An Abu Dhabi Lab Open-Sourced Six AI Models, Training Data Included. India Finished a Rail Project Twenty Years in the Making.

Two very different bets on AI infrastructure landed in the same week. One came from a university lab in Abu Dhabi that decided the way to compete with the giants was to give everything away for free. The other came from a Wall Street trading firm writing a $13 billion check for computing power. In India, meanwhile, a satellite startup closed the country’s biggest ever spacetech round, and a rail project that started before most of its own engineers had finished college finally opened end to end.

The most open AI model line yet doesn’t come from Silicon Valley

Most companies building frontier AI models treat the recipe as the crown jewels. You might get the finished model to download. You almost never get the training data, the code that built it, or the notes on how the whole thing was put together.

The Institute of Foundation Models, a research arm of Mohamed bin Zayed University of Artificial Intelligence based in Abu Dhabi, did the opposite on September 3. It released K2 Horizon, a family of six models ranging from 0.9 billion to 375 billion parameters, under the Apache 2.0 license, and included the weights, the training data, and the training code together. Researchers can not just use these models, they can see exactly how they were built and rebuild them from scratch if they want to check the work.

The models are not just open, they are competitive. IFM says its smallest models, the 0.9 billion, 3.7 billion and 7 billion parameter versions, post the strongest published scores at their respective sizes on a range of coding, math and reasoning benchmarks, ahead of other open models of similar size. The flagship 375 billion parameter model scored 47 on the Artificial Analysis Intelligence Index, a jump of roughly 30 points over IFM’s previous generation. The models are available now through Hugging Face and several inference providers, including Amazon Web Services.

IFM was only launched in May 2025. Less than a year and a half later, its models are treated as a serious open alternative to the more tightly held systems coming out of the largest AI labs, and the choice to publish the training data alongside the weights sets a higher bar for what “open” is supposed to mean in this industry.

A trading firm just bet $13 billion on AI’s plumbing

While IFM was giving models away, a very different kind of AI story played out around data center company Crusoe. According to Bloomberg, Crusoe signed a five-year cloud contract worth roughly $13 billion with Jane Street, the quantitative trading firm, to supply clusters of GPUs and other AI infrastructure. It is the most valuable customer deal Crusoe has landed for its cloud business.

The Jane Street contract helped Crusoe close a $3 billion funding round at a $30 billion valuation, according to TechCrunch and Tech Funding News, up from just over $10 billion a year earlier. The round was co-led by Atreides Management and Valor Equity Partners, with Mubadala Capital, the investment arm of Abu Dhabi’s sovereign wealth fund, also taking part. Crusoe already counts Meta, Microsoft and OpenAI among its customers.

What makes the deal notable is who is writing the check. Jane Street is not an AI company. It is a trading firm that needs raw computing power for its own models, and its willingness to commit $13 billion over five years is a sign of how far AI compute demand has spread beyond the companies actually building chatbots and image generators. The infrastructure underneath AI has become its own market, and firms with deep pockets and no interest in shipping a consumer product are becoming some of its biggest customers.

India’s satellite champion just raised more money than any spacetech startup in the country’s history

Pixxel, the Bengaluru based satellite imaging company, closed a $100 million Series C funding round on September 7, the largest single funding round any spacetech company in India has raised, according to Business Standard and Tech Times. The round was co-led by Temasek and Seraphim, with new investors 360 ONE Asset and IMM Investment joining existing backers Radical Ventures and growX Ventures. The raise takes Pixxel’s total funding to $195 million and values the company at roughly $400 million.

Pixxel builds hyperspectral imaging satellites, the kind that can pick apart the chemical signature of what they are looking at from orbit rather than just taking a picture. The company already holds contracts with NASA and the United States’ National Reconnaissance Office, has won multiple defence innovation challenges from India’s Ministry of Defence, and has unveiled plans for a satellite that would test running AI data centers in orbit. The new funding is earmarked for its next generation Honeybee satellite constellation and for expanding into synthetic aperture radar and ultra high resolution optical imaging.

Founded in 2019, Pixxel has grown from a college project into a company competing for contracts with the biggest space agencies in the world, and this round is the clearest sign yet that Indian investors and sovereign funds are willing to bet serious money on India producing more of that kind of company.

The freight corridor twenty years in the making finally runs end to end

On September 8, Prime Minister Narendra Modi dedicated the final three sections of the Western Dedicated Freight Corridor at an event in Vadodara, Gujarat, completing a 2,843 kilometer freight rail network that stretches from Jawaharlal Nehru Port near Mumbai to Dadri near Delhi, with a parallel Eastern corridor running from Punjab to West Bengal. The three sections dedicated on Monday, covering 326 route kilometers between Sanand and the Jawaharlal Nehru Port Terminal, cost more than 20,700 crore rupees to build, according to multiple Indian outlets covering the event.

The project was first announced in the 2005 to 2006 railway budget. Building it took two decades of land acquisition, engineering and, at points, funding delays, work that outlasted several changes of government along the way. The finished corridor is fully electrified and built to run double stack container trains, meaning two shipping containers stacked on top of each other, on tracks separate from India’s crowded passenger network. That separation matters because it means freight no longer has to wait its turn behind passenger trains, cutting the time it takes goods to move across the country.

“Dedicated Freight Corridor is the lifeline of a rapidly developing India,” Modi said at the event, according to Punjab Kesari, adding that farmers moving perishable produce like flowers, vegetables and fruit would benefit from goods reaching distant markets faster before they spoil.

None of these four stories share an industry or a country, but they share a shape. A lab in Abu Dhabi gave away its work so anyone could check it. A trading firm bought computing power it doesn’t build itself. A satellite company raised money to prove Indian engineering can compete for NASA contracts. And a rail line that took twenty years to finish exists now so that a truckload of mangoes gets to market a little faster. Different scales, same instinct: build the thing under the thing, and worry about who gets the credit later.