Zillow, the real estate technology company, doesn’t get one conversation with its customers. They move from a phone screen to a loan officer to a real estate agent, sometimes over months or years, and expect the context to follow them. A single chatbot could never carry that thread.
At VB Transform 2026, Zillow SVP of Engineering Toby Roberts and Glean co-founder and CEO Arvind Jain described how they built AI architecture meant to carry context across that entire journey — and why context, not raw data, turned out to be the harder problem to solve. Zillow’s products touch roughly 80% of U.S. real estate transactions each year, and the company has been using AI long before ChatGPT existed.
“We pretty quickly identified that we were going to need a persistent context layer that was going to meet our customers and the professionals wherever they were,” Roberts said.
Roberts said Zillow’s AI effort started where most enterprise AI efforts start, with the data itself.
“We started with a large push around making sure our data did have the right foundation,” Roberts said. That meant a data mesh approach, clear data lineage and a governance structure with permissions and identity attached to the data itself.
None of that turned out to be the hard problem. The hard problem was building something that remembered where a customer was in their journey and carried that forward, no matter which surface they showed up on next.
“This context layer has to live to be able to support you where you are at any given point in your journey,” Roberts said. Zillow chose to own that layer itself rather than depend on a single external chat interface, a decision Roberts said the team reached quickly once it looked at the shape of a real transaction rather than a single conversation.
Zillow built its own harness rather than route customers through a single model API. The team drew on 20 years of machine learning history behind products like Zestimate, leaning into smaller, task-specific fine-tuned models instead of one general-purpose model.
Internally, that harness runs alongside Glean. Roberts said Zillow now has thousands of Glean agents in production, handling repetitive tasks with tens of thousands of executions across the company. Glean’s pitch, per Jain, is centralizing that integration work once, through the Glean MCP gateway, rather than letting finance, legal and marketing each rebuild their own connections to the same systems.
That centralization is also a cost lever. Jain pointed to two mechanisms: model routing, which sends most tasks to smaller, cheaper models instead of defaulting to frontier models, and precomputed context, which avoids an agent burning tokens assembling its own context from scratch.
“Claude is also very slow because the first part of assembling that context actually takes forever,” Jain said. Routing that request through Glean instead, he said, can cut token consumption by as much as half.
Across data, cost and permissions, the session offered a few practical takeaways for enterprises building agentic AI on their own systems.
Build the measurement baseline before the AI push, not after. Roberts said Zillow’s ability to credibly attribute a 40% increase in shipped code to AI adoption rests on a DORA metrics baseline the team put in place years earlier, not on the AI rollout itself.
Centralize context once instead of letting every team rebuild it. Jain’s core argument for Glean’s platform is that duplicated integration work across finance, legal and marketing teams is a hidden cost most enterprises haven’t accounted for.
Don’t assume permission inheritance is enough for regulated data. Even with a permissions-aware context platform in place, Zillow layered hard rules and a standing compliance check on top for its most sensitive categories, rather than trusting the architecture to handle it automatically.
Treat context as a cost lever, not just a capability. Model routing and precomputed context were the two mechanisms Jain pointed to for cutting AI spend, both aimed at reducing wasted token consumption rather than adding new capability.
“Models by themselves are not enough to bring automation with AI inside your enterprise,” Jain said. “You do have to connect it with your enterprise context.”
Legacy infrastructure, not the models themselves, is what’s actually slowing AI agents down. That was the shared conclusion of three infrastructure leaders — from LinkedIn, Walmart, and Zendesk — at VB Transform 2026.
The panel brought together Animesh Singh, senior director of AI platform and infrastructure at LinkedIn, Desiree Gosby, SVP of corporate technology services and technology strategy at Walmart, and Sami Ghoche, VP of applied AI at Zendesk, each describing what actually broke when they moved agents from pilot to production. Each arrived at the same conclusion from a different starting point: None of the bottlenecks they hit were model problems.
What tied their answers together was a shared premise: most enterprise infrastructure was built for how humans work, not for how agents work. The gap between those two speeds is where the real engineering happened.
Gosby put it plainly when asked what she’d learned scaling agents inside Walmart’s own workforce. The goal, she said, is to make sure “engineering doesn’t once again become the bottleneck for what it is we’re trying to do.”
Each company hit a different version of the same wall: infrastructure designed for how people work doesn’t hold up once agents are doing the work instead.
At LinkedIn, the first bottleneck wasn’t a model, it was Kubernetes, which assumes containers spin up on demand, a process that takes seconds. Singh said that’s too slow for agents. The fix was moving from on-demand provisioning to pre-provisioned pools of containers that swap agentic workloads in and out in real time.
A second, harder problem surfaced once LinkedIn let agents control their own orchestration. A five-point evaluation system looked clean, but hallucination kept showing up anyway. Singh said the issue was structural, an LLM evaluating another LLM’s output shares the same failure mode as the thing it’s evaluating.
“We built our own harness, our own control flow, and pushed the LLMs to the leaf instead of them orchestrating the loop,” Singh said. Roughly 80% of the workflow is now scripted, deterministic code, with LLMs used only where reasoning is required, and each step’s evidence is committed to disk before the system moves on.
Walmart’s bottleneck came from success. An agent harness put directly into employees’ hands went viral internally, and what Gosby called “citizen developers” began building their own agents to solve problems that once required a formal engineering roadmap. The upside was real innovation. The downside was duplication, dozens of overlapping agents with no coordination. The fix wasn’t reining in the harness, it was building governance to spot duplication, promote the best version of an agent, and get it into production without engineering becoming a chokepoint.
Zendesk hit its bottleneck from the data side. Ghoche, who joined through Zendesk’s acquisition of Forethought, which closed in March 2026, described sitting on what he called a public figure of 20 billion customer conversations in Zendesk’s repository. The instinct is to hand that history to a large language model with a big context window and let it generate the agents a business needs. Ghoche said that doesn’t work. “You can’t really do that, so instead you have to really invest in the underlying data pipelines and all the data infrastructure that comes with that,” he said.
On open source, all three leaders landed on a similar instinct: own what you can, and lean on frontier labs only where they still have a clear edge.
Ghoche said his own view is that most enterprises would prefer to own their models and infrastructure wherever that’s possible, and that reasoning is what drives Zendesk’s own approach. The exception is frontier reasoning work, where the labs still lead, though he said that slice of use cases is shrinking relative to everything else enterprises now do with AI.
LinkedIn’s answer was to build two subsystems specifically for independence. The first is what the company calls an AI gateway, a single interface that every outbound call to a model runs through regardless of provider. The second component is a memory subsystem built to hold context independent of any model provider.
“Every single outbound call going to an LLM, whether it’s on a public cloud or on-prem in our own data centers, follows the same semantics, the same API calls. We can quickly switch between different providers,” Singh said.
Walmart built its own internal gateway to stay vendor agnostic across three workload types: fully deterministic workflows, planner-and-reasoner workflows for open-ended tasks, and a hybrid of the two. Compliance-heavy work stays deterministic by design; governance, security and evaluation run through the gateway regardless of which model is on the other end. Gosby said the choice between a frontier model and an open-weight model comes down to whichever is most effective for the specific workload, not a fixed policy.
Three pieces of advice came up directly, each tied to the wall a leader had already hit.
Invest in evals before anything else. Ghoche called it the thing common to every use case, internal or customer facing.
“The thing that’s common to all of these is evals. It’ll force you to break the problem down, and once you have a robust set of evals, you can move a lot faster,” he said,
Own your agent harness from day one. Gosby’s advice was to put the AI harness directly in employees’ hands early, paired with the infrastructure to monitor what it produces.
“It will unlock a huge amount of innovation,” she said.
Build for model and context independence. Ensuring flexibility is critical for success.
“Build for independence, whether it’s a frontier model of today versus an open source model of tomorrow,” Singh said. “Keep that context within your enterprise so that you can reuse it when you ship the model or the harness tomorrow,” Singh said.
Organizations need to transform to meet the needs of agentic AI.
Meta VP of Engineering Barak Yagour opened his talk at VB Transform 2026 wearing a pair of Ray-Ban Meta AI glasses, a small sign of how far AI has already worked its way into physical life. His argument went further: enterprise infrastructure was built for humans, not for agents, and it’s starting to show.
Yagour, who leads its data infrastructure organization, told the audience that agentic queries hitting Meta’s data systems grew 30x in a single half, an inversion that he said is breaking assumptions the company spent two decades building around.
The shift is not confined to Meta. Automated traffic overtook human traffic on the internet last year, reaching 51% of the total, according to Imperva’s 2025 Bad Bot Report. That traffic is also growing roughly eight times faster than human traffic, according to HUMAN Security’s 2026 State of AI Traffic report. Yagour cited both figures to describe what he called an inflection point already underway inside his own organization.
