- AI KATANA
- Posts
- OpenAI and Microsoft Rewrite Mega-Deal to Clear IPO Path
OpenAI and Microsoft Rewrite Mega-Deal to Clear IPO Path
Also: Alibaba Says It Can Train Search Models for 12¢ on the Dollar
AI KATANA
May 12, 2025
Hello!
A whirlwind few days in AI saw OpenAI and Microsoft redraft their multibillion-dollar pact to clear a path to an IPO while still guaranteeing Redmond first dibs on future models; Amazon rolled out its “Vulcan” warehouse robot and, crucially, a tuition-free program to turn pickers into robotics techs; Washington drama flared as the US Copyright Office chief was abruptly fired a day after she questioned AI training practices; Alibaba researchers claimed an 88 per cent cut in search-model training costs, signalling China’s big tech is ruthlessly hunting efficiency; regulators quietly closed a headline investigation into Scale AI, easing pressure on one of the sector’s most valuable startups; and industry chiefs warned senators that without faster permits for power lines and wider access to federal data, the US risks ceding the AI lead. Dive in for the details, tool launches, fresh funding, and one sweet Candy Crush twist.
Sliced just for you:
🤝 OpenAI and Microsoft Rewrite Mega-Deal to Clear IPO Path
🦾 Amazon’s ‘Vulcan’ Robot Hints at Tomorrow’s Warehouse Jobs
🚫 White House Ousts Copyright Boss After AI Training Clash
🏯 Alibaba Says It Can Train Search Models for 12¢ on the Dollar
🛡️ Labor Dept Drops Probe Into Scale AI’s Worker Practices
⚡ Tech Titans Urge Senate to Fast-Track Permits for AI-Hungry Power
OpenAI and its primary backer are renegotiating a partnership that has already poured more than $13 billion into the ChatGPT creator. Talks centre on swapping a smaller Microsoft equity stake for extended access to any breakthrough AI developed after 2030, smoothing OpenAI’s ambition to float shares while keeping Azure first in line for premium models. The discussion also updates a 2019 revenue-sharing contract and follows OpenAI’s plan to shrink the cut it sends Microsoft as part of an ongoing corporate restructuring. Analysts say the move signals both firms’ confidence that blockbuster, capital-hungry models will keep coming—and that neither wants a messy, exclusive tie-up to spook regulators ahead of a public listing. Investors now frame the revised deal as the last major hurdle before bankers can sketch an S-1.
Amazon gave reporters a tour of its new “Vulcan” picker, a dexterous arm that can “feel” soft items and reach awkward shelves. The headline wasn’t the robot—it was the retraining pipeline: select associates are already earning Amazon-funded certificates to become robotics technicians, safety analysts, and system calibrators. Management says robots now touch 75 % of orders, creating “hundreds of new categories of work.” Labour economists view the program as a bell-weather: rather than full automation, retail logistics may blend bots with upskilled humans who maintain, audit, and fine-tune them. Critics counter that the tech still removes far more roles than it creates and pushes injury risks onto remaining staff. Either way, Vulcan shows how quickly embodied AI is graduating from pilot to production.
Hours after a scathing report suggested limits on fair-use defences for data scraping, the US President dismissed Register of Copyrights Shira Perlmutter. Lawmakers allege the firing is retaliation for her reluctance to “rubber-stamp” unrestricted dataset harvesting for model training. The sudden move has ignited debate over whether copyright enforcement—or the lack of it—could shape America’s AI edge. Policy insiders fear a chilling effect on regulators just as lawsuits over AI-generated works pile up. Meanwhile, creators’ groups celebrate Perlmutter’s stand, saying unchecked scraping imperils entire industries. Observers expect Congress to fast-track hearings on where the training-data line should be drawn.
Researchers at the Chinese e-commerce giant unveiled “ZeroSearch,” a simulation-driven approach that pares real-time query costs by 88 %. Instead of pinging external engines, models learn via synthetic search sessions generated from proprietary knowledge graphs. The team claims comparable relevance with drastically lower API bills—a potent pitch for domestic firms squeezed by tight budgets and export-control worries. If reproducible, the method could slash the entry fee for specialised vertical search bots and further localise China’s AI stack. Investors note the timing: Beijing is pushing for “AI self-reliance,” and cheaper training fits the mandate. Skeptics, however, question how well simulated clicks reflect messy user intent.
A year-long federal investigation into wage and overtime compliance at data-labeling powerhouse Scale AI has been closed with no penalties, according to sources briefed on the matter. The company—valued near $14 billion and supplying datasets to OpenAI, Meta and the Pentagon—had faced scrutiny over whether crowd annotators were mis-classified. Resolution removes an overhang as Scale AI accelerates expansion, including new Middle East offices and deeper government contracts. Labour advocates still urge thorough audits of gig-style labeling across the industry, calling the closure “a warning, not a clean bill of health.”
In written testimony for a Commerce Committee hearing titled “Winning the AI Race,” leaders from Microsoft, OpenAI, AMD and CoreWeave warned lawmakers that US data-center demand could triple electricity loads within three years. They called for streamlined permitting for new transmission lines, predictable renewable-energy incentives, and an open-data mandate to let startups train on federally held troves. Without action, they argued, grid bottlenecks and data hoarding could drive developers offshore. The hearing spotlights a new policy flashpoint: AI’s carbon-and-kilowatt appetite may now be as strategically sensitive as chips.
🛠️ AI tools updates
Mistral AI has introduced a new suite of enterprise-grade tools, including the Medium 3 model family and Le Chat Enterprise chatbot, aiming to provide efficient and cost-effective AI alternatives to offerings from industry giants like OpenAI and Google. The Medium 3 lineup—available in Small, Standard, and Large versions—offers high performance with low computational demands, capable of handling up to 128,000-token context windows for complex tasks. Le Chat Enterprise is a privacy-focused business assistant designed for professional rather than consumer use. Supported by €385 million in funding and a team of ex-DeepMind and Meta talent, Mistral is blending open-source innovation with commercial strategy. If the company's performance claims are validated, its tools could reshape enterprise AI by shifting competition toward practical value over hype.
Amazon has unveiled a new generative AI-powered tool called Enhance My Listing (EML), designed to help sellers automatically update and optimize existing product listings with minimal effort. Building on the success of previous listing creation tools, which over 900,000 sellers now use, EML allows sellers to refine product content using AI-generated recommendations for titles, attributes, and descriptions based on customer insights and shopping trends. Sellers can input simple descriptions, images, or URLs, and the system generates rich, Amazon-style listings, improving discoverability and engagement—often accepted with little to no edits. Integrated with Amazon Bedrock, EML provides up-to-date, personalized suggestions that sellers can review or modify, streamlining the process of maintaining relevant and appealing listings. This launch follows the rollout of other AI tools like Project Amelia, personalized recommendations, and automated A+ content generation, all aiming to empower sellers and enhance their operational efficiency within Amazon’s marketplace ecosystem.
💵 Venture Capital updates
London-based SaaS provider Ascendx Cloud secured $60 million in equity and $50 million in debt to expand its AI-enhanced customer-data tools across North America and fund acquisitions. The raise underscores investor appetite for enterprise AI platforms that promise immediate revenue lift.
A startup still in stealth mode scored a Series A after its unnamed diffusion model outperformed DALL-E and Midjourney on a leading benchmark. The funds will scale training compute and hire researchers ahead of a public reveal later this year.
🫡 Meme of the day
⭐️ Generative AI image of the day
Before you go, Candy Crush Levels Up with AI-Tuned Puzzles
