Sabrina Ortiz

Improving productivity may seem like a logical goal of embedding AI into the enterprise, but that simply won't cut it.

While 81% of enterprises report having used AI to speed up decision-making, only 25% report that it has meaningfully transformed the business, according to Deloitte's State of AI in the Enterprise report, which surveyed 3,500 global business leaders. This gap matters because, as Jim Rowan, head of AI at Deloitte, explains, real ROI for enterprises is only possible when meaningful transformation is achieved.

"The real ROI comes not only from doing existing work faster, but from redesigning work altogether," Rowan told The Deep View. "The report shows that AI is already helping most companies move faster. The next test is whether they can use that momentum to change how they operate, make decisions, and create value."

However, many of the companies reported that they have failed to implement AI in more meaningful ways than just productivity gains. Deloitte found that, while 60% of the companies surveyed have redesigned jobs around AI, only 15% reported to have redesigned jobs extensively or completely. Furthermore, only 15% report a significant fundamental shift in decision-making authority as a result of AI.

The Wharton School of the University of Pennsylvania's fourth annual report on the state of AI found similar results regarding this disconnect in tracking real AI ROI: Around 40% of respondents reported measuring AI returns by hours saved. Net-new revenue was the least utilized metric for success, with 27% reporting that they track this. However, the departments tracking returns by productivity are doing better comparatively.

Despite the drawbacks, the Deloitte report found revenue gains from AI rose to 28%, up from 20% in the January report. Even though this number is encouraging, Rowan said that it is not a guarantee that returns will compound automatically. Rather, it reemphasizes the need to redesign operations around AI.

"The organizations most likely to scale successful use cases across the enterprise treat AI as a transformation initiative rather than a technology rollout, redesigning workflows, establishing governance structures, and building accountability mechanisms," added Rowan. "Without addressing those challenges, organizations risk reaching a plateau where AI increases activity and spending without producing proportional business value."

Our Deeper View

The transformation AI is bringing to businesses resembles earlier shifts, like moving to the cloud, installing Wi-Fi, or upgrading employees to laptops. But because AI seems akin to any other software, organizations aren’t taking the precautions they would when handing an employee a new laptop: training them and teaching them how to actually use an unfamiliar technology. Instead, many assume that simply giving employees access to AI is a shortcut to returns. As a result, report after report finds that organizations aren’t unlocking AI’s full potential. This is especially concerning given how much money enterprises are pouring into AI tools and upgrades. If that spending doesn’t start producing a measurable return soon, enterprises may start to pull back on their AI investments, potentially shaking a major pillar of revenue that frontier labs are counting on.

For everyone who has ever spent 10 minutes searching for the perfect eye-roll GIF, OpenAI has a solution for you.

On Wednesday, OpenAI launched its next-generation image generation model: Images 2.6. Though the previous generation was already capable of creating realistic images — so realistic that it tricks on average 50% of The Deep View's readers in the "AI or Not" poll at the end of the newsletter every day — this model includes a number of updates, such as quicker use of reference images and creating GIFs from text.

To create GIFs, users can enter a prompt or image to generate an animated GIF, then customize it by adjusting the speed, editing individual frames, and sharing it The reference image feature seems tailored to marketers, who can insert brand assets, mascots, product visuals, or simple sketches to bring a GIF concept to life—leveraging the potential for social-media virality.

"Marketers and creators can animate a brand mascot for a campaign, create reactions and memes for social media, or use animated storyboards to explore character movement before video production," said OpenAI in the release.

The image reference feature lets users save reference images of themselves, pets, and characters to reuse in new creations. Though this seems like an attempt to make its image model more of a mainstay for users in everyday conversations, it also has a practical component for businesses, which can now more easily create new content from brand assets.

OpenAI also touts that image creation is up to 50% faster than before, and users can refine images with greater precision. Developers also have the option to choose from GPT‑Image‑2.6 Flare and Sunburst, with OpenAI describing the differences as:

• With quality set to low, Flare can generate images in as little as six seconds.
• GPT‑Image‑2.6 Sunburst is aimed at higher-quality image generation.
• Sunburst also has greater image consistency across edits when quality is set to medium or high.