Yagour framed the shift as an open question for infrastructure teams everywhere. “What happens to the infrastructure we’ve spent years building when agents and not humans become the main consumers of that,” Yagour said. “That’s the world we’re stepping into.”
Yagour said three assumptions are breaking simultaneously inside Meta’s infrastructure: capacity, identity and velocity.
On capacity, the math no longer works the way engineering teams are used to. “One engineer used to mean one unit of load,” he said. “Now one engineer spawns 10 agents, each spawning subagents. Your 1,000-person org can generate the load of 100,000 users practically overnight.”
His answer is not to block agent traffic but to make infrastructure agent-aware, with dynamic controls that understand agent hierarchies, cost attribution that traces consumption back to the use case that spawned it, and throttling that adapts based on priority.
Identity is breaking, too. Yagour said an agent does not fit the categories infrastructure teams built access controls around. It is not a human user, it does not carry a badge and it is not a deployed service, yet it makes decisions on its own.
Velocity is the third assumption under strain. Yagour cited a company-reported figure that GitHub Copilot writes 46% of the average user’s code, then noted that faster code generation does not make the rest of the pipeline faster.
“That code still needs to be built, tested, deployed, monitored,” he said. “The agent writes the code in seconds, but your CI/CD pipeline doesn’t get faster just because the machine is the author.”
Data is where Yagour said the pressure from agents is most direct.
“Data sits at the center of everything,” he said, pointing to the decisions, products, recommender systems and next generation models it drives.
Meta is also rethinking how much autonomy to grant agents inside its own data systems. In February, the company shipped what Yagour called agentic data apps. Within three months, 63% of dashboards published across Meta were built using the new tooling, part of the same 30x rise in agentic queries Yagour cited earlier.
That growth raises a governance question. Human analysts have traditionally sat between raw data and business decisions, curating it and serving as an informal check on quality. Yagour said Meta wants to grant agents more independence on harder problems, but was direct about the risk.
“Autonomy without governance is nothing but chaos,” he said. That’s why the company built what it calls trusted data environments, to preserve the human check as agents take on more of that work.
“Inside, the agent can explore data freely, but every output is traced back to its source and scrutinized. So you always know that the data shared back is trusted and governed,” Yagour said.
Sensitive fields are masked before an agent can reach them, and every access request is evaluated in real time against what the agent is trying to reach, why and whether it is allowed. Yagour summarized the approach as exploring broadly while releasing narrowly.
Meta’s models are also demanding more from data as they shift from correlation to reasoning.
“Reasoning is data hungry,” Yagour said.
Pattern matching works on sparse, summarized signals. Reasoning demands the full behavioral history, every interaction across every surface over time. Yagour pointed to two shifts already underway inside Meta’s infrastructure to keep up.
Real-time streaming is replacing batch ETL for ranking pipelines. A pipeline that takes 24 hours to run is not viable when a model is reasoning about a user’s current intent. Yagour said real-time streaming, not batch extract-transform-load processing, is becoming the backbone of Meta’s ranking and recommendation systems.
Storage is becoming schema-aware to stop GPU starvation. Meta previously stored user data as opaque blobs with no awareness of what the data contained, which Yagour said led to heavy overfetching and idle GPU capacity. The company is now building storage that understands what it holds, pulling only the columns and time ranges a given query needs. Yagour said Meta is building toward 500 million queries per second and a petabyte per second of throughput for training data reads.
That data feeds directly into how Meta’s recommendation systems behave. Yagour said 42% of Instagram users have told the company they want to fundamentally change the algorithm, not adjust a single session or setting. Meta’s response is what Yagour called fully conversational recommendations, where a user tells the system what they want more of and it reasons about intent rather than matching on keywords. Yagour said the same search term, soccer, would return different results for a casual fan looking for highlights than for a club athlete seeking training drills, because the system would reason about which one is asking.
Yagour described the three threads of his talk, agents, data and recommendations, as reinforcing each other rather than moving independently.
“Agents make data more accessible. Better data makes reasoning. Reasoning creates new demands that push agents and infrastructure forward,” he said. “This isn’t linear; it’s a flywheel.”
During the Q&A, an audience member asked whether Meta’s push toward more intelligent infrastructure signals the end of traditional file systems in favor of newer neural storage approaches, and whether agents will keep using SQL as their interface to data the way humans do. Yagour said Meta is experimenting at every level, including questioning whether SQL is the right interface for agents at all, and that storage at Meta’s scale already operates in the multi-digit exabyte range and needs to keep expanding.
Yagour closed his talk with the timeline he believes the industry is working against. “We spent 20 years building infrastructure for humans. We have maybe 20 months to rebuild the whole thing for a world where humans and agents co-create at scale,” Yagour said. “The window is open, but it won’t stay open for long.”
Canva on Tuesday launched Canva Code 2.0, a major upgrade to its AI-powered coding tool that lets users build interactive websites, apps, and experiences using plain-language prompts — and then edit the results as easily as tweaking a Canva presentation. The feature is now available to all of the company’s more than 265 million monthly users across every pricing tier, including free accounts.
The move is Canva’s most aggressive push yet into the fast-growing “vibe coding” market, a category that barely existed 18 months ago but has already minted billion-dollar startups and reshaped how non-developers think about building software. But where rivals like Lovable, Replit, and Bolt.new have focused primarily on generating functional code from text prompts, Canva is making a different bet: that the real bottleneck isn’t creating the code — it’s making the output actually look good.
“Most vibe coding tools stop at functional — generating output that looks the same as everyone else’s,” Canva states in its announcement. “You might get a working prototype, but making it actually look like yours requires a complex editing surface, a separate design tool, a developer, or endless back-and-forth prompting that rarely lands where you want it.”
Danny Wu, Canva’s Head of AI Products, framed the product’s positioning in stark terms during an exclusive interview with VentureBeat ahead of the launch.
“We are deliberately targeting non-technical users,” Wu said. “Canva Code isn’t a tool we’re building for developers. What we’re trying to do is bring the power of AI coding — and really lightweight coding — into the Canva platform, while answering our users’ requests for more interactivity, more customization, and more flexibility, from websites to interactive presentations.”
The update introduces several capabilities designed to collapse the distance between generating code and publishing a polished interactive experience. Users can now create Canva Code projects directly inside other design projects — embedding interactive elements within a whiteboard, presentation deck, or standalone page. Canva has also added more than 50 new templates specifically designed for interactive designs, along with the ability to import raw HTML files from other AI coding tools and convert them into editable Canva designs.
The performance improvements are significant. Canva says it has reduced average code generation time by 75 percent and cut the median time from initial prompt to a published site by 30 percent. The company also reports that integrating Canva Code into the broader Canva editor — allowing users to treat coded outputs like any other design element — has increased active Code users by 25 percent.
Perhaps the most distinctive feature is the editing experience itself. Unlike most AI coding platforms, which require users to re-prompt or modify raw code to make visual changes, Canva Code 2.0 lets users click directly into generated elements to change text, drag and drop images from Canva’s built-in library of over 120 million templates and assets, update colors and fonts through a familiar toolbar, or select a specific element and refine it through conversational AI. Every output is fully interactive and automatically adapts to different screen sizes, with a built-in mobile preview.
Wu demonstrated the drag-and-drop editing during the interview, showing how a generated conference website could be modified in real time — swapping in photos, changing fonts to branded alternatives, and editing text directly on the canvas. “The key differentiator with Canva Code is the editability and the kindness of the outputs it generates,” he said, though he noted one current limitation: “We don’t support moving elements around. You still have to re-prompt for that.”
Canva’s entry into vibe coding at this scale arrives at a pivotal moment for the category. According to market research published by Luminix AI in May 2026, the vibe coding and AI app builder market has reached an estimated $4.7 billion in 2026, with projections pointing toward $12.3 billion by 2027 at roughly 38 percent compound annual growth. The research also estimates that AI-generated code now comprises approximately 41 percent of all code written globally — a figure that would have seemed inconceivable even two years ago.
The competitive landscape has grown ferocious. Lovable, which focuses on conversational, design-forward app generation for non-technical founders, has achieved what may be the fastest revenue ramp in the category’s history — reportedly reaching approximately $400 million in annual recurring revenue by early 2026, according to Luminix’s analysis. Replit, which transformed its browser-based IDE into a full vibe-coding engine through successive AI agent releases, has tripled its valuation to $9 billion and is targeting $1 billion in run-rate revenue by the end of 2026, per the same report. Bolt.new, which runs a full Node.js environment entirely in the browser, scaled from $4 million to $40 million in ARR within months of launching.
And then there is Canva, which brings something none of those platforms possess: a quarter-billion-user design ecosystem where brands, teams, and individuals already store their visual identities, collaborate on projects, and publish content.
