Research — JULY 25, 2025
Generative AI Digest: Significant investment momentum
June brought significant investment momentum, highlighted by Meta Platforms Inc.'s backing of data labeling startup Scale AI Inc. Among other standout moves were the $80 million acquisition of eight-person coding startup Base44 Inc. and a massive $2 billion seed round for Thinking Machines Lab Inc. The latter's business model remains unclear beyond its well-regarded leadership team.
On June 6, Apple Inc.'s AI research division released a notable paper, "The Illusion of Thinking," which took aim at the "reasoning" capabilities of current language models. The publication has sharply divided opinion, with its suggestion that reasoning models faced limitations that meant they were unable to deal with highly complex problems. Large language model (LLM) skeptics have used it to bludgeon enthusiasts, while some critical researchers suggest the paper merely serves a strategic purpose — that by casting doubt on the entire field, Apple addresses the perception that it is trailing. The timing of the paper was certainly interesting, front-running Apple's annual Worldwide Developers Conference event. While many new AI-enabled capabilities were showcased, reports noted that the Apple Intelligence vision presented at the keynote in 2024 has yet to be fully realized.
Product releases and updates
In late May, Odyssey unveiled its interactive video technology for the first time. The startup, founded by autonomous vehicle veterans Oliver Cameron and Jeff Hawke, creates dynamic video content that responds to user interaction. To date, Odyssey has raised a total of $27 million in funding, including an $18 million series A round led by EQT Ventures. The most notable aspect of the showcase was its "real-time" generation capabilities, with the startup reportedly able to generate a new frame in 40 milliseconds.
Apple showcased AI-powered image analysis technology, a feature that will become available with iOS 26, alongside live translation capabilities and an AI workout coach, at Apple's WWDC. Much of the focus around the event was the lack of emphasis on Siri, Apple's long-standing "personal intelligence" assistant, with many commentators having expected upgrades to take center stage.
Midjourney Inc. has launched its V1 Video model, an image-to-video capability that complements its popular image-generation models. Access to the model requires a subscription and is available exclusively through Midjourney's Discord desktop app.
Advanced Micro Devices Inc. announced the Instinct MI350 Series GPUs, including the MI350X and MI355X, alongside benchmarks suggesting significant improvements in AI compute and inference performance. It also previewed its next-generation Helios AI rack built on upcoming MI400 Series GPUs and Zen 6-based EPYC Venice CPUs, and announced a new version of open-source AI software stack ROCm, ROCm 7.
Mistral AI SAS announced the release of Magistral, the startup's first reasoning model, with a smaller open-source version or a larger commercialized "enterprise" variant. The most notable strength of Magistral Medium appears to be the speed of its reasoning process.
In June, another high-performing, low-cost large language model emerged from a Chinese startup, MiniMax-M1. This open-weight model is presented as having strong performance in long-context reasoning and agentic tool use. Its standout feature, a 1 million-token context window, has drawn significant attention, matching the upper limit offered by Google LLC's Gemini 2.5 Pro.
Beijing-based Moonshot AI launched a research agent, Kimi-Researcher, which it suggests was trained through a new end-to-end reinforcement learning approach. The agent appears to perform well against the infamously challenging Humanity's Last Exam benchmark.
Black Forest Labs Inc. has launched FLUX.1 Kontext, a family of models for generating and editing images using text and image prompts. This new model set enables in-context image generation, as well as enhanced character and asset consistency across iterative generations. The FLUX.1 Kontext [pro] model is presented as offering faster editing, while the company suggests the FLUX.1 Kontext [max] model has enhanced prompt adherence and typography. In addition, FLUX.1 Kontext [dev], an open variant, was announced in private beta, which can be deployed locally.
Hitachi Everflex AI Data Hub as a Service was announced by Hitachi Vantara. It is a managed service for AI data preparation and governance in the hybrid cloud built on Virtual Storage Platform One and VSP 360. It consists of a data lake house with integrated tools for AI, business intelligence and data management.
Reasoning model o3-pro was released by OpenAI LLC, with this "upgraded" variant given access to tools, allowing the model to analyze files and access the web, for example. Paid users were also treated to an enhanced Advanced Voice mode in ChatGPT, now able to translate between languages — reportedly with improved cadence.
International Business Machines Corp. announced that it will be more deeply integrating Guardium AI Security with watsonx.governance to offer a more unified approach to managing AI security and compliance. New features, developed with AllTrue.ai, are designed to allow detection of where AI is in use across cloud and embedded systems, triggering automated governance workflows. Additional updates include automated red teaming and customizable security policies designed to mitigate risks like data leakage and code injection.
In late May, web browser company Opera Ltd. announced an "agentic" browser, Opera Neon. The browser is designed to allow actions to take place autonomously, with the company suggesting it could be used to write a website or do your shopping on your behalf.
