Seven Developments Shaping AI and Deeptech

On September 15, TypeSafe AI introduced Jev, its first “System One” model. Jev does not generate free-form prose; it returns structured choices, scores and probabilities for defined tasks such as classification, request routing and tool selection. Its architecture generates outputs in parallel within a predefined structure. At launch, TypeSafe published pricing of $0.042 per million input tokens, with no charge for output tokens, and reported response times of 70–500 milliseconds. These performance figures reflect the company’s evaluations, and the structural guarantee concerns output format, which should be assessed separately from decision accuracy. Jev is designed for software workflows with defined outputs, while general chat and drafting fall outside its stated purpose. Lower decision costs could make AI economical to embed in more business processes, with orchestration, evaluation and specialised workflow design helping translate inexpensive decisions into dependable automation. Read the original announcement.

On September 7, Bengaluru-based Pixxel announced a $100 million Series C, co-led by Temasek and Seraphim, bringing its total funding to $195 million. The investment will support its next-generation Honeybee hyperspectral constellation, synthetic aperture radar and ultra-high-resolution optical imaging, the Aurora Earth-intelligence platform, sovereign Earth-observation programmes, and manufacturing capacity in India and the United States. Honeybee’s first satellite is expected to launch in 2027, extending observations into shortwave infrared. Pixxel says it has deployed six Firefly satellites and launched Aurora; it also reports contracts with NASA and the US National Reconnaissance Office, wins in India’s iDEX challenges, and selection to lead development of India’s first public-private Earth-observation constellation of 12 satellites under IN-SPACe. The expansion links proprietary satellite data with software and government programmes, while recurring customer demand and execution across these planned systems will determine the durability of the business. Read the original announcement.

On September 18, L&T Semiconductor Technologies unveiled 40 products at SEMICON India, including its first 1200-volt silicon-carbide MOSFET platform for EV fast chargers, microgrids, solid-state transformers and traction inverters. The portfolio also spans secure identity, surveillance, communications and compute, including a vision-camera system-on-chip with on-device AI. L&T announced the tape-out of an India-designed BLDC motor controller, meaning the design has been submitted for fabrication. Separately, L&T Semiconductor and Tata Electronics agreed to explore fabrication and packaging opportunities for L&T’s current and future products, alongside design-to-fabrication optimisation. The product launch and partnership address two parts of India’s semiconductor opportunity: designing differentiated chips and developing domestic manufacturing pathways. The partnership remains exploratory, and the products will still need to progress through fabrication, qualification and commercial adoption. Read L&T’s product announcement and its partnership announcement with Tata Electronics.

On September 8, Google DeepMind released AlphaGenome Atlas, a research resource containing predictions for the molecular effects of nine billion possible single-letter changes in human DNA. The one-petabyte dataset covers protein-coding and non-coding regions, and its AlphaGenome Variant Impact score combines predictions from AlphaGenome and AlphaMissense to help researchers prioritise variants. DeepMind says external collaborators used the resource to identify a variant in unsolved rare-disease research and experimentally verified the predicted effect on RNA splicing. In a separate analysis of more than 54,000 UK Biobank participants, a researcher reported finding 22% more associations involving rare non-coding variants by grouping variants according to their predicted molecular effects. Available for academic research through a website portal and API, the Atlas may help researchers focus laboratory work on promising candidates. Its predictions remain research tools that require experimental validation, with commercial differentiation likely to depend on proprietary data, experimental capabilities and validated biological insights. Read the original research announcement.

On September 15, Agility Robotics unveiled Digit 5, a humanoid robot designed for industrial work near people. Its safety architecture uses human detection, visual and auditory signals, and an independent safety controller; Agility is also partnering with NVIDIA’s Halos for Robotics platform and IGX Thor. Announced hardware upgrades include a 50-pound payload, 40% more than its predecessor, a 90-minute runtime with a nine-minute recharge, and interchangeable grippers. Digit 5 builds on more than 65,000 hours of operation logged by Digit 4, including customer deployments. Agility reports more than $300 million in multi-year customer orders for Digit 5 as of May 2026, subject to contractual milestones. Early access is expected in the first half of 2027, with general availability planned by year-end. The safety systems, field experience and customer commitments address important requirements for industrial deployment, while sustained throughput and integration with existing facilities will shape adoption. Read the original announcement.

On September 1, Fervo Energy announced a 396-megawatt power purchase agreement with Google for its Cape Station enhanced geothermal project in Utah, which is expected to come online in 2028. The electricity is designed to support a potential Google data centre, and the agreement gives Google an option to expand its purchases by approximately 600 megawatts, potentially bringing the total to nearly one gigawatt by June 2030. Fervo describes the power as carbon-free and designed for around-the-clock supply. Final data-centre plans remain subject to engineering feasibility, state and local approvals, and commercial conditions, while the additional capacity depends on project development. The agreement gives enhanced geothermal a major prospective customer and demonstrates how long-term power contracts can support development of firm electricity for data-intensive infrastructure. It also creates potential demand for drilling technology, subsurface sensing, power equipment and project development. Read the original announcement.

On September 24, Solcoa announced $75 million in financing for Solcoa One, its first commercial rare-earth metalmaking plant. The financing comprises $45 million in equity led by Bain Capital Ventures and $30 million in debt and equipment financing anchored by J.P. Morgan. The company expects the plant to begin operating in July 2027 with annual capacity of 500 tonnes. Solcoa says its Alameda operation is already producing and shipping more than 10 tonnes of magnet-grade metal annually, and that China controls 95% of global rare-earth metallisation capacity. Metallisation converts refined rare-earth oxides into metals used in high-performance magnets for electric vehicles, robotics and defence systems. The company also points to US defence sourcing rules that expand from January 1, 2027, increasing the importance of compliant supply. If commissioned and operated at its planned scale, Solcoa One would add significant rare-earth metal production outside China; actual output and economics remain to be demonstrated. Read the original announcement.