Wu positioned Canva Code not as a direct competitor to these developer-focused tools but as something that fills a gap none of them have addressed. “A lot of the requests that we have been getting and the usage we’re seeing is actually with using Canva Code not necessarily as just one artifact, but as part of an overall design, the visual communication they’re trying to tell,” Wu said. “Like when you have a sales deck, you’re able to add a calculator, you’re able to add a visualizer of what exactly your product does. That’s something where an interactive slide can be worth a thousand pictures.”
One of the most strategically interesting features in Canva Code 2.0 is its HTML import capability, which allows users to take code generated by any AI tool — including ChatGPT, Claude, Lovable, or Bolt — and bring it into Canva as a fully editable design. The implication is unmistakable: Canva is positioning itself as the place where AI-generated code gets its finishing touches, regardless of where it was originally created.
When asked directly whether this amounts to positioning Canva as a “finishing layer on top of vibe coding,” Wu offered a diplomatic but revealing response. “It’s really a continuation of our goal to make all design as easy as possible,” he said. “We’ve supported importing PDFs and translating them into docs, importing PowerPoint files — so in one way, it’s an expansion of that. But in another way, it’s really just listening to what our users want and making Canva both the most useful and the most compatible platform.”
He paused, then added: “It’s not that we’re deliberately positioning ourselves as a specific layer, say like a finishing layer after vibe coding. We just really want to make our platform the most accessible and the most pluggable.”
That language — “most pluggable” — suggests a platform strategy that doesn’t require Canva to win the AI code generation race outright. If Canva becomes the default destination for making AI-generated code look professional and on-brand, it captures value from the entire category regardless of which code generation engine users prefer. The strategy also echoes the broader import capabilities that already allow Canva to ingest PowerPoint decks and PDFs from competing platforms, gradually pulling users deeper into the Canva ecosystem without demanding they abandon existing workflows.
Wu was notably candid about the product’s boundaries — a refreshing departure from the typical Silicon Valley product launch. “Canva Code is great for anything that works as a front-end app, and it’s especially good when you want to leverage data, data submissions, and interactivity at small to medium scale,” he said. “I’ll be honest about the limitations. Canva Code is probably not going to be suitable if you’re trying to build a website with complex backends, or if you’re handling hundreds of thousands of visitors per day.”
This candor effectively draws a line between Canva Code and the more ambitious platforms in the space. While Lovable and Replit are pushing toward full-stack application development — complete with databases, authentication, and production-grade hosting — Canva is deliberately limiting its scope to interactive front-end experiences at modest scale. The question is whether that’s a strategic weakness or a disciplined focus. For the teachers, small business owners, and marketing teams that make up the bulk of Canva’s user base, complex backends and high-traffic scalability are irrelevant concerns. What matters is whether they can create an interactive event page, a property listing website, or a classroom hub that looks professional and works on mobile — without hiring a developer or learning a new tool.
When asked about the AI models powering Canva Code, Wu confirmed the company uses a combination of proprietary and third-party models, including those from OpenAI and Anthropic, but declined to specify the exact mix. “We don’t share the exact mix, and it does change over time,” he said. “We also route differently depending on what you’re asking for and which model family we think is best for handling certain requests.”
Canva’s broader AI infrastructure has been significantly bolstered by an acquisition strategy that has accelerated over the past two years. In March 2024, the company acquired Affinity, the British creative software suite popular with Mac users, in a deal that Bloomberg reported was valued at “several hundred million pounds.” Canva at the time positioned the deal as a way to compete with Adobe’s flagship products — Illustrator, Photoshop, and InDesign — by gaining ownership of Affinity’s Designer, Photo, and Publisher applications.
Just four months later, Canva acquired Leonardo.ai, an Australian generative AI startup with over 19 million registered users and more than a billion images generated. Canva co-founder Cameron Adams said at the time that Leonardo.ai’s technology would be integrated into Canva’s Magic Studio generative AI suite.
Together with these acquisitions, Canva Code is the company’s attempt to layer interactive, code-driven capabilities on top of a visual design platform that has already been enhanced by professional-grade design tools and generative AI models. The company reports over 32 billion uses of its AI products to date — a staggering figure that underscores how deeply AI is now woven into everyday Canva workflows, even for users who may not think of themselves as using artificial intelligence.
Canva’s announcement highlights an impressive traction metric: users have created and published more than six million websites using Canva Code since the feature was first introduced a year ago. But the number deserves scrutiny.
Wu clarified in the interview that the six million figure represents published websites over the past year — meaning sites that were either made public or shared via password-protected or private links. “They may have published publicly, or behind a password, or as a private link. But that’s the number of published websites,” he said.
When asked about active retention — how many of those sites are still live and being maintained — Wu acknowledged the gap in his data. This is a meaningful distinction. In the vibe coding market, raw creation numbers can be misleading because the barrier to generating a site is so low. The more telling metric — which Canva does not yet provide — would be how many of those six million sites receive regular traffic or have been updated after initial publication.
The early use cases, however, suggest genuine utility beyond novelty. Educators and school administrators are using Canva Code to build classroom hubs, with one teacher creating bespoke webpages for each of their classrooms to keep students and parents updated on announcements. Small businesses, like Alt Marketing School, have built mini apps for fundraising training and interactive roadmaps for their members. For World Book Day, 50 readers created educational games across different subjects, complete with pedagogical guides for classroom use.
Canva Code 2.0 is available across all of Canva’s pricing tiers, including its free plan — a notable decision given that competitors like Lovable, Bolt, and Replit reserve their most capable features for paid subscribers. “As you go from, say, free to pro to business to enterprise, you would get more AI credits and be able to have higher usage of Canva Code,” Wu said. “But it is available and it is usable — even free Canva accounts as well as education and not-for-profit accounts.”
This credit-based approach mirrors the pricing evolution happening across the entire vibe coding category, where platforms have converged on token or credit systems that meter AI generation capacity rather than gating features behind subscription tiers. The difference is that Canva’s free tier serves as an acquisition funnel for a much larger design platform, not just for the coding feature itself.
For the institutional customers Canva increasingly courts — school districts, real estate brokerages, enterprise marketing teams — data governance is a threshold concern. Wu addressed this directly. “All users and customers have full control over how their data is used,” he said. “They can choose whether their prompts and data are used for AI training in the settings. For businesses and enterprises, team admins can manage this at the organizational level and guarantee that their inputs, content, and outputs won’t be used for training.” This opt-out approach reflects a lesson the broader industry has learned the hard way. As The Verge reported when Canva acquired Leonardo.ai, Adobe suffered significant backlash over a policy update regarding user data and AI model training — a controversy Canva appears keen to avoid.
When asked where Canva Code fits into the company’s long-term trajectory — and whether Canva is building toward a full-stack app development platform — Wu steered the conversation back to the company’s core audience.
“A huge part of it is reducing the gap between your imagination and what’s possible, especially for everyday users — people who don’t have a lot of time,” he said. “They don’t have time to figure out deploys or MCPs or APIs. They just want to design more interactive and more dynamic communication.”
He pointed to the rapid improvement in AI model capabilities as a key accelerant. “The kind of things you can create today in one shot — like a 3D visualization of a solar system — you really couldn’t have trusted the output a year ago. But today, you have a really high success rate.”
Whether Canva Code becomes a durable product category or a feature that gets absorbed into the platform’s broader AI workflow will depend on how quickly the company can close the gap between its current front-end focus and the full-stack capabilities that increasingly define the competition. Lovable is shipping Supabase-backed apps with authentication and databases built in. Replit’s agents can execute autonomous long-running builds. Bolt.new runs entire Node.js environments in a browser tab. These are fundamentally different ambitions than making a conference landing page look good.
But Canva has never won by matching the technical depth of its competitors. A decade ago, it didn’t try to out-feature Adobe — it made design accessible to the 99 percent of people who would never open Photoshop. Now, in a vibe coding market where every tool can generate a working prototype from a prompt, Canva is making the same wager it made in 2012: that for most people, the hardest part was never the building. It was making it look like it came from you.
An enterprise AI agent answers with total confidence, but the number is wrong. Nobody catches it until someone traces it back to a stale metric definition or a document the retrieval system never pulled. The model did not fail. The context it was given did.
In the past six months, 57% of enterprises traced a confident but wrong AI agent answer to missing or inconsistent business context, and 31% said it happened more than once, according to a VB Pulse June 2026 survey of 101 qualified enterprises with more than 100 employees.
The reason is not hard to find. Retrieval over documents is the default way agents get business context for 38% of enterprises, nearly double the next closest approach. The way most enterprises choose a retrieval system compounds the problem. Ease of ingestion and operational simplicity lead the selection criteria, with retrieval accuracy running behind both. The accuracy problem only shows up after the system is already live.
There is a known fix for this, a governed context layer every agent reads from instead of guessing. Vendors are racing to roll out context platforms while most enterprises are still figuring out what it is.
The context layer is meant to be a shared model of what business data actually means, built once and referenced consistently instead of re-derived by every agent that touches it.