This release builds on OpenAI's success with image generation, with a billion images per week across ChatGPT, Codex, and the API, or more than half a billion images created per day. In addition to the API, Images 2.6 is also available in ChatGPT and Codex. Ultimately, it's a bet that people will go beyond entertainment and find real productivity use cases.

"People are using it almost in a productivity-like fashion to create visual diagrams of what they would like to see," said Adam Fry, Consumer Product Lead for ChatGPT at a media event earlier this week. "And I think as it comes down even further, the number of images that people are creating and the number of use cases is going to grow a lot."

Our Deeper View

It is interesting to see OpenAI double down on image generation, since it is clearly aimed at the general public as well as working professionals. However, with an impending IPO, it would have been reasonable for OpenAI to take a page out of Anthropic's book and double down on enterprise use cases. Still, with image generation, it is catering to one of the most entertaining consumer use cases — despite it being so compute-heavy. My bet is that OpenAI relies on virality because, realistically, every time it launches another advanced image-generation tool, something goes viral. Racking up these moments of virality, OpenAI can try to make ChatGPT Images as synonymous with image generation as ChatGPT already is with the term "chatbot." That level of brand awareness on its own is something you can't really put a price on.

OpenAI is applying the power of its frontier models to its long awaited hardware efforts.

On Wednesday, OpenAI published a blog post detailing an experiment in which the company used GPT-6 Astra to design a new device where users can talk to Codex. OpenAI explained that the model designed every component without direct human editing, including the functional requirements, enclosure, the electrical components and printed circuit board, the user interface, and software.

Two team members designed the test device with Astra as a side project in under a week, and a contract manufacturer built a small batch of prototypes in under a month. The lab notes that getting it ready for manufacturing still required rounds of review and interaction, as Astra did not account for details such as the manufacturer's production limits, but it is using the identified gaps to improve its frontier model.

OpenAI also claims its hardware and infrastructure teams use Codex and ChatGPT in many capacities, including technical tasks like creating and revising computer-aided design models, and creative ones, such as preparing manufacturing drawings spanning both robotics and infrastructure. And the results have started to show:

• Since February, model usage on OpenAI's hardware teams has grown 23-fold to 212 million tokens per employee per day.
• On BenchCAD 2.0, a 3D modeling benchmark, a newer internal model achieved an average score of 0.875, compared with 0.594 for GPT‑6 Astra at its highest reasoning setting.

The company is putting its engineering capabilities to the test by working with a small group of industry collaborators to evaluate them on real engineering problems. OpenAI announced an expanded collaboration with Chip Ganassi Racing and a new research collaboration with Mazda to explore how OpenAI models can help with new vehicle design. The lab also has a program for companies or individuals who want to participate in its physical engineering research efforts.

Our Deeper View

Typically, when people think about AI and hardware, they think of consumer devices being polluted by unnecessary AI features or bloatware. However, this paper highlights another possibility: AI can actually help accelerate the buildout of entirely new hardware. It is particularly interesting to hear how its own hardware teams are using it to build, as there haven't been many updates since the company originally announced it was collaborating with Apple's Jony Ive on a product, other than the Apple lawsuit alleging that the iPhone maker's employees systematically defected with trade secrets during and after being hired by OpenAI. When OpenAI finally releases its device, it will be interesting to learn how much AI assisted the development. Ideally, incorporating AI models into the development process would help OpenAI push the envelope beyond what would have been possible without them. Ultimately, I don't think there's a better endorsement for AI models in engineering than OpenAI developing a truly cutting-edge device and then openly sharing how AI made that possible.

Data centers' bad reputation has come for the ballot box.

AI infrastructure has emerged as one of the most politically-charged issues of the November midterm elections, with candidates leaning into issues that US voters across the political spectrum feel strongly about. In an NBC News exit poll, 56% of voters said data centers were a “crucial or major factor” in their vote for Congress, while only 13% said the issue played no role at all.

And anti-data center sentiment showed up strongly in the results. Time reported that Democratic candidates spent more than $100 million on broadcast advertising that painted Republican opponents as pro-Big Tech and data centers, allowing these candidates to win key races in Ohio, Kansas and Iowa.