Pure Storage Inc. recently announced a series of new arrays at its //Accelerate annual conference, targeting AI workloads. These include FlashBlade//S R2 (the vendor's certified storage for NVIDIA DGX BasePOD and DGX SuperPOD), FlashArray//ST and FlashArray//XL R5, which supplement Pure's FlashBlade//EXA array (announced earlier this year). The vendor also unveiled Enterprise Data Cloud, a framework for unifying data in Pure environments that includes a unified data plane and intelligent control plane.
Funding and M&A
Meta has poured $15 billion into Scale AI, a data-labeling startup, and brought co-founder Alexandr Wang onboard. The move is seen as an effort to regain momentum in the AI race after the lukewarm response to its Llama 4 LLM. The deal values Scale at $29 billion, twice its valuation from the previous year. With Scale AI working closely with many of Meta's LLM competitors, questions are being asked about whether these relationships will hold.
Wix.com Ltd. acquired Base44 for $80 million on June 18. Wix provides a website builder service. The acquisition attracted attention because Base44, a developer of autonomous coding tools, reportedly had only eight employees.
AI chip startup Snowcap Compute announced that it has raised $23 million in a round of funding led by new investor Playground Global. The round included participation from Cambium Capital Management and Vsquared Ventures Management.
Counsel AI Corp., a legal AI startup that does business as Harvey AI, has secured an additional $300 million in series E funding at a $5 billion valuation. The round was co-led by Kleiner Perkins and Coatue and follows its $300 million round announced in February, bringing total funding to $827 million.
Work AI platform Glean Technologies Inc. announced its latest $150 million funding round, bringing the company's valuation to $7.2 billion. The latest funding round was led by Wellington Management, with participation from additional investors including Khosla Ventures, Bicycle Capital, Archerman Capital and Geodesic Capital. The startup has its roots in enterprise search but has broadened its value proposition to building and managing AI agents and assistants.
Former OpenAI CTO Mira Murati's new AI startup, Thinking Machines Lab, has raised $2 billion in seed funding, valuing the company at $10 billion. The initial funding round, led by Andreessen Horowitz, is one of the largest in Silicon Valley's history, which is particularly notable in the context of Thinking Machines Lab not having a recognized product or even having showcased any technology.
Autonomous coding startup Anysphere Inc. has raised $105 million in series B funding. The startup suggests the funding will help expand its team and further invest in its popular Cursor tool, which the press release suggests is now being used by millions of programmers.
AI data infrastructure startup Typedef emerged from stealth with a $5.5 million seed round led by Pear VC, supported by Verissimo Ventures, Monochrome, Tokyo Black and a group of angels. The startup suggests it is focusing on rebuilding data pipelines to better account for new requirements around tokenization, chunking, prompt formats, model routing and agent orchestration, which are operations it suggests fit poorly into existing architectures.
Legal AI Wordsmith AI Ltd. announced a $25 million series A led by Index Ventures, bringing total funding to $30 million. The company plans to expand its team in Edinburgh, London, New York and across Europe.
Politics and regulations
The Memphis branch of the civil rights group National Association for the Advancement of Colored People released an open letter to Memphis Light Gas and Water related to xAI's Colossus datacenter facility. The letter related to pollutants from 15 gas turbines used to power the site, and the impact of these pollutants on the local neighborhood. This letter drew into focus a long-standing discussion about the powering of datacenters and negative externalities.
In June, Walt Disney Co. and Universal filed a landmark lawsuit against Midjourney with the US District Court for the Central District of California. The studios allege that Midjourney ignored repeated cease-and-desist requests and that its technology continued to generate and distribute unauthorized images of copyrighted characters, including Darth Vader, Homer Simpson, the Minions and others. The complaint describes Midjourney as a "copyright free-rider" and accuses it of "calculated and willful" infringement.
A moratorium on state regulation of artificial intelligence was approved as part of the current US administration's "Big Beautiful Bill," passing a key procedural hurdle. However, the provision continues to be heavily scrutinized, with several prominent Republicans, including some who had voted for the bill, opposed to its inclusion. As currently written, it appears to block states from enforcing any laws that limit or regulate AI systems unless they forgo all access to federal broadband funding.
Swedish Prime Minister Ulf Kristersson reportedly took aim at the EU's AI regulations, noting in a meeting with Swedish lawmakers that he would raise the issue of the rules' complexity and lack of common standards with other EU leaders. This contribution is particularly notable as there has been an indication that the European Commission may be open to pausing implementation.
This article was published by S&P Global Market Intelligence and not by S&P Global Ratings, which is a separately managed division of S&P Global.
S&P Global Market Intelligence 451 Research is a technology research group within S&P Global Market Intelligence.
Location
Products & Offerings
Segment