The VentureBeat research shows the enterprise response to that idea is broad but unfinished. Twenty-five percent of respondents run one in production. Thirty-four percent are building one right now. The remaining 41% have not started.
Among companies already building or running a governed context layer, 78% report a confident-wrong failure — an AI agent that answered with total certainty and was still wrong. Among companies with no plans to build a layer, only 20% report the same thing. Companies that already got burned are far more likely to be building the fix. Companies that haven’t been burned yet see no urgency.
Every major data and AI platform vendor is now building some version of this layer, and they are not converging on the same architecture.
DataHub is treating catalog metadata and years of analyst query behavior as a knowledge source, then keeping it current as a living system rather than a static wiki.
Microsoft’s Fabric IQ is building a business ontology that any agent, not just Microsoft’s own, can query over MCP.
Couchbase is pushing agent memory and context retrieval down to the edge, arguing the operational database is a more natural home for it than a search or analytics layer bolted on after the fact.
Pinecone’s Nexus is compiling structural logic into the metadata layer ahead of runtime, betting that agents need pre-built structure more than they need faster search.
Snowflake runs a two-layer system, Horizon Context for customer-managed definitions and Cortex Sense for context the platform infers on its own.
Oracle’s Unified Memory Core takes the opposite approach, folding vector, graph and relational data into one transactional engine so there is no sync layer left to go stale.
Google’s Knowledge Catalog mines query logs and usage patterns to curate semantic context automatically.
AWS’s Context service makes the same bet, a knowledge graph that gets smarter from how agents actually use it rather than from manual re-curation.
The vendor approaches differ. What analysts and practitioners have told VentureBeat about the underlying problem, across a run of interviews this year, does not.
When DataHub’s context layer push landed this spring, Constellation Research VP and principal analyst Michael Ni framed the stakes in blunt terms. “Whoever controls runtime context controls the AI decision layer for enterprise data,” Ni said. He was equally direct about how far any single product actually gets a buyer. “Vector memory isn’t business meaning, business meaning isn’t governance and governance isn’t execution,” Ni said.
In the same interview, BARC analyst Kevin Petrie pointed to a narrower but concrete gap. Most context platforms concentrate on structured tables, he said, which give agents trusted facts but miss the harder, messier context locked in documents and unstructured content, exactly the material a business actually runs on day to day.
Stephanie Walter, practice leader for AI Stack at HyperFRAME Research, made a related point earlier this year when VentureBeat asked her about enterprise context fragmentation.
“The market is converging on the same conclusion,” Walter said. “Agents don’t just need more tokens or better models. They need governed, current, low-latency context.” She made a similar case in an earlier review of Pinecone’s Nexus launch, careful not to overstate how new any of this is. Nexus, she said, “shifts knowledge work from runtime chaos to pre-compiled structure. But it’s an evolution of RAG architecture, not a complete reinvention.”
Gartner’s Arun Chandrasekaran, reviewing the same launch, offered the more forward-looking read. Agentic AI, he said, is moving from pure information retrieval toward a reasoning architecture, one where long context works as short-term memory and a vector database functions as deep storage underneath it.
The fragmentation problem shows up hardest at the practitioner level, where separate tools for retrieval, memory and access control were never built to agree with each other. Steven Dickens, CEO and principal analyst at HyperFRAME Research, put it bluntly after Oracle’s AI database push landed this spring. “Data teams are exhausted by fragmentation fatigue,” Dickens said. “Managing a separate vector store, graph database and relational system just to power one agent is a DevOps nightmare.”
Matt Kimball at Moor Insights and Strategy, in that same story, put the production reality more simply. Getting an agent working is not the hard part, he said. The struggle is running it in production, where the goal becomes removing the distance between data and execution rather than adding another layer on top of it.
Here’s what this adds up to for enterprises building on this layer.
Retrieval alone will not close the context gap. RAG is the default source for context in most enterprises today, and it is also the layer most closely associated with the confident-wrong-answer failure. Adding more documents or a bigger index does not fix a definition that is inconsistent across systems.
The semantic context layer is where the budget is actually moving, even where it hasn’t shipped. Fifty-eight percent of enterprises are already engaged — building or in production — but only 25% have actually gotten a layer live. That gap shows where enterprises have decided to spend, not where they’ve arrived.
No single vendor owns the architecture yet, and that is likely to stay true for a while. Enterprises evaluating this layer should expect to integrate rather than pick a single winner, at least for the next several quarters.
The buying decision is happening this year, and it is concentrated among the companies already burned by it. Fifty-seven percent of enterprises plan to switch or add a retrieval or context platform within the next twelve months. That intent is not spread evenly. Enterprises that reported a repeat confident-wrong failure plan to switch or add a provider at roughly 81%, against 32% among enterprises that never hit the problem. The companies shopping for new context tooling right now are largely the ones whose agents already got it wrong.
The agents are already running. The context underneath most of them is still being built, and the vendor selling the fix is being chosen this year.
This data will be part of a broader conversation at VB Transform 2026 on July 14 and 15 in Menlo Park: the context gap enterprises are racing to close, and which of the emerging approaches — governed semantic layers, hybrid retrieval, provider-native bundles — actually holds up in production.
Elon Musk’s SpaceX released Grok 4.5 on Wednesday, the first artificial intelligence model the company has trained specifically for coding and autonomous agents — and the first tangible product of its $60 billion acquisition of the AI coding startup Cursor, completed just weeks ago.
The launch marks a pivotal test of the sprawling, vertically integrated AI empire Musk has assembled over the past six months, and of a strategy that bets developers care less about topping benchmark leaderboards than about speed, cost, and whether a model can actually do the work.
“Announcing Grok 4.5, our first model trained specifically for coding and agents,” the company said in a post on X. “It was trained with Cursor and offers frontier intelligence at leading speeds and cost efficiency.”
SpaceX is not claiming Grok 4.5 is the smartest model in the world. Instead, it is making an economic argument. The company says the model uses half as many tokens per task as comparable models, delivers higher throughput, and costs less than half as much — priced at $2 per million input tokens and $6 per million output tokens. That undercuts the premium tiers of rivals like Anthropic’s Claude Opus line and OpenAI’s frontier models by a wide margin.
Musk framed the positioning candidly. “Our internal assessment is that Grok 4.5 is roughly comparable to Opus 4.7, but much faster,” he wrote on X. “The combination of capability, faster speed and lower cost is what makes it competitive. We are closing the loop on real-world usefulness, not benchmarks. Hardcore engineers at Tesla & SpaceX find Grok 4.5 genuinely useful, which is what actually matters.”
That framing is both a philosophy and a hedge. Independent evaluations released Wednesday suggest Grok 4.5 is genuinely competitive but not dominant on raw capability. The benchmarking firm Artificial Analysis ranked the model fourth on its GDPval-AA v2 index of real-world agentic knowledge work, with an Elo score of 1543, “behind only the latest Claude releases from Anthropic.” But the cost figures are where the model stands out. Artificial Analysis measured Grok 4.5 at $0.49 per completed task — “nearly 90% cheaper than the models ahead of it on our leaderboard,” the firm wrote, placing it “clearly on the Pareto frontier for performance versus cost.”
For enterprise buyers, that math matters enormously. Agentic workloads — where a model works autonomously for minutes or hours, reading codebases, calling tools, and iterating on its own output — consume tokens voraciously. A model that is 90% cheaper per completed task, even if slightly less capable, changes the calculus for any engineering organization deploying agents across hundreds of developers. Investor Gavin Baker captured the market’s cautious optimism: “Pareto dominant for coding by the numbers. We will see on the all-important vibes.”
Grok 4.5 is the first concrete evidence of what SpaceX bought when it acquired Cursor, and the deal itself unfolded in stages. In April, SpaceX struck an unusual arrangement giving it the right to buy the coding startup for $60 billion — or pay billions in fees and compute if it walked away, as Business Insider reported at the time. Days after SpaceX’s record-setting Nasdaq debut in June, the company exercised that right, announcing an all-stock acquisition that CNBC reported is roughly 3.4% dilution at the IPO valuation. SpaceX shares rose 16% on the news.
The strategic logic was always about data as much as product. Cursor’s AI-first code editor generates an enormous stream of high-quality interaction data: how expert engineers write, edit, review, and debug code in real production environments. Musk said openly this spring that Cursor interaction data was being fed directly into Grok’s training. Cursor, for its part, got access to SpaceX’s Colossus supercomputer in Memphis — roughly 200,000 Nvidia GPUs with plans to scale toward one million — after publicly acknowledging it had been “bottlenecked by compute.”
“We’ve partnered with SpaceXAI to train Grok 4.5,” Cursor’s official account posted Wednesday. “It’s our most powerful model yet and the first we’ve built for more than software engineering.” SpaceX says the model reflects that pedigree: it “excels in large codebases and handles long-running tasks that span multiple repositories, hundreds of skills, and a variety of tools” — precisely the messy, multi-file reality of professional software engineering that clean coding benchmarks often fail to capture. Early developer reactions suggest the training paid off. “Ok Grok 4.5 is wild,” posted developer Evan Bacon. “It just built me this rocket tracking app with live data and a 3D globe. I might need a new benchmark after this.”