Beyond the candidates themselves, dozens of local ballot measures across the country asked whether to limit or permit data center development in their communities. Redbud Consulting, a polling and data strategy firm, tracks these measures in its Data Centers on the Ballot 2026 Results.

Because the questions were worded differently for each municipality, Redbud's tracker sorts each outcome by its effect: “Restrict” for results that limit data centers and “Allow” for results that permit them. As of the time of publishing this article:

• "Restrict" results had won 38 of the 39 measures with complete counts; "Allow" measures had only won one.
• While votes are still being counted, six elections have "restrict" results leading and two elections have posted no results as of yet, "Restrict" results have 77% of the vote.

Several bills to restrict the building of these facilities were already in place prior to the election, such as New York Gov. Kathy Hochul, who has since been re-elected, implementing the nation's first statewide moratorium on hyperscale data centers of 50 megawatts or more for a year back in July. The results of the midterms could turn that trickle of regulation in the US into a flood. For instance, Texas's first Democratic governor in more than 30 years Gina Hinojosa campaigned on a statewide data center moratorium.

That, in turn, could stymie the multi-billion-dollar buildout that frontier model labs are counting on to achieve the bombastic growth targets they've promised to their investors.

Our Deeper View

As it stands, the frontier AI labs are likely questioning where they go from here. The reality is, with so much on the line ahead of major IPOs, these labs aren't going to stop their relentless pursuit of compute, even as many of them call to pace frontier development. And despite the public abhorrence of these facilities, as well as the political and regulatory threats, these labs don't necessarily need to stop. Rather, they just have to become more strategic. For one, domestic data center buildouts could simply be made smaller to fly under the radar of regulators, as evidenced by reports that OpenAI and Anthropic are eyeing smaller data center facilities, such as 20 to 30 megawatt facilities. These companies can also start building out more compute overseas, such as Google's $15 billion initiative to build out data centers in Malaysia. And of course, with access to the mountains of investor cash that these companies have, don't be surprised to see an increase in pro-data center lobbying as we approach the 2028 election, which is already an industry practice.

Not long ago, ChatGPT hit a wall the moment a question fell outside its training data. Today, it searches the web to answer almost anything, and that shift is winning over users.

At a fireside chat in New York City, Adam Fry, ChatGPT’s consumer product lead, shared that an internal OpenAI analysis found ChatGPT now handles roughly one-fifth as many queries as Google, which has dominated search for decades. To put that in perspective, Google dominates about 90% of the worldwide search market, while its closest rival, Bing, has just over 5%. That means ChatGPT’s query volume would be more than three times Bing’s.

"What people came to ChatGPT to start, if you go back about two years, was really like writing help and coding help, and those wouldn't be things that you really search for, but what we've seen over the last year is people have really sort of grown into all these other topics," added Fry.

That growth is indicative of ChatGPT's overall popularity, which has grown from one billion to 1.2 billion users in the last two months. Other features OpenAI has developed to continue to improve the ChatGPT experience have also continued to gain traction, such as voice mode, allowing users to naturally hold conversations with ChatGPT, which now has 250 million weekly active users, and image generation, which generates 1.5 billion images per week.

While the Voice Mode experience isn't widely adopted, with only 20% of users relying on it, Fry explained it is important for the team to continue to build on it as users who start using voice mode continue to use voice mode and dictation for interacting with ChatGPT. A big bottleneck, Fry explains, is simply the user's ability to discover the tool, which OpenAI hopes to improve. Additionally, on Tuesday, OpenAI announced that Voice Mode is expanding to the ChatGPT Finances experience.

The improvements don't end there, including the long-awaited change in the works: finally combining the "Chat" and "Work" tabs. Currently, the two tabs add a layer of friction to make users choose between a conversation and a task, which can be confusing. Fry shared that OpenAI plans to bring them together into one interface, with the model inferring whether you want a quick reply or a longer, agentic task. TechCrunch has already reported that OpenAI is rolling out this redesigned experience to select users for feedback before it is rolled out widely, meaning that it could be coming to users soon.

