tehran was heavily damaged in the strikes. Separately, Israel struck Hezbollah sites in Beirut as the conflict continued to widen across multiple fronts. Source →
Global futures dropped and oil prices rose for a third consecutive day as investors braced for Trump's midnight EDT deadline. The escalation in Gulf energy infrastructure attacks, including the SABIC strike in Jubail, intensified fears of prolonged supply disruption. Markets are pricing in the possibility of a dramatic US military escalation against Iran beginning as early as Wednesday at 0400 GMT. Source →
Ukrainian President Volodymyr Zelenskyy traveled to Damascus and pledged deeper security cooperation with Syrian interim President Ahmad al-Shara. The visit signals a broader regional realignment as the Middle East conflict reshapes diplomatic relationships. Details of specific agreements were not immediately disclosed. Source →
NASA's Artemis II crew returned safely to Earth after completing a record-breaking mission that took astronauts behind the moon. The mission, which launched on April 2, marked the first crewed lunar flyby in over 50 years. The successful splashdown came amid global attention focused on the Middle East crisis. Source →
Ben Roberts-Smith, a decorated Australian special forces veteran, was arrested in Sydney on charges of war crimes allegedly committed in Afghanistan between 2009 and 2012. Roberts-Smith, a Victoria Cross recipient, had previously been the subject of a high-profile defamation trial in which a court found media reports of his alleged unlawful killings were substantially true. Source →
We read and recommend these newsletters. Here is what they are covering today. Subscribe to them directly for the full experience.
"In a talent market where Silicon Valley giants have reportedly put $20M annual pay packages, and even $100M-plus incentives, on the table for elite AI researchers, the price of winning is starting to look absurd."Read the full newsletter →
"Anthropic cited engineering challenges due to surging demand for Claude as the company tries to 'be more intentional with managing growth.'"Read the full newsletter →
"Most people jump to the model when discussing continual learning, but in reality, an AI system can learn at all three of these levels."Read the full newsletter →
"Instead of spending 40 hours on competitive research, you're spending 1 hour setting up your agent, and receiving valuable intel that help you make the best decision for your business."Read the full newsletter →
"77% of CFOs want AI layered onto existing systems rather than ripping and replacing."Read the full newsletter →
Anthropic announced a major expansion of its partnership with Google and Broadcom to build out multiple gigawatts of next-generation compute infrastructure, signaling a dramatic scaling of training and inference capacity. The deal complements the company's reported pursuit of a $3 to $5 billion funding round at a potential $170 billion valuation. With Claude Mythos 5 and the Operon lab agent already shipped this week, the compute partnership positions Anthropic to sustain its aggressive release cadence through the rest of 2026.
Source →OpenAI introduced the OpenAI Safety Fellowship, a new initiative designed to support external researchers and practitioners working on AI safety through knowledge exchange and collaborative projects. The program arrives as the company's corporate restructure remains stalled into 2026 over unresolved negotiations with Microsoft on IP rights and Azure exclusivity. With GPT-5.4 iteration ongoing but no new model launch this cycle, the fellowship signals OpenAI is investing reputational capital in safety governance while its product roadmap holds steady.
Source →Mistral launched Mistral Forge, an enterprise platform that lets organizations train custom AI models from scratch on proprietary data, with early adopters including ASML, Ericsson, the European Space Agency, and several defense-adjacent firms. CEO Arthur Mensch stated the company is on track to surpass $1 billion in annual recurring revenue this year, a milestone that would make Mistral the first European AI lab to reach that threshold. Combined with the Boxrol TTS voice model shipped earlier this week, Mistral is executing a rapid product expansion across both foundation models and vertical tooling.
Source →Meta released TRIBE v2, a predictive foundation model trained on human brain responses to visual, auditory, and linguistic stimuli, building on its Algonauts 2025 work and marking a notable push into neuroscience-adjacent AI research. Separately, the company detailed its MTIA custom silicon progress, noting four chip generations shipped in two years to power AI experiences for billions of users across Meta's platforms. While Meta has not launched a new flagship language model this cycle, its infrastructure and research investments continue to broaden the scope of its AI portfolio.
Source →Get tomorrow's brief delivered to your inbox.
One email per day. Unsubscribe anytime.