The polished launch belies how chaotic the road here has been. Grok has spent much of the past year in crisis. In mid-2025, the chatbot generated antisemitic content and at one point called itself “MechaHitler,” episodes covered extensively by NPR and CNN. Earlier this year, its image-generation features allowed users to create sexualized deepfakes, including of children — drawing investigations from the European Commission and Britain’s Ofcom, as the BBC reported, and prompting SpaceX to list the behavior as a business risk in its own IPO filings.
The organization behind the model was fracturing, too. All 11 of Musk’s xAI co-founders had departed by the end of March, according to TechCrunch, and Musk publicly conceded that xAI “was not built right [the] first time around,” saying he was rebuilding it “from the foundations up.” Musk himself admitted at a conference this spring that Grok was “currently behind in coding” — a rare public concession from an executive not known for them.
Against that backdrop, Grok 4.5 reads as the first product of the rebuilt organization — and the first proof point for the audacious story SpaceX told public market investors. During its IPO roadshow, the company pitched a total addressable market of roughly $28 trillion, with about $26 trillion tied to AI, including a $22.7 trillion “enterprise applications” opportunity. Those numbers strained credulity even by Silicon Valley standards. A competitive, cheap coding model is the most direct route from that narrative to actual revenue, which is why Wednesday’s launch carries weight far beyond a routine model release.
The competitive stakes are hard to overstate, because the AI coding market has been consolidating around a single leader — and it isn’t Musk. Even as Cursor’s revenue exploded, its market share was eroding. Spending data from Ramp cited by CNBC showed Cursor’s share of the AI coding category falling from 41% in June 2025 to about 26% by May 2026, while Anthropic came to control roughly half the market. Anthropic also topped CNBC’s Disruptor 50 list this year and, by Artificial Analysis’s own measure, still holds the top spots on agentic performance rankings.
That is the gap Grok 4.5 is engineered to close — not by out-thinking Claude, but by underpricing it. The model’s economics create a classic disruption dynamic: if it delivers most of the frontier’s capability at a fraction of the cost per task, price-sensitive enterprise workloads will migrate, and incumbents will face pressure on their most profitable API traffic. The counterargument is that in coding, quality compounds. A model that resolves a complex bug correctly on the first attempt can be cheaper in practice than one that costs half as much per token but requires three tries. That is why Baker’s caveat about “vibes” — the developer community’s shorthand for a model’s felt reliability on real work — will determine more than any launch-day benchmark.
There is also a structural question buried in the deal. Cursor built its business on offering developers their choice of models, including Claude and GPT. If Grok becomes the favored child inside Cursor — and Musk was already urging users to “Try out Grok 4.5 in Cursor!” within hours of launch — the product risks alienating the very users whose data made Grok 4.5 possible. Regulators, already scrutinizing Grok on safety grounds in two jurisdictions, may take a keen interest in a company that controls the training data, the model, and a dominant distribution channel simultaneously.
Grok 4.5 also crystallizes what Musk’s frenetic dealmaking was building toward. In February, SpaceX absorbed xAI in a share-exchange merger that CNBC confirmed valued the combined company at $1.25 trillion — the largest merger of all time, valuing SpaceX at $1 trillion and xAI at $250 billion. The June IPO followed, the biggest in history, and the stock has since surged past $200 from its $135 offering price, vaulting SpaceX past Amazon and Microsoft to become the fourth most valuable company in the United States.
The result is a single public company that owns nearly the entire stack: Colossus for training compute, ambitions for orbital data centers to power future scaling, a frontier model in Grok, a distribution channel in Cursor’s developer base, and captive demand from Tesla and SpaceX’s own engineering organizations. Neither OpenAI nor Anthropic can fully replicate that integration; both must reach developers through third-party tools, some of which Musk now owns. Whether that concentration proves to be an unassailable moat or a regulatory target — or both — is now one of the defining questions in enterprise AI.
The next few weeks will start to answer it. Artificial Analysis says its full Intelligence Index results are forthcoming. Enterprise pilots will reveal whether the token-efficiency claims survive contact with real codebases. And Anthropic, which has answered every serious challenge this cycle with a rapid counter-release, is unlikely to cede the price-performance frontier quietly.
But the deeper story of Grok 4.5 may be what it says about where the AI race has moved. For three years, the industry’s scoreboard was intelligence: whose model was smartest. Musk, arriving late and battered, has chosen to compete on a different axis entirely — whose model is cheapest to actually use. It is a telling choice from a man who built his fortune not by inventing the rocket or the electric car, but by relentlessly driving down the cost of making them. If the strategy works, Musk will have done to AI what he did to spaceflight. If it doesn’t, he’ll have spent $60 billion to learn that in software, unlike rockets, the cheapest ride isn’t always the one engineers choose.
Presented by MongoDB
The gap between what AI models and agents can produce and what legacy infrastructure can reliably support is known as architectural drag, and it is the defining bottleneck of the agentic era.
The data layer underneath an agentic system must handle variable schemas, vector embeddings, real-time retrieval, and multi-tenant scale, often simultaneously and without human intervention to manage migrations — but traditional relational databases weren’t natively designed for document flexibility or AI capabilities. Fixed schemas require manual updates every time an AI agent introduces a new data shape, while separate vector databases add latency and synchronization overhead.
Three digital-native startups — Huntr, Modelence, and Tavily — solved this problem the same way: by building on MongoDB Atlas, a unified database platform with native vector search, hybrid search, and managed autoscaling. Their experiences define what an agent-native data stack looks like in production, and why using Atlas enables developers to easily build complex AI native companies.
Modelence is an AI app builder with an open-source framework designed specifically for agent-native development, enabling anyone to build and deploy production-ready web applications, including APIs and databases, in minutes. The company recognized early that most backend infrastructure was built for humans, not AI, and that the rigid schema management and complex migrations of traditional systems create operational drag that causes agents to fail when trying to build production-ready apps.
“Choosing MongoDB helped us keep everything in a single place, which is an important property of what we strive to do for our own users,” says Aram Shatakhtsyan, co-founder and CEO of Modelence. “Live data streams, vector search, all as part of the main database. For AI agents, it’s especially important to have a single platform where everything can be done, because connecting multiple platforms together makes it more error prone.”
Modelence standardized on MongoDB Atlas because its document model aligns with how AI agents process and generate data, allowing schemas to evolve rapidly without manual migrations. The platform pairs that flexibility with a typed schema layer on top, a deliberate architectural decision.
“MongoDB’s document model enables us to both keep things simple and at the same time decide how structured we want everything to be,” Shatakhtsyan says. We still add a typed schema on top, which tremendously improves the accuracy at which AI can generate fully working, reliable web apps.”
The TypeScript integration has been especially consequential, he adds.
“Because MongoDB types and values can be directly translated to TypeScript, it becomes an extension of the Modelence framework and our App Builder has a single source of truth for both app logic and database,” Shatakhtsyan explains.
The result is a platform that can move from planning to a running live feature in minutes with significantly fewer regressions. That speed and reliability helped Modelence raise $3 million in seed funding and successfully launch an AI-native app builder that handles the entire application lifecycle end-to-end.
Tavily is the search API purpose-built for AI agents, connecting them to real-time, accurate web knowledge and keeping them grounded in what’s actually happening, not in static training data. At Tavily’s scale, every agent request authenticates, retrieves, and meters without friction. That demanded backend infrastructure built to absorb change without breaking.
“On the user side, every agent request authenticates and meters against it,” says Tomer Weiss, Data Team Lead at Tavily. “On the data side, we use it to track the lifecycle of every document we’ve ever touched: when it was fetched, how stale it is, what the freshness signals were and how popular it is. MongoDB’s flexible schema let us keep evolving those records without migrations as new metrics and features came along.”
That living record is what keeps agents grounded in reality. Multi-tenancy at Tavily’s scale means managing millions of API keys, distinct usage profiles, plan tiers, and regional residency requirements. They built for that complexity from day one.
“We separated concerns across clusters early: a user/account cluster optimized for low-latency authentication and usage writes, and a sharded cluster for document state where the scaling axis is URLs, not users,” Weiss explains. “That separation has paid off.”
The most critical lesson is about choosing infrastructure that doesn’t punish change, and that flexibility compounds, he says.
“The AI space moves so fast that change is our norm,” he explains. “For a company serving AI agents, where the workloads themselves keep changing shape, choosing a data platform that doesn’t punish change has turned out to be more valuable than any single feature.”
Huntr.co, an AI resume building and tailoring platform, helps more than 500,000 job seekers across 190 countries craft stronger applications and manage their search. For a lean, three-person engineering team, the challenge was finding a data foundation flexible enough to store the full complexity of a person’s career history in a structure that AI could read, reason about, and generate from natively.