Our Deeper View

While ChatGPT's growth is likely due to the constant improvements that OpenAI makes to the user experience, it's also worth mentioning a major change that could have had the opposite effect: ads. Realistically, nobody likes when ads are shoehorned into a platform, but OpenAI's rollout of ads in ChatGPT has not seemed to hurt its user numbers. As a matter of fact, Fry explained that many people are finding it useful in certain settings, such as for hotel bookings, as it's helping them find what they need more quickly. Part of this success, and what OpenAI has done really well, is putting a lot of thought into user intent, privacy and ad placement integrity, as my colleague Nat Rubio-Licht explains in a recent feature. For instance, the company claims it does not share any of your data with the ad companies, and ads are only shown on queries that seem appropriate, not those of a sensitive matter. Ultimately, OpenAI understands what rivals like Meta seem to be missing: User trust is hard-won and easily lost, and no amount of ad revenue is worth that cost. Getting ads right by nailing user trust and utility from the start only serves to aid OpenAI on its path to upending Google's search dominance and becoming the go-to for consumer-facing AI.

Voice is often the easiest way to interact with AI, but only if the model can understand what you actually mean, including how you said it. Hume AI thinks it has the answer.

On Monday, Hume AI unveiled its Hume Audio Decision API, built on its audio-native decision model. It uses the same architecture Jev and other similar models use to take the world by storm—instead of just predicting the next token like a large language model, decision models can answer many structured questions in parallel and show confidence levels. In voice interactions, this means an AI system doesn't have to rely only on a transcript to interpret users' needs; instead, it can understand intent, emotion, and more.

“There is a lot of opportunity to be able to leverage that raw audio and do really interesting things from an interrogation perspective to understand that, so that you can then figure out where to go build your agents, how to go build your agents, where they would perform well," Andrew Ettinger, Hume CEO, told The Deep View.

I got to see the model in action pre-release, which put it into perspective. Chris Ren, a product lead at Hume, spoke to the model in a demo and, in real time, it was able to give feedback on different questions such as how many speakers were in the room, the language used, the tone of voice, the gender of the speaker, whether there was music playing in the background and more, alongside a confidence threshold for each response. Developers can, of course, add their own questions to best fit their applications in formats that include "yes/no", "choice", "scale" and "span".

To measure its accuracy, Hume tested the model against real human testers and, in initial testing, it matched a human rating panel's collective judgments more closely (70.7%) than any individual rater (56.5%). Hume also reported:

• Using average accuracy-vs.-speed across leading benchmarks, Hume's Audio Decision Model outperformed every other model on the market.
• Hume's Audio Decision Model's average score over 10 leading benchmarks was 82%.
• The closest competitors scored as follows: Gemini 3.8 Flash (78%), Qwen 3.8-Omni-Flash (76%), and Inkling-Small (67%).

There is also a bulk option for analyzing many existing audio recordings. In a demo, there were three hours of simulated patient–physician recordings with eight or nine questions about the conversations queued. Hume explained that, in this case, the model could track the patient's emotions and identify signals such as distress that a traditional transcript would miss.

Our Deeper View

In the agentic-first world we live in, predicting the next token often falls short because agents' actions depend on instantaneous input from several simultaneous queues. That is one reason TypeSafe AI's Jev model made such a popular debut, along with its impressive cost cuts that make it much cheaper than leading models from frontier labs. Its release also couldn't have come at a better time, as the industry pushes toward efficiency and lower costs. OpenAI followed suit, releasing its own Decisions API to customers shortly after. Hume is uniquely positioned to capitalize on the moment, with deep expertise in emotionally aware models and interfaces, an area with many practical use cases for developers to create engaging experiences for users.

Soaring demand for data centers and growing constraints on Earth's resources have pushed the industry to consider a new frontier: space.

Google has been a major proponent of this movement with Project Suncatcher, a research mission designed to determine whether low Earth orbit is a suitable host for scalable, solar-powered data centers. After launching it's first prototype satellite with SpaceX in early October, the company is celebrating the first ever-AI response from a data center-class TPU in orbit.