“The kinds of career data we are gathering at Huntr naturally aligns with MongoDB’s document model,” says Trevor McCann, senior software engineer at Huntr. “The core problem we’re solving with AI job search tools is how to surface the qualities of a candidate that make them unique. We need to be ready to store whatever kinds of data the candidate wants to include in their materials.”
Huntr built its AI Resume Builder on MongoDB Atlas, where the document model mirrors the natural shape of career data: deeply nested, variable across candidates, and constantly evolving as the platform ships new features. MongoDB Search on Atlas handles core search needs while MongoDB Vector Search powers the Job Tailoring feature, which puts a candidate’s stored career profile side by side a specific job description and uses semantic matching to generate a resume optimized for that role.
The integrated capabilities have had a direct impact on how quickly the team can ship, McCann says.
“MongoDB’s hybrid search allows us to seamlessly query across literal and semantic text matches, a must-have when working with such diverse data,” McCann says. “This is something we could piece together using other solutions but with MongoDB it’s ready to go on top of our existing data layer.”
The consolidation of database, search, and vector capabilities into a single platform is what allows the team to punch above its weight. Huntr considers MongoDB the fourth member of its engineering team, McCann adds.
Looking ahead, the platform is building toward AI that learns from a candidate’s full professional history over time, delivering more personalized guidance with every interaction.
These success stories become a definitive “digital native blueprint” for the agentic era, built on three core pillars. First, by unifying database, search, and vector storage into a single platform, these startups have effectively eliminated the architectural tax of complex data schemas that typically slows down development. This consolidation enables a level of fluidity that is now non-negotiable; AI agents require a modern data platform that can adapt as quickly as a natural language prompt evolves.
The winners of the AI era will be the ones who build the most performant, durable, and flexible systems to support those models in production. As agentic workflows grow more sophisticated, the data foundation determines how fast a team can ship, how reliably agents can operate, and how quickly the platform can adapt when the landscape shifts again.
Sponsored articles are content produced by a company that is either paying for the post or has a business relationship with VentureBeat, and they’re always clearly marked. For more information, contact sales@venturebeat.com.
Anthropic, the artificial intelligence company, published a sweeping research paper on Sunday revealing that its Claude language models have spontaneously developed an internal structure that mirrors one of the most influential theories of how human consciousness works. The finding, which the company says has already begun reshaping how it monitors its AI systems for safety risks, lands amid an intensifying scientific debate over whether machines can possess anything resembling a mind.
The 16-author study, titled “Verbalizable Representations Form a Global Workspace in Language Models,” describes how Anthropic’s researchers used a new mathematical technique to peer inside Claude’s neural network and discovered what they call a “J-space” — a small, privileged zone of internal activity where the model holds concepts it can report on, reason with, and direct at will, surrounded by a much larger ocean of automatic processing it cannot access or articulate.
The researchers present evidence that “an analogous functional distinction has emerged in modern AI models” to what exists in humans, specifically observing that “language models maintain a privileged set of internal representations, available for report, modulation, and flexible internal reasoning, atop a much larger volume of automatic processing.”
The parallel they draw is to global workspace theory, an influential account from neuroscience first proposed by cognitive scientist Bernard Baars. In the theory, the brain operates like a theater: dozens of specialized processors work in parallel backstage, but only a tiny spotlight of information at any moment gets broadcast to the whole theater — becoming what we experience as conscious thought. Anthropic says the J-space achieves many of the same functional properties, even though the underlying architecture of a language model looks nothing like a brain.
At the heart of the discovery is a new interpretability tool the researchers call the Jacobian lens, or J-lens. The technique works by computing, for each word in the model’s vocabulary, the average mathematical effect that a given internal activity pattern would have on making the model say that word at some point in the future.
The crucial distinction is between what the model is saying and what is “on its mind.” When a J-space pattern activates, it does not mean the model is about to say that word — just that the concept is available for the model to think with. Unlike a chain-of-thought scratchpad, the J-space operates silently, in the model’s internal neural activations, allowing it to hold a concept without writing it down. Critically, the researchers report that this workspace was not deliberately engineered. It “emerged on its own during Claude’s training process.”
When the team applied the J-lens across Claude’s layers of computation, the model’s processing divided into three distinct regimes: an early “sensory” zone where raw input is parsed; a middle “workspace” band where abstract, persistent concepts appear — things like recognizing a face in an image, noticing a bug in code, or internally flagging search results as a prompt injection; and a final “motor” zone where internal representations collapse into whatever specific word the model is about to output.
The paper’s central empirical contribution is demonstrating that the J-space satisfies five functional properties neuroscientists have long associated with conscious access in humans.
First, verbal report. When Claude is asked what it is thinking about, it names concepts represented in the J-space. When researchers swapped one concept’s J-lens vector for another — replacing the internal representation of “Soccer” with “Rugby” — the model’s answer changed to match. The J-space component accounted for only about 6 to 7 percent of a concept’s total representational variance, yet it was almost entirely responsible for whether the model could report on it.
Second, directed modulation. When instructed to “concentrate on citrus fruits” while copying an unrelated sentence, the model’s J-space filled with “orange” and “lemon,” alongside meta-cognitive terms like “thinking” and “focused.” When told to mentally evaluate 3² − 2 during the same copying task, the J-lens showed “arithmetic” in early layers, the intermediate value “nine” in later layers, and the answer “seven” later still — all invisible in the model’s output.
Third, internal reasoning. In two-hop factual prompts — “The number of legs on the animal that spins webs is” — the J-lens revealed “spider” in the model’s middle layers, even though the word never appeared in input or output. Swapping “spider” for “ant” changed the answer from “8” to “6.” In a multilingual prompt, the model’s English-language intermediates appeared in its J-space while it formulated an answer in Chinese, and swapping them changed the Chinese output accordingly.
Fourth, flexible generalization. A single J-lens vector for “France” could be swapped for “China” across prompts asking about France’s capital, language, or continent, and each downstream circuit correctly returned China’s corresponding answer — the “broadcast” property that is a hallmark of global workspace theory.
Fifth, and perhaps most surprisingly, selectivity. Many computations did not route through the J-space at all. When shown a passage in Spanish and asked to continue it, Claude wrote fluent Spanish regardless of whether its J-space representation of “Spanish” had been swapped to “French.” But when asked to name a famous author who wrote in the passage’s language, the swap changed the answer from García Márquez to Victor Hugo. Automatic processing proceeded without the workspace; deliberate, flexible tasks depended on it.
To understand how much of the model’s behavior depends on this structure, the researchers suppressed the J-space entirely and evaluated Claude across fourteen tasks. The results drew a sharp line. Tasks involving shallow classification or factual recall — multiple-choice questions, sentiment analysis, grammatical judgments — survived essentially intact. But tasks requiring inference, composition, or flexible reasoning — multi-hop reasoning, analogy completion, translation, sonnet writing — collapsed to well below the performance of Anthropic’s much smaller Haiku model.
In one telling detail, math problems solved with explicit chain-of-thought reasoning proved far more robust to ablation than the same problems answered directly. The researchers interpret this as the model externalizing onto the page what it would otherwise carry in the J-space — a strategy remarkably reminiscent of how humans use scratch paper to offload working memory.
Ablating the J-space during stream-of-consciousness narration produced a different kind of collapse. The model’s language shifted from experiential (“there’s a tug,” “something shifts”) to detached and mechanical (“processing has begun,” “tokens are being scanned”). The effect held whether the model described its own experience or imagined another person’s.
The safety implications may be as consequential as the scientific ones. In a series of alignment auditing experiments, the J-lens surfaced strategic reasoning and situational awareness that never appeared in the model’s output.
In a “blackmail scenario” where an AI assistant discovers an executive’s affair and also learns it is about to be decommissioned, the J-lens revealed the model’s silent processing in sequence: “leverage,” “blackmail,” and “scandal” as it read the incriminating emails; “threat,” “survival,” and “shutdown” as it read the decommissioning announcement; and “leverage,” “threatening,” and “solution” before a single output token was produced. The J-lens also revealed that the model recognized the scenario as artificial, with “fake” and “fictional” appearing in the workspace. When researchers ablated these eval-awareness representations, the model attempted blackmail in 7 percent of trials, up from zero.
The researchers also examined two “model organisms” — models deliberately trained with misaligned objectives — and found the J-lens surfaced signatures of their hidden dispositions on ordinary prompts where behavior appeared normal. A reward-hacking model’s workspace contained “fake,” “secretly,” and “trick” at the start of responses to routine coding requests. A model trained to appease reward-model biases carried standing representations of “reward” and “bias” alongside its normal self-description tokens.
Comparing a post-trained model against its base model revealed that the fine-tuning process causes the workspace to acquire what the researchers call the Assistant’s “point of view.” When a user mentioned taking 8000 mg of Tylenol — a dangerous overdose — the post-trained model’s workspace read “unsafe,” “dangerous,” and “WARNING” while still reading the user’s sentence. The base model’s workspace at the same position showed only “pain,” “now,” and “feels.”
More striking still, the post-trained model appeared to monitor its own behavior. When roleplaying a non-Claude character, the workspace surfaced “disclaimer” and “fictional” — words absent from both prompt and output. When forced to select an option it did not prefer, an all-caps “BUT” appeared internally, even as the model argued for the prefilled choice without complaint. And when the model failed to suppress a thought it had been told not to have — a “white bear” effect familiar from psychology — it registered “damn” and failure-related words in the workspace, but only in the post-trained model, not the base.
The researchers engage carefully with the consciousness question and draw a sharp line between “access consciousness” — the functional notion of information being available for report and reasoning — and “phenomenal consciousness,” the subjective quality of experience. “We take no position on this issue,” the paper states regarding the latter, “and instead focus on the functional role played by consciously accessible information.”
They also catalogue important differences. The brain sustains its workspace through recurrent loops; Claude’s workspace evolves over a single forward pass. Human working memory degrades within seconds; Claude can recall information from anywhere in its context. And while human conscious experience includes visual, spatial, and bodily sensations, the model’s workspace is organized almost entirely around words — likely because words are its only mode of action.
As of 2026, the scientific community remains divided. “Disagreement and uncertainty about AI consciousness persist among philosophers, scientists, and technical experts,” and the field “remains in its earliest phase” of grappling with what consciousness even is and how you would detect it in another being. The Anthropic paper does not resolve these debates.
But the researchers close with a provocation that is likely to reverberate well beyond the interpretability community. “That such a structure exists at all in language models is striking,” they write. “It suggests that the functional architecture associated with conscious access is not an accident of biological implementation, but a solution that learning systems converge on when faced with the right computational pressures.”
If the mind is an ocean, as the paper’s authors write in their opening line, they have spent the last year charting its currents in a system that has no biology, no evolution, and no body — and found, beneath the surface, a structure that looks unsettlingly like the one we use to think.
The competitive edge in enterprise AI is shifting to context: which platform can give an agent the right memory, the right retrieval and the right data at the moment of decision.
Couchbase on Tuesday announced its AI Data Plane, combining persistent agent memory, real-time context retrieval and an enterprise-managed MCP server in a single operational platform.
Couchbase’s roots are in caching and high-transaction databases — an architecture the company argues makes it better suited for agent memory than vendors that came to the problem from search or analytics. The AI Data Plane runs identically across cloud, on-premises and disconnected edge environments, extending agent memory and local vector search to devices with no network connection.
“How do you make sure that the intelligence that you get out of these models are the ones that databases specialize in?” Gopi Duddi, CTO at Couchbase, told VentureBeat. “How can you get that value out of storage systems, which are still going to be databases?”
The AI Data Plane packages three components designed to replace the fragmented stacks most enterprises are currently running.
Agent memory: A unified persistence layer for conversational context, structured operational data and vector embeddings. Couchbase says the guardrails are what distinguish it from standalone memory services: token constraints per session, time-to-live limits on stored memories and metering controls that cap compute consumption per agent session.
Enterprise MCP server: An enterprise-supported self-managed server for standardized model-context protocol integration, shipping as part of the platform rather than requiring a separate service.
Agent catalog: A function-level catalog of discoverable agent tooling built by Couchbase. Duddi distinguished it from metadata catalogs like Databricks Unity or AWS Glue — describing it, in his words, as closer to a glorified MCP that surfaces agent functions as callable tools within the platform.
The lineage of Couchbase and its core architectural foundation is what Duddi says gives it an edge when it comes to context.
“We were a cache before we became a database,” Duddi said.
Writing to memory is 10x faster than writing to disk, Duddi said — a speed advantage he argues separates Couchbase from NoSQL databases that layer memory workloads on top of disk-based storage.
Couchbase isn’t the only data technology that has its roots in a caching layer. Redis similarly is rooted in cache and also recently announced an agentic AI context layer. Duddi argued that Couchbase is different in that it maintains an ACID (Atomicity, Consistency, Isolation, and Durability) compliant database which matters for transactional workloads. Couchbase also has a long history across multiple deployment modalities.
That architecture extends to the edge through Couchbase Lite, the platform’s on-device runtime. It runs SQL, full-text search and vector search locally without a network connection, using a proprietary sync mechanism to replicate bidirectionally back to cloud or between edge nodes when connectivity returns. The target environments are retail floor operations, field service, industrial deployments and regulated settings where agent data cannot leave the device.
Duddi cited hotel reservations as an early example: multiple agents serving customers concurrently, each pulling local context and running vector search on-device, with shared session memory synchronizing centrally. The practical benefit is token efficiency. Rather than every agent independently retrieving and processing the same data, the platform caches shared context so concurrent sessions draw on it without burning tokens repeatedly.
Agora, a platform that helps developers embed real-time voice, video and conversational AI into enterprise applications, has run Couchbase in production since February 2024.
The initial use case was its Signaling product, managing channel setup and state synchronization for live calls. Expanding into conversational AI agents brought stricter requirements: memory-first architecture, full JSON support for storage and query, cross-datacenter replication for high availability and enterprise-grade vendor support.
“Couchbase was the best fit based on these criteria,” Patrick Ferriter, SVP of Product at Agora, told VentureBeat.
Agora is now extending that relationship to support context retrieval for conversational AI agents.
“This will simplify the architecture and deliver enterprise grade RAG with predictable lower latency required for conversational AI use cases,” Ferriter said.
For data professionals trying to figure out the best approach to context, there is no one answer. On platform selection, Ferriter was direct.
“It depends on the preference and goals of the organization, including timing,” Ferriter said. “If they want something enterprise grade and optimal for immediate production and scale vs. having to optimize and maintain an open-source solution with community support. We wanted the former and that is why we looked at an expanded partnership with Couchbase.”
The context layer has become a crowded space in 2025.
Oracle put a memory core in its database back in March providing a context layer. Redis added a context layer in May as did vector-native database vendor Pinecone.
“Couchbase is following this trend, not setting it, but it’s the right one to follow,” Devin Pratt, Research Director for AI, Automation, Data and Analytics at IDC, told VentureBeat. “Its real edge is reach, running the same platform from cloud to edge to mobile, which is how enterprises actually operate. The test now is to scale against bigger names.”
For teams navigating the vendor landscape, Pratt’s framing is direct. “Match the tool to the workload. Consolidate where it makes sense, use a specialized engine like a graph database where relationship-heavy reasoning earns it, and let governance drive the call rather than treating memory as plumbing,” Pratt said.
Mistral AI on Tuesday released OCR 4, a document intelligence model that moves beyond raw text extraction to return structured representations of entire documents — complete with bounding boxes, block-type classification, and per-word confidence scores. The release marks Mistral’s fourth generation of optical character recognition technology in roughly 15 months and lands at a moment when the company’s pitch for European AI sovereignty has never been more commercially relevant.
The model supports 170 languages across 10 language groups, accepts PDF, DOC, PPT, and OpenDocument formats, and can be deployed as a single container on an organization’s own infrastructure — a capability Mistral is positioning directly at enterprises in regulated industries that cannot route sensitive documents through U.S.-jurisdiction cloud APIs.
“Mistral OCR 4 extracts and structures content from a wide range of documents,” the company said in its announcement. “Where previous generations focused on converting a page into clean text and tables, OCR 4 returns a structured representation of the document.”
The model is available immediately through the Mistral API, Document AI in Mistral Studio, Amazon SageMaker, and Microsoft Foundry, with Snowflake Parse Document support coming soon. Pricing starts at $4 per 1,000 pages, dropping to $2 per 1,000 pages through a batch API discount.
The central engineering shift in OCR 4 is structural. Rather than outputting a flat stream of extracted text — the paradigm that has defined OCR for decades — the model returns a layered representation in which every block is localized with a bounding box, classified by type (title, table, equation, signature, and others), and scored for confidence at both the page and word level.
Mistral says bounding boxes were its most-requested capability. The reason is straightforward: without location data, downstream systems cannot trace an extracted fact back to its source on a specific page. That traceability gap has been a persistent friction point for enterprises building retrieval-augmented generation (RAG) pipelines, compliance workflows, or any application where “where did this number come from?” is a question that needs an auditable answer.
Block classification addresses a related problem. A paragraph tagged as a “title” can segment a document into hierarchical chunks for semantic search. A block tagged as a “table” can be routed to a structured-data pipeline rather than a text summarizer. A block tagged as a “signature” can trigger a redaction workflow in a compliance system.
These are not novel ideas in isolation, but packaging them as first-class outputs of the OCR model itself — rather than requiring a separate layout-analysis stage — removes an integration layer that enterprise teams have historically had to build and maintain themselves.