Google shared that, during a ground-station pass, it uplinked its first prompt to the onboard Gemma 4 open model and asked it to generate a haiku about being in space. Then, as the satellites passed Svalbard, Norway, it shared the downlinked poem. Next steps include verifying how the TPUs perform under thermal stress and radiation exposure in orbit, but the fact that the chips could execute prompts in space is an important step in the research moonshot.

Supporters of building data centers in space cite natural limitations of building AI infrastructure on Earth, including physical space, local power grid capacity, clean water, and the disruption to communities that don't want it. However, many debate whether data centers in space are a viable alternative, as low Earth orbit is also becoming increasingly populated.

Currently, SpaceX has about 11,000 active satellites in low Earth orbit and is part of a growing network that includes Amazon Leo, as well as Qianfan and Guowang in China. Low Earth orbit is also home to some longer-standing fixtures, such as the International Space Station. And this number is only bound to increase, following reports that Russian President Vladimir Putin is developing an alternative to Starlink after praising Elon Musk as a "genius."

While physical space constraints aren't necessarily a concern in space, experts are also concerned about the availability of safe, collision-free orbital paths. This is exacerbated by the fact that orbital data centers would require an unprecedented number of satellites, on the order of hundreds of thousands to a million new devices.

Additionally, space debris is a growing problem that can have long-term impacts. According to Kessler syndrome, as fast-moving objects collide, they shatter into thousands of high-velocity fragments, causing more collisions and debris and reaching a point where certain orbital areas become unusable for generations. A recent Bloomberg report cites another concern: catastrophic conflict in space, in which people attack targets with missiles, cyberattacks, or other means. In that scenario, the potential for orbital debris becomes even greater.

Our Deeper View

This represents an outlandish solution to a very real problem: Amid unprecedented AI demand, the industry is running out of data centers and the resources to power them. But it also raises bigger questions: How much AI infrastructure do we actually need? Additionally, when do we seek another path, whether that’s making models more optimized and efficient or simply slowing down until the infrastructure can catch up? It isn’t just about running out of space and the world’s resources — It’s also about the hundreds of billions of dollars being poured into this buildout. We may one day realize much of it wasn’t necessary, and that shifting our focus to creating more efficient models themselves could have taken the industry farther.

Open-source AI models are gaining traction for their control, lower costs, and optimization options. Though China leads the way in this market, Mistral is mounting a challenge.

On Thursday, Mistral released the open weights of Mistral Large 4, its flagship, 1-trillion-parameter Le Chonk model. When the company first announced the model in October, it was in public preview. In the three weeks since, Mistral kept improving it with reinforcement learning and customer feedback. Through those improvements, Mistral pushed the model to benchmark breakthroughs, including claiming it's the "best open agentic coding model in the world" and "state-of-the-art in cyber defense capabilities among open models."

To support those claims, the company shared some of its benchmark breakthroughs:

• ML4 ranks as the top open model in the world for agentic coding on Artificial Analysis’ Coding Agent Index.
• On Terminal-Bench 4.0, a benchmark that measures how well agents can complete real-world technical tasks using a computer terminal, ML4 climbed from 28% to 41%, tying GLM-5.3 as the best model for coding agents.
• It also tied MiMo V2 6-Pro and surpassed DeepSeek-V4.1-Flash 1017 and GLM-5.3 Flash on the Artificial Analysis Cyber Index, or CyberGym-E2E, a cybersecurity benchmark that tests AI agents.

But ML4's advanced capabilities aren't the only thing that make it attractive to enterprises, Marjorie Janiewicz, Mistral's chief revenue officer, told The Deep View. Its open-source nature offers several perks, having control over when to stop using your model (as opposed to it being sunset), being able to build defenses that are more deeply integrated into your database, and often being able to save on costs.

"Before, you had to compromise between performance and control," said Janiewicz. "With this release, you can have access to frontier performance and then control that AI. It's the best of both worlds."