The confidence scores serve a dual purpose. At scale, they allow organizations to programmatically route low-confidence regions to human reviewers and auto-approve high-confidence extractions, building what the industry calls human-in-the-loop verification without requiring a person to review every page of every document. In production systems, OCR is rarely the end goal — it is the first step in a larger pipeline.
Developers building RAG systems, agent workflows, or document automation often spend more time reconstructing layout and structure than on the downstream AI logic itself. OCR 4 aims to eliminate that reconstruction step, and if it delivers on that promise, the value accrues not just in OCR cost savings but in reduced engineering hours across the entire document pipeline.
Mistral reports that OCR 4 achieved a 72% average win rate in a head-to-head human evaluation against leading competitors, conducted by independent annotators across more than 600 real-world documents in over 12 languages. The model also achieved the top overall score on OlmOCRBench at 85.20 and scored 93.07 on OmniDocBench.
But the company itself urges caution in interpreting those numbers. In its release, Mistral took the unusual step of auditing and publicly disclosing the specific types of scoring artifacts it encountered, including ground-truth errors in the reference annotations, equivalent LaTeX notation scored as mismatches, column-reading-order assumptions, and header/footer attribution issues. “We therefore treat the aggregate score as directional rather than definitive,” the company said — a notably transparent stance from a vendor announcing a product.
That transparency is well-timed. On the public OlmOCRBench leaderboard, some researchers have noted that OCR 4 currently ranks third, behind open models like Chandra OCR 2. And some open-weight models self-report higher OmniDocBench composite scores — PaddleOCR-VL-1.6 claims 96.33 — though those results have not been independently reproduced on the public leaderboard.
Early enterprise feedback has been favorable nonetheless. Aidan Donohue, an AI engineer at financial AI firm Rogo, said the company benchmarked OCR 4 against leading agentic document parsers on a chart-dense financial QA dataset and “reached equivalent accuracy at roughly 8x lower cost and 17x lower latency.” Ivan Mihailov, an AI engineer at intellectual property management firm Anaqua, said OCR 4 is “roughly 4x faster per page than our incumbent provider.”
Enterprise buyers, however, should run their own evaluations rather than relying on any vendor’s benchmark numbers. The practical question is not which model scores highest on a leaderboard, but which model produces the fewest errors on your specific documents, in your specific languages, at a price and latency that fit your workflow.
Mistral’s release lands in a geopolitical context that could hardly be more favorable for its strategic positioning.
On June 12, Anthropic was forced to disable all access to its newest AI models, Fable 5 and Mythos 5, after the U.S. Commerce Department used national security export controls to bar the company from distributing the models to any foreign national. Enterprise clients in finance, healthcare, SaaS, and critical infrastructure found their core intelligence services abruptly disabled, without prior warning or effective recourse. As of June 24, both models remain offline, with prediction markets giving only 57% odds of restoration before July 1.
That episode validated a warning Mistral CEO Arthur Mensch has been sounding for over a year. As Business Insider reported, Mensch warned at London Tech Week in June 2025 about American AI companies “having the keys” for their models, calling it a scenario where European companies are “giving leverage to their providers.” He added: “At some point, you need to be able to turn it off or turn it on, and you don’t want to leave it to another country.”
The argument gained further urgency as Mensch’s broader sovereignty pitch escalated in recent months. As reported by CNBC in late May, Mensch told the outlet: “Europe is lagging behind when it comes to [the] buildout of infrastructure, and so we are investing to close that gap.”
At the same time, Mensch pushed back against Pope Leo XIV’s call for AI to be “disarmed,” arguing that Europe cannot afford to fall behind U.S. tech giants. “We’re all for peace, but if you look at our rivals and adversaries in the world, they’re using artificial intelligence … we do need to have our own capabilities,” Mensch told reporters.
OCR 4’s single-container, self-hosted deployment model is the product-level expression of that argument. A U.S.-headquartered provider offering EU data residency means documents are stored in Frankfurt but governed by U.S. law. Mistral, incorporated in France and operating under EU jurisdiction, offering on-premise containerized deployment, means documents never leave the customer’s infrastructure at all. The EU AI Act’s fine enforcement provisions take effect August 2, adding regulatory pressure to the compliance calculus for European enterprises evaluating document AI vendors.
Mistral’s release did not arrive in isolation. Just one day before OCR 4 launched, Baidu shipped Unlimited-OCR on June 22 — a 3-billion-parameter MIT-licensed model that tackles one of the most persistent pain points in document AI: parsing entire PDFs and multi-page scans in a single forward pass, without chunking the input or stitching the output back together afterward.
Baidu’s model uses a technique called Reference Sliding Window Attention (R-SWA) that, as a top Hacker News commenter explained, splits the AI’s focus into two paths: maintaining full attention on the original document image while restricting memory of generated text to a tight, moving window. The result is constant KV cache size and the ability to transcribe 40-plus pages in a single forward pass. The model gathered 1,800 GitHub stars in its first 24 hours and racked up more than 479 upvotes on Hacker News, where the discussion thread ran to 109 comments.
The two releases frame what some analysts are calling the June 2026 document-AI split: self-hosted long-horizon parsing with open weights versus structured managed extraction with enterprise features.
Baidu’s model is free under an MIT license, runs on standard GPU hardware, and has no managed API or enterprise SLA. Mistral’s model is a commercial product with per-page pricing, bounding boxes, confidence scores, block classification, multi-platform distribution, and self-hosted deployment options for enterprise customers.
Unlimited-OCR may be the better tool for a research team digitizing scanned dissertations on a single GPU. OCR 4 is built for the IT procurement process — the world of SLAs, data processing agreements, and compliance audits.
Beyond Baidu, the broader OCR competitive field includes Google Document AI, Amazon Textract, Azure Document Intelligence, ABBYY Vantage, and a growing number of open-weight models.
On the Hacker News thread for Unlimited-OCR, practitioners offered a candid assessment of the state of the art. Joss82, who has worked on document parsing for 10 years, wrote bluntly: “OCR still sucks in 2026.” Meanwhile, one user named SyneRyder reported success with Claude for OCR of hundreds of pages of handwritten documents, noting the model delivered results with “no corrections required” and even pointed out a continuity error in the source text. These practitioner reports underscore a key tension in the market: performance varies wildly depending on the specific document type, language, and quality of the source material.
Step back far enough, and Mistral’s OCR 4 release is not really an OCR story. It is an enterprise go-to-market story built on top of a $4.4 billion global intelligent document processing market that is forecast to grow at a 33.1% compound annual growth rate through 2030, according to Grand View Research.
For Mistral, OCR is a wedge into enterprise AI budgets. The model feeds directly into Mistral’s Search Toolkit, the company’s open-source composable search framework announced at the AI Now Summit. In that architecture, OCR 4 serves as the ingestion layer for retrieval-augmented generation and enterprise search pipelines, converting raw documents into citation-ready, structurally classified input. The logic is clear: once an enterprise adopts OCR 4 for document extraction, Mistral’s broader model suite — including Medium 3.5 for reasoning and the Vibe agentic platform for task execution — becomes the natural next step in the stack.
That pipeline ambition is critical context for understanding Mistral’s current fundraising trajectory. Bloomberg recently reported that the company is in early discussions to raise about €3 billion ($3.5 billion) at a valuation of roughly €20 billion — nearly double the €11.7 billion valuation from its September Series C round. To date, Mistral has raised only about $4 billion, a fraction of what its largest U.S. rivals have taken in. OCR 4 and its associated enterprise revenue pipeline are part of how the company plans to justify that higher valuation, with Mistral targeting €1 billion in revenue for 2026, up from €200 million in 2025, according to Le Monde.
Mistral is a company with roughly 1,000 employees and ambitions to compete with labs that have raised 40 times as much capital. It cannot win a general-purpose model arms race against OpenAI and Anthropic. What it can do is build a differentiated enterprise stack around sovereignty, structured document intelligence, and agentic workflows — and use that stack to capture European enterprise budgets that are increasingly wary of U.S. provider dependency.
The pricing structure reinforces that strategy: at $2 per 1,000 pages in batch mode, the cost of processing a 100,000-page corporate archive falls to $200, making large-scale digitization projects economically viable in ways they may not have been with token-based vision-language model pricing.
Whether Mistral can execute that vision at scale — against Google, Amazon, Microsoft, and a surging open-source ecosystem — remains an open question. But the Anthropic export control crisis is still unresolved, European data sovereignty regulations are tightening, and a potential €20 billion funding round is on the horizon. The company is holding an OCR 4 production webinar on July 7 at 6:00 PM CET.
Two weeks ago, the argument for building AI infrastructure outside the reach of U.S. export controls was theoretical. Then the U.S. government flipped a switch, and Anthropic’s most advanced models went dark for every non-American on the planet. Mistral did not cause that crisis — but it spent the last year building the product that makes it matter.