While other labs also use reinforcement learning to improve their models, Mistral said it invests more heavily in it than its open-weight competitors, creating increasingly difficult cyber and coding problems in-house for this training, thereby making it especially proficient at the tasks that matter most to enterprises. The company said "end-to-end control of (its) own infrastructure and scientific innovations" also allows it to deliver these results at a fraction of the compute and capital of the closed model labs.

Our Deeper View

Enterprises have driven most of the profit for AI companies for a long time, but it's more evident than ever after Anthropic's IPO filing this week. Enterprises could benefit from open weights for many reasons, but perhaps the most interesting is that these models are being tailored to specific enterprise tasks, such as coding in ML4's case. When talking to Janiewicz, she highlighted the difference between simply improving productivity, as a chatbot can do for employees, and actual workload transformation: "We think there's a very vast and bright future for open-weight models that are focused on enterprise transformation and ROI rather than productivity."

When Amazon unveiled this year's Kindle lineup without AI, the omission was surprising. A few weeks into testing the device, it feels like a quiet act of genius.

The Kindle Colorsoft is a more premium take on the Paperwhite, with the biggest difference being that it can display color while keeping the e-ink experience. This year’s edition tackles one of the biggest trade-offs of color e-ink displays: contrast. The redesigned display includes a light guide with nitride LEDs that enhances color and boosts brightness without washing out details, an oxide backplane with custom waveforms that delivers higher contrast in both color and black-and-white content, and reverse printing technology that enables edge-to-edge color across the screen.

While black-on-white still isn't as crisp as on non-color alternatives, this device offers a significant improvement, allowing users to highlight text in different colors and see book covers in color. The Kindle Colorsoft is also lighter and thinner than predecessors, which is an added bonus for reading on commutes. The flush, edge-to-edge glass display layer is also a nice upgrade to create a more seamless experience.

Additionally, I haven't had to charge my Kindle once since I started reviewing it, reading over 300 pages across multiple sessions. Amazon claims up to eight weeks of battery life on a single charge.

Most importantly, my time with the Kindle made me a believer. I was a holdout for a long time, as my foldable smartphones have an 8-inch screen available right in my pocket, so I didn't want to carry yet another device for reading with a 7-inch display that could only fulfill one use. However, having a device that was completely removed from all notifications and tools, AI chatbots included, made all of my curiosities have to wait until I was finished reading, which I found incredibly valuable.

Until now, I’ve been e-reading on an iPad decked out with a matte screen protector, Magic Keyboard, and Apple Pencil, the device effectively doubling as a mini laptop. Naturally, this came with many distractions, an issue which I hadn't realized before testing the Kindle. It also made me the target customer for the Alexa 12 Pro and its pencil, keyboard folio, and matte screen. Using it showed me what my iPad setup was missing: Kindle Sanctuary.

Designed to mimic a real Kindle, Kindle Sanctuary silences notifications, reduces blue light, and stretches battery life to up to 20 hours. The distractions are still a tap of the home button away, but it makes staying in the flow a bit easier, and again, I didn’t miss AI as part of the experience. If anything, I wish I could dial back the Alexa integration. I’m not sold on the Alexa widgets cluttering the top-left and bottom-right corners of my home screen.

Testing the tablet has left me with a few thoughts. For one, I never want a screen that is not matte again (a sentiment I also felt with the iPhone Duo). Additionally, it is a solid tablet at a very reasonable price point of $550 for 128GB, and Kindle Sanctuary needs to be brought to the app across any device.

Our Deeper View

My biggest takeaway: as AI integrations spread across more and more consumer devices, some things are better left simple, and there’s real beauty in many of the products that fit that description. Smart photo frames, for instance, still look like photo frames and do little more than display photos. That includes the Amazon Alexa Photo Frame that was just released. This is a corner of the industry that manufacturers repeatedly get wrong: because many of these devices were “smart” to begin with, piling on more intelligence can feel almost inevitable. The Kindle Sanctuary stands out because its approach could be applied to almost any device, including smartphones. While I’d still want a voice assistant like Siri or Gemini available on my phone at all times, I’d like certain parts of my everyday interactions to be AI-free, and I’d like that boundary to be enforced more firmly. A social media app that served as a sanctuary from AI editing, for example, sounds incredible and honestly necessary.